hack/tools/scripts/llm-key-hunter/results/_cache/content/cheng20050714_litscout-rs.json

{"repo": "cheng20050714/litscout-rs", "files": {"config.example.toml": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1785745357.613245, "n": 827, "h": "deb67d9f8ebe0499", "c": "# Reference configuration for LitScout-RS.\n# The current CLI reads these values from command-line flags and environment\n# variables; this file documents a stable project configuration profile.\n\n[run]\ngithub_limit = 10\narxiv_limit = 10\nacademic_extra = false\nacademic_limit = 10\noutput = \"reports\"\ncache_dir = \".litscout-cache\"\ncache_ttl_hours = 24\nsession_dir = \"sessions\"\ntags_file = \"tags.example.toml\"\ntimeout_secs = 30\nenrich = false\n\n[github]\ntoken_env = \"GITHUB_TOKEN\"\n\n[semantic_scholar]\napi_key_env = \"SEMANTIC_SCHOLAR_API_KEY\"\n\n[openalex]\napi_key_env = \"OPENALEX_API_KEY\"\n\n[crossref]\nmailto_env = \"CROSSREF_MAILTO\"\n\n[llm.deepseek]\nenabled = false\napi_key_env = \"DEEPSEEK_API_KEY\"\nbase_url = \"https://api.deepseek.com\"\nmain_model = \"deepseek-v4-pro\"\nside_model = \"deepseek-v4-flash\"\nmax_tokens = 4096\ntimeout_secs = 30\n"}, "b8f76fdd9fdfaaf0957a9293477eb193d392c98f": {"t": 1785750660.5172215, "n": 827, "h": "deb67d9f8ebe0499", "c": "# Reference configuration for LitScout-RS.\n# The current CLI reads these values from command-line flags and environment\n# variables; this file documents a stable project configuration profile.\n\n[run]\ngithub_limit = 10\narxiv_limit = 10\nacademic_extra = false\nacademic_limit = 10\noutput = \"reports\"\ncache_dir = \".litscout-cache\"\ncache_ttl_hours = 24\nsession_dir = \"sessions\"\ntags_file = \"tags.example.toml\"\ntimeout_secs = 30\nenrich = false\n\n[github]\ntoken_env = \"GITHUB_TOKEN\"\n\n[semantic_scholar]\napi_key_env = \"SEMANTIC_SCHOLAR_API_KEY\"\n\n[openalex]\napi_key_env = \"OPENALEX_API_KEY\"\n\n[crossref]\nmailto_env = \"CROSSREF_MAILTO\"\n\n[llm.deepseek]\nenabled = false\napi_key_env = \"DEEPSEEK_API_KEY\"\nbase_url = \"https://api.deepseek.com\"\nmain_model = \"deepseek-v4-pro\"\nside_model = \"deepseek-v4-flash\"\nmax_tokens = 4096\ntimeout_secs = 30\n"}, "89599bc6feef2b153f24871f481493e359e6ff87": {"t": 1785750666.2890484, "n": 739, "h": "d325bc293add82bc", "c": "# Reference configuration for LitScout-RS.\n# The current CLI reads these values from command-line flags and environment\n# variables; this file documents a stable project configuration profile.\n\n[run]\ngithub_limit = 10\narxiv_limit = 10\nacademic_extra = false\nacademic_limit = 10\noutput = \"reports\"\ncache_dir = \".litscout-cache\"\ncache_ttl_hours = 24\nsession_dir = \"sessions\"\ntags_file = \"tags.example.toml\"\ntimeout_secs = 30\nenrich = false\n\n[github]\ntoken_env = \"GITHUB_TOKEN\"\n\n[semantic_scholar]\napi_key_env = \"SEMANTIC_SCHOLAR_API_KEY\"\n\n[llm.deepseek]\nenabled = false\napi_key_env = \"DEEPSEEK_API_KEY\"\nbase_url = \"https://api.deepseek.com\"\nmain_model = \"deepseek-v4-pro\"\nside_model = \"deepseek-v4-flash\"\nmax_tokens = 4096\ntimeout_secs = 30\n"}, "75250f5469aba3f88ce699f257a03aa1287fee25": {"t": 1785750667.3604565, "n": 635, "h": "ec987d5297f27b2e", "c": "# Reference configuration for LitScout-RS.\n# The current CLI reads these values from command-line flags and environment\n# variables; this file documents a stable project configuration profile.\n\n[run]\ngithub_limit = 10\narxiv_limit = 10\noutput = \"reports\"\ncache_dir = \".litscout-cache\"\ncache_ttl_hours = 24\nsession_dir = \"sessions\"\ntags_file = \"tags.example.toml\"\ntimeout_secs = 30\nenrich = false\n\n[github]\ntoken_env = \"GITHUB_TOKEN\"\n\n[llm.deepseek]\nenabled = false\napi_key_env = \"DEEPSEEK_API_KEY\"\nbase_url = \"https://api.deepseek.com\"\nmain_model = \"deepseek-v4-pro\"\nside_model = \"deepseek-v4-flash\"\nmax_tokens = 4096\ntimeout_secs = 30\n"}}, "src/server/routes.rs": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1785745842.0671983, "n": 23034, "h": "d939f3203833caea", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::pin::Pin;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\n\nuse crate::agent::orchestrator;\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::reading::models::{NoteQuality, PaperChatMessage, ReadingStatus, TextCoverage};\nuse crate::reading::{\n    fetcher as full_text_fetcher, library as reading_library, note_writer, paper_chat,\n};\n\nuse super::dto::{\n    AddReadingLibraryItemRequest, BranchRunRequest, ChatStreamEvent, CheckpointListResponse,\n    CitationAuditResponse, ContinueStatefulRunRequest, CoverageResponse, CreateStatefulRunRequest,\n    EvidenceResponse, FrontendConfig, GenerateReadingNoteRequest, PaperChatRequest,\n    ReadingLibraryItemResponse, ReadingLibraryResponse, ReportTranslateRequest,\n    ReportTranslateResponse, ReviseStatefulPlanRequest, StatefulRunResponse,\n    StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/report/translate\", post(translate_report))\n        .route(\"/api/library\", get(list_reading_library))\n        .route(\"/api/library/items\", post(add_reading_library_item))\n        .route(\n            \"/api/library/items/{paper_key}\",\n            get(get_reading_library_item).delete(delete_reading_library_item),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/generate-note\",\n            post(generate_reading_note),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/chat/stream\",\n            post(chat_with_paper_stream),\n        )\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    req: Option<Json<ContinueStatefulRunRequest>>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            effective_app_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            effective_llm_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn list_reading_library(\n    State(state): State<AppState>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn add_reading_library_item(\n    State(state): State<AppState>,\n    Json(req): Json<AddReadingLibraryItemRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::add_item(&state.app_config, req.run_id, req.evidence).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn get_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn delete_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    reading_library::delete_item(&state.app_config, &paper_key).await?;\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn generate_reading_note(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<GenerateReadingNoteRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let mut item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.error = None;\n    item.note_quality = None;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let fetch_item = item.clone();\n    let callback_config = state.app_config.clone();\n    let callback_paper_key = paper_key.clone();\n    let mut status_callback = move |status: ReadingStatus| -> Pin<\n        Box<dyn std::future::Future<Output = Result<()>> + Send>,\n    > {\n        let app_config = callback_config.clone();\n        let paper_key = callback_paper_key.clone();\n        Box::pin(async move {\n            let mut item = reading_library::get_item(&app_config, &paper_key).await?;\n            item.status = status;\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&app_config, &item).await\n        })\n    };\n\n    let text = match full_text_fetcher::fetch_full_text(\n        &state.app_config,\n        &fetch_item,\n        Some(&mut status_callback),\n    )\n    .await\n    {\n        Ok(text) => text,\n        Err(err) => {\n            item.status = ReadingStatus::TextFailed;\n            item.error = Some(err.to_string());\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&state.app_config, &item).await?;\n            return Ok(Json(ReadingLibraryItemResponse { item }));\n        }\n    };\n    item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.text = Some(text.text);\n    item.text_source_url = Some(text.source_url);\n    item.text_coverage = Some(text.coverage);\n    item.text_meta = Some(text.meta);\n    if matches!(\n        item.text_coverage,\n        Some(TextCoverage::Failed | TextCoverage::AbstractOnly)\n    ) {\n        item.status = ReadingStatus::TextFailed;\n        item.error = Some(\"全文获取失败，未生成详细笔记。\".to_string());\n        item.updated_at = chrono::Utc::now();\n        reading_library::save_item(&state.app_config, &item).await?;\n        return Ok(Json(ReadingLibraryItemResponse { item }));\n    }\n\n    item.status = ReadingStatus::TextReady;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    item.status = ReadingStatus::GeneratingNote;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let llm_config = effective_llm_config(&state, &req.config);\n    match note_writer::generate_note(&item, &llm_config).await {\n        Ok(note) => {\n            item.note = Some(note);\n            item.note_quality = Some(NoteQuality::FullText);\n            item.status = ReadingStatus::Ready;\n            item.error = None;\n        }\n        Err(err) => {\n            item.status = ReadingStatus::NoteFailed;\n            item.error = Some(err.to_string());\n        }\n    }\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn chat_with_paper_stream(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<PaperChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let mut item = match reading_library::get_item(&state.app_config, &paper_key).await {\n            Ok(item) => item,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let llm_config = effective_llm_config(&state, &req.config);\n        let request = match paper_chat::build_paper_chat_request(&item, &req.question, &llm_config)\n        {\n            Ok(request) => request,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let config = match DeepSeekConfig::from_llm_config(&llm_config) {\n            Some(config) => config,\n            None => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: \"论文追问需要 DeepSeek API Key。\".to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let client = match DeepSeekClient::new(config) {\n            Ok(client) => client,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match client\n            .chat_completions_stream_text(request, Some(delta_tx))\n            .await\n        {\n            Ok(answer) => {\n                let now = chrono::Utc::now();\n                item.chat_history.push(PaperChatMessage {\n                    role: \"user\".to_string(),\n                    content: req.question,\n                    created_at: now,\n                });\n                item.chat_history.push(PaperChatMessage {\n                    role: \"assistant\".to_string(),\n                    content: answer,\n                    created_at: chrono::Utc::now(),\n                });\n                item.updated_at = chrono::Utc::now();\n                let _ = reading_library::save_item(&state.app_config, &item).await;\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    if let Some(key) = non_empty(frontend.semantic_scholar_api_key.as_deref()) {\n        app_config.semantic_scholar_api_key = Some(key.to_string());\n    }\n    if let Some(key) = non_empty(frontend.openalex_api_key.as_deref()) {\n        app_config.openalex_api_key = Some(key.to_string());\n    }\n    if let Some(mailto) = non_empty(frontend.crossref_mailto.as_deref()) {\n        app_config.crossref_mailto = Some(mailto.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            openalex_api_key: None,\n            crossref_mailto: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "f0e381df64f2448f45cfc7566949703029fd6216": {"t": 1785750659.9460387, "n": 23034, "h": "d939f3203833caea", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::pin::Pin;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\n\nuse crate::agent::orchestrator;\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::reading::models::{NoteQuality, PaperChatMessage, ReadingStatus, TextCoverage};\nuse crate::reading::{\n    fetcher as full_text_fetcher, library as reading_library, note_writer, paper_chat,\n};\n\nuse super::dto::{\n    AddReadingLibraryItemRequest, BranchRunRequest, ChatStreamEvent, CheckpointListResponse,\n    CitationAuditResponse, ContinueStatefulRunRequest, CoverageResponse, CreateStatefulRunRequest,\n    EvidenceResponse, FrontendConfig, GenerateReadingNoteRequest, PaperChatRequest,\n    ReadingLibraryItemResponse, ReadingLibraryResponse, ReportTranslateRequest,\n    ReportTranslateResponse, ReviseStatefulPlanRequest, StatefulRunResponse,\n    StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/report/translate\", post(translate_report))\n        .route(\"/api/library\", get(list_reading_library))\n        .route(\"/api/library/items\", post(add_reading_library_item))\n        .route(\n            \"/api/library/items/{paper_key}\",\n            get(get_reading_library_item).delete(delete_reading_library_item),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/generate-note\",\n            post(generate_reading_note),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/chat/stream\",\n            post(chat_with_paper_stream),\n        )\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    req: Option<Json<ContinueStatefulRunRequest>>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            effective_app_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            effective_llm_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn list_reading_library(\n    State(state): State<AppState>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn add_reading_library_item(\n    State(state): State<AppState>,\n    Json(req): Json<AddReadingLibraryItemRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::add_item(&state.app_config, req.run_id, req.evidence).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn get_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn delete_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    reading_library::delete_item(&state.app_config, &paper_key).await?;\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn generate_reading_note(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<GenerateReadingNoteRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let mut item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.error = None;\n    item.note_quality = None;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let fetch_item = item.clone();\n    let callback_config = state.app_config.clone();\n    let callback_paper_key = paper_key.clone();\n    let mut status_callback = move |status: ReadingStatus| -> Pin<\n        Box<dyn std::future::Future<Output = Result<()>> + Send>,\n    > {\n        let app_config = callback_config.clone();\n        let paper_key = callback_paper_key.clone();\n        Box::pin(async move {\n            let mut item = reading_library::get_item(&app_config, &paper_key).await?;\n            item.status = status;\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&app_config, &item).await\n        })\n    };\n\n    let text = match full_text_fetcher::fetch_full_text(\n        &state.app_config,\n        &fetch_item,\n        Some(&mut status_callback),\n    )\n    .await\n    {\n        Ok(text) => text,\n        Err(err) => {\n            item.status = ReadingStatus::TextFailed;\n            item.error = Some(err.to_string());\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&state.app_config, &item).await?;\n            return Ok(Json(ReadingLibraryItemResponse { item }));\n        }\n    };\n    item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.text = Some(text.text);\n    item.text_source_url = Some(text.source_url);\n    item.text_coverage = Some(text.coverage);\n    item.text_meta = Some(text.meta);\n    if matches!(\n        item.text_coverage,\n        Some(TextCoverage::Failed | TextCoverage::AbstractOnly)\n    ) {\n        item.status = ReadingStatus::TextFailed;\n        item.error = Some(\"全文获取失败，未生成详细笔记。\".to_string());\n        item.updated_at = chrono::Utc::now();\n        reading_library::save_item(&state.app_config, &item).await?;\n        return Ok(Json(ReadingLibraryItemResponse { item }));\n    }\n\n    item.status = ReadingStatus::TextReady;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    item.status = ReadingStatus::GeneratingNote;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let llm_config = effective_llm_config(&state, &req.config);\n    match note_writer::generate_note(&item, &llm_config).await {\n        Ok(note) => {\n            item.note = Some(note);\n            item.note_quality = Some(NoteQuality::FullText);\n            item.status = ReadingStatus::Ready;\n            item.error = None;\n        }\n        Err(err) => {\n            item.status = ReadingStatus::NoteFailed;\n            item.error = Some(err.to_string());\n        }\n    }\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn chat_with_paper_stream(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<PaperChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let mut item = match reading_library::get_item(&state.app_config, &paper_key).await {\n            Ok(item) => item,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let llm_config = effective_llm_config(&state, &req.config);\n        let request = match paper_chat::build_paper_chat_request(&item, &req.question, &llm_config)\n        {\n            Ok(request) => request,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let config = match DeepSeekConfig::from_llm_config(&llm_config) {\n            Some(config) => config,\n            None => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: \"论文追问需要 DeepSeek API Key。\".to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let client = match DeepSeekClient::new(config) {\n            Ok(client) => client,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match client\n            .chat_completions_stream_text(request, Some(delta_tx))\n            .await\n        {\n            Ok(answer) => {\n                let now = chrono::Utc::now();\n                item.chat_history.push(PaperChatMessage {\n                    role: \"user\".to_string(),\n                    content: req.question,\n                    created_at: now,\n                });\n                item.chat_history.push(PaperChatMessage {\n                    role: \"assistant\".to_string(),\n                    content: answer,\n                    created_at: chrono::Utc::now(),\n                });\n                item.updated_at = chrono::Utc::now();\n                let _ = reading_library::save_item(&state.app_config, &item).await;\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    if let Some(key) = non_empty(frontend.semantic_scholar_api_key.as_deref()) {\n        app_config.semantic_scholar_api_key = Some(key.to_string());\n    }\n    if let Some(key) = non_empty(frontend.openalex_api_key.as_deref()) {\n        app_config.openalex_api_key = Some(key.to_string());\n    }\n    if let Some(mailto) = non_empty(frontend.crossref_mailto.as_deref()) {\n        app_config.crossref_mailto = Some(mailto.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            openalex_api_key: None,\n            crossref_mailto: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "89599bc6feef2b153f24871f481493e359e6ff87": {"t": 1785750661.1822298, "n": 25061, "h": "3bc7f741c37daf5c", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::pin::Pin;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\n\nuse crate::agent::{orchestrator, report_chat};\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::reading::models::{NoteQuality, PaperChatMessage, ReadingStatus, TextCoverage};\nuse crate::reading::{\n    fetcher as full_text_fetcher, library as reading_library, note_writer, paper_chat,\n};\n\nuse super::dto::{\n    AddReadingLibraryItemRequest, BranchRunRequest, ChatStreamEvent, CheckpointListResponse,\n    CitationAuditResponse, ContinueStatefulRunRequest, CoverageResponse, CreateStatefulRunRequest,\n    EvidenceResponse, FrontendConfig, GenerateReadingNoteRequest, PaperChatRequest,\n    ReadingLibraryItemResponse, ReadingLibraryResponse, ReportChatRequest, ReportChatResponse,\n    ReportTranslateRequest, ReportTranslateResponse, ReviseStatefulPlanRequest,\n    StatefulRunResponse, StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .route(\"/api/library\", get(list_reading_library))\n        .route(\"/api/library/items\", post(add_reading_library_item))\n        .route(\n            \"/api/library/items/{paper_key}\",\n            get(get_reading_library_item).delete(delete_reading_library_item),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/generate-note\",\n            post(generate_reading_note),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/chat/stream\",\n            post(chat_with_paper_stream),\n        )\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    req: Option<Json<ContinueStatefulRunRequest>>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            effective_app_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            effective_llm_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n            delta_tx,\n        )\n        .await;\n        match result {\n            Ok(_) => {\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn list_reading_library(\n    State(state): State<AppState>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn add_reading_library_item(\n    State(state): State<AppState>,\n    Json(req): Json<AddReadingLibraryItemRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::add_item(&state.app_config, req.run_id, req.evidence).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn get_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn delete_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    reading_library::delete_item(&state.app_config, &paper_key).await?;\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn generate_reading_note(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<GenerateReadingNoteRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let mut item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.error = None;\n    item.note_quality = None;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let fetch_item = item.clone();\n    let callback_config = state.app_config.clone();\n    let callback_paper_key = paper_key.clone();\n    let mut status_callback = move |status: ReadingStatus| -> Pin<\n        Box<dyn std::future::Future<Output = Result<()>> + Send>,\n    > {\n        let app_config = callback_config.clone();\n        let paper_key = callback_paper_key.clone();\n        Box::pin(async move {\n            let mut item = reading_library::get_item(&app_config, &paper_key).await?;\n            item.status = status;\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&app_config, &item).await\n        })\n    };\n\n    let text = match full_text_fetcher::fetch_full_text(\n        &state.app_config,\n        &fetch_item,\n        Some(&mut status_callback),\n    )\n    .await\n    {\n        Ok(text) => text,\n        Err(err) => {\n            item.status = ReadingStatus::TextFailed;\n            item.error = Some(err.to_string());\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&state.app_config, &item).await?;\n            return Ok(Json(ReadingLibraryItemResponse { item }));\n        }\n    };\n    item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.text = Some(text.text);\n    item.text_source_url = Some(text.source_url);\n    item.text_coverage = Some(text.coverage);\n    item.text_meta = Some(text.meta);\n    if matches!(\n        item.text_coverage,\n        Some(TextCoverage::Failed | TextCoverage::AbstractOnly)\n    ) {\n        item.status = ReadingStatus::TextFailed;\n        item.error = Some(\"全文获取失败，未生成详细笔记。\".to_string());\n        item.updated_at = chrono::Utc::now();\n        reading_library::save_item(&state.app_config, &item).await?;\n        return Ok(Json(ReadingLibraryItemResponse { item }));\n    }\n\n    item.status = ReadingStatus::TextReady;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    item.status = ReadingStatus::GeneratingNote;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let llm_config = effective_llm_config(&state, &req.config);\n    match note_writer::generate_note(&item, &llm_config).await {\n        Ok(note) => {\n            item.note = Some(note);\n            item.note_quality = Some(NoteQuality::FullText);\n            item.status = ReadingStatus::Ready;\n            item.error = None;\n        }\n        Err(err) => {\n            item.status = ReadingStatus::NoteFailed;\n            item.error = Some(err.to_string());\n        }\n    }\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn chat_with_paper_stream(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<PaperChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let mut item = match reading_library::get_item(&state.app_config, &paper_key).await {\n            Ok(item) => item,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let llm_config = effective_llm_config(&state, &req.config);\n        let request = match paper_chat::build_paper_chat_request(&item, &req.question, &llm_config)\n        {\n            Ok(request) => request,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let config = match DeepSeekConfig::from_llm_config(&llm_config) {\n            Some(config) => config,\n            None => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: \"论文追问需要 DeepSeek API Key。\".to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let client = match DeepSeekClient::new(config) {\n            Ok(client) => client,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match client\n            .chat_completions_stream_text(request, Some(delta_tx))\n            .await\n        {\n            Ok(answer) => {\n                let now = chrono::Utc::now();\n                item.chat_history.push(PaperChatMessage {\n                    role: \"user\".to_string(),\n                    content: req.question,\n                    created_at: now,\n                });\n                item.chat_history.push(PaperChatMessage {\n                    role: \"assistant\".to_string(),\n                    content: answer,\n                    created_at: chrono::Utc::now(),\n                });\n                item.updated_at = chrono::Utc::now();\n                let _ = reading_library::save_item(&state.app_config, &item).await;\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    if let Some(key) = non_empty(frontend.semantic_scholar_api_key.as_deref()) {\n        app_config.semantic_scholar_api_key = Some(key.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "638e49c5b42c5e902379c23b653519ba0ed59bda": {"t": 1785750666.4291215, "n": 24861, "h": "da2ddf37c6b0e9d7", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::pin::Pin;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\n\nuse crate::agent::{orchestrator, report_chat};\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::reading::models::{NoteQuality, PaperChatMessage, ReadingStatus, TextCoverage};\nuse crate::reading::{\n    fetcher as full_text_fetcher, library as reading_library, note_writer, paper_chat,\n};\n\nuse super::dto::{\n    AddReadingLibraryItemRequest, BranchRunRequest, ChatStreamEvent, CheckpointListResponse,\n    CitationAuditResponse, ContinueStatefulRunRequest, CoverageResponse, CreateStatefulRunRequest,\n    EvidenceResponse, FrontendConfig, GenerateReadingNoteRequest, PaperChatRequest,\n    ReadingLibraryItemResponse, ReadingLibraryResponse, ReportChatRequest, ReportChatResponse,\n    ReportTranslateRequest, ReportTranslateResponse, ReviseStatefulPlanRequest,\n    StatefulRunResponse, StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .route(\"/api/library\", get(list_reading_library))\n        .route(\"/api/library/items\", post(add_reading_library_item))\n        .route(\n            \"/api/library/items/{paper_key}\",\n            get(get_reading_library_item).delete(delete_reading_library_item),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/generate-note\",\n            post(generate_reading_note),\n        )\n        .route(\n            \"/api/library/items/{paper_key}/chat/stream\",\n            post(chat_with_paper_stream),\n        )\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    req: Option<Json<ContinueStatefulRunRequest>>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            effective_app_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            effective_llm_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n            delta_tx,\n        )\n        .await;\n        match result {\n            Ok(_) => {\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn list_reading_library(\n    State(state): State<AppState>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn add_reading_library_item(\n    State(state): State<AppState>,\n    Json(req): Json<AddReadingLibraryItemRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::add_item(&state.app_config, req.run_id, req.evidence).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn get_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn delete_reading_library_item(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n) -> Result<Json<ReadingLibraryResponse>> {\n    reading_library::delete_item(&state.app_config, &paper_key).await?;\n    let items = reading_library::list_items(&state.app_config).await?;\n    Ok(Json(ReadingLibraryResponse { items }))\n}\n\nasync fn generate_reading_note(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<GenerateReadingNoteRequest>,\n) -> Result<Json<ReadingLibraryItemResponse>> {\n    let mut item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.error = None;\n    item.note_quality = None;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let fetch_item = item.clone();\n    let callback_config = state.app_config.clone();\n    let callback_paper_key = paper_key.clone();\n    let mut status_callback = move |status: ReadingStatus| -> Pin<\n        Box<dyn std::future::Future<Output = Result<()>> + Send>,\n    > {\n        let app_config = callback_config.clone();\n        let paper_key = callback_paper_key.clone();\n        Box::pin(async move {\n            let mut item = reading_library::get_item(&app_config, &paper_key).await?;\n            item.status = status;\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&app_config, &item).await\n        })\n    };\n\n    let text = match full_text_fetcher::fetch_full_text(\n        &state.app_config,\n        &fetch_item,\n        Some(&mut status_callback),\n    )\n    .await\n    {\n        Ok(text) => text,\n        Err(err) => {\n            item.status = ReadingStatus::TextFailed;\n            item.error = Some(err.to_string());\n            item.updated_at = chrono::Utc::now();\n            reading_library::save_item(&state.app_config, &item).await?;\n            return Ok(Json(ReadingLibraryItemResponse { item }));\n        }\n    };\n    item = reading_library::get_item(&state.app_config, &paper_key).await?;\n    item.text = Some(text.text);\n    item.text_source_url = Some(text.source_url);\n    item.text_coverage = Some(text.coverage);\n    item.text_meta = Some(text.meta);\n    if matches!(\n        item.text_coverage,\n        Some(TextCoverage::Failed | TextCoverage::AbstractOnly)\n    ) {\n        item.status = ReadingStatus::TextFailed;\n        item.error = Some(\"全文获取失败，未生成详细笔记。\".to_string());\n        item.updated_at = chrono::Utc::now();\n        reading_library::save_item(&state.app_config, &item).await?;\n        return Ok(Json(ReadingLibraryItemResponse { item }));\n    }\n\n    item.status = ReadingStatus::TextReady;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    item.status = ReadingStatus::GeneratingNote;\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n\n    let llm_config = effective_llm_config(&state, &req.config);\n    match note_writer::generate_note(&item, &llm_config).await {\n        Ok(note) => {\n            item.note = Some(note);\n            item.note_quality = Some(NoteQuality::FullText);\n            item.status = ReadingStatus::Ready;\n            item.error = None;\n        }\n        Err(err) => {\n            item.status = ReadingStatus::NoteFailed;\n            item.error = Some(err.to_string());\n        }\n    }\n    item.updated_at = chrono::Utc::now();\n    reading_library::save_item(&state.app_config, &item).await?;\n    Ok(Json(ReadingLibraryItemResponse { item }))\n}\n\nasync fn chat_with_paper_stream(\n    State(state): State<AppState>,\n    Path(paper_key): Path<String>,\n    Json(req): Json<PaperChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let mut item = match reading_library::get_item(&state.app_config, &paper_key).await {\n            Ok(item) => item,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let llm_config = effective_llm_config(&state, &req.config);\n        let request = match paper_chat::build_paper_chat_request(&item, &req.question, &llm_config)\n        {\n            Ok(request) => request,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let config = match DeepSeekConfig::from_llm_config(&llm_config) {\n            Some(config) => config,\n            None => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: \"论文追问需要 DeepSeek API Key。\".to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let client = match DeepSeekClient::new(config) {\n            Ok(client) => client,\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n                return;\n            }\n        };\n        let (delta_tx, mut delta_rx) = mpsc::channel::<String>(32);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(text) = delta_rx.recv().await {\n                if forward_tx\n                    .send(ChatStreamEvent::Delta { text })\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match client\n            .chat_completions_stream_text(request, Some(delta_tx))\n            .await\n        {\n            Ok(answer) => {\n                let now = chrono::Utc::now();\n                item.chat_history.push(PaperChatMessage {\n                    role: \"user\".to_string(),\n                    content: req.question,\n                    created_at: now,\n                });\n                item.chat_history.push(PaperChatMessage {\n                    role: \"assistant\".to_string(),\n                    content: answer,\n                    created_at: chrono::Utc::now(),\n                });\n                item.updated_at = chrono::Utc::now();\n                let _ = reading_library::save_item(&state.app_config, &item).await;\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "bccd8c422afb8962e69d703d2ddcbdac277a8101": {"t": 1785750667.539814, "n": 16543, "h": "04895b929b6d17dd", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\n\nuse crate::agent::{orchestrator, report_chat};\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\n\nuse super::dto::{\n    BranchRunRequest, ChatStreamEvent, CheckpointListResponse, CitationAuditResponse,\n    ContinueStatefulRunRequest, CoverageResponse, CreateStatefulRunRequest, EvidenceResponse,\n    FrontendConfig, ReportChatRequest, ReportChatResponse, ReportTranslateRequest,\n    ReportTranslateResponse, ReviseStatefulPlanRequest, StatefulFollowupRequest,\n    StatefulFollowupResponse, StatefulRunResponse, StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/follow-up\", post(followup_stateful_run))\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    req: Option<Json<ContinueStatefulRunRequest>>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            effective_app_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            effective_llm_config(\n                &state,\n                req.as_ref()\n                    .map(|Json(req)| &req.config)\n                    .unwrap_or(&FrontendConfig::default()),\n            ),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn followup_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<StatefulFollowupRequest>,\n) -> Result<Json<StatefulFollowupResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let route = orchestrator::route_followup_for_run(\n        &req.question,\n        run.evidence_memory.as_ref(),\n        run.report_markdown.as_deref(),\n    );\n    Ok(Json(StatefulFollowupResponse { run_id, route }))\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n        )\n        .await;\n\n        match result {\n            Ok(answer) => {\n                for chunk in markdown_chunks(&answer) {\n                    if tx\n                        .send(ChatStreamEvent::Delta { text: chunk })\n                        .await\n                        .is_err()\n                    {\n                        return;\n                    }\n                    tokio::time::sleep(Duration::from_millis(45)).await;\n                }\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn markdown_chunks(answer: &str) -> Vec<String> {\n    let mut chunks = answer\n        .split_inclusive('\\n')\n        .map(str::to_string)\n        .filter(|chunk| !chunk.is_empty())\n        .collect::<Vec<_>>();\n    if chunks.is_empty() && !answer.is_empty() {\n        chunks.push(answer.to_string());\n    }\n    chunks\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "1d352f50bc41fcab4e021216bca591d876481974": {"t": 1785750668.7556002, "n": 23315, "h": "508035d99f858294", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\nuse uuid::Uuid;\n\nuse crate::agent::{orchestrator, planner, report_chat};\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::model::{SearchAspect, SearchPlan, SearchQuery};\nuse crate::workflow;\n\nuse super::dto::{\n    BranchRunRequest, ChatStreamEvent, CheckpointListResponse, CitationAuditResponse,\n    CoverageResponse, CreateStatefulRunRequest, EvidenceResponse, FrontendConfig, PlanRequest,\n    PlanResponse, PlanReviseRequest, ReportChatRequest, ReportChatResponse, ReportTranslateRequest,\n    ReportTranslateResponse, ReviseStatefulPlanRequest, RunRequest, RunResponse, RunStreamEvent,\n    StatefulFollowupRequest, StatefulFollowupResponse, StatefulRunResponse, StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/plan\", post(create_plan))\n        .route(\"/api/plan/revise\", post(revise_plan))\n        .route(\"/api/run\", post(run_research))\n        .route(\"/api/run/stream\", post(run_research_stream))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/follow-up\", post(followup_stateful_run))\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanRequest>,\n) -> Result<Json<PlanResponse>> {\n    let plan = planner::generate_chinese_plan(\n        &req.topic,\n        req.github_limit,\n        req.arxiv_limit,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    let mut response = PlanResponse::from(plan);\n    response.language = req.language;\n    Ok(Json(response))\n}\n\nasync fn revise_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanReviseRequest>,\n) -> Result<Json<PlanResponse>> {\n    let mut current_plan = planner::PlanOutput::from(req.current_plan);\n    current_plan.plan_id = req.plan_id;\n    let revised = planner::revise_chinese_plan(\n        &current_plan,\n        &req.user_feedback,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(PlanResponse::from(revised)))\n}\n\nasync fn run_research(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Result<Json<RunResponse>> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_for_report(query, search_plan, app_config, llm_config).await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n\n    Ok(Json(RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n        report_markdown,\n    }))\n}\n\nasync fn run_research_stream(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<RunStreamEvent>(64);\n\n    tokio::spawn(async move {\n        let stream_result = run_research_for_stream(state, req, tx.clone()).await;\n        if let Err(err) = stream_result {\n            let _ = tx\n                .send(RunStreamEvent::RunFailed {\n                    error: err.to_string(),\n                })\n                .await;\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            RunStreamEvent::Workflow(workflow_event) => match workflow_event {\n                workflow::WorkflowEvent::FetchStarted { .. } => \"fetch_started\",\n                workflow::WorkflowEvent::SourceFinished { .. } => \"source_finished\",\n                workflow::WorkflowEvent::RankingFinished { .. } => \"ranking_finished\",\n                workflow::WorkflowEvent::ClassificationFinished { .. } => \"classification_finished\",\n                workflow::WorkflowEvent::SynthesisStarted => \"synthesis_started\",\n                workflow::WorkflowEvent::QualityWarning { .. } => \"quality_warning\",\n                workflow::WorkflowEvent::ReportReady { .. } => \"workflow_report_ready\",\n            },\n            RunStreamEvent::ReportReady(_) => \"report_ready\",\n            RunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn run_research_for_stream(\n    state: AppState,\n    req: RunRequest,\n    tx: mpsc::Sender<RunStreamEvent>,\n) -> Result<()> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_events(query, search_plan, app_config, llm_config, |event| {\n            let tx = tx.clone();\n            tokio::spawn(async move {\n                let _ = tx.send(RunStreamEvent::Workflow(event)).await;\n            });\n        })\n        .await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n    let response = RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        report_markdown,\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n    };\n    tx.send(RunStreamEvent::ReportReady(response))\n        .await\n        .map_err(|_| AppError::Workflow(\"failed to send report_ready event\".to_string()))?;\n    Ok(())\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            (*state.app_config).clone(),\n            (*state.llm_config).clone(),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunReady(Box::new(response)))\n                    .await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by_key(|checkpoint| checkpoint.created_at);\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn followup_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<StatefulFollowupRequest>,\n) -> Result<Json<StatefulFollowupResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let route = orchestrator::route_followup_for_run(\n        &req.question,\n        run.evidence_memory.as_ref(),\n        run.report_markdown.as_deref(),\n    );\n    Ok(Json(StatefulFollowupResponse { run_id, route }))\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n        )\n        .await;\n\n        match result {\n            Ok(answer) => {\n                for chunk in markdown_chunks(&answer) {\n                    if tx\n                        .send(ChatStreamEvent::Delta { text: chunk })\n                        .await\n                        .is_err()\n                    {\n                        return;\n                    }\n                    tokio::time::sleep(Duration::from_millis(45)).await;\n                }\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn markdown_chunks(answer: &str) -> Vec<String> {\n    let mut chunks = answer\n        .split_inclusive('\\n')\n        .map(str::to_string)\n        .filter(|chunk| !chunk.is_empty())\n        .collect::<Vec<_>>();\n    if chunks.is_empty() && !answer.is_empty() {\n        chunks.push(answer.to_string());\n    }\n    chunks\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\nfn to_search_plan(plan: planner::PlanOutput) -> SearchPlan {\n    SearchPlan {\n        original_topic: plan.original_topic,\n        llm_generated: plan.llm_generated,\n        aspects: plan\n            .aspects\n            .into_iter()\n            .map(|aspect| SearchAspect {\n                name: aspect.name_zh,\n                github_query: aspect.github_query,\n                arxiv_query: aspect.arxiv_query,\n                github_limit: aspect.github_limit,\n                arxiv_limit: aspect.arxiv_limit,\n                rationale: Some(aspect.rationale_zh),\n            })\n            .collect(),\n    }\n}\n\nfn total_github_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.github_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\nfn total_arxiv_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.arxiv_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "7c0c9f55db1e84d4039b524e6a5c6712e65c973e": {"t": 1785750670.0229423, "n": 23211, "h": "a57ee88fad2392c9", "c": "use axum::{\n    extract::{Path, State},\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\nuse uuid::Uuid;\n\nuse crate::agent::{orchestrator, planner, report_chat};\nuse crate::checkpoint;\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::model::{SearchAspect, SearchPlan, SearchQuery};\nuse crate::workflow;\n\nuse super::dto::{\n    BranchRunRequest, ChatStreamEvent, CheckpointListResponse, CitationAuditResponse,\n    CoverageResponse, CreateStatefulRunRequest, EvidenceResponse, FrontendConfig, PlanRequest,\n    PlanResponse, PlanReviseRequest, ReportChatRequest, ReportChatResponse, ReportTranslateRequest,\n    ReportTranslateResponse, ReviseStatefulPlanRequest, RunRequest, RunResponse, RunStreamEvent,\n    StatefulFollowupRequest, StatefulFollowupResponse, StatefulRunResponse, StatefulRunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/plan\", post(create_plan))\n        .route(\"/api/plan/revise\", post(revise_plan))\n        .route(\"/api/run\", post(run_research))\n        .route(\"/api/run/stream\", post(run_research_stream))\n        .route(\"/api/runs\", post(create_stateful_run))\n        .route(\"/api/runs/{run_id}\", get(get_stateful_run))\n        .route(\"/api/runs/{run_id}/events\", get(stream_stateful_run_events))\n        .route(\n            \"/api/runs/{run_id}/approve-plan\",\n            post(continue_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/continue\", post(continue_stateful_run))\n        .route(\"/api/runs/{run_id}/revise-plan\", post(revise_stateful_plan))\n        .route(\"/api/runs/{run_id}/evidence\", get(get_stateful_evidence))\n        .route(\"/api/runs/{run_id}/coverage\", get(get_stateful_coverage))\n        .route(\n            \"/api/runs/{run_id}/citation-audit\",\n            get(get_stateful_citation_audit),\n        )\n        .route(\n            \"/api/runs/{run_id}/checkpoints\",\n            get(list_stateful_checkpoints),\n        )\n        .route(\n            \"/api/runs/{run_id}/branch-from-checkpoint\",\n            post(branch_stateful_run),\n        )\n        .route(\"/api/runs/{run_id}/follow-up\", post(followup_stateful_run))\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanRequest>,\n) -> Result<Json<PlanResponse>> {\n    let plan = planner::generate_chinese_plan(\n        &req.topic,\n        req.github_limit,\n        req.arxiv_limit,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    let mut response = PlanResponse::from(plan);\n    response.language = req.language;\n    Ok(Json(response))\n}\n\nasync fn revise_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanReviseRequest>,\n) -> Result<Json<PlanResponse>> {\n    let mut current_plan = planner::PlanOutput::from(req.current_plan);\n    current_plan.plan_id = req.plan_id;\n    let revised = planner::revise_chinese_plan(\n        &current_plan,\n        &req.user_feedback,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(PlanResponse::from(revised)))\n}\n\nasync fn run_research(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Result<Json<RunResponse>> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_for_report(query, search_plan, app_config, llm_config).await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n\n    Ok(Json(RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n        report_markdown,\n    }))\n}\n\nasync fn run_research_stream(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<RunStreamEvent>(64);\n\n    tokio::spawn(async move {\n        let stream_result = run_research_for_stream(state, req, tx.clone()).await;\n        if let Err(err) = stream_result {\n            let _ = tx\n                .send(RunStreamEvent::RunFailed {\n                    error: err.to_string(),\n                })\n                .await;\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            RunStreamEvent::Workflow(workflow_event) => match workflow_event {\n                workflow::WorkflowEvent::FetchStarted { .. } => \"fetch_started\",\n                workflow::WorkflowEvent::SourceFinished { .. } => \"source_finished\",\n                workflow::WorkflowEvent::RankingFinished { .. } => \"ranking_finished\",\n                workflow::WorkflowEvent::ClassificationFinished { .. } => \"classification_finished\",\n                workflow::WorkflowEvent::SynthesisStarted => \"synthesis_started\",\n                workflow::WorkflowEvent::QualityWarning { .. } => \"quality_warning\",\n                workflow::WorkflowEvent::ReportReady { .. } => \"workflow_report_ready\",\n            },\n            RunStreamEvent::ReportReady(_) => \"report_ready\",\n            RunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn run_research_for_stream(\n    state: AppState,\n    req: RunRequest,\n    tx: mpsc::Sender<RunStreamEvent>,\n) -> Result<()> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_events(query, search_plan, app_config, llm_config, |event| {\n            let tx = tx.clone();\n            tokio::spawn(async move {\n                let _ = tx.send(RunStreamEvent::Workflow(event)).await;\n            });\n        })\n        .await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n    let response = RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        report_markdown,\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n    };\n    tx.send(RunStreamEvent::ReportReady(response))\n        .await\n        .map_err(|_| AppError::Workflow(\"failed to send report_ready event\".to_string()))?;\n    Ok(())\n}\n\nasync fn create_stateful_run(\n    State(state): State<AppState>,\n    Json(req): Json<CreateStatefulRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let run = orchestrator::create_run(req.topic, app_config, llm_config, req.policy).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn stream_stateful_run_events(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(16);\n    tokio::spawn(async move {\n        match orchestrator::load_run(&state.app_config, &run_id).await {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx.send(StatefulRunStreamEvent::RunReady(response)).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn continue_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<StatefulRunStreamEvent>(64);\n    tokio::spawn(async move {\n        let (event_tx, mut event_rx) = mpsc::channel(64);\n        let forward_tx = tx.clone();\n        tokio::spawn(async move {\n            while let Some(event) = event_rx.recv().await {\n                if forward_tx\n                    .send(StatefulRunStreamEvent::Agent(event))\n                    .await\n                    .is_err()\n                {\n                    break;\n                }\n            }\n        });\n        match orchestrator::continue_run(\n            &run_id,\n            (*state.app_config).clone(),\n            (*state.llm_config).clone(),\n            Some(event_tx),\n        )\n        .await\n        {\n            Ok(run) => {\n                let response = StatefulRunResponse {\n                    run_id: run.run_id.clone(),\n                    topic: run.topic.clone(),\n                    state: run.state.clone(),\n                    run,\n                };\n                let _ = tx.send(StatefulRunStreamEvent::RunReady(response)).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(StatefulRunStreamEvent::RunFailed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            StatefulRunStreamEvent::Agent(_) => \"agent\",\n            StatefulRunStreamEvent::RunReady(_) => \"run_ready\",\n            StatefulRunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn revise_stateful_plan(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<ReviseStatefulPlanRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::revise_plan(\n        &state.app_config,\n        &run_id,\n        req.chapters,\n        req.query_portfolio,\n        req.user_feedback,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn get_stateful_evidence(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<EvidenceResponse>> {\n    let evidence_memory = orchestrator::evidence_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(EvidenceResponse {\n        run_id,\n        evidence_memory,\n    }))\n}\n\nasync fn get_stateful_coverage(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CoverageResponse>> {\n    let coverage_report = orchestrator::coverage_for_run(&state.app_config, &run_id).await?;\n    Ok(Json(CoverageResponse {\n        run_id,\n        coverage_report,\n    }))\n}\n\nasync fn get_stateful_citation_audit(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CitationAuditResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let citation_audit = run\n        .citation_audit\n        .ok_or_else(|| AppError::Workflow(\"run has no CitationAuditReport yet\".to_string()))?;\n    Ok(Json(CitationAuditResponse {\n        run_id,\n        citation_audit,\n    }))\n}\n\nasync fn list_stateful_checkpoints(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n) -> Result<Json<CheckpointListResponse>> {\n    let mut checkpoints = Vec::new();\n    let checkpoint_dir = state\n        .app_config\n        .session_dir\n        .join(&run_id)\n        .join(\"checkpoints\");\n    if let Ok(mut entries) = tokio::fs::read_dir(checkpoint_dir).await {\n        while let Some(entry) = entries.next_entry().await? {\n            let body = tokio::fs::read_to_string(entry.path()).await?;\n            let snapshot: checkpoint::CheckpointSnapshot = serde_json::from_str(&body)?;\n            checkpoints.push(snapshot.checkpoint);\n        }\n    }\n    checkpoints.sort_by(|a, b| a.created_at.cmp(&b.created_at));\n    Ok(Json(CheckpointListResponse {\n        run_id,\n        checkpoints,\n    }))\n}\n\nasync fn branch_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<BranchRunRequest>,\n) -> Result<Json<StatefulRunResponse>> {\n    let run = orchestrator::branch_from_plan_ready(\n        (*state.app_config).clone(),\n        &run_id,\n        &req.checkpoint_id,\n    )\n    .await?;\n    Ok(Json(StatefulRunResponse {\n        run_id: run.run_id.clone(),\n        topic: run.topic.clone(),\n        state: run.state.clone(),\n        run,\n    }))\n}\n\nasync fn followup_stateful_run(\n    State(state): State<AppState>,\n    Path(run_id): Path<String>,\n    Json(req): Json<StatefulFollowupRequest>,\n) -> Result<Json<StatefulFollowupResponse>> {\n    let run = orchestrator::load_run(&state.app_config, &run_id).await?;\n    let route = orchestrator::route_followup_for_run(\n        &req.question,\n        run.evidence_memory.as_ref(),\n        run.report_markdown.as_deref(),\n    );\n    Ok(Json(StatefulFollowupResponse { run_id, route }))\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n        )\n        .await;\n\n        match result {\n            Ok(answer) => {\n                for chunk in markdown_chunks(&answer) {\n                    if tx\n                        .send(ChatStreamEvent::Delta { text: chunk })\n                        .await\n                        .is_err()\n                    {\n                        return;\n                    }\n                    tokio::time::sleep(Duration::from_millis(45)).await;\n                }\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn markdown_chunks(answer: &str) -> Vec<String> {\n    let mut chunks = answer\n        .split_inclusive('\\n')\n        .map(str::to_string)\n        .filter(|chunk| !chunk.is_empty())\n        .collect::<Vec<_>>();\n    if chunks.is_empty() && !answer.is_empty() {\n        chunks.push(answer.to_string());\n    }\n    chunks\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\nfn to_search_plan(plan: planner::PlanOutput) -> SearchPlan {\n    SearchPlan {\n        original_topic: plan.original_topic,\n        llm_generated: plan.llm_generated,\n        aspects: plan\n            .aspects\n            .into_iter()\n            .map(|aspect| SearchAspect {\n                name: aspect.name_zh,\n                github_query: aspect.github_query,\n                arxiv_query: aspect.arxiv_query,\n                github_limit: aspect.github_limit,\n                arxiv_limit: aspect.arxiv_limit,\n                rationale: Some(aspect.rationale_zh),\n            })\n            .collect(),\n    }\n}\n\nfn total_github_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.github_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\nfn total_arxiv_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.arxiv_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}, "bc832b32501006b47bc177689b760973d026efe2": {"t": 1785750672.1664362, "n": 13942, "h": "293a00c4017b7a8b", "c": "use axum::{\n    extract::State,\n    response::sse::{Event, KeepAlive, Sse},\n    routing::{get, post},\n    Json, Router,\n};\nuse serde::Serialize;\nuse std::convert::Infallible;\nuse std::time::Duration;\nuse tokio::sync::mpsc;\nuse tokio_stream::{wrappers::ReceiverStream, Stream, StreamExt};\nuse tower_http::cors::{Any, CorsLayer};\nuse tower_http::services::ServeDir;\nuse uuid::Uuid;\n\nuse crate::agent::{planner, report_chat};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{DeepSeekClient, DeepSeekConfig};\nuse crate::model::{SearchAspect, SearchPlan, SearchQuery};\nuse crate::workflow;\n\nuse super::dto::{\n    ChatStreamEvent, FrontendConfig, PlanRequest, PlanResponse, PlanReviseRequest,\n    ReportChatRequest, ReportChatResponse, ReportTranslateRequest, ReportTranslateResponse,\n    RunRequest, RunResponse, RunStreamEvent,\n};\nuse super::state::AppState;\n\npub fn build_router(state: AppState) -> Router {\n    let cors = CorsLayer::new()\n        .allow_origin(Any)\n        .allow_methods(Any)\n        .allow_headers(Any);\n\n    Router::new()\n        .route(\"/api/health\", get(health))\n        .route(\"/api/plan\", post(create_plan))\n        .route(\"/api/plan/revise\", post(revise_plan))\n        .route(\"/api/run\", post(run_research))\n        .route(\"/api/run/stream\", post(run_research_stream))\n        .route(\"/api/report/chat\", post(chat_with_report))\n        .route(\"/api/report/chat/stream\", post(chat_with_report_stream))\n        .route(\"/api/report/translate\", post(translate_report))\n        .fallback_service(ServeDir::new(\"web/dist\"))\n        .layer(cors)\n        .with_state(state)\n}\n\nasync fn health(State(state): State<AppState>) -> Json<HealthResponse> {\n    Json(HealthResponse {\n        status: \"ok\",\n        service: \"litscout-rs\",\n        llm_enabled: state.llm_config.enabled,\n        github_token_configured: state.app_config.github_token.is_some(),\n    })\n}\n\nasync fn create_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanRequest>,\n) -> Result<Json<PlanResponse>> {\n    let plan = planner::generate_chinese_plan(\n        &req.topic,\n        req.github_limit,\n        req.arxiv_limit,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    let mut response = PlanResponse::from(plan);\n    response.language = req.language;\n    Ok(Json(response))\n}\n\nasync fn revise_plan(\n    State(state): State<AppState>,\n    Json(req): Json<PlanReviseRequest>,\n) -> Result<Json<PlanResponse>> {\n    let mut current_plan = planner::PlanOutput::from(req.current_plan);\n    current_plan.plan_id = req.plan_id;\n    let revised = planner::revise_chinese_plan(\n        &current_plan,\n        &req.user_feedback,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(PlanResponse::from(revised)))\n}\n\nasync fn run_research(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Result<Json<RunResponse>> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_for_report(query, search_plan, app_config, llm_config).await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n\n    Ok(Json(RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n        report_markdown,\n    }))\n}\n\nasync fn run_research_stream(\n    State(state): State<AppState>,\n    Json(req): Json<RunRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<RunStreamEvent>(64);\n\n    tokio::spawn(async move {\n        let stream_result = run_research_for_stream(state, req, tx.clone()).await;\n        if let Err(err) = stream_result {\n            let _ = tx\n                .send(RunStreamEvent::RunFailed {\n                    error: err.to_string(),\n                })\n                .await;\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match &event {\n            RunStreamEvent::Workflow(workflow_event) => match workflow_event {\n                workflow::WorkflowEvent::FetchStarted { .. } => \"fetch_started\",\n                workflow::WorkflowEvent::SourceFinished { .. } => \"source_finished\",\n                workflow::WorkflowEvent::RankingFinished { .. } => \"ranking_finished\",\n                workflow::WorkflowEvent::ClassificationFinished { .. } => \"classification_finished\",\n                workflow::WorkflowEvent::SynthesisStarted => \"synthesis_started\",\n                workflow::WorkflowEvent::QualityWarning { .. } => \"quality_warning\",\n                workflow::WorkflowEvent::ReportReady { .. } => \"workflow_report_ready\",\n            },\n            RunStreamEvent::ReportReady(_) => \"report_ready\",\n            RunStreamEvent::RunFailed { .. } => \"run_failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"run_failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn run_research_for_stream(\n    state: AppState,\n    req: RunRequest,\n    tx: mpsc::Sender<RunStreamEvent>,\n) -> Result<()> {\n    let plan_id = req.plan_id.clone();\n    let language = req.language.clone();\n    let plan_output = planner::PlanOutput::from(req.current_plan);\n    let query = SearchQuery {\n        topic: plan_output.original_topic.clone(),\n        github_limit: total_github_limit(&plan_output),\n        arxiv_limit: total_arxiv_limit(&plan_output),\n    };\n    let search_plan = to_search_plan(plan_output);\n    let app_config = effective_app_config(&state, &req.config);\n    let llm_config = effective_llm_config(&state, &req.config);\n    let result =\n        workflow::run_with_plan_events(query, search_plan, app_config, llm_config, |event| {\n            let tx = tx.clone();\n            tokio::spawn(async move {\n                let _ = tx.send(RunStreamEvent::Workflow(event)).await;\n            });\n        })\n        .await?;\n    let report_markdown = tokio::fs::read_to_string(&result.output_path).await?;\n    let response = RunResponse {\n        session_id: format!(\"{language}-{plan_id}-{}\", Uuid::new_v4()),\n        output_report: result.output_path.display().to_string(),\n        session_path: result\n            .session_path\n            .as_ref()\n            .map(|path| path.display().to_string()),\n        report_markdown,\n        warnings: result.report.quality.warnings.clone(),\n        citations: result.report.citations.citations.clone(),\n        ranked_items: result.report.ranked_items.clone(),\n        quality: result.report.quality.clone(),\n    };\n    tx.send(RunStreamEvent::ReportReady(response))\n        .await\n        .map_err(|_| AppError::Workflow(\"failed to send report_ready event\".to_string()))?;\n    Ok(())\n}\n\nasync fn chat_with_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Result<Json<ReportChatResponse>> {\n    let answer = report_chat::answer_report_question(\n        &req.report_markdown,\n        &req.question,\n        &effective_llm_config(&state, &req.config),\n    )\n    .await?;\n    Ok(Json(ReportChatResponse { answer }))\n}\n\nasync fn chat_with_report_stream(\n    State(state): State<AppState>,\n    Json(req): Json<ReportChatRequest>,\n) -> Sse<impl Stream<Item = std::result::Result<Event, Infallible>>> {\n    let (tx, rx) = mpsc::channel::<ChatStreamEvent>(32);\n\n    tokio::spawn(async move {\n        let result = report_chat::answer_report_question_streaming(\n            &req.report_markdown,\n            &req.question,\n            &effective_llm_config(&state, &req.config),\n        )\n        .await;\n\n        match result {\n            Ok(answer) => {\n                for chunk in markdown_chunks(&answer) {\n                    if tx\n                        .send(ChatStreamEvent::Delta { text: chunk })\n                        .await\n                        .is_err()\n                    {\n                        return;\n                    }\n                    tokio::time::sleep(Duration::from_millis(45)).await;\n                }\n                let _ = tx.send(ChatStreamEvent::Done).await;\n            }\n            Err(err) => {\n                let _ = tx\n                    .send(ChatStreamEvent::Failed {\n                        error: err.to_string(),\n                    })\n                    .await;\n            }\n        }\n    });\n\n    let stream = ReceiverStream::new(rx).map(|event| {\n        let event_name = match event {\n            ChatStreamEvent::Delta { .. } => \"delta\",\n            ChatStreamEvent::Done => \"done\",\n            ChatStreamEvent::Failed { .. } => \"failed\",\n        };\n        Ok(Event::default()\n            .event(event_name)\n            .json_data(event)\n            .unwrap_or_else(|err| Event::default().event(\"failed\").data(err.to_string())))\n    });\n\n    Sse::new(stream).keep_alive(KeepAlive::new().interval(Duration::from_secs(15)))\n}\n\nasync fn translate_report(\n    State(state): State<AppState>,\n    Json(req): Json<ReportTranslateRequest>,\n) -> Result<Json<ReportTranslateResponse>> {\n    let llm_config = effective_llm_config(&state, &req.config);\n    let config = DeepSeekConfig::from_llm_config(&llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"翻译报告需要 DeepSeek API Key，请在阶段 1 配置或设置 DEEPSEEK_API_KEY。\".to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let translated = client\n        .translate_report_to_chinese(&req.report_markdown)\n        .await?;\n    Ok(Json(ReportTranslateResponse {\n        translated_markdown: translated,\n    }))\n}\n\nfn markdown_chunks(answer: &str) -> Vec<String> {\n    let mut chunks = answer\n        .split_inclusive('\\n')\n        .map(str::to_string)\n        .filter(|chunk| !chunk.is_empty())\n        .collect::<Vec<_>>();\n    if chunks.is_empty() && !answer.is_empty() {\n        chunks.push(answer.to_string());\n    }\n    chunks\n}\n\nfn effective_app_config(state: &AppState, frontend: &FrontendConfig) -> AppConfig {\n    let mut app_config = (*state.app_config).clone();\n    if let Some(token) = non_empty(frontend.github_token.as_deref()) {\n        app_config.github_token = Some(token.to_string());\n    }\n    app_config\n}\n\nfn effective_llm_config(state: &AppState, frontend: &FrontendConfig) -> LlmConfig {\n    let mut llm_config = (*state.llm_config).clone();\n    if let Some(api_key) = non_empty(frontend.deepseek_api_key.as_deref()) {\n        llm_config.enabled = true;\n        llm_config.api_key = Some(api_key.to_string());\n    }\n    if let Some(base_url) = non_empty(frontend.deepseek_base_url.as_deref()) {\n        llm_config.base_url = Some(base_url.to_string());\n    }\n    if let Some(model) = non_empty(frontend.deepseek_model.as_deref()) {\n        llm_config.main_model = model.to_string();\n    }\n    if let Some(side_model) = non_empty(frontend.deepseek_side_model.as_deref()) {\n        llm_config.side_model = Some(side_model.to_string());\n    }\n    llm_config\n}\n\nfn non_empty(value: Option<&str>) -> Option<&str> {\n    value.map(str::trim).filter(|value| !value.is_empty())\n}\n\nfn to_search_plan(plan: planner::PlanOutput) -> SearchPlan {\n    SearchPlan {\n        original_topic: plan.original_topic,\n        llm_generated: plan.llm_generated,\n        aspects: plan\n            .aspects\n            .into_iter()\n            .map(|aspect| SearchAspect {\n                name: aspect.name_zh,\n                github_query: aspect.github_query,\n                arxiv_query: aspect.arxiv_query,\n                github_limit: aspect.github_limit,\n                arxiv_limit: aspect.arxiv_limit,\n                rationale: Some(aspect.rationale_zh),\n            })\n            .collect(),\n    }\n}\n\nfn total_github_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.github_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\nfn total_arxiv_limit(plan: &planner::PlanOutput) -> usize {\n    plan.aspects\n        .iter()\n        .map(|aspect| aspect.arxiv_limit)\n        .sum::<usize>()\n        .max(1)\n}\n\n#[derive(Debug, Serialize)]\nstruct HealthResponse {\n    status: &'static str,\n    service: &'static str,\n    llm_enabled: bool,\n    github_token_configured: bool,\n}\n\n#[cfg(test)]\nmod tests {\n    use std::path::PathBuf;\n\n    use super::build_router;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::server::state::AppState;\n\n    #[test]\n    fn builds_router_for_health_endpoint() {\n        let state = AppState::new(app_config(), LlmConfig::from_env(false, 30));\n\n        let _router = build_router(state);\n    }\n\n    fn app_config() -> AppConfig {\n        AppConfig {\n            github_token: None,\n            output: PathBuf::from(\"reports\"),\n            cache_dir: PathBuf::from(\".litscout-cache\"),\n            session_dir: PathBuf::from(\"sessions\"),\n            tags_file: None,\n            use_cache: true,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        }\n    }\n}\n"}}, "src/agent/writer.rs": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1785745857.1112325, "n": 50022, "h": "1c25fd4fdd794826", "c": "use crate::config::LlmConfig;\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{\n    strip_json_fence, ChatCompletionsRequest, ChatMessage, DeepSeekClient, DeepSeekConfig,\n    ResponseFormat,\n};\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft, SourceKind,\n};\nuse chrono::Utc;\nuse serde::{Deserialize, Serialize};\nuse std::collections::BTreeSet;\nuse tracing::{debug, warn};\n\nconst MAX_EVIDENCE_PER_BATCH: usize = 3;\nconst MAX_BATCH_CONTEXT_CHARS: usize = 3_500;\nconst MAX_BATCH_PARAGRAPHS: usize = 1;\nconst MAX_CHAPTER_PARAGRAPHS: usize = 4;\nconst MAX_WRITER_EVIDENCE_PER_CHAPTER: usize = 8;\nconst WRITER_GITHUB_EVIDENCE_CAP: usize = 4;\nconst WRITER_ARXIV_EVIDENCE_CAP: usize = 4;\nconst WRITER_ACADEMIC_INDEX_EVIDENCE_CAP: usize = 3;\nconst WRITER_BIBLIOGRAPHY_EVIDENCE_CAP: usize = 1;\nconst BATCH_FINDINGS_TARGET_CHARS: usize = 180;\nconst BATCH_PARAGRAPH_TARGET_CHARS: usize = 220;\nconst CHAPTER_PARAGRAPH_TARGET_CHARS: usize = 260;\nconst GLOBAL_SUMMARY_TARGET_CHARS: usize = 420;\nconst WRITER_TITLE_MAX_CHARS: usize = 180;\nconst WRITER_EVIDENCE_NOTE_MAX_CHARS: usize = 220;\nconst WRITER_GITHUB_SNIPPET_MAX_CHARS: usize = 900;\nconst WRITER_ACADEMIC_SNIPPET_MAX_CHARS: usize = 700;\nconst WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS: usize = 200;\nconst REPAIR_PREVIOUS_OUTPUT_MAX_CHARS: usize = 2_400;\nconst WRITER_MIN_OUTPUT_TOKENS: usize = 4096;\nconst WRITER_MAX_OUTPUT_TOKENS: usize = 8192;\n\npub async fn draft_report_with_llm(\n    topic: &str,\n    chapters: &[ChapterNode],\n    memory: &EvidenceMemory,\n    llm_config: &LlmConfig,\n) -> Result<ReportDraft> {\n    let deepseek_config = DeepSeekConfig::from_llm_config(llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"Writer LLM requires an enabled DeepSeek config with API key; refusing to generate a template report.\"\n                .to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(deepseek_config.clone())?;\n    let mut chapter_drafts = Vec::with_capacity(chapters.len());\n\n    for chapter in chapters {\n        let evidence_items = memory.by_chapter(&chapter.id);\n        if evidence_items.is_empty() {\n            chapter_drafts.push(empty_chapter_draft(chapter));\n            continue;\n        }\n\n        let draft = draft_chapter_with_llm(chapter, &evidence_items, &client, &deepseek_config)\n            .await\n            .map_err(|err| {\n                AppError::Llm(format!(\n                    \"Writer LLM failed for chapter `{}`; refusing template fallback: {err}\",\n                    chapter.title_zh\n                ))\n            })?;\n        chapter_drafts.push(draft);\n    }\n\n    let global_summary_zh = synthesize_global_summary_with_llm(\n        topic,\n        &chapter_drafts,\n        memory.items.len(),\n        &client,\n        &deepseek_config,\n    )\n    .await\n    .map_err(|err| {\n        AppError::Llm(format!(\n            \"Writer global summary failed; refusing deterministic fallback: {err}\"\n        ))\n    })?;\n\n    Ok(ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    })\n}\n\nasync fn draft_chapter_with_llm(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let selected_evidence = select_evidence_for_writing(evidence_items);\n    let batches = batch_chapter_evidence(&selected_evidence);\n    debug!(\n        \"Writer selected {}/{} evidence items and split chapter `{}` into {} evidence batches\",\n        selected_evidence.len(),\n        evidence_items.len(),\n        chapter.title_zh,\n        batches.len()\n    );\n\n    let mut batch_drafts = Vec::with_capacity(batches.len());\n    for batch in &batches {\n        let draft = draft_batch_with_llm(chapter, batch, batches.len(), client, config).await?;\n        batch_drafts.push(draft);\n    }\n\n    synthesize_chapter_from_batches(chapter, &selected_evidence, &batch_drafts, client, config)\n        .await\n}\n\nfn empty_chapter_draft(chapter: &ChapterNode) -> ChapterDraft {\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs: vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未获得与章节计划直接匹配的 GitHub/arXiv 证据，因此不生成分析段落。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }],\n    }\n}\n\n#[derive(Debug, Clone)]\nstruct EvidenceBatch {\n    batch_id: usize,\n    evidence_items: Vec<EvidenceItem>,\n    context_chars: usize,\n}\n\n#[derive(Debug, Clone, Serialize)]\nstruct BatchDraft {\n    batch_id: usize,\n    findings_summary_zh: String,\n    paragraphs: Vec<ParagraphWithCitations>,\n}\n\nfn batch_chapter_evidence(evidence_items: &[EvidenceItem]) -> Vec<EvidenceBatch> {\n    let mut batches = Vec::new();\n    let mut current_items = Vec::new();\n    let mut current_chars = 0usize;\n\n    for item in evidence_items {\n        let item_chars = evidence_context_chars(item);\n        let count_limit_reached = current_items.len() >= MAX_EVIDENCE_PER_BATCH;\n        let char_limit_reached =\n            !current_items.is_empty() && current_chars + item_chars > MAX_BATCH_CONTEXT_CHARS;\n\n        if count_limit_reached || char_limit_reached {\n            batches.push(EvidenceBatch {\n                batch_id: batches.len() + 1,\n                evidence_items: current_items,\n                context_chars: current_chars,\n            });\n            current_items = Vec::new();\n            current_chars = 0;\n        }\n\n        current_chars += item_chars;\n        current_items.push(item.clone());\n    }\n\n    if !current_items.is_empty() {\n        batches.push(EvidenceBatch {\n            batch_id: batches.len() + 1,\n            evidence_items: current_items,\n            context_chars: current_chars,\n        });\n    }\n\n    batches\n}\n\nfn select_evidence_for_writing(evidence_items: &[EvidenceItem]) -> Vec<EvidenceItem> {\n    let mut selected = Vec::new();\n    let has_stronger_sources = evidence_items\n        .iter()\n        .any(|item| item.source_kind != SourceKind::Bibliography);\n\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::Arxiv,\n        WRITER_ARXIV_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::AcademicIndex,\n        WRITER_ACADEMIC_INDEX_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::GitHub,\n        WRITER_GITHUB_EVIDENCE_CAP,\n    );\n    if !has_stronger_sources {\n        append_evidence_by_kind(\n            evidence_items,\n            &mut selected,\n            SourceKind::Bibliography,\n            WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        );\n    }\n\n    selected\n}\n\nfn append_evidence_by_kind(\n    evidence_items: &[EvidenceItem],\n    selected: &mut Vec<EvidenceItem>,\n    source_kind: SourceKind,\n    source_cap: usize,\n) {\n    let mut source_count = 0usize;\n    for item in evidence_items\n        .iter()\n        .filter(|item| item.source_kind == source_kind)\n    {\n        if selected.len() >= MAX_WRITER_EVIDENCE_PER_CHAPTER || source_count >= source_cap {\n            break;\n        }\n        selected.push(item.clone());\n        source_count += 1;\n    }\n}\n\nfn evidence_context_chars(item: &EvidenceItem) -> usize {\n    let input = EvidenceSourceInput::from(item);\n    input.evidence_id.chars().count()\n        + input.title.chars().count()\n        + input.url.chars().count()\n        + input.evidence_note_zh.chars().count()\n        + input.evidence_snippet.chars().count()\n}\n\nfn excerpt_chars(text: &str, max_chars: usize) -> String {\n    let trimmed = text.trim();\n    if trimmed.chars().count() <= max_chars {\n        return trimmed.to_string();\n    }\n\n    let mut out = trimmed.chars().take(max_chars).collect::<String>();\n    out.push_str(\"...\");\n    out\n}\n\nfn parse_batch_draft_content(content: &str, batch: &EvidenceBatch) -> Result<BatchDraft> {\n    let output =\n        serde_json::from_str::<BatchDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid BatchDraft JSON: {err}\"\n            ))\n        })?;\n    let findings_summary_zh = output.findings_summary_zh.trim().to_string();\n    if findings_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer batch response contained empty findings_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&findings_summary_zh)?;\n    let paragraphs = parse_paragraph_outputs(\n        output.paragraphs,\n        &batch.evidence_items,\n        MAX_BATCH_PARAGRAPHS,\n    )?;\n\n    Ok(BatchDraft {\n        batch_id: batch.batch_id,\n        findings_summary_zh,\n        paragraphs,\n    })\n}\n\nasync fn draft_batch_with_llm(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<BatchDraft> {\n    let request = build_batch_draft_request(chapter, batch, total_batches, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_batch_draft_content(content, batch) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM JSON validation failed for chapter `{}` batch {}; requesting one repair: {first_err}\",\n                chapter.title_zh, batch.batch_id\n            );\n            let repair_request = build_batch_draft_repair_request(\n                chapter,\n                batch,\n                total_batches,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_batch_draft_content(repaired_content, batch)\n        }\n    }\n}\n\nfn build_batch_draft_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请为下面章节的第 {}/{} 个证据批次生成局部分析草稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.3),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_batch_draft_repair_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次第 {}/{} 个证据批次的输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 evidence：\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nasync fn synthesize_chapter_from_batches(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let request = build_chapter_synthesis_request(chapter, evidence_items, batch_drafts, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_chapter_draft_content(content, chapter, evidence_items) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM chapter synthesis validation failed for chapter `{}`; requesting one repair: {first_err}\",\n                chapter.title_zh\n            );\n            let repair_request = build_chapter_synthesis_repair_request(\n                chapter,\n                evidence_items,\n                batch_drafts,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_chapter_draft_content(repaired_content, chapter, evidence_items)\n        }\n    }\n}\n\nfn batch_writer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的技术报告撰稿人，正在为 `{chapter_title}` 章节处理一个 evidence batch。\\n\\\n规则：\\n\\\n1. 使用中文撰写。\\n\\\n2. findings_summary_zh 不超过 {BATCH_FINDINGS_TARGET_CHARS} 个汉字，只总结本批证据的关键发现、共性和边界。\\n\\\n3. paragraphs 输出 1-{MAX_BATCH_PARAGRAPHS} 段，每段不超过 {BATCH_PARAGRAPH_TARGET_CHARS} 个汉字；提炼项目/论文简介、核心亮点或设计思想，以及与章节问题的具体关联。\\n\\\n4. 可合并多个来源做对比分析，但每个段落必须引用至少一个 evidence_id。\\n\\\n5. 只能使用用户提供的 evidence_id 和 URL，不得新增来源、不得编造事实。\\n\\\n6. 不要只重复原文，不要输出“来源链接”模板句。\\n\\\n7. 只返回 JSON：{{\\\"findings_summary_zh\\\":\\\"...\\\",\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nfn writer_max_tokens(config: &DeepSeekConfig) -> u32 {\n    config\n        .max_tokens\n        .max(WRITER_MIN_OUTPUT_TOKENS)\n        .min(WRITER_MAX_OUTPUT_TOKENS)\n        .min(u32::MAX as usize) as u32\n}\n\nfn build_chapter_synthesis_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于已验证的 batch findings 合成章节终稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_chapter_synthesis_repair_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次章节综合输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 batch findings：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn chapter_synthesizer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的章节综合编辑，正在合成 `{chapter_title}` 章节终稿。\\n\\\n规则：\\n\\\n1. 使用中文撰写自然、分析性的技术报告段落。\\n\\\n2. 只能使用用户提供的 batch findings、batch paragraphs 和 allowed_evidence_ids。\\n\\\n3. 合并重复信息，按主题组织，不要机械罗列每条来源。\\n\\\n4. 已有 evidence 时，不得用“尚未收集到足够证据”替代分析内容。\\n\\\n5. 每个段落必须引用至少一个 allowed_evidence_ids 中的 evidence_id。\\n\\\n6. 不得新增来源、URL 或 evidence_id，不得编造事实。\\n\\\n7. 输出 1-{MAX_CHAPTER_PARAGRAPHS} 段，每段不超过 {CHAPTER_PARAGRAPH_TARGET_CHARS} 个汉字。\\n\\\n8. 不要输出“来源链接”模板句或“是本章节的关键来源之一”。\\n\\\n9. 只返回 JSON：{{\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nasync fn synthesize_global_summary_with_llm(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<String> {\n    let request = build_global_summary_request(topic, chapter_drafts, evidence_count, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_global_summary_content(content) {\n        Ok(summary) => Ok(summary),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM global summary validation failed; requesting one repair: {first_err}\"\n            );\n            let repair_request = build_global_summary_repair_request(\n                topic,\n                chapter_drafts,\n                evidence_count,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_global_summary_content(repaired_content)\n        }\n    }\n}\n\nfn build_global_summary_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于章节终稿生成报告的全局摘要。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_global_summary_repair_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次全局摘要输出，使其满足 JSON schema 和写作约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n章节终稿上下文：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn parse_global_summary_content(content: &str) -> Result<String> {\n    let output =\n        serde_json::from_str::<GlobalSummaryOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid GlobalSummary JSON: {err}\"\n            ))\n        })?;\n    let global_summary_zh = output.global_summary_zh.trim().to_string();\n    if global_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer global summary response contained empty global_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&global_summary_zh)?;\n\n    Ok(global_summary_zh)\n}\n\nfn global_synthesizer_system_prompt() -> String {\n    format!(\n        \"你是 LitScout-RS 的报告总编辑，负责基于章节终稿生成全局摘要。\\n\\\n规则：\\n\\\n1. 使用中文撰写 1-2 段自然的全局摘要，总长度不超过 {GLOBAL_SUMMARY_TARGET_CHARS} 个汉字。\\n\\\n2. 概括报告覆盖范围、主要发现、证据边界和不足。\\n\\\n3. 只能使用用户提供的章节终稿和 citation id，不得新增来源、URL 或事实。\\n\\\n4. 不要输出固定模板句，不要写“来源链接”，不要写“是本章节的关键来源之一”。\\n\\\n5. 不要把引用账本逐条改写成摘要。\\n\\\n6. 只返回 JSON：{{\\\"global_summary_zh\\\":\\\"...\\\"}}。\"\n    )\n}\n\nfn parse_chapter_draft_content(\n    content: &str,\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n) -> Result<ChapterDraft> {\n    let output =\n        serde_json::from_str::<ChapterDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid ChapterDraft JSON: {err}\"\n            ))\n        })?;\n    let paragraphs =\n        parse_paragraph_outputs(output.paragraphs, evidence_items, MAX_CHAPTER_PARAGRAPHS)?;\n\n    Ok(ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    })\n}\n\nfn parse_paragraph_outputs(\n    paragraph_outputs: Vec<ChapterParagraphOutput>,\n    evidence_items: &[EvidenceItem],\n    max_paragraphs: usize,\n) -> Result<Vec<ParagraphWithCitations>> {\n    let allowed_ids = evidence_items\n        .iter()\n        .map(|item| item.evidence_id.as_str())\n        .collect::<BTreeSet<_>>();\n    let mut paragraphs = Vec::new();\n\n    for paragraph in paragraph_outputs.into_iter().take(max_paragraphs) {\n        let text_zh = paragraph.text_zh.trim().to_string();\n        if text_zh.is_empty() {\n            continue;\n        }\n        reject_template_residue(&text_zh)?;\n        let cited_evidence_ids = paragraph\n            .cited_evidence_ids\n            .into_iter()\n            .filter_map(|id| {\n                let trimmed = id.trim();\n                (!trimmed.is_empty()).then(|| trimmed.to_string())\n            })\n            .collect::<Vec<_>>();\n        if cited_evidence_ids.is_empty() {\n            return Err(AppError::Llm(\n                \"Writer paragraph did not cite any evidence ID\".to_string(),\n            ));\n        }\n        for id in &cited_evidence_ids {\n            if !allowed_ids.contains(id.as_str()) {\n                return Err(AppError::Llm(format!(\n                    \"Writer paragraph referenced unknown evidence id `{id}`\"\n                )));\n            }\n        }\n        paragraphs.push(ParagraphWithCitations {\n            text_zh,\n            cited_evidence_ids,\n        });\n    }\n\n    if paragraphs.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer response contained no usable paragraphs\".to_string(),\n        ));\n    }\n\n    Ok(paragraphs)\n}\n\nfn reject_template_residue(text: &str) -> Result<()> {\n    for marker in [\n        \"是本章节的关键来源之一\",\n        \"来源链接：\",\n        \"提供了与主题相关的证据\",\n    ] {\n        if text.contains(marker) {\n            return Err(AppError::Llm(format!(\n                \"Writer output still contains legacy template marker `{marker}`\"\n            )));\n        }\n    }\n    Ok(())\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push(' ');\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[derive(Debug, Serialize)]\nstruct BatchDraftInput {\n    chapter_title: String,\n    research_question: String,\n    batch_id: usize,\n    total_batches: usize,\n    batch_context_chars: usize,\n    evidence_sources: Vec<EvidenceSourceInput>,\n}\n\n#[derive(Debug, Serialize)]\nstruct ChapterSynthesisInput {\n    chapter_title: String,\n    research_question: String,\n    allowed_evidence_ids: Vec<String>,\n    batch_drafts: Vec<BatchDraft>,\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryInput {\n    topic: String,\n    chapter_count: usize,\n    evidence_count: usize,\n    chapters: Vec<GlobalSummaryChapterInput>,\n}\n\nimpl GlobalSummaryInput {\n    fn from_report(topic: &str, chapter_drafts: &[ChapterDraft], evidence_count: usize) -> Self {\n        Self {\n            topic: topic.to_string(),\n            chapter_count: chapter_drafts.len(),\n            evidence_count,\n            chapters: chapter_drafts\n                .iter()\n                .map(GlobalSummaryChapterInput::from)\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryChapterInput {\n    chapter_id: String,\n    title_zh: String,\n    paragraph_count: usize,\n    cited_evidence_ids: Vec<String>,\n    paragraph_summaries: Vec<String>,\n}\n\nimpl From<&ChapterDraft> for GlobalSummaryChapterInput {\n    fn from(chapter: &ChapterDraft) -> Self {\n        let cited_evidence_ids = chapter\n            .paragraphs\n            .iter()\n            .flat_map(|paragraph| paragraph.cited_evidence_ids.iter().cloned())\n            .collect::<BTreeSet<_>>()\n            .into_iter()\n            .collect();\n        Self {\n            chapter_id: chapter.chapter_id.clone(),\n            title_zh: chapter.title_zh.clone(),\n            paragraph_count: chapter.paragraphs.len(),\n            cited_evidence_ids,\n            paragraph_summaries: chapter\n                .paragraphs\n                .iter()\n                .map(|paragraph| excerpt_chars(&paragraph.text_zh, 500))\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct EvidenceSourceInput {\n    evidence_id: String,\n    title: String,\n    url: String,\n    source_kind: SourceKind,\n    evidence_note_zh: String,\n    evidence_snippet: String,\n}\n\nimpl From<&EvidenceItem> for EvidenceSourceInput {\n    fn from(item: &EvidenceItem) -> Self {\n        Self {\n            evidence_id: item.evidence_id.clone(),\n            title: excerpt_chars(&item.title, WRITER_TITLE_MAX_CHARS),\n            url: item.url.clone(),\n            source_kind: item.source_kind,\n            evidence_note_zh: excerpt_chars(&item.evidence_note_zh, WRITER_EVIDENCE_NOTE_MAX_CHARS),\n            evidence_snippet: excerpt_chars(\n                &item.evidence_snippet,\n                writer_snippet_max_chars(item.source_kind),\n            ),\n        }\n    }\n}\n\nfn writer_snippet_max_chars(source_kind: SourceKind) -> usize {\n    match source_kind {\n        SourceKind::GitHub => WRITER_GITHUB_SNIPPET_MAX_CHARS,\n        SourceKind::Arxiv | SourceKind::AcademicIndex => WRITER_ACADEMIC_SNIPPET_MAX_CHARS,\n        SourceKind::Bibliography => WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS,\n    }\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterDraftOutput {\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct BatchDraftOutput {\n    #[serde(default)]\n    findings_summary_zh: String,\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct GlobalSummaryOutput {\n    #[serde(default)]\n    global_summary_zh: String,\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterParagraphOutput {\n    text_zh: String,\n    #[serde(default)]\n    cited_evidence_ids: Vec<String>,\n}\n\n#[cfg(test)]\nmod tests {\n    use chrono::{DateTime, Utc};\n\n    use crate::llm::deepseek::DeepSeekConfig;\n    use crate::model::{\n        ChapterDraft, ChapterNode, EvidenceItem, EvidenceMemory, ParagraphWithCitations, SourceKind,\n    };\n\n    use super::{\n        batch_chapter_evidence, build_batch_draft_request, build_chapter_synthesis_request,\n        build_global_summary_request, draft_report_with_llm, empty_chapter_draft,\n        parse_batch_draft_content, parse_chapter_draft_content, parse_global_summary_content,\n        select_evidence_for_writing, writer_max_tokens, BatchDraft, MAX_BATCH_CONTEXT_CHARS,\n        MAX_EVIDENCE_PER_BATCH, MAX_WRITER_EVIDENCE_PER_CHAPTER, WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        WRITER_MIN_OUTPUT_TOKENS,\n    };\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = empty_chapter_draft(&chapters[0]);\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert!(draft.paragraphs[0]\n            .text_zh\n            .contains(\"尚未获得与章节计划直接匹配\"));\n    }\n\n    #[tokio::test]\n    async fn llm_writer_refuses_to_template_when_llm_is_disabled() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n            selection_report: Default::default(),\n        };\n\n        let err = draft_report_with_llm(\n            \"Rust Agent\",\n            &chapters,\n            &memory,\n            &crate::config::LlmConfig::from_env(false, 30),\n        )\n        .await\n        .expect_err(\"disabled LLM should not produce a template report\");\n\n        assert!(err\n            .to_string()\n            .contains(\"refusing to generate a template report\"));\n    }\n\n    #[tokio::test]\n    #[ignore = \"requires a real DeepSeek API key in DEEPSEEK_API_KEY\"]\n    async fn writer_real_deepseek_smoke() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n            selection_report: Default::default(),\n        };\n        let config = crate::config::LlmConfig::from_env(true, 120);\n\n        assert!(\n            config.api_key.is_some(),\n            \"DEEPSEEK_API_KEY must be set for this ignored smoke test\"\n        );\n\n        let draft = draft_report_with_llm(\"TTS 入门级项目简单调研\", &chapters, &memory, &config)\n            .await\n            .expect(\"real DeepSeek writer smoke should complete\");\n\n        assert!(!draft.global_summary_zh.trim().is_empty());\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(!draft.chapters[0].paragraphs.is_empty());\n        assert!(draft.chapters[0].paragraphs[0]\n            .cited_evidence_ids\n            .contains(&\"ev-C1\".to_string()));\n        println!(\n            \"writer_real_deepseek_smoke ok: global_summary_chars={}, chapter_paragraphs={}\",\n            draft.global_summary_zh.chars().count(),\n            draft.chapters[0].paragraphs.len()\n        );\n    }\n\n    #[test]\n    fn parses_llm_chapter_draft_json() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了从 README 提炼出的 ReAct 主循环和工具调用结构，适合作为入门实现分析对象。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect(\"valid chapter JSON should parse\");\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n        assert!(draft.paragraphs[0].text_zh.contains(\"ReAct\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_unknown_evidence_id() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"引用了不存在的证据。\",\n                    \"cited_evidence_ids\": [\"ev-C9\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_template_residue() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"acme/rust-agent 是本章节的关键来源之一。来源链接：https://github.com/acme/rust-agent\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batches_small_chapter_evidence_into_one_batch() {\n        let evidence = vec![\n            sample_evidence(\"ev-C1\", \"ch-1\"),\n            sample_evidence(\"ev-C2\", \"ch-1\"),\n            sample_evidence(\"ev-C3\", \"ch-1\"),\n        ];\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 1);\n        assert_eq!(batches[0].batch_id, 1);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert!(batches[0].context_chars > 0);\n    }\n\n    #[test]\n    fn selects_bounded_mixed_evidence_for_writing() {\n        let mut evidence = Vec::new();\n        evidence.extend((1..=5).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-A{index}\"), \"ch-1\", SourceKind::Arxiv)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-S{index}\"), \"ch-1\", SourceKind::AcademicIndex)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-G{index}\"), \"ch-1\", SourceKind::GitHub)\n        }));\n        evidence.extend((1..=2).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n        }));\n\n        let selected = select_evidence_for_writing(&evidence);\n        let bibliography_count = selected\n            .iter()\n            .filter(|item| item.source_kind == SourceKind::Bibliography)\n            .count();\n\n        assert_eq!(selected.len(), MAX_WRITER_EVIDENCE_PER_CHAPTER);\n        assert_eq!(bibliography_count, 0);\n        assert_eq!(selected[0].source_kind, SourceKind::Arxiv);\n        assert_eq!(selected[4].source_kind, SourceKind::AcademicIndex);\n        assert_eq!(selected[7].source_kind, SourceKind::GitHub);\n    }\n\n    #[test]\n    fn selects_bibliography_only_when_no_stronger_sources_exist() {\n        let evidence = (1..=3)\n            .map(|index| {\n                sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n            })\n            .collect::<Vec<_>>();\n\n        let selected = select_evidence_for_writing(&evidence);\n\n        assert_eq!(selected.len(), WRITER_BIBLIOGRAPHY_EVIDENCE_CAP);\n        assert_eq!(selected[0].source_kind, SourceKind::Bibliography);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_count() {\n        let evidence = (1..=9)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 3);\n        assert_eq!(batches[0].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[1].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].batch_id, 3);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_sanitized_context_chars() {\n        let mut first = sample_evidence(\"ev-C1\", \"ch-1\");\n        first.evidence_snippet = \"a\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut second = sample_evidence(\"ev-C2\", \"ch-1\");\n        second.evidence_snippet = \"b\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut third = sample_evidence(\"ev-C3\", \"ch-1\");\n        third.evidence_snippet = \"c\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut fourth = sample_evidence(\"ev-C4\", \"ch-1\");\n        fourth.evidence_snippet = \"d\".repeat(MAX_BATCH_CONTEXT_CHARS);\n\n        let batches = batch_chapter_evidence(&[first, second, third, fourth]);\n\n        assert_eq!(batches.len(), 2);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert_eq!(batches[1].evidence_items.len(), 1);\n        assert!(batches[0].context_chars <= MAX_BATCH_CONTEXT_CHARS);\n    }\n\n    #[test]\n    fn batch_request_truncates_long_github_snippets_for_writer_context() {\n        let chapter = sample_chapter();\n        let mut evidence = sample_evidence(\"ev-C1\", \"ch-1\");\n        evidence.evidence_snippet = format!(\"{}{}\", \"A\".repeat(1500), \"UNTRUNCATED_TAIL\");\n        let batches = batch_chapter_evidence(&[evidence]);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[0],\n            batches.len(),\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(&\"A\".repeat(900)));\n        assert!(!user_message.contains(\"UNTRUNCATED_TAIL\"));\n    }\n\n    #[test]\n    fn parses_llm_batch_draft_json() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据说明了项目 README 中的 ReAct 主循环、工具注册和入门示例。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目的 README 将 ReAct 主循环和工具注册拆成清晰的入门实现，适合用于说明基础 agent 架构。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft =\n            parse_batch_draft_content(content, &batches[0]).expect(\"valid batch JSON should parse\");\n\n        assert_eq!(draft.batch_id, 1);\n        assert!(draft.findings_summary_zh.contains(\"ReAct\"));\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_empty_findings() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"empty findings should fail\");\n\n        assert!(err.to_string().contains(\"empty findings_summary_zh\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_unknown_evidence_id() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据有一个段落引用了不存在的证据。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"这里引用了不存在的 evidence id。\",\n                    \"cited_evidence_ids\": [\"ev-C2\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_template_residue() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"acme/rust-agent 是本章节的关键来源之一。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batch_request_only_contains_current_batch_evidence() {\n        let chapter = sample_chapter();\n        let evidence = (1..=5)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n        let batches = batch_chapter_evidence(&evidence);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[1],\n            batches.len(),\n            &sample_deepseek_config(2048),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"\\\"batch_id\\\": 2\"));\n        assert!(user_message.contains(\"ev-C5\"));\n        assert!(!user_message.contains(\"ev-C1\"));\n        assert_eq!(request.max_tokens, Some(WRITER_MIN_OUTPUT_TOKENS as u32));\n    }\n\n    #[test]\n    fn writer_max_tokens_raises_too_small_config_for_reasoning_models() {\n        let config = sample_deepseek_config(32);\n\n        assert_eq!(writer_max_tokens(&config), WRITER_MIN_OUTPUT_TOKENS as u32);\n    }\n\n    #[test]\n    fn chapter_synthesis_request_uses_batch_findings_instead_of_raw_evidence() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batch_drafts = vec![BatchDraft {\n            batch_id: 1,\n            findings_summary_zh: \"该批证据说明 README 中的 ReAct 主循环适合作为入门样例。\"\n                .to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该项目展示了 ReAct 主循环和工具注册。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_chapter_synthesis_request(\n            &chapter,\n            &evidence,\n            &batch_drafts,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"batch_drafts\"));\n        assert!(user_message.contains(\"allowed_evidence_ids\"));\n        assert!(user_message.contains(\"ReAct 主循环适合作为入门样例\"));\n        assert!(!user_message.contains(\"evidence_snippet\"));\n        assert!(!user_message.contains(\"README explains ReAct loop\"));\n    }\n\n    #[test]\n    fn parses_llm_global_summary_json() {\n        let content = r#\"{\n            \"global_summary_zh\": \"本报告显示，可控 TTS 入门调研应同时关注基础项目、说话人控制和模型框架边界；现有证据主要来自 GitHub README 与 arXiv 摘要，因此结论应限定在这些来源范围内。\"\n        }\"#;\n\n        let summary =\n            parse_global_summary_content(content).expect(\"valid global summary should parse\");\n\n        assert!(summary.contains(\"可控 TTS\"));\n        assert!(summary.contains(\"来源范围\"));\n    }\n\n    #[test]\n    fn rejects_llm_global_summary_with_template_residue() {\n        let content = r#\"{\n            \"global_summary_zh\": \"harinandanmv/text-to-speech 是本章节的关键来源之一。\"\n        }\"#;\n\n        let err =\n            parse_global_summary_content(content).expect_err(\"legacy template marker should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn global_summary_request_uses_chapter_drafts() {\n        let chapter_drafts = vec![ChapterDraft {\n            chapter_id: \"ch-1\".to_string(),\n            title_zh: \"可控 TTS 基础模型\".to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该章节围绕基础 TTS 模型、说话人控制和入门项目进行综合分析。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_global_summary_request(\n            \"TTS 入门级项目简单调研\",\n            &chapter_drafts,\n            1,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"TTS 入门级项目简单调研\"));\n        assert!(user_message.contains(\"可控 TTS 基础模型\"));\n        assert!(user_message.contains(\"ev-C1\"));\n        assert!(user_message.contains(\"paragraph_summaries\"));\n    }\n\n    fn sample_chapter() -> ChapterNode {\n        ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }\n    }\n\n    fn sample_evidence(evidence_id: &str, chapter_id: &str) -> EvidenceItem {\n        sample_evidence_with_kind(evidence_id, chapter_id, SourceKind::GitHub)\n    }\n\n    fn sample_evidence_with_kind(\n        evidence_id: &str,\n        chapter_id: &str,\n        source_kind: SourceKind,\n    ) -> EvidenceItem {\n        EvidenceItem {\n            evidence_id: evidence_id.to_string(),\n            source_item_id: format!(\"{source_kind:?}:acme/rust-agent\"),\n            citation_id: \"C1\".to_string(),\n            chapter_ids: vec![chapter_id.to_string()],\n            query_attempt_ids: vec![\"gh-1\".to_string()],\n            source_kind,\n            title: \"acme/rust-agent\".to_string(),\n            url: \"https://github.com/acme/rust-agent\".to_string(),\n            evidence_note_zh: \"GitHub 仓库 `acme/rust-agent`：Rust agent framework\".to_string(),\n            evidence_snippet: \"README explains ReAct loop, tool registry, and examples.\"\n                .to_string(),\n            support_score: None,\n        }\n    }\n\n    fn sample_deepseek_config(max_tokens: usize) -> DeepSeekConfig {\n        DeepSeekConfig {\n            api_key: \"sk-test\".to_string(),\n            base_url: \"https://api.deepseek.com\".to_string(),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: \"deepseek-v4-flash\".to_string(),\n            max_tokens,\n            timeout_secs: 120,\n        }\n    }\n\n    fn _dt() -> DateTime<Utc> {\n        DateTime::parse_from_rfc3339(\"2026-05-30T12:00:00Z\")\n            .unwrap()\n            .with_timezone(&Utc)\n    }\n}\n"}, "b8f76fdd9fdfaaf0957a9293477eb193d392c98f": {"t": 1785750659.2451828, "n": 50022, "h": "1c25fd4fdd794826", "c": "use crate::config::LlmConfig;\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{\n    strip_json_fence, ChatCompletionsRequest, ChatMessage, DeepSeekClient, DeepSeekConfig,\n    ResponseFormat,\n};\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft, SourceKind,\n};\nuse chrono::Utc;\nuse serde::{Deserialize, Serialize};\nuse std::collections::BTreeSet;\nuse tracing::{debug, warn};\n\nconst MAX_EVIDENCE_PER_BATCH: usize = 3;\nconst MAX_BATCH_CONTEXT_CHARS: usize = 3_500;\nconst MAX_BATCH_PARAGRAPHS: usize = 1;\nconst MAX_CHAPTER_PARAGRAPHS: usize = 4;\nconst MAX_WRITER_EVIDENCE_PER_CHAPTER: usize = 8;\nconst WRITER_GITHUB_EVIDENCE_CAP: usize = 4;\nconst WRITER_ARXIV_EVIDENCE_CAP: usize = 4;\nconst WRITER_ACADEMIC_INDEX_EVIDENCE_CAP: usize = 3;\nconst WRITER_BIBLIOGRAPHY_EVIDENCE_CAP: usize = 1;\nconst BATCH_FINDINGS_TARGET_CHARS: usize = 180;\nconst BATCH_PARAGRAPH_TARGET_CHARS: usize = 220;\nconst CHAPTER_PARAGRAPH_TARGET_CHARS: usize = 260;\nconst GLOBAL_SUMMARY_TARGET_CHARS: usize = 420;\nconst WRITER_TITLE_MAX_CHARS: usize = 180;\nconst WRITER_EVIDENCE_NOTE_MAX_CHARS: usize = 220;\nconst WRITER_GITHUB_SNIPPET_MAX_CHARS: usize = 900;\nconst WRITER_ACADEMIC_SNIPPET_MAX_CHARS: usize = 700;\nconst WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS: usize = 200;\nconst REPAIR_PREVIOUS_OUTPUT_MAX_CHARS: usize = 2_400;\nconst WRITER_MIN_OUTPUT_TOKENS: usize = 4096;\nconst WRITER_MAX_OUTPUT_TOKENS: usize = 8192;\n\npub async fn draft_report_with_llm(\n    topic: &str,\n    chapters: &[ChapterNode],\n    memory: &EvidenceMemory,\n    llm_config: &LlmConfig,\n) -> Result<ReportDraft> {\n    let deepseek_config = DeepSeekConfig::from_llm_config(llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"Writer LLM requires an enabled DeepSeek config with API key; refusing to generate a template report.\"\n                .to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(deepseek_config.clone())?;\n    let mut chapter_drafts = Vec::with_capacity(chapters.len());\n\n    for chapter in chapters {\n        let evidence_items = memory.by_chapter(&chapter.id);\n        if evidence_items.is_empty() {\n            chapter_drafts.push(empty_chapter_draft(chapter));\n            continue;\n        }\n\n        let draft = draft_chapter_with_llm(chapter, &evidence_items, &client, &deepseek_config)\n            .await\n            .map_err(|err| {\n                AppError::Llm(format!(\n                    \"Writer LLM failed for chapter `{}`; refusing template fallback: {err}\",\n                    chapter.title_zh\n                ))\n            })?;\n        chapter_drafts.push(draft);\n    }\n\n    let global_summary_zh = synthesize_global_summary_with_llm(\n        topic,\n        &chapter_drafts,\n        memory.items.len(),\n        &client,\n        &deepseek_config,\n    )\n    .await\n    .map_err(|err| {\n        AppError::Llm(format!(\n            \"Writer global summary failed; refusing deterministic fallback: {err}\"\n        ))\n    })?;\n\n    Ok(ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    })\n}\n\nasync fn draft_chapter_with_llm(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let selected_evidence = select_evidence_for_writing(evidence_items);\n    let batches = batch_chapter_evidence(&selected_evidence);\n    debug!(\n        \"Writer selected {}/{} evidence items and split chapter `{}` into {} evidence batches\",\n        selected_evidence.len(),\n        evidence_items.len(),\n        chapter.title_zh,\n        batches.len()\n    );\n\n    let mut batch_drafts = Vec::with_capacity(batches.len());\n    for batch in &batches {\n        let draft = draft_batch_with_llm(chapter, batch, batches.len(), client, config).await?;\n        batch_drafts.push(draft);\n    }\n\n    synthesize_chapter_from_batches(chapter, &selected_evidence, &batch_drafts, client, config)\n        .await\n}\n\nfn empty_chapter_draft(chapter: &ChapterNode) -> ChapterDraft {\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs: vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未获得与章节计划直接匹配的 GitHub/arXiv 证据，因此不生成分析段落。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }],\n    }\n}\n\n#[derive(Debug, Clone)]\nstruct EvidenceBatch {\n    batch_id: usize,\n    evidence_items: Vec<EvidenceItem>,\n    context_chars: usize,\n}\n\n#[derive(Debug, Clone, Serialize)]\nstruct BatchDraft {\n    batch_id: usize,\n    findings_summary_zh: String,\n    paragraphs: Vec<ParagraphWithCitations>,\n}\n\nfn batch_chapter_evidence(evidence_items: &[EvidenceItem]) -> Vec<EvidenceBatch> {\n    let mut batches = Vec::new();\n    let mut current_items = Vec::new();\n    let mut current_chars = 0usize;\n\n    for item in evidence_items {\n        let item_chars = evidence_context_chars(item);\n        let count_limit_reached = current_items.len() >= MAX_EVIDENCE_PER_BATCH;\n        let char_limit_reached =\n            !current_items.is_empty() && current_chars + item_chars > MAX_BATCH_CONTEXT_CHARS;\n\n        if count_limit_reached || char_limit_reached {\n            batches.push(EvidenceBatch {\n                batch_id: batches.len() + 1,\n                evidence_items: current_items,\n                context_chars: current_chars,\n            });\n            current_items = Vec::new();\n            current_chars = 0;\n        }\n\n        current_chars += item_chars;\n        current_items.push(item.clone());\n    }\n\n    if !current_items.is_empty() {\n        batches.push(EvidenceBatch {\n            batch_id: batches.len() + 1,\n            evidence_items: current_items,\n            context_chars: current_chars,\n        });\n    }\n\n    batches\n}\n\nfn select_evidence_for_writing(evidence_items: &[EvidenceItem]) -> Vec<EvidenceItem> {\n    let mut selected = Vec::new();\n    let has_stronger_sources = evidence_items\n        .iter()\n        .any(|item| item.source_kind != SourceKind::Bibliography);\n\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::Arxiv,\n        WRITER_ARXIV_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::AcademicIndex,\n        WRITER_ACADEMIC_INDEX_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::GitHub,\n        WRITER_GITHUB_EVIDENCE_CAP,\n    );\n    if !has_stronger_sources {\n        append_evidence_by_kind(\n            evidence_items,\n            &mut selected,\n            SourceKind::Bibliography,\n            WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        );\n    }\n\n    selected\n}\n\nfn append_evidence_by_kind(\n    evidence_items: &[EvidenceItem],\n    selected: &mut Vec<EvidenceItem>,\n    source_kind: SourceKind,\n    source_cap: usize,\n) {\n    let mut source_count = 0usize;\n    for item in evidence_items\n        .iter()\n        .filter(|item| item.source_kind == source_kind)\n    {\n        if selected.len() >= MAX_WRITER_EVIDENCE_PER_CHAPTER || source_count >= source_cap {\n            break;\n        }\n        selected.push(item.clone());\n        source_count += 1;\n    }\n}\n\nfn evidence_context_chars(item: &EvidenceItem) -> usize {\n    let input = EvidenceSourceInput::from(item);\n    input.evidence_id.chars().count()\n        + input.title.chars().count()\n        + input.url.chars().count()\n        + input.evidence_note_zh.chars().count()\n        + input.evidence_snippet.chars().count()\n}\n\nfn excerpt_chars(text: &str, max_chars: usize) -> String {\n    let trimmed = text.trim();\n    if trimmed.chars().count() <= max_chars {\n        return trimmed.to_string();\n    }\n\n    let mut out = trimmed.chars().take(max_chars).collect::<String>();\n    out.push_str(\"...\");\n    out\n}\n\nfn parse_batch_draft_content(content: &str, batch: &EvidenceBatch) -> Result<BatchDraft> {\n    let output =\n        serde_json::from_str::<BatchDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid BatchDraft JSON: {err}\"\n            ))\n        })?;\n    let findings_summary_zh = output.findings_summary_zh.trim().to_string();\n    if findings_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer batch response contained empty findings_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&findings_summary_zh)?;\n    let paragraphs = parse_paragraph_outputs(\n        output.paragraphs,\n        &batch.evidence_items,\n        MAX_BATCH_PARAGRAPHS,\n    )?;\n\n    Ok(BatchDraft {\n        batch_id: batch.batch_id,\n        findings_summary_zh,\n        paragraphs,\n    })\n}\n\nasync fn draft_batch_with_llm(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<BatchDraft> {\n    let request = build_batch_draft_request(chapter, batch, total_batches, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_batch_draft_content(content, batch) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM JSON validation failed for chapter `{}` batch {}; requesting one repair: {first_err}\",\n                chapter.title_zh, batch.batch_id\n            );\n            let repair_request = build_batch_draft_repair_request(\n                chapter,\n                batch,\n                total_batches,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_batch_draft_content(repaired_content, batch)\n        }\n    }\n}\n\nfn build_batch_draft_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请为下面章节的第 {}/{} 个证据批次生成局部分析草稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.3),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_batch_draft_repair_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次第 {}/{} 个证据批次的输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 evidence：\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nasync fn synthesize_chapter_from_batches(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let request = build_chapter_synthesis_request(chapter, evidence_items, batch_drafts, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_chapter_draft_content(content, chapter, evidence_items) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM chapter synthesis validation failed for chapter `{}`; requesting one repair: {first_err}\",\n                chapter.title_zh\n            );\n            let repair_request = build_chapter_synthesis_repair_request(\n                chapter,\n                evidence_items,\n                batch_drafts,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_chapter_draft_content(repaired_content, chapter, evidence_items)\n        }\n    }\n}\n\nfn batch_writer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的技术报告撰稿人，正在为 `{chapter_title}` 章节处理一个 evidence batch。\\n\\\n规则：\\n\\\n1. 使用中文撰写。\\n\\\n2. findings_summary_zh 不超过 {BATCH_FINDINGS_TARGET_CHARS} 个汉字，只总结本批证据的关键发现、共性和边界。\\n\\\n3. paragraphs 输出 1-{MAX_BATCH_PARAGRAPHS} 段，每段不超过 {BATCH_PARAGRAPH_TARGET_CHARS} 个汉字；提炼项目/论文简介、核心亮点或设计思想，以及与章节问题的具体关联。\\n\\\n4. 可合并多个来源做对比分析，但每个段落必须引用至少一个 evidence_id。\\n\\\n5. 只能使用用户提供的 evidence_id 和 URL，不得新增来源、不得编造事实。\\n\\\n6. 不要只重复原文，不要输出“来源链接”模板句。\\n\\\n7. 只返回 JSON：{{\\\"findings_summary_zh\\\":\\\"...\\\",\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nfn writer_max_tokens(config: &DeepSeekConfig) -> u32 {\n    config\n        .max_tokens\n        .max(WRITER_MIN_OUTPUT_TOKENS)\n        .min(WRITER_MAX_OUTPUT_TOKENS)\n        .min(u32::MAX as usize) as u32\n}\n\nfn build_chapter_synthesis_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于已验证的 batch findings 合成章节终稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_chapter_synthesis_repair_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次章节综合输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 batch findings：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn chapter_synthesizer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的章节综合编辑，正在合成 `{chapter_title}` 章节终稿。\\n\\\n规则：\\n\\\n1. 使用中文撰写自然、分析性的技术报告段落。\\n\\\n2. 只能使用用户提供的 batch findings、batch paragraphs 和 allowed_evidence_ids。\\n\\\n3. 合并重复信息，按主题组织，不要机械罗列每条来源。\\n\\\n4. 已有 evidence 时，不得用“尚未收集到足够证据”替代分析内容。\\n\\\n5. 每个段落必须引用至少一个 allowed_evidence_ids 中的 evidence_id。\\n\\\n6. 不得新增来源、URL 或 evidence_id，不得编造事实。\\n\\\n7. 输出 1-{MAX_CHAPTER_PARAGRAPHS} 段，每段不超过 {CHAPTER_PARAGRAPH_TARGET_CHARS} 个汉字。\\n\\\n8. 不要输出“来源链接”模板句或“是本章节的关键来源之一”。\\n\\\n9. 只返回 JSON：{{\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nasync fn synthesize_global_summary_with_llm(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<String> {\n    let request = build_global_summary_request(topic, chapter_drafts, evidence_count, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_global_summary_content(content) {\n        Ok(summary) => Ok(summary),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM global summary validation failed; requesting one repair: {first_err}\"\n            );\n            let repair_request = build_global_summary_repair_request(\n                topic,\n                chapter_drafts,\n                evidence_count,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_global_summary_content(repaired_content)\n        }\n    }\n}\n\nfn build_global_summary_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于章节终稿生成报告的全局摘要。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_global_summary_repair_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次全局摘要输出，使其满足 JSON schema 和写作约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n章节终稿上下文：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn parse_global_summary_content(content: &str) -> Result<String> {\n    let output =\n        serde_json::from_str::<GlobalSummaryOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid GlobalSummary JSON: {err}\"\n            ))\n        })?;\n    let global_summary_zh = output.global_summary_zh.trim().to_string();\n    if global_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer global summary response contained empty global_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&global_summary_zh)?;\n\n    Ok(global_summary_zh)\n}\n\nfn global_synthesizer_system_prompt() -> String {\n    format!(\n        \"你是 LitScout-RS 的报告总编辑，负责基于章节终稿生成全局摘要。\\n\\\n规则：\\n\\\n1. 使用中文撰写 1-2 段自然的全局摘要，总长度不超过 {GLOBAL_SUMMARY_TARGET_CHARS} 个汉字。\\n\\\n2. 概括报告覆盖范围、主要发现、证据边界和不足。\\n\\\n3. 只能使用用户提供的章节终稿和 citation id，不得新增来源、URL 或事实。\\n\\\n4. 不要输出固定模板句，不要写“来源链接”，不要写“是本章节的关键来源之一”。\\n\\\n5. 不要把引用账本逐条改写成摘要。\\n\\\n6. 只返回 JSON：{{\\\"global_summary_zh\\\":\\\"...\\\"}}。\"\n    )\n}\n\nfn parse_chapter_draft_content(\n    content: &str,\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n) -> Result<ChapterDraft> {\n    let output =\n        serde_json::from_str::<ChapterDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid ChapterDraft JSON: {err}\"\n            ))\n        })?;\n    let paragraphs =\n        parse_paragraph_outputs(output.paragraphs, evidence_items, MAX_CHAPTER_PARAGRAPHS)?;\n\n    Ok(ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    })\n}\n\nfn parse_paragraph_outputs(\n    paragraph_outputs: Vec<ChapterParagraphOutput>,\n    evidence_items: &[EvidenceItem],\n    max_paragraphs: usize,\n) -> Result<Vec<ParagraphWithCitations>> {\n    let allowed_ids = evidence_items\n        .iter()\n        .map(|item| item.evidence_id.as_str())\n        .collect::<BTreeSet<_>>();\n    let mut paragraphs = Vec::new();\n\n    for paragraph in paragraph_outputs.into_iter().take(max_paragraphs) {\n        let text_zh = paragraph.text_zh.trim().to_string();\n        if text_zh.is_empty() {\n            continue;\n        }\n        reject_template_residue(&text_zh)?;\n        let cited_evidence_ids = paragraph\n            .cited_evidence_ids\n            .into_iter()\n            .filter_map(|id| {\n                let trimmed = id.trim();\n                (!trimmed.is_empty()).then(|| trimmed.to_string())\n            })\n            .collect::<Vec<_>>();\n        if cited_evidence_ids.is_empty() {\n            return Err(AppError::Llm(\n                \"Writer paragraph did not cite any evidence ID\".to_string(),\n            ));\n        }\n        for id in &cited_evidence_ids {\n            if !allowed_ids.contains(id.as_str()) {\n                return Err(AppError::Llm(format!(\n                    \"Writer paragraph referenced unknown evidence id `{id}`\"\n                )));\n            }\n        }\n        paragraphs.push(ParagraphWithCitations {\n            text_zh,\n            cited_evidence_ids,\n        });\n    }\n\n    if paragraphs.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer response contained no usable paragraphs\".to_string(),\n        ));\n    }\n\n    Ok(paragraphs)\n}\n\nfn reject_template_residue(text: &str) -> Result<()> {\n    for marker in [\n        \"是本章节的关键来源之一\",\n        \"来源链接：\",\n        \"提供了与主题相关的证据\",\n    ] {\n        if text.contains(marker) {\n            return Err(AppError::Llm(format!(\n                \"Writer output still contains legacy template marker `{marker}`\"\n            )));\n        }\n    }\n    Ok(())\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push(' ');\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[derive(Debug, Serialize)]\nstruct BatchDraftInput {\n    chapter_title: String,\n    research_question: String,\n    batch_id: usize,\n    total_batches: usize,\n    batch_context_chars: usize,\n    evidence_sources: Vec<EvidenceSourceInput>,\n}\n\n#[derive(Debug, Serialize)]\nstruct ChapterSynthesisInput {\n    chapter_title: String,\n    research_question: String,\n    allowed_evidence_ids: Vec<String>,\n    batch_drafts: Vec<BatchDraft>,\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryInput {\n    topic: String,\n    chapter_count: usize,\n    evidence_count: usize,\n    chapters: Vec<GlobalSummaryChapterInput>,\n}\n\nimpl GlobalSummaryInput {\n    fn from_report(topic: &str, chapter_drafts: &[ChapterDraft], evidence_count: usize) -> Self {\n        Self {\n            topic: topic.to_string(),\n            chapter_count: chapter_drafts.len(),\n            evidence_count,\n            chapters: chapter_drafts\n                .iter()\n                .map(GlobalSummaryChapterInput::from)\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryChapterInput {\n    chapter_id: String,\n    title_zh: String,\n    paragraph_count: usize,\n    cited_evidence_ids: Vec<String>,\n    paragraph_summaries: Vec<String>,\n}\n\nimpl From<&ChapterDraft> for GlobalSummaryChapterInput {\n    fn from(chapter: &ChapterDraft) -> Self {\n        let cited_evidence_ids = chapter\n            .paragraphs\n            .iter()\n            .flat_map(|paragraph| paragraph.cited_evidence_ids.iter().cloned())\n            .collect::<BTreeSet<_>>()\n            .into_iter()\n            .collect();\n        Self {\n            chapter_id: chapter.chapter_id.clone(),\n            title_zh: chapter.title_zh.clone(),\n            paragraph_count: chapter.paragraphs.len(),\n            cited_evidence_ids,\n            paragraph_summaries: chapter\n                .paragraphs\n                .iter()\n                .map(|paragraph| excerpt_chars(&paragraph.text_zh, 500))\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct EvidenceSourceInput {\n    evidence_id: String,\n    title: String,\n    url: String,\n    source_kind: SourceKind,\n    evidence_note_zh: String,\n    evidence_snippet: String,\n}\n\nimpl From<&EvidenceItem> for EvidenceSourceInput {\n    fn from(item: &EvidenceItem) -> Self {\n        Self {\n            evidence_id: item.evidence_id.clone(),\n            title: excerpt_chars(&item.title, WRITER_TITLE_MAX_CHARS),\n            url: item.url.clone(),\n            source_kind: item.source_kind,\n            evidence_note_zh: excerpt_chars(&item.evidence_note_zh, WRITER_EVIDENCE_NOTE_MAX_CHARS),\n            evidence_snippet: excerpt_chars(\n                &item.evidence_snippet,\n                writer_snippet_max_chars(item.source_kind),\n            ),\n        }\n    }\n}\n\nfn writer_snippet_max_chars(source_kind: SourceKind) -> usize {\n    match source_kind {\n        SourceKind::GitHub => WRITER_GITHUB_SNIPPET_MAX_CHARS,\n        SourceKind::Arxiv | SourceKind::AcademicIndex => WRITER_ACADEMIC_SNIPPET_MAX_CHARS,\n        SourceKind::Bibliography => WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS,\n    }\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterDraftOutput {\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct BatchDraftOutput {\n    #[serde(default)]\n    findings_summary_zh: String,\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct GlobalSummaryOutput {\n    #[serde(default)]\n    global_summary_zh: String,\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterParagraphOutput {\n    text_zh: String,\n    #[serde(default)]\n    cited_evidence_ids: Vec<String>,\n}\n\n#[cfg(test)]\nmod tests {\n    use chrono::{DateTime, Utc};\n\n    use crate::llm::deepseek::DeepSeekConfig;\n    use crate::model::{\n        ChapterDraft, ChapterNode, EvidenceItem, EvidenceMemory, ParagraphWithCitations, SourceKind,\n    };\n\n    use super::{\n        batch_chapter_evidence, build_batch_draft_request, build_chapter_synthesis_request,\n        build_global_summary_request, draft_report_with_llm, empty_chapter_draft,\n        parse_batch_draft_content, parse_chapter_draft_content, parse_global_summary_content,\n        select_evidence_for_writing, writer_max_tokens, BatchDraft, MAX_BATCH_CONTEXT_CHARS,\n        MAX_EVIDENCE_PER_BATCH, MAX_WRITER_EVIDENCE_PER_CHAPTER, WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        WRITER_MIN_OUTPUT_TOKENS,\n    };\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = empty_chapter_draft(&chapters[0]);\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert!(draft.paragraphs[0]\n            .text_zh\n            .contains(\"尚未获得与章节计划直接匹配\"));\n    }\n\n    #[tokio::test]\n    async fn llm_writer_refuses_to_template_when_llm_is_disabled() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n            selection_report: Default::default(),\n        };\n\n        let err = draft_report_with_llm(\n            \"Rust Agent\",\n            &chapters,\n            &memory,\n            &crate::config::LlmConfig::from_env(false, 30),\n        )\n        .await\n        .expect_err(\"disabled LLM should not produce a template report\");\n\n        assert!(err\n            .to_string()\n            .contains(\"refusing to generate a template report\"));\n    }\n\n    #[tokio::test]\n    #[ignore = \"requires a real DeepSeek API key in DEEPSEEK_API_KEY\"]\n    async fn writer_real_deepseek_smoke() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n            selection_report: Default::default(),\n        };\n        let config = crate::config::LlmConfig::from_env(true, 120);\n\n        assert!(\n            config.api_key.is_some(),\n            \"DEEPSEEK_API_KEY must be set for this ignored smoke test\"\n        );\n\n        let draft = draft_report_with_llm(\"TTS 入门级项目简单调研\", &chapters, &memory, &config)\n            .await\n            .expect(\"real DeepSeek writer smoke should complete\");\n\n        assert!(!draft.global_summary_zh.trim().is_empty());\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(!draft.chapters[0].paragraphs.is_empty());\n        assert!(draft.chapters[0].paragraphs[0]\n            .cited_evidence_ids\n            .contains(&\"ev-C1\".to_string()));\n        println!(\n            \"writer_real_deepseek_smoke ok: global_summary_chars={}, chapter_paragraphs={}\",\n            draft.global_summary_zh.chars().count(),\n            draft.chapters[0].paragraphs.len()\n        );\n    }\n\n    #[test]\n    fn parses_llm_chapter_draft_json() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了从 README 提炼出的 ReAct 主循环和工具调用结构，适合作为入门实现分析对象。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect(\"valid chapter JSON should parse\");\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n        assert!(draft.paragraphs[0].text_zh.contains(\"ReAct\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_unknown_evidence_id() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"引用了不存在的证据。\",\n                    \"cited_evidence_ids\": [\"ev-C9\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_template_residue() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"acme/rust-agent 是本章节的关键来源之一。来源链接：https://github.com/acme/rust-agent\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batches_small_chapter_evidence_into_one_batch() {\n        let evidence = vec![\n            sample_evidence(\"ev-C1\", \"ch-1\"),\n            sample_evidence(\"ev-C2\", \"ch-1\"),\n            sample_evidence(\"ev-C3\", \"ch-1\"),\n        ];\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 1);\n        assert_eq!(batches[0].batch_id, 1);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert!(batches[0].context_chars > 0);\n    }\n\n    #[test]\n    fn selects_bounded_mixed_evidence_for_writing() {\n        let mut evidence = Vec::new();\n        evidence.extend((1..=5).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-A{index}\"), \"ch-1\", SourceKind::Arxiv)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-S{index}\"), \"ch-1\", SourceKind::AcademicIndex)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-G{index}\"), \"ch-1\", SourceKind::GitHub)\n        }));\n        evidence.extend((1..=2).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n        }));\n\n        let selected = select_evidence_for_writing(&evidence);\n        let bibliography_count = selected\n            .iter()\n            .filter(|item| item.source_kind == SourceKind::Bibliography)\n            .count();\n\n        assert_eq!(selected.len(), MAX_WRITER_EVIDENCE_PER_CHAPTER);\n        assert_eq!(bibliography_count, 0);\n        assert_eq!(selected[0].source_kind, SourceKind::Arxiv);\n        assert_eq!(selected[4].source_kind, SourceKind::AcademicIndex);\n        assert_eq!(selected[7].source_kind, SourceKind::GitHub);\n    }\n\n    #[test]\n    fn selects_bibliography_only_when_no_stronger_sources_exist() {\n        let evidence = (1..=3)\n            .map(|index| {\n                sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n            })\n            .collect::<Vec<_>>();\n\n        let selected = select_evidence_for_writing(&evidence);\n\n        assert_eq!(selected.len(), WRITER_BIBLIOGRAPHY_EVIDENCE_CAP);\n        assert_eq!(selected[0].source_kind, SourceKind::Bibliography);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_count() {\n        let evidence = (1..=9)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 3);\n        assert_eq!(batches[0].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[1].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].batch_id, 3);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_sanitized_context_chars() {\n        let mut first = sample_evidence(\"ev-C1\", \"ch-1\");\n        first.evidence_snippet = \"a\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut second = sample_evidence(\"ev-C2\", \"ch-1\");\n        second.evidence_snippet = \"b\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut third = sample_evidence(\"ev-C3\", \"ch-1\");\n        third.evidence_snippet = \"c\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut fourth = sample_evidence(\"ev-C4\", \"ch-1\");\n        fourth.evidence_snippet = \"d\".repeat(MAX_BATCH_CONTEXT_CHARS);\n\n        let batches = batch_chapter_evidence(&[first, second, third, fourth]);\n\n        assert_eq!(batches.len(), 2);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert_eq!(batches[1].evidence_items.len(), 1);\n        assert!(batches[0].context_chars <= MAX_BATCH_CONTEXT_CHARS);\n    }\n\n    #[test]\n    fn batch_request_truncates_long_github_snippets_for_writer_context() {\n        let chapter = sample_chapter();\n        let mut evidence = sample_evidence(\"ev-C1\", \"ch-1\");\n        evidence.evidence_snippet = format!(\"{}{}\", \"A\".repeat(1500), \"UNTRUNCATED_TAIL\");\n        let batches = batch_chapter_evidence(&[evidence]);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[0],\n            batches.len(),\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(&\"A\".repeat(900)));\n        assert!(!user_message.contains(\"UNTRUNCATED_TAIL\"));\n    }\n\n    #[test]\n    fn parses_llm_batch_draft_json() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据说明了项目 README 中的 ReAct 主循环、工具注册和入门示例。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目的 README 将 ReAct 主循环和工具注册拆成清晰的入门实现，适合用于说明基础 agent 架构。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft =\n            parse_batch_draft_content(content, &batches[0]).expect(\"valid batch JSON should parse\");\n\n        assert_eq!(draft.batch_id, 1);\n        assert!(draft.findings_summary_zh.contains(\"ReAct\"));\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_empty_findings() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"empty findings should fail\");\n\n        assert!(err.to_string().contains(\"empty findings_summary_zh\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_unknown_evidence_id() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据有一个段落引用了不存在的证据。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"这里引用了不存在的 evidence id。\",\n                    \"cited_evidence_ids\": [\"ev-C2\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_template_residue() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"acme/rust-agent 是本章节的关键来源之一。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batch_request_only_contains_current_batch_evidence() {\n        let chapter = sample_chapter();\n        let evidence = (1..=5)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n        let batches = batch_chapter_evidence(&evidence);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[1],\n            batches.len(),\n            &sample_deepseek_config(2048),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"\\\"batch_id\\\": 2\"));\n        assert!(user_message.contains(\"ev-C5\"));\n        assert!(!user_message.contains(\"ev-C1\"));\n        assert_eq!(request.max_tokens, Some(WRITER_MIN_OUTPUT_TOKENS as u32));\n    }\n\n    #[test]\n    fn writer_max_tokens_raises_too_small_config_for_reasoning_models() {\n        let config = sample_deepseek_config(32);\n\n        assert_eq!(writer_max_tokens(&config), WRITER_MIN_OUTPUT_TOKENS as u32);\n    }\n\n    #[test]\n    fn chapter_synthesis_request_uses_batch_findings_instead_of_raw_evidence() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batch_drafts = vec![BatchDraft {\n            batch_id: 1,\n            findings_summary_zh: \"该批证据说明 README 中的 ReAct 主循环适合作为入门样例。\"\n                .to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该项目展示了 ReAct 主循环和工具注册。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_chapter_synthesis_request(\n            &chapter,\n            &evidence,\n            &batch_drafts,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"batch_drafts\"));\n        assert!(user_message.contains(\"allowed_evidence_ids\"));\n        assert!(user_message.contains(\"ReAct 主循环适合作为入门样例\"));\n        assert!(!user_message.contains(\"evidence_snippet\"));\n        assert!(!user_message.contains(\"README explains ReAct loop\"));\n    }\n\n    #[test]\n    fn parses_llm_global_summary_json() {\n        let content = r#\"{\n            \"global_summary_zh\": \"本报告显示，可控 TTS 入门调研应同时关注基础项目、说话人控制和模型框架边界；现有证据主要来自 GitHub README 与 arXiv 摘要，因此结论应限定在这些来源范围内。\"\n        }\"#;\n\n        let summary =\n            parse_global_summary_content(content).expect(\"valid global summary should parse\");\n\n        assert!(summary.contains(\"可控 TTS\"));\n        assert!(summary.contains(\"来源范围\"));\n    }\n\n    #[test]\n    fn rejects_llm_global_summary_with_template_residue() {\n        let content = r#\"{\n            \"global_summary_zh\": \"harinandanmv/text-to-speech 是本章节的关键来源之一。\"\n        }\"#;\n\n        let err =\n            parse_global_summary_content(content).expect_err(\"legacy template marker should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn global_summary_request_uses_chapter_drafts() {\n        let chapter_drafts = vec![ChapterDraft {\n            chapter_id: \"ch-1\".to_string(),\n            title_zh: \"可控 TTS 基础模型\".to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该章节围绕基础 TTS 模型、说话人控制和入门项目进行综合分析。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_global_summary_request(\n            \"TTS 入门级项目简单调研\",\n            &chapter_drafts,\n            1,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"TTS 入门级项目简单调研\"));\n        assert!(user_message.contains(\"可控 TTS 基础模型\"));\n        assert!(user_message.contains(\"ev-C1\"));\n        assert!(user_message.contains(\"paragraph_summaries\"));\n    }\n\n    fn sample_chapter() -> ChapterNode {\n        ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }\n    }\n\n    fn sample_evidence(evidence_id: &str, chapter_id: &str) -> EvidenceItem {\n        sample_evidence_with_kind(evidence_id, chapter_id, SourceKind::GitHub)\n    }\n\n    fn sample_evidence_with_kind(\n        evidence_id: &str,\n        chapter_id: &str,\n        source_kind: SourceKind,\n    ) -> EvidenceItem {\n        EvidenceItem {\n            evidence_id: evidence_id.to_string(),\n            source_item_id: format!(\"{source_kind:?}:acme/rust-agent\"),\n            citation_id: \"C1\".to_string(),\n            chapter_ids: vec![chapter_id.to_string()],\n            query_attempt_ids: vec![\"gh-1\".to_string()],\n            source_kind,\n            title: \"acme/rust-agent\".to_string(),\n            url: \"https://github.com/acme/rust-agent\".to_string(),\n            evidence_note_zh: \"GitHub 仓库 `acme/rust-agent`：Rust agent framework\".to_string(),\n            evidence_snippet: \"README explains ReAct loop, tool registry, and examples.\"\n                .to_string(),\n            support_score: None,\n        }\n    }\n\n    fn sample_deepseek_config(max_tokens: usize) -> DeepSeekConfig {\n        DeepSeekConfig {\n            api_key: \"sk-test\".to_string(),\n            base_url: \"https://api.deepseek.com\".to_string(),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: \"deepseek-v4-flash\".to_string(),\n            max_tokens,\n            timeout_secs: 120,\n        }\n    }\n\n    fn _dt() -> DateTime<Utc> {\n        DateTime::parse_from_rfc3339(\"2026-05-30T12:00:00Z\")\n            .unwrap()\n            .with_timezone(&Utc)\n    }\n}\n"}, "f0e381df64f2448f45cfc7566949703029fd6216": {"t": 1785750660.7074842, "n": 49922, "h": "76c4cc62091cd5a3", "c": "use crate::config::LlmConfig;\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{\n    strip_json_fence, ChatCompletionsRequest, ChatMessage, DeepSeekClient, DeepSeekConfig,\n    ResponseFormat,\n};\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft, SourceKind,\n};\nuse chrono::Utc;\nuse serde::{Deserialize, Serialize};\nuse std::collections::BTreeSet;\nuse tracing::{debug, warn};\n\nconst MAX_EVIDENCE_PER_BATCH: usize = 3;\nconst MAX_BATCH_CONTEXT_CHARS: usize = 3_500;\nconst MAX_BATCH_PARAGRAPHS: usize = 1;\nconst MAX_CHAPTER_PARAGRAPHS: usize = 4;\nconst MAX_WRITER_EVIDENCE_PER_CHAPTER: usize = 8;\nconst WRITER_GITHUB_EVIDENCE_CAP: usize = 4;\nconst WRITER_ARXIV_EVIDENCE_CAP: usize = 4;\nconst WRITER_ACADEMIC_INDEX_EVIDENCE_CAP: usize = 3;\nconst WRITER_BIBLIOGRAPHY_EVIDENCE_CAP: usize = 1;\nconst BATCH_FINDINGS_TARGET_CHARS: usize = 180;\nconst BATCH_PARAGRAPH_TARGET_CHARS: usize = 220;\nconst CHAPTER_PARAGRAPH_TARGET_CHARS: usize = 260;\nconst GLOBAL_SUMMARY_TARGET_CHARS: usize = 420;\nconst WRITER_TITLE_MAX_CHARS: usize = 180;\nconst WRITER_EVIDENCE_NOTE_MAX_CHARS: usize = 220;\nconst WRITER_GITHUB_SNIPPET_MAX_CHARS: usize = 900;\nconst WRITER_ACADEMIC_SNIPPET_MAX_CHARS: usize = 700;\nconst WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS: usize = 200;\nconst REPAIR_PREVIOUS_OUTPUT_MAX_CHARS: usize = 2_400;\nconst WRITER_MIN_OUTPUT_TOKENS: usize = 4096;\nconst WRITER_MAX_OUTPUT_TOKENS: usize = 8192;\n\npub async fn draft_report_with_llm(\n    topic: &str,\n    chapters: &[ChapterNode],\n    memory: &EvidenceMemory,\n    llm_config: &LlmConfig,\n) -> Result<ReportDraft> {\n    let deepseek_config = DeepSeekConfig::from_llm_config(llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"Writer LLM requires an enabled DeepSeek config with API key; refusing to generate a template report.\"\n                .to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(deepseek_config.clone())?;\n    let mut chapter_drafts = Vec::with_capacity(chapters.len());\n\n    for chapter in chapters {\n        let evidence_items = memory.by_chapter(&chapter.id);\n        if evidence_items.is_empty() {\n            chapter_drafts.push(empty_chapter_draft(chapter));\n            continue;\n        }\n\n        let draft = draft_chapter_with_llm(chapter, &evidence_items, &client, &deepseek_config)\n            .await\n            .map_err(|err| {\n                AppError::Llm(format!(\n                    \"Writer LLM failed for chapter `{}`; refusing template fallback: {err}\",\n                    chapter.title_zh\n                ))\n            })?;\n        chapter_drafts.push(draft);\n    }\n\n    let global_summary_zh = synthesize_global_summary_with_llm(\n        topic,\n        &chapter_drafts,\n        memory.items.len(),\n        &client,\n        &deepseek_config,\n    )\n    .await\n    .map_err(|err| {\n        AppError::Llm(format!(\n            \"Writer global summary failed; refusing deterministic fallback: {err}\"\n        ))\n    })?;\n\n    Ok(ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    })\n}\n\nasync fn draft_chapter_with_llm(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let selected_evidence = select_evidence_for_writing(evidence_items);\n    let batches = batch_chapter_evidence(&selected_evidence);\n    debug!(\n        \"Writer selected {}/{} evidence items and split chapter `{}` into {} evidence batches\",\n        selected_evidence.len(),\n        evidence_items.len(),\n        chapter.title_zh,\n        batches.len()\n    );\n\n    let mut batch_drafts = Vec::with_capacity(batches.len());\n    for batch in &batches {\n        let draft = draft_batch_with_llm(chapter, batch, batches.len(), client, config).await?;\n        batch_drafts.push(draft);\n    }\n\n    synthesize_chapter_from_batches(chapter, &selected_evidence, &batch_drafts, client, config)\n        .await\n}\n\nfn empty_chapter_draft(chapter: &ChapterNode) -> ChapterDraft {\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs: vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未获得与章节计划直接匹配的 GitHub/arXiv 证据，因此不生成分析段落。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }],\n    }\n}\n\n#[derive(Debug, Clone)]\nstruct EvidenceBatch {\n    batch_id: usize,\n    evidence_items: Vec<EvidenceItem>,\n    context_chars: usize,\n}\n\n#[derive(Debug, Clone, Serialize)]\nstruct BatchDraft {\n    batch_id: usize,\n    findings_summary_zh: String,\n    paragraphs: Vec<ParagraphWithCitations>,\n}\n\nfn batch_chapter_evidence(evidence_items: &[EvidenceItem]) -> Vec<EvidenceBatch> {\n    let mut batches = Vec::new();\n    let mut current_items = Vec::new();\n    let mut current_chars = 0usize;\n\n    for item in evidence_items {\n        let item_chars = evidence_context_chars(item);\n        let count_limit_reached = current_items.len() >= MAX_EVIDENCE_PER_BATCH;\n        let char_limit_reached =\n            !current_items.is_empty() && current_chars + item_chars > MAX_BATCH_CONTEXT_CHARS;\n\n        if count_limit_reached || char_limit_reached {\n            batches.push(EvidenceBatch {\n                batch_id: batches.len() + 1,\n                evidence_items: current_items,\n                context_chars: current_chars,\n            });\n            current_items = Vec::new();\n            current_chars = 0;\n        }\n\n        current_chars += item_chars;\n        current_items.push(item.clone());\n    }\n\n    if !current_items.is_empty() {\n        batches.push(EvidenceBatch {\n            batch_id: batches.len() + 1,\n            evidence_items: current_items,\n            context_chars: current_chars,\n        });\n    }\n\n    batches\n}\n\nfn select_evidence_for_writing(evidence_items: &[EvidenceItem]) -> Vec<EvidenceItem> {\n    let mut selected = Vec::new();\n    let has_stronger_sources = evidence_items\n        .iter()\n        .any(|item| item.source_kind != SourceKind::Bibliography);\n\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::Arxiv,\n        WRITER_ARXIV_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::AcademicIndex,\n        WRITER_ACADEMIC_INDEX_EVIDENCE_CAP,\n    );\n    append_evidence_by_kind(\n        evidence_items,\n        &mut selected,\n        SourceKind::GitHub,\n        WRITER_GITHUB_EVIDENCE_CAP,\n    );\n    if !has_stronger_sources {\n        append_evidence_by_kind(\n            evidence_items,\n            &mut selected,\n            SourceKind::Bibliography,\n            WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        );\n    }\n\n    selected\n}\n\nfn append_evidence_by_kind(\n    evidence_items: &[EvidenceItem],\n    selected: &mut Vec<EvidenceItem>,\n    source_kind: SourceKind,\n    source_cap: usize,\n) {\n    let mut source_count = 0usize;\n    for item in evidence_items\n        .iter()\n        .filter(|item| item.source_kind == source_kind)\n    {\n        if selected.len() >= MAX_WRITER_EVIDENCE_PER_CHAPTER || source_count >= source_cap {\n            break;\n        }\n        selected.push(item.clone());\n        source_count += 1;\n    }\n}\n\nfn evidence_context_chars(item: &EvidenceItem) -> usize {\n    let input = EvidenceSourceInput::from(item);\n    input.evidence_id.chars().count()\n        + input.title.chars().count()\n        + input.url.chars().count()\n        + input.evidence_note_zh.chars().count()\n        + input.evidence_snippet.chars().count()\n}\n\nfn excerpt_chars(text: &str, max_chars: usize) -> String {\n    let trimmed = text.trim();\n    if trimmed.chars().count() <= max_chars {\n        return trimmed.to_string();\n    }\n\n    let mut out = trimmed.chars().take(max_chars).collect::<String>();\n    out.push_str(\"...\");\n    out\n}\n\nfn parse_batch_draft_content(content: &str, batch: &EvidenceBatch) -> Result<BatchDraft> {\n    let output =\n        serde_json::from_str::<BatchDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid BatchDraft JSON: {err}\"\n            ))\n        })?;\n    let findings_summary_zh = output.findings_summary_zh.trim().to_string();\n    if findings_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer batch response contained empty findings_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&findings_summary_zh)?;\n    let paragraphs = parse_paragraph_outputs(\n        output.paragraphs,\n        &batch.evidence_items,\n        MAX_BATCH_PARAGRAPHS,\n    )?;\n\n    Ok(BatchDraft {\n        batch_id: batch.batch_id,\n        findings_summary_zh,\n        paragraphs,\n    })\n}\n\nasync fn draft_batch_with_llm(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<BatchDraft> {\n    let request = build_batch_draft_request(chapter, batch, total_batches, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_batch_draft_content(content, batch) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM JSON validation failed for chapter `{}` batch {}; requesting one repair: {first_err}\",\n                chapter.title_zh, batch.batch_id\n            );\n            let repair_request = build_batch_draft_repair_request(\n                chapter,\n                batch,\n                total_batches,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_batch_draft_content(repaired_content, batch)\n        }\n    }\n}\n\nfn build_batch_draft_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请为下面章节的第 {}/{} 个证据批次生成局部分析草稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.3),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_batch_draft_repair_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次第 {}/{} 个证据批次的输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 evidence：\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nasync fn synthesize_chapter_from_batches(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let request = build_chapter_synthesis_request(chapter, evidence_items, batch_drafts, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_chapter_draft_content(content, chapter, evidence_items) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM chapter synthesis validation failed for chapter `{}`; requesting one repair: {first_err}\",\n                chapter.title_zh\n            );\n            let repair_request = build_chapter_synthesis_repair_request(\n                chapter,\n                evidence_items,\n                batch_drafts,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_chapter_draft_content(repaired_content, chapter, evidence_items)\n        }\n    }\n}\n\nfn batch_writer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的技术报告撰稿人，正在为 `{chapter_title}` 章节处理一个 evidence batch。\\n\\\n规则：\\n\\\n1. 使用中文撰写。\\n\\\n2. findings_summary_zh 不超过 {BATCH_FINDINGS_TARGET_CHARS} 个汉字，只总结本批证据的关键发现、共性和边界。\\n\\\n3. paragraphs 输出 1-{MAX_BATCH_PARAGRAPHS} 段，每段不超过 {BATCH_PARAGRAPH_TARGET_CHARS} 个汉字；提炼项目/论文简介、核心亮点或设计思想，以及与章节问题的具体关联。\\n\\\n4. 可合并多个来源做对比分析，但每个段落必须引用至少一个 evidence_id。\\n\\\n5. 只能使用用户提供的 evidence_id 和 URL，不得新增来源、不得编造事实。\\n\\\n6. 不要只重复原文，不要输出“来源链接”模板句。\\n\\\n7. 只返回 JSON：{{\\\"findings_summary_zh\\\":\\\"...\\\",\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nfn writer_max_tokens(config: &DeepSeekConfig) -> u32 {\n    config\n        .max_tokens\n        .max(WRITER_MIN_OUTPUT_TOKENS)\n        .min(WRITER_MAX_OUTPUT_TOKENS)\n        .min(u32::MAX as usize) as u32\n}\n\nfn build_chapter_synthesis_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于已验证的 batch findings 合成章节终稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_chapter_synthesis_repair_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次章节综合输出，使其满足 JSON schema、引用约束和长度约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n允许使用的 batch findings：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn chapter_synthesizer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的章节综合编辑，正在合成 `{chapter_title}` 章节终稿。\\n\\\n规则：\\n\\\n1. 使用中文撰写自然、分析性的技术报告段落。\\n\\\n2. 只能使用用户提供的 batch findings、batch paragraphs 和 allowed_evidence_ids。\\n\\\n3. 合并重复信息，按主题组织，不要机械罗列每条来源。\\n\\\n4. 已有 evidence 时，不得用“尚未收集到足够证据”替代分析内容。\\n\\\n5. 每个段落必须引用至少一个 allowed_evidence_ids 中的 evidence_id。\\n\\\n6. 不得新增来源、URL 或 evidence_id，不得编造事实。\\n\\\n7. 输出 1-{MAX_CHAPTER_PARAGRAPHS} 段，每段不超过 {CHAPTER_PARAGRAPH_TARGET_CHARS} 个汉字。\\n\\\n8. 不要输出“来源链接”模板句或“是本章节的关键来源之一”。\\n\\\n9. 只返回 JSON：{{\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nasync fn synthesize_global_summary_with_llm(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<String> {\n    let request = build_global_summary_request(topic, chapter_drafts, evidence_count, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_global_summary_content(content) {\n        Ok(summary) => Ok(summary),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM global summary validation failed; requesting one repair: {first_err}\"\n            );\n            let repair_request = build_global_summary_repair_request(\n                topic,\n                chapter_drafts,\n                evidence_count,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_global_summary_content(repaired_content)\n        }\n    }\n}\n\nfn build_global_summary_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于章节终稿生成报告的全局摘要。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_global_summary_repair_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n    let previous_excerpt = excerpt_chars(previous_output, REPAIR_PREVIOUS_OUTPUT_MAX_CHARS);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次全局摘要输出，使其满足 JSON schema 和写作约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出节选：\\n{previous_excerpt}\\n\\n章节终稿上下文：\\n{context_json}\",\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn parse_global_summary_content(content: &str) -> Result<String> {\n    let output =\n        serde_json::from_str::<GlobalSummaryOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid GlobalSummary JSON: {err}\"\n            ))\n        })?;\n    let global_summary_zh = output.global_summary_zh.trim().to_string();\n    if global_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer global summary response contained empty global_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&global_summary_zh)?;\n\n    Ok(global_summary_zh)\n}\n\nfn global_synthesizer_system_prompt() -> String {\n    format!(\n        \"你是 LitScout-RS 的报告总编辑，负责基于章节终稿生成全局摘要。\\n\\\n规则：\\n\\\n1. 使用中文撰写 1-2 段自然的全局摘要，总长度不超过 {GLOBAL_SUMMARY_TARGET_CHARS} 个汉字。\\n\\\n2. 概括报告覆盖范围、主要发现、证据边界和不足。\\n\\\n3. 只能使用用户提供的章节终稿和 citation id，不得新增来源、URL 或事实。\\n\\\n4. 不要输出固定模板句，不要写“来源链接”，不要写“是本章节的关键来源之一”。\\n\\\n5. 不要把引用账本逐条改写成摘要。\\n\\\n6. 只返回 JSON：{{\\\"global_summary_zh\\\":\\\"...\\\"}}。\"\n    )\n}\n\nfn parse_chapter_draft_content(\n    content: &str,\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n) -> Result<ChapterDraft> {\n    let output =\n        serde_json::from_str::<ChapterDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid ChapterDraft JSON: {err}\"\n            ))\n        })?;\n    let paragraphs =\n        parse_paragraph_outputs(output.paragraphs, evidence_items, MAX_CHAPTER_PARAGRAPHS)?;\n\n    Ok(ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    })\n}\n\nfn parse_paragraph_outputs(\n    paragraph_outputs: Vec<ChapterParagraphOutput>,\n    evidence_items: &[EvidenceItem],\n    max_paragraphs: usize,\n) -> Result<Vec<ParagraphWithCitations>> {\n    let allowed_ids = evidence_items\n        .iter()\n        .map(|item| item.evidence_id.as_str())\n        .collect::<BTreeSet<_>>();\n    let mut paragraphs = Vec::new();\n\n    for paragraph in paragraph_outputs.into_iter().take(max_paragraphs) {\n        let text_zh = paragraph.text_zh.trim().to_string();\n        if text_zh.is_empty() {\n            continue;\n        }\n        reject_template_residue(&text_zh)?;\n        let cited_evidence_ids = paragraph\n            .cited_evidence_ids\n            .into_iter()\n            .filter_map(|id| {\n                let trimmed = id.trim();\n                (!trimmed.is_empty()).then(|| trimmed.to_string())\n            })\n            .collect::<Vec<_>>();\n        if cited_evidence_ids.is_empty() {\n            return Err(AppError::Llm(\n                \"Writer paragraph did not cite any evidence ID\".to_string(),\n            ));\n        }\n        for id in &cited_evidence_ids {\n            if !allowed_ids.contains(id.as_str()) {\n                return Err(AppError::Llm(format!(\n                    \"Writer paragraph referenced unknown evidence id `{id}`\"\n                )));\n            }\n        }\n        paragraphs.push(ParagraphWithCitations {\n            text_zh,\n            cited_evidence_ids,\n        });\n    }\n\n    if paragraphs.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer response contained no usable paragraphs\".to_string(),\n        ));\n    }\n\n    Ok(paragraphs)\n}\n\nfn reject_template_residue(text: &str) -> Result<()> {\n    for marker in [\n        \"是本章节的关键来源之一\",\n        \"来源链接：\",\n        \"提供了与主题相关的证据\",\n    ] {\n        if text.contains(marker) {\n            return Err(AppError::Llm(format!(\n                \"Writer output still contains legacy template marker `{marker}`\"\n            )));\n        }\n    }\n    Ok(())\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push(' ');\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[derive(Debug, Serialize)]\nstruct BatchDraftInput {\n    chapter_title: String,\n    research_question: String,\n    batch_id: usize,\n    total_batches: usize,\n    batch_context_chars: usize,\n    evidence_sources: Vec<EvidenceSourceInput>,\n}\n\n#[derive(Debug, Serialize)]\nstruct ChapterSynthesisInput {\n    chapter_title: String,\n    research_question: String,\n    allowed_evidence_ids: Vec<String>,\n    batch_drafts: Vec<BatchDraft>,\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryInput {\n    topic: String,\n    chapter_count: usize,\n    evidence_count: usize,\n    chapters: Vec<GlobalSummaryChapterInput>,\n}\n\nimpl GlobalSummaryInput {\n    fn from_report(topic: &str, chapter_drafts: &[ChapterDraft], evidence_count: usize) -> Self {\n        Self {\n            topic: topic.to_string(),\n            chapter_count: chapter_drafts.len(),\n            evidence_count,\n            chapters: chapter_drafts\n                .iter()\n                .map(GlobalSummaryChapterInput::from)\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryChapterInput {\n    chapter_id: String,\n    title_zh: String,\n    paragraph_count: usize,\n    cited_evidence_ids: Vec<String>,\n    paragraph_summaries: Vec<String>,\n}\n\nimpl From<&ChapterDraft> for GlobalSummaryChapterInput {\n    fn from(chapter: &ChapterDraft) -> Self {\n        let cited_evidence_ids = chapter\n            .paragraphs\n            .iter()\n            .flat_map(|paragraph| paragraph.cited_evidence_ids.iter().cloned())\n            .collect::<BTreeSet<_>>()\n            .into_iter()\n            .collect();\n        Self {\n            chapter_id: chapter.chapter_id.clone(),\n            title_zh: chapter.title_zh.clone(),\n            paragraph_count: chapter.paragraphs.len(),\n            cited_evidence_ids,\n            paragraph_summaries: chapter\n                .paragraphs\n                .iter()\n                .map(|paragraph| excerpt_chars(&paragraph.text_zh, 500))\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct EvidenceSourceInput {\n    evidence_id: String,\n    title: String,\n    url: String,\n    source_kind: SourceKind,\n    evidence_note_zh: String,\n    evidence_snippet: String,\n}\n\nimpl From<&EvidenceItem> for EvidenceSourceInput {\n    fn from(item: &EvidenceItem) -> Self {\n        Self {\n            evidence_id: item.evidence_id.clone(),\n            title: excerpt_chars(&item.title, WRITER_TITLE_MAX_CHARS),\n            url: item.url.clone(),\n            source_kind: item.source_kind,\n            evidence_note_zh: excerpt_chars(&item.evidence_note_zh, WRITER_EVIDENCE_NOTE_MAX_CHARS),\n            evidence_snippet: excerpt_chars(\n                &item.evidence_snippet,\n                writer_snippet_max_chars(item.source_kind),\n            ),\n        }\n    }\n}\n\nfn writer_snippet_max_chars(source_kind: SourceKind) -> usize {\n    match source_kind {\n        SourceKind::GitHub => WRITER_GITHUB_SNIPPET_MAX_CHARS,\n        SourceKind::Arxiv | SourceKind::AcademicIndex => WRITER_ACADEMIC_SNIPPET_MAX_CHARS,\n        SourceKind::Bibliography => WRITER_BIBLIOGRAPHY_SNIPPET_MAX_CHARS,\n    }\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterDraftOutput {\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct BatchDraftOutput {\n    #[serde(default)]\n    findings_summary_zh: String,\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct GlobalSummaryOutput {\n    #[serde(default)]\n    global_summary_zh: String,\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterParagraphOutput {\n    text_zh: String,\n    #[serde(default)]\n    cited_evidence_ids: Vec<String>,\n}\n\n#[cfg(test)]\nmod tests {\n    use chrono::{DateTime, Utc};\n\n    use crate::llm::deepseek::DeepSeekConfig;\n    use crate::model::{\n        ChapterDraft, ChapterNode, EvidenceItem, EvidenceMemory, ParagraphWithCitations, SourceKind,\n    };\n\n    use super::{\n        batch_chapter_evidence, build_batch_draft_request, build_chapter_synthesis_request,\n        build_global_summary_request, draft_report_with_llm, empty_chapter_draft,\n        parse_batch_draft_content, parse_chapter_draft_content, parse_global_summary_content,\n        select_evidence_for_writing, writer_max_tokens, BatchDraft, MAX_BATCH_CONTEXT_CHARS,\n        MAX_EVIDENCE_PER_BATCH, MAX_WRITER_EVIDENCE_PER_CHAPTER, WRITER_BIBLIOGRAPHY_EVIDENCE_CAP,\n        WRITER_MIN_OUTPUT_TOKENS,\n    };\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = empty_chapter_draft(&chapters[0]);\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert!(draft.paragraphs[0]\n            .text_zh\n            .contains(\"尚未获得与章节计划直接匹配\"));\n    }\n\n    #[tokio::test]\n    async fn llm_writer_refuses_to_template_when_llm_is_disabled() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n        };\n\n        let err = draft_report_with_llm(\n            \"Rust Agent\",\n            &chapters,\n            &memory,\n            &crate::config::LlmConfig::from_env(false, 30),\n        )\n        .await\n        .expect_err(\"disabled LLM should not produce a template report\");\n\n        assert!(err\n            .to_string()\n            .contains(\"refusing to generate a template report\"));\n    }\n\n    #[tokio::test]\n    #[ignore = \"requires a real DeepSeek API key in DEEPSEEK_API_KEY\"]\n    async fn writer_real_deepseek_smoke() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n        };\n        let config = crate::config::LlmConfig::from_env(true, 120);\n\n        assert!(\n            config.api_key.is_some(),\n            \"DEEPSEEK_API_KEY must be set for this ignored smoke test\"\n        );\n\n        let draft = draft_report_with_llm(\"TTS 入门级项目简单调研\", &chapters, &memory, &config)\n            .await\n            .expect(\"real DeepSeek writer smoke should complete\");\n\n        assert!(!draft.global_summary_zh.trim().is_empty());\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(!draft.chapters[0].paragraphs.is_empty());\n        assert!(draft.chapters[0].paragraphs[0]\n            .cited_evidence_ids\n            .contains(&\"ev-C1\".to_string()));\n        println!(\n            \"writer_real_deepseek_smoke ok: global_summary_chars={}, chapter_paragraphs={}\",\n            draft.global_summary_zh.chars().count(),\n            draft.chapters[0].paragraphs.len()\n        );\n    }\n\n    #[test]\n    fn parses_llm_chapter_draft_json() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了从 README 提炼出的 ReAct 主循环和工具调用结构，适合作为入门实现分析对象。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect(\"valid chapter JSON should parse\");\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n        assert!(draft.paragraphs[0].text_zh.contains(\"ReAct\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_unknown_evidence_id() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"引用了不存在的证据。\",\n                    \"cited_evidence_ids\": [\"ev-C9\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_template_residue() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"acme/rust-agent 是本章节的关键来源之一。来源链接：https://github.com/acme/rust-agent\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batches_small_chapter_evidence_into_one_batch() {\n        let evidence = vec![\n            sample_evidence(\"ev-C1\", \"ch-1\"),\n            sample_evidence(\"ev-C2\", \"ch-1\"),\n            sample_evidence(\"ev-C3\", \"ch-1\"),\n        ];\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 1);\n        assert_eq!(batches[0].batch_id, 1);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert!(batches[0].context_chars > 0);\n    }\n\n    #[test]\n    fn selects_bounded_mixed_evidence_for_writing() {\n        let mut evidence = Vec::new();\n        evidence.extend((1..=5).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-A{index}\"), \"ch-1\", SourceKind::Arxiv)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-S{index}\"), \"ch-1\", SourceKind::AcademicIndex)\n        }));\n        evidence.extend((1..=4).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-G{index}\"), \"ch-1\", SourceKind::GitHub)\n        }));\n        evidence.extend((1..=2).map(|index| {\n            sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n        }));\n\n        let selected = select_evidence_for_writing(&evidence);\n        let bibliography_count = selected\n            .iter()\n            .filter(|item| item.source_kind == SourceKind::Bibliography)\n            .count();\n\n        assert_eq!(selected.len(), MAX_WRITER_EVIDENCE_PER_CHAPTER);\n        assert_eq!(bibliography_count, 0);\n        assert_eq!(selected[0].source_kind, SourceKind::Arxiv);\n        assert_eq!(selected[4].source_kind, SourceKind::AcademicIndex);\n        assert_eq!(selected[7].source_kind, SourceKind::GitHub);\n    }\n\n    #[test]\n    fn selects_bibliography_only_when_no_stronger_sources_exist() {\n        let evidence = (1..=3)\n            .map(|index| {\n                sample_evidence_with_kind(&format!(\"ev-B{index}\"), \"ch-1\", SourceKind::Bibliography)\n            })\n            .collect::<Vec<_>>();\n\n        let selected = select_evidence_for_writing(&evidence);\n\n        assert_eq!(selected.len(), WRITER_BIBLIOGRAPHY_EVIDENCE_CAP);\n        assert_eq!(selected[0].source_kind, SourceKind::Bibliography);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_count() {\n        let evidence = (1..=9)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 3);\n        assert_eq!(batches[0].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[1].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].batch_id, 3);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_sanitized_context_chars() {\n        let mut first = sample_evidence(\"ev-C1\", \"ch-1\");\n        first.evidence_snippet = \"a\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut second = sample_evidence(\"ev-C2\", \"ch-1\");\n        second.evidence_snippet = \"b\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut third = sample_evidence(\"ev-C3\", \"ch-1\");\n        third.evidence_snippet = \"c\".repeat(MAX_BATCH_CONTEXT_CHARS);\n        let mut fourth = sample_evidence(\"ev-C4\", \"ch-1\");\n        fourth.evidence_snippet = \"d\".repeat(MAX_BATCH_CONTEXT_CHARS);\n\n        let batches = batch_chapter_evidence(&[first, second, third, fourth]);\n\n        assert_eq!(batches.len(), 2);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert_eq!(batches[1].evidence_items.len(), 1);\n        assert!(batches[0].context_chars <= MAX_BATCH_CONTEXT_CHARS);\n    }\n\n    #[test]\n    fn batch_request_truncates_long_github_snippets_for_writer_context() {\n        let chapter = sample_chapter();\n        let mut evidence = sample_evidence(\"ev-C1\", \"ch-1\");\n        evidence.evidence_snippet = format!(\"{}{}\", \"A\".repeat(1500), \"UNTRUNCATED_TAIL\");\n        let batches = batch_chapter_evidence(&[evidence]);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[0],\n            batches.len(),\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(&\"A\".repeat(900)));\n        assert!(!user_message.contains(\"UNTRUNCATED_TAIL\"));\n    }\n\n    #[test]\n    fn parses_llm_batch_draft_json() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据说明了项目 README 中的 ReAct 主循环、工具注册和入门示例。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目的 README 将 ReAct 主循环和工具注册拆成清晰的入门实现，适合用于说明基础 agent 架构。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft =\n            parse_batch_draft_content(content, &batches[0]).expect(\"valid batch JSON should parse\");\n\n        assert_eq!(draft.batch_id, 1);\n        assert!(draft.findings_summary_zh.contains(\"ReAct\"));\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_empty_findings() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"empty findings should fail\");\n\n        assert!(err.to_string().contains(\"empty findings_summary_zh\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_unknown_evidence_id() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据有一个段落引用了不存在的证据。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"这里引用了不存在的 evidence id。\",\n                    \"cited_evidence_ids\": [\"ev-C2\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_template_residue() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"acme/rust-agent 是本章节的关键来源之一。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batch_request_only_contains_current_batch_evidence() {\n        let chapter = sample_chapter();\n        let evidence = (1..=5)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n        let batches = batch_chapter_evidence(&evidence);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[1],\n            batches.len(),\n            &sample_deepseek_config(2048),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"\\\"batch_id\\\": 2\"));\n        assert!(user_message.contains(\"ev-C5\"));\n        assert!(!user_message.contains(\"ev-C1\"));\n        assert_eq!(request.max_tokens, Some(WRITER_MIN_OUTPUT_TOKENS as u32));\n    }\n\n    #[test]\n    fn writer_max_tokens_raises_too_small_config_for_reasoning_models() {\n        let config = sample_deepseek_config(32);\n\n        assert_eq!(writer_max_tokens(&config), WRITER_MIN_OUTPUT_TOKENS as u32);\n    }\n\n    #[test]\n    fn chapter_synthesis_request_uses_batch_findings_instead_of_raw_evidence() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batch_drafts = vec![BatchDraft {\n            batch_id: 1,\n            findings_summary_zh: \"该批证据说明 README 中的 ReAct 主循环适合作为入门样例。\"\n                .to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该项目展示了 ReAct 主循环和工具注册。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_chapter_synthesis_request(\n            &chapter,\n            &evidence,\n            &batch_drafts,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"batch_drafts\"));\n        assert!(user_message.contains(\"allowed_evidence_ids\"));\n        assert!(user_message.contains(\"ReAct 主循环适合作为入门样例\"));\n        assert!(!user_message.contains(\"evidence_snippet\"));\n        assert!(!user_message.contains(\"README explains ReAct loop\"));\n    }\n\n    #[test]\n    fn parses_llm_global_summary_json() {\n        let content = r#\"{\n            \"global_summary_zh\": \"本报告显示，可控 TTS 入门调研应同时关注基础项目、说话人控制和模型框架边界；现有证据主要来自 GitHub README 与 arXiv 摘要，因此结论应限定在这些来源范围内。\"\n        }\"#;\n\n        let summary =\n            parse_global_summary_content(content).expect(\"valid global summary should parse\");\n\n        assert!(summary.contains(\"可控 TTS\"));\n        assert!(summary.contains(\"来源范围\"));\n    }\n\n    #[test]\n    fn rejects_llm_global_summary_with_template_residue() {\n        let content = r#\"{\n            \"global_summary_zh\": \"harinandanmv/text-to-speech 是本章节的关键来源之一。\"\n        }\"#;\n\n        let err =\n            parse_global_summary_content(content).expect_err(\"legacy template marker should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn global_summary_request_uses_chapter_drafts() {\n        let chapter_drafts = vec![ChapterDraft {\n            chapter_id: \"ch-1\".to_string(),\n            title_zh: \"可控 TTS 基础模型\".to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该章节围绕基础 TTS 模型、说话人控制和入门项目进行综合分析。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_global_summary_request(\n            \"TTS 入门级项目简单调研\",\n            &chapter_drafts,\n            1,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"TTS 入门级项目简单调研\"));\n        assert!(user_message.contains(\"可控 TTS 基础模型\"));\n        assert!(user_message.contains(\"ev-C1\"));\n        assert!(user_message.contains(\"paragraph_summaries\"));\n    }\n\n    fn sample_chapter() -> ChapterNode {\n        ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }\n    }\n\n    fn sample_evidence(evidence_id: &str, chapter_id: &str) -> EvidenceItem {\n        sample_evidence_with_kind(evidence_id, chapter_id, SourceKind::GitHub)\n    }\n\n    fn sample_evidence_with_kind(\n        evidence_id: &str,\n        chapter_id: &str,\n        source_kind: SourceKind,\n    ) -> EvidenceItem {\n        EvidenceItem {\n            evidence_id: evidence_id.to_string(),\n            source_item_id: format!(\"{source_kind:?}:acme/rust-agent\"),\n            citation_id: \"C1\".to_string(),\n            chapter_ids: vec![chapter_id.to_string()],\n            query_attempt_ids: vec![\"gh-1\".to_string()],\n            source_kind,\n            title: \"acme/rust-agent\".to_string(),\n            url: \"https://github.com/acme/rust-agent\".to_string(),\n            evidence_note_zh: \"GitHub 仓库 `acme/rust-agent`：Rust agent framework\".to_string(),\n            evidence_snippet: \"README explains ReAct loop, tool registry, and examples.\"\n                .to_string(),\n            support_score: None,\n        }\n    }\n\n    fn sample_deepseek_config(max_tokens: usize) -> DeepSeekConfig {\n        DeepSeekConfig {\n            api_key: \"sk-test\".to_string(),\n            base_url: \"https://api.deepseek.com\".to_string(),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: \"deepseek-v4-flash\".to_string(),\n            max_tokens,\n            timeout_secs: 120,\n        }\n    }\n\n    fn _dt() -> DateTime<Utc> {\n        DateTime::parse_from_rfc3339(\"2026-05-30T12:00:00Z\")\n            .unwrap()\n            .with_timezone(&Utc)\n    }\n}\n"}, "bccd8c422afb8962e69d703d2ddcbdac277a8101": {"t": 1785750666.5143292, "n": 43384, "h": "e05aa6561b6b8927", "c": "use crate::config::LlmConfig;\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{\n    strip_json_fence, ChatCompletionsRequest, ChatMessage, DeepSeekClient, DeepSeekConfig,\n    ResponseFormat,\n};\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft, SourceKind,\n};\nuse chrono::Utc;\nuse serde::{Deserialize, Serialize};\nuse std::collections::BTreeSet;\nuse tracing::{debug, warn};\n\nconst MAX_EVIDENCE_PER_BATCH: usize = 4;\nconst MAX_BATCH_CONTEXT_CHARS: usize = 10_000;\nconst MAX_BATCH_PARAGRAPHS: usize = 4;\nconst MAX_CHAPTER_PARAGRAPHS: usize = 6;\nconst WRITER_MIN_OUTPUT_TOKENS: usize = 4096;\nconst WRITER_MAX_OUTPUT_TOKENS: usize = 8192;\n\npub async fn draft_report_with_llm(\n    topic: &str,\n    chapters: &[ChapterNode],\n    memory: &EvidenceMemory,\n    llm_config: &LlmConfig,\n) -> Result<ReportDraft> {\n    let deepseek_config = DeepSeekConfig::from_llm_config(llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"Writer LLM requires an enabled DeepSeek config with API key; refusing to generate a template report.\"\n                .to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(deepseek_config.clone())?;\n    let mut chapter_drafts = Vec::with_capacity(chapters.len());\n\n    for chapter in chapters {\n        let evidence_items = memory.by_chapter(&chapter.id);\n        if evidence_items.is_empty() {\n            chapter_drafts.push(empty_chapter_draft(chapter));\n            continue;\n        }\n\n        let draft = draft_chapter_with_llm(chapter, &evidence_items, &client, &deepseek_config)\n            .await\n            .map_err(|err| {\n                AppError::Llm(format!(\n                    \"Writer LLM failed for chapter `{}`; refusing template fallback: {err}\",\n                    chapter.title_zh\n                ))\n            })?;\n        chapter_drafts.push(draft);\n    }\n\n    let global_summary_zh = synthesize_global_summary_with_llm(\n        topic,\n        &chapter_drafts,\n        memory.items.len(),\n        &client,\n        &deepseek_config,\n    )\n    .await\n    .map_err(|err| {\n        AppError::Llm(format!(\n            \"Writer global summary failed; refusing deterministic fallback: {err}\"\n        ))\n    })?;\n\n    Ok(ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    })\n}\n\nasync fn draft_chapter_with_llm(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let batches = batch_chapter_evidence(evidence_items);\n    debug!(\n        \"Writer split chapter `{}` into {} evidence batches\",\n        chapter.title_zh,\n        batches.len()\n    );\n\n    let mut batch_drafts = Vec::with_capacity(batches.len());\n    for batch in &batches {\n        let draft = draft_batch_with_llm(chapter, batch, batches.len(), client, config).await?;\n        batch_drafts.push(draft);\n    }\n\n    synthesize_chapter_from_batches(chapter, evidence_items, &batch_drafts, client, config).await\n}\n\nfn empty_chapter_draft(chapter: &ChapterNode) -> ChapterDraft {\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs: vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未获得与章节计划直接匹配的 GitHub/arXiv 证据，因此不生成分析段落。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }],\n    }\n}\n\n#[derive(Debug, Clone)]\nstruct EvidenceBatch {\n    batch_id: usize,\n    evidence_items: Vec<EvidenceItem>,\n    context_chars: usize,\n}\n\n#[derive(Debug, Clone, Serialize)]\nstruct BatchDraft {\n    batch_id: usize,\n    findings_summary_zh: String,\n    paragraphs: Vec<ParagraphWithCitations>,\n}\n\nfn batch_chapter_evidence(evidence_items: &[EvidenceItem]) -> Vec<EvidenceBatch> {\n    let mut batches = Vec::new();\n    let mut current_items = Vec::new();\n    let mut current_chars = 0usize;\n\n    for item in evidence_items {\n        let item_chars = evidence_context_chars(item);\n        let count_limit_reached = current_items.len() >= MAX_EVIDENCE_PER_BATCH;\n        let char_limit_reached =\n            !current_items.is_empty() && current_chars + item_chars > MAX_BATCH_CONTEXT_CHARS;\n\n        if count_limit_reached || char_limit_reached {\n            batches.push(EvidenceBatch {\n                batch_id: batches.len() + 1,\n                evidence_items: current_items,\n                context_chars: current_chars,\n            });\n            current_items = Vec::new();\n            current_chars = 0;\n        }\n\n        current_chars += item_chars;\n        current_items.push(item.clone());\n    }\n\n    if !current_items.is_empty() {\n        batches.push(EvidenceBatch {\n            batch_id: batches.len() + 1,\n            evidence_items: current_items,\n            context_chars: current_chars,\n        });\n    }\n\n    batches\n}\n\nfn evidence_context_chars(item: &EvidenceItem) -> usize {\n    item.evidence_id.chars().count()\n        + item.title.chars().count()\n        + item.url.chars().count()\n        + item.evidence_note_zh.chars().count()\n        + item.evidence_snippet.chars().count()\n}\n\nfn excerpt_chars(text: &str, max_chars: usize) -> String {\n    let trimmed = text.trim();\n    if trimmed.chars().count() <= max_chars {\n        return trimmed.to_string();\n    }\n\n    let mut out = trimmed.chars().take(max_chars).collect::<String>();\n    out.push_str(\"...\");\n    out\n}\n\nfn parse_batch_draft_content(content: &str, batch: &EvidenceBatch) -> Result<BatchDraft> {\n    let output =\n        serde_json::from_str::<BatchDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid BatchDraft JSON: {err}\"\n            ))\n        })?;\n    let findings_summary_zh = output.findings_summary_zh.trim().to_string();\n    if findings_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer batch response contained empty findings_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&findings_summary_zh)?;\n    let paragraphs = parse_paragraph_outputs(\n        output.paragraphs,\n        &batch.evidence_items,\n        MAX_BATCH_PARAGRAPHS,\n    )?;\n\n    Ok(BatchDraft {\n        batch_id: batch.batch_id,\n        findings_summary_zh,\n        paragraphs,\n    })\n}\n\nasync fn draft_batch_with_llm(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<BatchDraft> {\n    let request = build_batch_draft_request(chapter, batch, total_batches, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_batch_draft_content(content, batch) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM JSON validation failed for chapter `{}` batch {}; requesting one repair: {first_err}\",\n                chapter.title_zh, batch.batch_id\n            );\n            let repair_request = build_batch_draft_repair_request(\n                chapter,\n                batch,\n                total_batches,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_batch_draft_content(repaired_content, batch)\n        }\n    }\n}\n\nfn build_batch_draft_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请为下面章节的第 {}/{} 个证据批次生成局部分析草稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.3),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_batch_draft_repair_request(\n    chapter: &ChapterNode,\n    batch: &EvidenceBatch,\n    total_batches: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = BatchDraftInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        batch_id: batch.batch_id,\n        total_batches,\n        batch_context_chars: batch.context_chars,\n        evidence_sources: batch\n            .evidence_items\n            .iter()\n            .map(EvidenceSourceInput::from)\n            .collect(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: batch_writer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次第 {}/{} 个证据批次的输出，使其满足 JSON schema 和引用约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出：\\n{previous_output}\\n\\n允许使用的 evidence：\\n{context_json}\",\n                    batch.batch_id, total_batches\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nasync fn synthesize_chapter_from_batches(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<ChapterDraft> {\n    let request = build_chapter_synthesis_request(chapter, evidence_items, batch_drafts, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_chapter_draft_content(content, chapter, evidence_items) {\n        Ok(draft) => Ok(draft),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM chapter synthesis validation failed for chapter `{}`; requesting one repair: {first_err}\",\n                chapter.title_zh\n            );\n            let repair_request = build_chapter_synthesis_repair_request(\n                chapter,\n                evidence_items,\n                batch_drafts,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_chapter_draft_content(repaired_content, chapter, evidence_items)\n        }\n    }\n}\n\nfn batch_writer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的技术报告撰稿人，正在为 `{chapter_title}` 章节处理一个 evidence batch。\\n\\\n规则：\\n\\\n1. 使用中文撰写。\\n\\\n2. 先在 findings_summary_zh 中总结本批证据的关键发现、共性和边界。\\n\\\n3. 再对本批来源提炼项目/论文简介、核心亮点或设计思想，以及与章节问题的具体关联。\\n\\\n4. 可合并多个来源做对比分析，但每个段落必须引用至少一个 evidence_id。\\n\\\n5. 只能使用用户提供的 evidence_id 和 URL，不得新增来源、不得编造事实。\\n\\\n6. 不要只重复原文，不要输出“来源链接”模板句。\\n\\\n7. 只返回 JSON：{{\\\"findings_summary_zh\\\":\\\"...\\\",\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nfn writer_max_tokens(config: &DeepSeekConfig) -> u32 {\n    config\n        .max_tokens\n        .max(WRITER_MIN_OUTPUT_TOKENS)\n        .min(WRITER_MAX_OUTPUT_TOKENS)\n        .min(u32::MAX as usize) as u32\n}\n\nfn build_chapter_synthesis_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于已验证的 batch findings 合成章节终稿。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_chapter_synthesis_repair_request(\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n    batch_drafts: &[BatchDraft],\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = ChapterSynthesisInput {\n        chapter_title: chapter.title_zh.clone(),\n        research_question: chapter.research_question.clone(),\n        allowed_evidence_ids: evidence_items\n            .iter()\n            .map(|item| item.evidence_id.clone())\n            .collect(),\n        batch_drafts: batch_drafts.to_vec(),\n    };\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: chapter_synthesizer_system_prompt(&chapter.title_zh),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次章节综合输出，使其满足 JSON schema 和引用约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出：\\n{previous_output}\\n\\n允许使用的 batch findings：\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn chapter_synthesizer_system_prompt(chapter_title: &str) -> String {\n    format!(\n        \"你是 LitScout-RS 的章节综合编辑，正在合成 `{chapter_title}` 章节终稿。\\n\\\n规则：\\n\\\n1. 使用中文撰写自然、分析性的技术报告段落。\\n\\\n2. 只能使用用户提供的 batch findings、batch paragraphs 和 allowed_evidence_ids。\\n\\\n3. 合并重复信息，按主题组织，不要机械罗列每条来源。\\n\\\n4. 已有 evidence 时，不得用“尚未收集到足够证据”替代分析内容。\\n\\\n5. 每个段落必须引用至少一个 allowed_evidence_ids 中的 evidence_id。\\n\\\n6. 不得新增来源、URL 或 evidence_id，不得编造事实。\\n\\\n7. 不要输出“来源链接”模板句或“是本章节的关键来源之一”。\\n\\\n8. 只返回 JSON：{{\\\"paragraphs\\\":[{{\\\"text_zh\\\":\\\"...\\\",\\\"cited_evidence_ids\\\":[\\\"ev-C1\\\"]}}]}}。\"\n    )\n}\n\nasync fn synthesize_global_summary_with_llm(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    client: &DeepSeekClient,\n    config: &DeepSeekConfig,\n) -> Result<String> {\n    let request = build_global_summary_request(topic, chapter_drafts, evidence_count, config)?;\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n\n    match parse_global_summary_content(content) {\n        Ok(summary) => Ok(summary),\n        Err(first_err) => {\n            warn!(\n                \"Writer LLM global summary validation failed; requesting one repair: {first_err}\"\n            );\n            let repair_request = build_global_summary_repair_request(\n                topic,\n                chapter_drafts,\n                evidence_count,\n                config,\n                content,\n                &first_err.to_string(),\n            )?;\n            let repaired_response = client.chat_completions_with_retry(repair_request).await?;\n            let repaired_content = repaired_response.first_content()?;\n            parse_global_summary_content(repaired_content)\n        }\n    }\n}\n\nfn build_global_summary_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请基于章节终稿生成报告的全局摘要。只返回 JSON 对象，不要输出 Markdown 代码围栏之外的解释。\\n\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.25),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn build_global_summary_repair_request(\n    topic: &str,\n    chapter_drafts: &[ChapterDraft],\n    evidence_count: usize,\n    config: &DeepSeekConfig,\n    previous_output: &str,\n    validation_error: &str,\n) -> Result<ChatCompletionsRequest> {\n    let context = GlobalSummaryInput::from_report(topic, chapter_drafts, evidence_count);\n    let context_json = serde_json::to_string_pretty(&context)?;\n    let max_tokens = writer_max_tokens(config);\n\n    Ok(ChatCompletionsRequest {\n        model: config.main_model.clone(),\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: global_synthesizer_system_prompt(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content: format!(\n                    \"请修复上一次全局摘要输出，使其满足 JSON schema 和写作约束。只返回修复后的 JSON 对象。\\n\\n校验错误：\\n{validation_error}\\n\\n上一次输出：\\n{previous_output}\\n\\n章节终稿上下文：\\n{context_json}\"\n                ),\n            },\n        ],\n        temperature: Some(0.1),\n        max_tokens: Some(max_tokens),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    })\n}\n\nfn parse_global_summary_content(content: &str) -> Result<String> {\n    let output =\n        serde_json::from_str::<GlobalSummaryOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid GlobalSummary JSON: {err}\"\n            ))\n        })?;\n    let global_summary_zh = output.global_summary_zh.trim().to_string();\n    if global_summary_zh.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer global summary response contained empty global_summary_zh\".to_string(),\n        ));\n    }\n    reject_template_residue(&global_summary_zh)?;\n\n    Ok(global_summary_zh)\n}\n\nfn global_synthesizer_system_prompt() -> String {\n    \"你是 LitScout-RS 的报告总编辑，负责基于章节终稿生成全局摘要。\\n\\\n规则：\\n\\\n1. 使用中文撰写 1-2 段自然的全局摘要。\\n\\\n2. 概括报告覆盖范围、主要发现、证据边界和不足。\\n\\\n3. 只能使用用户提供的章节终稿和 citation id，不得新增来源、URL 或事实。\\n\\\n4. 不要输出固定模板句，不要写“来源链接”，不要写“是本章节的关键来源之一”。\\n\\\n5. 不要把引用账本逐条改写成摘要。\\n\\\n6. 只返回 JSON：{\\\"global_summary_zh\\\":\\\"...\\\"}。\"\n        .to_string()\n}\n\nfn parse_chapter_draft_content(\n    content: &str,\n    chapter: &ChapterNode,\n    evidence_items: &[EvidenceItem],\n) -> Result<ChapterDraft> {\n    let output =\n        serde_json::from_str::<ChapterDraftOutput>(strip_json_fence(content)).map_err(|err| {\n            AppError::Llm(format!(\n                \"Writer response was not valid ChapterDraft JSON: {err}\"\n            ))\n        })?;\n    let paragraphs =\n        parse_paragraph_outputs(output.paragraphs, evidence_items, MAX_CHAPTER_PARAGRAPHS)?;\n\n    Ok(ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    })\n}\n\nfn parse_paragraph_outputs(\n    paragraph_outputs: Vec<ChapterParagraphOutput>,\n    evidence_items: &[EvidenceItem],\n    max_paragraphs: usize,\n) -> Result<Vec<ParagraphWithCitations>> {\n    let allowed_ids = evidence_items\n        .iter()\n        .map(|item| item.evidence_id.as_str())\n        .collect::<BTreeSet<_>>();\n    let mut paragraphs = Vec::new();\n\n    for paragraph in paragraph_outputs.into_iter().take(max_paragraphs) {\n        let text_zh = paragraph.text_zh.trim().to_string();\n        if text_zh.is_empty() {\n            continue;\n        }\n        reject_template_residue(&text_zh)?;\n        let cited_evidence_ids = paragraph\n            .cited_evidence_ids\n            .into_iter()\n            .filter_map(|id| {\n                let trimmed = id.trim();\n                (!trimmed.is_empty()).then(|| trimmed.to_string())\n            })\n            .collect::<Vec<_>>();\n        if cited_evidence_ids.is_empty() {\n            return Err(AppError::Llm(\n                \"Writer paragraph did not cite any evidence ID\".to_string(),\n            ));\n        }\n        for id in &cited_evidence_ids {\n            if !allowed_ids.contains(id.as_str()) {\n                return Err(AppError::Llm(format!(\n                    \"Writer paragraph referenced unknown evidence id `{id}`\"\n                )));\n            }\n        }\n        paragraphs.push(ParagraphWithCitations {\n            text_zh,\n            cited_evidence_ids,\n        });\n    }\n\n    if paragraphs.is_empty() {\n        return Err(AppError::Llm(\n            \"Writer response contained no usable paragraphs\".to_string(),\n        ));\n    }\n\n    Ok(paragraphs)\n}\n\nfn reject_template_residue(text: &str) -> Result<()> {\n    for marker in [\n        \"是本章节的关键来源之一\",\n        \"来源链接：\",\n        \"提供了与主题相关的证据\",\n    ] {\n        if text.contains(marker) {\n            return Err(AppError::Llm(format!(\n                \"Writer output still contains legacy template marker `{marker}`\"\n            )));\n        }\n    }\n    Ok(())\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push(' ');\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[derive(Debug, Serialize)]\nstruct BatchDraftInput {\n    chapter_title: String,\n    research_question: String,\n    batch_id: usize,\n    total_batches: usize,\n    batch_context_chars: usize,\n    evidence_sources: Vec<EvidenceSourceInput>,\n}\n\n#[derive(Debug, Serialize)]\nstruct ChapterSynthesisInput {\n    chapter_title: String,\n    research_question: String,\n    allowed_evidence_ids: Vec<String>,\n    batch_drafts: Vec<BatchDraft>,\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryInput {\n    topic: String,\n    chapter_count: usize,\n    evidence_count: usize,\n    chapters: Vec<GlobalSummaryChapterInput>,\n}\n\nimpl GlobalSummaryInput {\n    fn from_report(topic: &str, chapter_drafts: &[ChapterDraft], evidence_count: usize) -> Self {\n        Self {\n            topic: topic.to_string(),\n            chapter_count: chapter_drafts.len(),\n            evidence_count,\n            chapters: chapter_drafts\n                .iter()\n                .map(GlobalSummaryChapterInput::from)\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct GlobalSummaryChapterInput {\n    chapter_id: String,\n    title_zh: String,\n    paragraph_count: usize,\n    cited_evidence_ids: Vec<String>,\n    paragraph_summaries: Vec<String>,\n}\n\nimpl From<&ChapterDraft> for GlobalSummaryChapterInput {\n    fn from(chapter: &ChapterDraft) -> Self {\n        let cited_evidence_ids = chapter\n            .paragraphs\n            .iter()\n            .flat_map(|paragraph| paragraph.cited_evidence_ids.iter().cloned())\n            .collect::<BTreeSet<_>>()\n            .into_iter()\n            .collect();\n        Self {\n            chapter_id: chapter.chapter_id.clone(),\n            title_zh: chapter.title_zh.clone(),\n            paragraph_count: chapter.paragraphs.len(),\n            cited_evidence_ids,\n            paragraph_summaries: chapter\n                .paragraphs\n                .iter()\n                .map(|paragraph| excerpt_chars(&paragraph.text_zh, 500))\n                .collect(),\n        }\n    }\n}\n\n#[derive(Debug, Serialize)]\nstruct EvidenceSourceInput {\n    evidence_id: String,\n    title: String,\n    url: String,\n    source_kind: SourceKind,\n    evidence_note_zh: String,\n    evidence_snippet: String,\n}\n\nimpl From<&EvidenceItem> for EvidenceSourceInput {\n    fn from(item: &EvidenceItem) -> Self {\n        Self {\n            evidence_id: item.evidence_id.clone(),\n            title: item.title.clone(),\n            url: item.url.clone(),\n            source_kind: item.source_kind,\n            evidence_note_zh: item.evidence_note_zh.clone(),\n            evidence_snippet: item.evidence_snippet.clone(),\n        }\n    }\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterDraftOutput {\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct BatchDraftOutput {\n    #[serde(default)]\n    findings_summary_zh: String,\n    #[serde(default)]\n    paragraphs: Vec<ChapterParagraphOutput>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct GlobalSummaryOutput {\n    #[serde(default)]\n    global_summary_zh: String,\n}\n\n#[derive(Debug, Deserialize)]\nstruct ChapterParagraphOutput {\n    text_zh: String,\n    #[serde(default)]\n    cited_evidence_ids: Vec<String>,\n}\n\n#[cfg(test)]\nmod tests {\n    use chrono::{DateTime, Utc};\n\n    use crate::llm::deepseek::DeepSeekConfig;\n    use crate::model::{\n        ChapterDraft, ChapterNode, EvidenceItem, EvidenceMemory, ParagraphWithCitations, SourceKind,\n    };\n\n    use super::{\n        batch_chapter_evidence, build_batch_draft_request, build_chapter_synthesis_request,\n        build_global_summary_request, draft_report_with_llm, empty_chapter_draft,\n        parse_batch_draft_content, parse_chapter_draft_content, parse_global_summary_content,\n        writer_max_tokens, BatchDraft, MAX_BATCH_CONTEXT_CHARS, MAX_EVIDENCE_PER_BATCH,\n        WRITER_MIN_OUTPUT_TOKENS,\n    };\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = empty_chapter_draft(&chapters[0]);\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert!(draft.paragraphs[0]\n            .text_zh\n            .contains(\"尚未获得与章节计划直接匹配\"));\n    }\n\n    #[tokio::test]\n    async fn llm_writer_refuses_to_template_when_llm_is_disabled() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n        };\n\n        let err = draft_report_with_llm(\n            \"Rust Agent\",\n            &chapters,\n            &memory,\n            &crate::config::LlmConfig::from_env(false, 30),\n        )\n        .await\n        .expect_err(\"disabled LLM should not produce a template report\");\n\n        assert!(err\n            .to_string()\n            .contains(\"refusing to generate a template report\"));\n    }\n\n    #[tokio::test]\n    #[ignore = \"requires a real DeepSeek API key in DEEPSEEK_API_KEY\"]\n    async fn writer_real_deepseek_smoke() {\n        let chapters = vec![sample_chapter()];\n        let memory = EvidenceMemory {\n            items: vec![sample_evidence(\"ev-C1\", \"ch-1\")],\n            query_attempts: Vec::new(),\n            source_lineage: Vec::new(),\n        };\n        let config = crate::config::LlmConfig::from_env(true, 120);\n\n        assert!(\n            config.api_key.is_some(),\n            \"DEEPSEEK_API_KEY must be set for this ignored smoke test\"\n        );\n\n        let draft = draft_report_with_llm(\"TTS 入门级项目简单调研\", &chapters, &memory, &config)\n            .await\n            .expect(\"real DeepSeek writer smoke should complete\");\n\n        assert!(!draft.global_summary_zh.trim().is_empty());\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(!draft.chapters[0].paragraphs.is_empty());\n        assert!(draft.chapters[0].paragraphs[0]\n            .cited_evidence_ids\n            .contains(&\"ev-C1\".to_string()));\n        println!(\n            \"writer_real_deepseek_smoke ok: global_summary_chars={}, chapter_paragraphs={}\",\n            draft.global_summary_zh.chars().count(),\n            draft.chapters[0].paragraphs.len()\n        );\n    }\n\n    #[test]\n    fn parses_llm_chapter_draft_json() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了从 README 提炼出的 ReAct 主循环和工具调用结构，适合作为入门实现分析对象。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect(\"valid chapter JSON should parse\");\n\n        assert_eq!(draft.chapter_id, \"ch-1\");\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n        assert!(draft.paragraphs[0].text_zh.contains(\"ReAct\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_unknown_evidence_id() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"引用了不存在的证据。\",\n                    \"cited_evidence_ids\": [\"ev-C9\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_chapter_draft_with_template_residue() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let content = r#\"{\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"acme/rust-agent 是本章节的关键来源之一。来源链接：https://github.com/acme/rust-agent\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_chapter_draft_content(content, &chapter, &evidence)\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batches_small_chapter_evidence_into_one_batch() {\n        let evidence = vec![\n            sample_evidence(\"ev-C1\", \"ch-1\"),\n            sample_evidence(\"ev-C2\", \"ch-1\"),\n            sample_evidence(\"ev-C3\", \"ch-1\"),\n        ];\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 1);\n        assert_eq!(batches[0].batch_id, 1);\n        assert_eq!(batches[0].evidence_items.len(), 3);\n        assert!(batches[0].context_chars > 0);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_count() {\n        let evidence = (1..=9)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n\n        let batches = batch_chapter_evidence(&evidence);\n\n        assert_eq!(batches.len(), 3);\n        assert_eq!(batches[0].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[1].evidence_items.len(), MAX_EVIDENCE_PER_BATCH);\n        assert_eq!(batches[2].evidence_items.len(), 1);\n        assert_eq!(batches[2].batch_id, 3);\n    }\n\n    #[test]\n    fn batches_chapter_evidence_by_context_chars() {\n        let mut first = sample_evidence(\"ev-C1\", \"ch-1\");\n        first.evidence_snippet = \"a\".repeat(MAX_BATCH_CONTEXT_CHARS / 2 + 200);\n        let mut second = sample_evidence(\"ev-C2\", \"ch-1\");\n        second.evidence_snippet = \"b\".repeat(MAX_BATCH_CONTEXT_CHARS / 2 + 200);\n\n        let batches = batch_chapter_evidence(&[first, second]);\n\n        assert_eq!(batches.len(), 2);\n        assert_eq!(batches[0].evidence_items[0].evidence_id, \"ev-C1\");\n        assert_eq!(batches[1].evidence_items[0].evidence_id, \"ev-C2\");\n        assert!(batches[0].context_chars > MAX_BATCH_CONTEXT_CHARS / 2);\n    }\n\n    #[test]\n    fn parses_llm_batch_draft_json() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据说明了项目 README 中的 ReAct 主循环、工具注册和入门示例。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目的 README 将 ReAct 主循环和工具注册拆成清晰的入门实现，适合用于说明基础 agent 架构。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let draft =\n            parse_batch_draft_content(content, &batches[0]).expect(\"valid batch JSON should parse\");\n\n        assert_eq!(draft.batch_id, 1);\n        assert!(draft.findings_summary_zh.contains(\"ReAct\"));\n        assert_eq!(draft.paragraphs[0].cited_evidence_ids, vec![\"ev-C1\"]);\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_empty_findings() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"empty findings should fail\");\n\n        assert!(err.to_string().contains(\"empty findings_summary_zh\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_unknown_evidence_id() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"该批证据有一个段落引用了不存在的证据。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"这里引用了不存在的 evidence id。\",\n                    \"cited_evidence_ids\": [\"ev-C2\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"unknown evidence id should fail\");\n\n        assert!(err.to_string().contains(\"unknown evidence id\"));\n    }\n\n    #[test]\n    fn rejects_llm_batch_draft_with_template_residue() {\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batches = batch_chapter_evidence(&evidence);\n        let content = r#\"{\n            \"findings_summary_zh\": \"acme/rust-agent 是本章节的关键来源之一。\",\n            \"paragraphs\": [\n                {\n                    \"text_zh\": \"该项目展示了一个基础实现。\",\n                    \"cited_evidence_ids\": [\"ev-C1\"]\n                }\n            ]\n        }\"#;\n\n        let err = parse_batch_draft_content(content, &batches[0])\n            .expect_err(\"legacy template residue should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn batch_request_only_contains_current_batch_evidence() {\n        let chapter = sample_chapter();\n        let evidence = (1..=5)\n            .map(|index| sample_evidence(&format!(\"ev-C{index}\"), \"ch-1\"))\n            .collect::<Vec<_>>();\n        let batches = batch_chapter_evidence(&evidence);\n\n        let request = build_batch_draft_request(\n            &chapter,\n            &batches[1],\n            batches.len(),\n            &sample_deepseek_config(2048),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"\\\"batch_id\\\": 2\"));\n        assert!(user_message.contains(\"ev-C5\"));\n        assert!(!user_message.contains(\"ev-C1\"));\n        assert_eq!(request.max_tokens, Some(WRITER_MIN_OUTPUT_TOKENS as u32));\n    }\n\n    #[test]\n    fn writer_max_tokens_raises_too_small_config_for_reasoning_models() {\n        let config = sample_deepseek_config(32);\n\n        assert_eq!(writer_max_tokens(&config), WRITER_MIN_OUTPUT_TOKENS as u32);\n    }\n\n    #[test]\n    fn chapter_synthesis_request_uses_batch_findings_instead_of_raw_evidence() {\n        let chapter = sample_chapter();\n        let evidence = vec![sample_evidence(\"ev-C1\", \"ch-1\")];\n        let batch_drafts = vec![BatchDraft {\n            batch_id: 1,\n            findings_summary_zh: \"该批证据说明 README 中的 ReAct 主循环适合作为入门样例。\"\n                .to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该项目展示了 ReAct 主循环和工具注册。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_chapter_synthesis_request(\n            &chapter,\n            &evidence,\n            &batch_drafts,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"batch_drafts\"));\n        assert!(user_message.contains(\"allowed_evidence_ids\"));\n        assert!(user_message.contains(\"ReAct 主循环适合作为入门样例\"));\n        assert!(!user_message.contains(\"evidence_snippet\"));\n        assert!(!user_message.contains(\"README explains ReAct loop\"));\n    }\n\n    #[test]\n    fn parses_llm_global_summary_json() {\n        let content = r#\"{\n            \"global_summary_zh\": \"本报告显示，可控 TTS 入门调研应同时关注基础项目、说话人控制和模型框架边界；现有证据主要来自 GitHub README 与 arXiv 摘要，因此结论应限定在这些来源范围内。\"\n        }\"#;\n\n        let summary =\n            parse_global_summary_content(content).expect(\"valid global summary should parse\");\n\n        assert!(summary.contains(\"可控 TTS\"));\n        assert!(summary.contains(\"来源范围\"));\n    }\n\n    #[test]\n    fn rejects_llm_global_summary_with_template_residue() {\n        let content = r#\"{\n            \"global_summary_zh\": \"harinandanmv/text-to-speech 是本章节的关键来源之一。\"\n        }\"#;\n\n        let err =\n            parse_global_summary_content(content).expect_err(\"legacy template marker should fail\");\n\n        assert!(err.to_string().contains(\"legacy template marker\"));\n    }\n\n    #[test]\n    fn global_summary_request_uses_chapter_drafts() {\n        let chapter_drafts = vec![ChapterDraft {\n            chapter_id: \"ch-1\".to_string(),\n            title_zh: \"可控 TTS 基础模型\".to_string(),\n            paragraphs: vec![ParagraphWithCitations {\n                text_zh: \"该章节围绕基础 TTS 模型、说话人控制和入门项目进行综合分析。\".to_string(),\n                cited_evidence_ids: vec![\"ev-C1\".to_string()],\n            }],\n        }];\n\n        let request = build_global_summary_request(\n            \"TTS 入门级项目简单调研\",\n            &chapter_drafts,\n            1,\n            &sample_deepseek_config(4096),\n        )\n        .expect(\"request should build\");\n        let user_message = &request.messages[1].content;\n\n        assert!(user_message.contains(\"TTS 入门级项目简单调研\"));\n        assert!(user_message.contains(\"可控 TTS 基础模型\"));\n        assert!(user_message.contains(\"ev-C1\"));\n        assert!(user_message.contains(\"paragraph_summaries\"));\n    }\n\n    fn sample_chapter() -> ChapterNode {\n        ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }\n    }\n\n    fn sample_evidence(evidence_id: &str, chapter_id: &str) -> EvidenceItem {\n        EvidenceItem {\n            evidence_id: evidence_id.to_string(),\n            source_item_id: \"github:acme/rust-agent\".to_string(),\n            citation_id: \"C1\".to_string(),\n            chapter_ids: vec![chapter_id.to_string()],\n            query_attempt_ids: vec![\"gh-1\".to_string()],\n            source_kind: SourceKind::GitHub,\n            title: \"acme/rust-agent\".to_string(),\n            url: \"https://github.com/acme/rust-agent\".to_string(),\n            evidence_note_zh: \"GitHub 仓库 `acme/rust-agent`：Rust agent framework\".to_string(),\n            evidence_snippet: \"README explains ReAct loop, tool registry, and examples.\"\n                .to_string(),\n            support_score: None,\n        }\n    }\n\n    fn sample_deepseek_config(max_tokens: usize) -> DeepSeekConfig {\n        DeepSeekConfig {\n            api_key: \"sk-test\".to_string(),\n            base_url: \"https://api.deepseek.com\".to_string(),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: \"deepseek-v4-flash\".to_string(),\n            max_tokens,\n            timeout_secs: 120,\n        }\n    }\n\n    fn _dt() -> DateTime<Utc> {\n        DateTime::parse_from_rfc3339(\"2026-05-30T12:00:00Z\")\n            .unwrap()\n            .with_timezone(&Utc)\n    }\n}\n"}, "1d352f50bc41fcab4e021216bca591d876481974": {"t": 1785750667.6021082, "n": 3718, "h": "f933505dac8bd5ac", "c": "use chrono::Utc;\n\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft,\n};\n\npub fn draft_report(topic: &str, chapters: &[ChapterNode], memory: &EvidenceMemory) -> ReportDraft {\n    let chapter_drafts = chapters\n        .iter()\n        .map(|chapter| draft_chapter(chapter, &memory.by_chapter(&chapter.id)))\n        .collect::<Vec<_>>();\n    let global_summary_zh =\n        format!(\"本报告围绕 `{topic}`，基于当前程序已经抓取到的 GitHub 与 arXiv 证据生成。\");\n    ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    }\n}\n\nfn draft_chapter(chapter: &ChapterNode, evidence_items: &[EvidenceItem]) -> ChapterDraft {\n    let paragraphs = if evidence_items.is_empty() {\n        vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未收集到足够证据，建议查看 CoverageCritic 的缺口说明后决定是否增量研究。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }]\n    } else {\n        evidence_items\n            .iter()\n            .take(4)\n            .map(|item| ParagraphWithCitations {\n                text_zh: format!(\n                    \"`{}` 是本章节的关键来源之一。{} 来源链接：{}\",\n                    item.title, item.evidence_note_zh, item.url\n                ),\n                cited_evidence_ids: vec![item.evidence_id.clone()],\n            })\n            .collect()\n    };\n\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    }\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push(' ');\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[cfg(test)]\nmod tests {\n    use crate::model::{ChapterNode, EvidenceMemory};\n\n    use super::draft_report;\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = draft_report(\"Rust Agent\", &chapters, &EvidenceMemory::default());\n\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(draft.chapters[0].paragraphs[0]\n            .text_zh\n            .contains(\"尚未收集到足够证据\"));\n    }\n}\n"}, "7c0c9f55db1e84d4039b524e6a5c6712e65c973e": {"t": 1785750668.6996644, "n": 3722, "h": "c9c343cdf26f8b77", "c": "use chrono::Utc;\n\nuse crate::model::{\n    ChapterDraft, ChapterNode, CitationLedger, EvidenceItem, EvidenceMemory,\n    ParagraphWithCitations, ReportDraft,\n};\n\npub fn draft_report(topic: &str, chapters: &[ChapterNode], memory: &EvidenceMemory) -> ReportDraft {\n    let chapter_drafts = chapters\n        .iter()\n        .map(|chapter| draft_chapter(chapter, &memory.by_chapter(&chapter.id)))\n        .collect::<Vec<_>>();\n    let global_summary_zh =\n        format!(\"本报告围绕 `{topic}`，基于当前程序已经抓取到的 GitHub 与 arXiv 证据生成。\");\n    ReportDraft {\n        title_zh: format!(\"LitScout-RS 调研报告：{topic}\"),\n        chapters: chapter_drafts,\n        global_summary_zh,\n        written_at: Utc::now(),\n    }\n}\n\nfn draft_chapter(chapter: &ChapterNode, evidence_items: &[EvidenceItem]) -> ChapterDraft {\n    let paragraphs = if evidence_items.is_empty() {\n        vec![ParagraphWithCitations {\n            text_zh: format!(\n                \"当前章节 `{}` 尚未收集到足够证据，建议查看 CoverageCritic 的缺口说明后决定是否增量研究。\",\n                chapter.title_zh\n            ),\n            cited_evidence_ids: Vec::new(),\n        }]\n    } else {\n        evidence_items\n            .iter()\n            .take(4)\n            .map(|item| ParagraphWithCitations {\n                text_zh: format!(\n                    \"`{}` 是本章节的关键来源之一。{} 来源链接：{}\",\n                    item.title, item.evidence_note_zh, item.url\n                ),\n                cited_evidence_ids: vec![item.evidence_id.clone()],\n            })\n            .collect()\n    };\n\n    ChapterDraft {\n        chapter_id: chapter.id.clone(),\n        title_zh: chapter.title_zh.clone(),\n        paragraphs,\n    }\n}\n\npub fn render_report_markdown(draft: &ReportDraft, citations: &CitationLedger) -> String {\n    let mut out = String::new();\n    out.push_str(&format!(\"# {}\\n\\n\", draft.title_zh));\n    out.push_str(\"## 1. 全局摘要\\n\\n\");\n    out.push_str(&draft.global_summary_zh);\n    out.push_str(\"\\n\\n\");\n\n    for (index, chapter) in draft.chapters.iter().enumerate() {\n        out.push_str(&format!(\"## {}. {}\\n\\n\", index + 2, chapter.title_zh));\n        for paragraph in &chapter.paragraphs {\n            out.push_str(&paragraph.text_zh);\n            if !paragraph.cited_evidence_ids.is_empty() {\n                out.push_str(\" \");\n                out.push_str(\n                    &paragraph\n                        .cited_evidence_ids\n                        .iter()\n                        .map(|id| format!(\"`{id}`\"))\n                        .collect::<Vec<_>>()\n                        .join(\" \"),\n                );\n            }\n            out.push_str(\"\\n\\n\");\n        }\n    }\n\n    out.push_str(\"## 引用账本\\n\\n\");\n    if citations.citations.is_empty() {\n        out.push_str(\"- 暂无引用。\\n\");\n    } else {\n        for citation in &citations.citations {\n            out.push_str(&format!(\n                \"- `{}` [{}]({}) ({:?})\\n\",\n                citation.id, citation.title, citation.url, citation.source_kind\n            ));\n        }\n    }\n    out\n}\n\n#[cfg(test)]\nmod tests {\n    use crate::model::{ChapterNode, EvidenceMemory};\n\n    use super::draft_report;\n\n    #[test]\n    fn writer_drafts_empty_chapter_with_warning_text() {\n        let chapters = vec![ChapterNode {\n            id: \"ch-1\".to_string(),\n            parent_id: None,\n            title_zh: \"核心方向\".to_string(),\n            research_question: \"测试\".to_string(),\n            required_evidence_kinds: vec![\"github\".to_string()],\n            evidence_quota: 1,\n            sort_order: 1,\n        }];\n\n        let draft = draft_report(\"Rust Agent\", &chapters, &EvidenceMemory::default());\n\n        assert_eq!(draft.chapters.len(), 1);\n        assert!(draft.chapters[0].paragraphs[0]\n            .text_zh\n            .contains(\"尚未收集到足够证据\"));\n    }\n}\n"}}, "src/agent/orchestrator.rs": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1785746247.5815625, "n": 32597, "h": "d7a917273aee943f", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    academic_scout, arxiv_scout, citation_auditor, coverage_critic, evidence, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(\n        &trace,\n        &run.query_portfolio,\n        run.policy.academic_extra_enabled,\n    )\n    .await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let (academic_items, academic_attempts, academic_lineage) = if run.policy.academic_extra_enabled\n    {\n        academic_scout::scout_academic_extra(\n            &run.query_portfolio,\n            &app_config,\n            run.policy.academic_budget,\n            1,\n        )\n        .await?\n    } else {\n        (Vec::new(), Vec::new(), Vec::new())\n    };\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .chain(academic_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .chain(academic_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() && academic_items.is_empty() {\n        let message =\n            \"GitHub、arXiv 与已启用的扩展学术源均未返回可用结果，无法生成 stateful report。\"\n                .to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        academic_items,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let selected_total = evidence.memory.items.len();\n    let rejected_total = evidence.rejected_items.len();\n    let top_rejection_reasons = evidence\n        .selection_report\n        .rejection_reasons\n        .iter()\n        .take(3)\n        .map(|reason| format!(\"{}={}\", reason.reason, reason.count))\n        .collect::<Vec<_>>()\n        .join(\", \");\n    if run.policy.academic_extra_enabled\n        && evidence.selection_report.rejected_item_count > 0\n        && evidence.memory.items.iter().all(|item| {\n            !matches!(\n                item.source_kind,\n                crate::model::SourceKind::AcademicIndex | crate::model::SourceKind::Bibliography\n            )\n        })\n    {\n        run.warnings.push(\n            \"扩展学术源返回了候选结果，但 EvidenceQualityGate 未筛出可引用证据；报告将继续使用 GitHub/arXiv 证据。\"\n                .to_string(),\n        );\n    }\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceQualityGate ranked {ranked_total} item(s), accepted {selected_total}, rejected {rejected_total}, grouped {group_total}. Top rejection reasons: {}\",\n                if top_rejection_reasons.is_empty() {\n                    \"none\".to_string()\n                } else {\n                    top_rejection_reasons\n                }\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(\n    trace: &TraceWriter,\n    portfolio: &[QueryPortfolio],\n    include_academic_extra: bool,\n) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        if include_academic_extra {\n            for query in academic_queries_for_trace(item) {\n                for source in academic_scout::academic_source_names() {\n                    trace\n                        .append(&TraceEvent::ToolCallStarted {\n                            tool: source.to_string(),\n                            query: query.clone(),\n                            at: Utc::now(),\n                        })\n                        .await?;\n                }\n            }\n        }\n    }\n    Ok(())\n}\n\nfn academic_queries_for_trace(item: &QueryPortfolio) -> Vec<String> {\n    let mut queries = item.arxiv_queries.clone();\n    for query in &item.github_queries {\n        if !queries.iter().any(|existing| existing == query) {\n            queries.push(query.clone());\n        }\n    }\n    queries\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            openalex_api_key: None,\n            crossref_mailto: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "b8f76fdd9fdfaaf0957a9293477eb193d392c98f": {"t": 1785750657.6714153, "n": 32597, "h": "d7a917273aee943f", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    academic_scout, arxiv_scout, citation_auditor, coverage_critic, evidence, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(\n        &trace,\n        &run.query_portfolio,\n        run.policy.academic_extra_enabled,\n    )\n    .await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let (academic_items, academic_attempts, academic_lineage) = if run.policy.academic_extra_enabled\n    {\n        academic_scout::scout_academic_extra(\n            &run.query_portfolio,\n            &app_config,\n            run.policy.academic_budget,\n            1,\n        )\n        .await?\n    } else {\n        (Vec::new(), Vec::new(), Vec::new())\n    };\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .chain(academic_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .chain(academic_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() && academic_items.is_empty() {\n        let message =\n            \"GitHub、arXiv 与已启用的扩展学术源均未返回可用结果，无法生成 stateful report。\"\n                .to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        academic_items,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let selected_total = evidence.memory.items.len();\n    let rejected_total = evidence.rejected_items.len();\n    let top_rejection_reasons = evidence\n        .selection_report\n        .rejection_reasons\n        .iter()\n        .take(3)\n        .map(|reason| format!(\"{}={}\", reason.reason, reason.count))\n        .collect::<Vec<_>>()\n        .join(\", \");\n    if run.policy.academic_extra_enabled\n        && evidence.selection_report.rejected_item_count > 0\n        && evidence.memory.items.iter().all(|item| {\n            !matches!(\n                item.source_kind,\n                crate::model::SourceKind::AcademicIndex | crate::model::SourceKind::Bibliography\n            )\n        })\n    {\n        run.warnings.push(\n            \"扩展学术源返回了候选结果，但 EvidenceQualityGate 未筛出可引用证据；报告将继续使用 GitHub/arXiv 证据。\"\n                .to_string(),\n        );\n    }\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceQualityGate ranked {ranked_total} item(s), accepted {selected_total}, rejected {rejected_total}, grouped {group_total}. Top rejection reasons: {}\",\n                if top_rejection_reasons.is_empty() {\n                    \"none\".to_string()\n                } else {\n                    top_rejection_reasons\n                }\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(\n    trace: &TraceWriter,\n    portfolio: &[QueryPortfolio],\n    include_academic_extra: bool,\n) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        if include_academic_extra {\n            for query in academic_queries_for_trace(item) {\n                for source in academic_scout::academic_source_names() {\n                    trace\n                        .append(&TraceEvent::ToolCallStarted {\n                            tool: source.to_string(),\n                            query: query.clone(),\n                            at: Utc::now(),\n                        })\n                        .await?;\n                }\n            }\n        }\n    }\n    Ok(())\n}\n\nfn academic_queries_for_trace(item: &QueryPortfolio) -> Vec<String> {\n    let mut queries = item.arxiv_queries.clone();\n    for query in &item.github_queries {\n        if !queries.iter().any(|existing| existing == query) {\n            queries.push(query.clone());\n        }\n    }\n    queries\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            openalex_api_key: None,\n            crossref_mailto: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "f0e381df64f2448f45cfc7566949703029fd6216": {"t": 1785750658.9217067, "n": 31486, "h": "21e9ff2c1f1ca676", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    academic_scout, arxiv_scout, citation_auditor, coverage_critic, evidence, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(\n        &trace,\n        &run.query_portfolio,\n        run.policy.academic_extra_enabled,\n    )\n    .await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let (academic_items, academic_attempts, academic_lineage) = if run.policy.academic_extra_enabled\n    {\n        academic_scout::scout_academic_extra(\n            &run.query_portfolio,\n            &app_config,\n            run.policy.academic_budget,\n            1,\n        )\n        .await?\n    } else {\n        (Vec::new(), Vec::new(), Vec::new())\n    };\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .chain(academic_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .chain(academic_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() && academic_items.is_empty() {\n        let message =\n            \"GitHub、arXiv 与已启用的扩展学术源均未返回可用结果，无法生成 stateful report。\"\n                .to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        academic_items,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(\n    trace: &TraceWriter,\n    portfolio: &[QueryPortfolio],\n    include_academic_extra: bool,\n) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        if include_academic_extra {\n            for query in academic_queries_for_trace(item) {\n                for source in academic_scout::academic_source_names() {\n                    trace\n                        .append(&TraceEvent::ToolCallStarted {\n                            tool: source.to_string(),\n                            query: query.clone(),\n                            at: Utc::now(),\n                        })\n                        .await?;\n                }\n            }\n        }\n    }\n    Ok(())\n}\n\nfn academic_queries_for_trace(item: &QueryPortfolio) -> Vec<String> {\n    let mut queries = item.arxiv_queries.clone();\n    for query in &item.github_queries {\n        if !queries.iter().any(|existing| existing == query) {\n            queries.push(query.clone());\n        }\n    }\n    queries\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            openalex_api_key: None,\n            crossref_mailto: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "89599bc6feef2b153f24871f481493e359e6ff87": {"t": 1785750660.2229273, "n": 31563, "h": "2c1fee48845b3f72", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    academic_scout, arxiv_scout, citation_auditor, coverage_critic, evidence, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(\n        &trace,\n        &run.query_portfolio,\n        run.policy.academic_extra_enabled,\n    )\n    .await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let (academic_items, academic_attempts, academic_lineage) = if run.policy.academic_extra_enabled\n    {\n        academic_scout::scout_academic_extra(\n            &run.query_portfolio,\n            &app_config,\n            run.policy.academic_budget,\n            1,\n        )\n        .await?\n    } else {\n        (Vec::new(), Vec::new(), Vec::new())\n    };\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .chain(academic_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .chain(academic_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() && academic_items.is_empty() {\n        let message =\n            \"GitHub、arXiv 与已启用的扩展学术源均未返回可用结果，无法生成 stateful report。\"\n                .to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        academic_items,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(\n    trace: &TraceWriter,\n    portfolio: &[QueryPortfolio],\n    include_academic_extra: bool,\n) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        if include_academic_extra {\n            for query in academic_queries_for_trace(item) {\n                trace\n                    .append(&TraceEvent::ToolCallStarted {\n                        tool: \"semantic_scholar\".to_string(),\n                        query: query.clone(),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                trace\n                    .append(&TraceEvent::ToolCallStarted {\n                        tool: \"dblp\".to_string(),\n                        query,\n                        at: Utc::now(),\n                    })\n                    .await?;\n            }\n        }\n    }\n    Ok(())\n}\n\nfn academic_queries_for_trace(item: &QueryPortfolio) -> Vec<String> {\n    let mut queries = item.arxiv_queries.clone();\n    for query in &item.github_queries {\n        if !queries.iter().any(|existing| existing == query) {\n            queries.push(query.clone());\n        }\n    }\n    queries\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            semantic_scholar_api_key: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "638e49c5b42c5e902379c23b653519ba0ed59bda": {"t": 1785750661.467633, "n": 29922, "h": "2c1615c4dedf12ef", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    arxiv_scout, citation_auditor, coverage_critic, evidence, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(&trace, &run.query_portfolio).await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() {\n        let message = \"GitHub 与 arXiv 均未返回可用结果，无法生成 stateful report。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(trace: &TraceWriter, portfolio: &[QueryPortfolio]) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n    }\n    Ok(())\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "bccd8c422afb8962e69d703d2ddcbdac277a8101": {"t": 1785750666.4549646, "n": 30434, "h": "b5423b0d0473eced", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    arxiv_scout, citation_auditor, coverage_critic, evidence, followup_router, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(&trace, &run.query_portfolio).await?;\n\n    let mut github_config = app_config.clone();\n    github_config.enrich = app_config.enrich || run.policy.allow_github_enrich;\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &github_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts, github_lineage) = github_result?;\n    let (arxiv_papers, arxiv_attempts, arxiv_lineage) = arxiv_result?;\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    let source_lineage = github_lineage\n        .into_iter()\n        .chain(arxiv_lineage)\n        .collect::<Vec<_>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() {\n        let message = \"GitHub 与 arXiv 均未返回可用结果，无法生成 stateful report。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        query_attempts,\n        source_lineage,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let writer_input = serde_json::to_string(&(&run.topic, &run.chapters, &evidence.memory))?;\n    let writer_ctx = agent_context(\"writer\", writer_input.clone(), &run.policy, Vec::new());\n    if !llm_config_can_call(&llm_config) {\n        let message = \"Writer LLM 需要 DeepSeek API key；拒绝生成模板化最终报告。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::InvalidConfig(message));\n    }\n    let writer_model = llm_config.main_model.clone();\n    let writer_budget = TokenBudgetTracker::new();\n    writer_budget.before_llm_call(&writer_ctx)?;\n    trace_llm_started(&trace, \"writer\", writer_model, &writer_input).await?;\n    let draft_result =\n        writer::draft_report_with_llm(&run.topic, &run.chapters, &evidence.memory, &llm_config)\n            .await;\n    let draft = match draft_result {\n        Ok(draft) => draft,\n        Err(err) => {\n            let message = format!(\"Writer LLM 写作失败，拒绝生成模板化最终报告：{err}\");\n            run.warnings.push(message.clone());\n            trace\n                .append(&TraceEvent::QualityWarning {\n                    message: message.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n            transition(\n                run,\n                &trace,\n                &app_config,\n                ResearchRunState::Failed,\n                tx.as_ref(),\n            )\n            .await?;\n            let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n            trace_checkpoint(&trace, &checkpoint).await?;\n            send_checkpoint(tx.as_ref(), &checkpoint).await;\n            if let Some(tx) = tx.as_ref() {\n                let _ = tx\n                    .send(StatefulRunEvent::Failed {\n                        error: message.clone(),\n                    })\n                    .await;\n            }\n            return Err(AppError::Llm(message));\n        }\n    };\n    let draft_json = serde_json::to_string(&draft)?;\n    trace_llm_finished(&trace, \"writer\", &writer_input, &draft_json).await?;\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\npub fn route_followup_for_run(\n    question: &str,\n    memory: Option<&EvidenceMemory>,\n    report_markdown: Option<&str>,\n) -> followup_router::FollowupRoute {\n    match (memory, report_markdown) {\n        (Some(memory), Some(markdown)) => {\n            followup_router::route_followup(question, memory, markdown)\n        }\n        _ => followup_router::FollowupRoute::IncrementalResearchRequired {\n            reason: \"当前 run 尚未完成 evidence/report，无法基于现有证据回答。\".to_string(),\n        },\n    }\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(trace: &TraceWriter, portfolio: &[QueryPortfolio]) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n    }\n    Ok(())\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "1d352f50bc41fcab4e021216bca591d876481974": {"t": 1785750667.6862836, "n": 27615, "h": "d952740a326690ef", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    arxiv_scout, citation_auditor, coverage_critic, evidence, followup_router, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_agent_decision(\n        &trace,\n        \"scoper\",\n        &scoper_input,\n        &brief_json,\n        \"deterministic ResearchBrief scoping\",\n    )\n    .await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    let plan = if llm_config_can_call(&llm_config) {\n        token_budget.before_llm_call(&planner_ctx)?;\n        trace_llm_started(\n            &trace,\n            \"planner\",\n            side_model_name(&llm_config),\n            &planner_input,\n        )\n        .await?;\n        match planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await {\n            Ok(plan) => {\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n                plan\n            }\n            Err(err) => {\n                let reason = err.to_string();\n                trace\n                    .append(&TraceEvent::QualityWarning {\n                        message: format!(\"Planner LLM 调用失败，已降级为确定性默认计划：{reason}\"),\n                        at: Utc::now(),\n                    })\n                    .await?;\n                let plan = fallback_chapter_plan(&brief.topic, &run.policy, Some(&reason));\n                let plan_json = serde_json::to_string(&plan)?;\n                json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n                trace_agent_decision(\n                    &trace,\n                    \"planner_fallback\",\n                    &planner_input,\n                    &plan_json,\n                    \"planner LLM failed; deterministic default plan\",\n                )\n                .await?;\n                plan\n            }\n        }\n    } else {\n        let reason = llm_config\n            .enabled\n            .then_some(\"LLM 已启用但缺少 API key，已使用确定性默认计划。\");\n        let plan = fallback_chapter_plan(&brief.topic, &run.policy, reason);\n        let plan_json = serde_json::to_string(&plan)?;\n        json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n        trace_agent_decision(\n            &trace,\n            \"planner\",\n            &planner_input,\n            &plan_json,\n            \"deterministic default plan\",\n        )\n        .await?;\n        plan\n    };\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_agent_decision(\n        &trace,\n        \"plan_critic\",\n        &critic_input,\n        &critique_json,\n        \"deterministic plan consistency checks\",\n    )\n    .await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} real LLM request attempt(s); deterministic nodes are traced as AgentDecision events.\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    _llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(&trace, &run.query_portfolio).await?;\n\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &app_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts) = github_result?;\n    let (arxiv_papers, arxiv_attempts) = arxiv_result?;\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts)\n        .collect::<Vec<QueryAttempt>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() {\n        let message = \"GitHub 与 arXiv 均未返回可用结果，无法生成 stateful report。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        query_attempts,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let draft = writer::draft_report(&run.topic, &run.chapters, &evidence.memory);\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\npub fn route_followup_for_run(\n    question: &str,\n    memory: Option<&EvidenceMemory>,\n    report_markdown: Option<&str>,\n) -> followup_router::FollowupRoute {\n    match (memory, report_markdown) {\n        (Some(memory), Some(markdown)) => {\n            followup_router::route_followup(question, memory, markdown)\n        }\n        _ => followup_router::FollowupRoute::IncrementalResearchRequired {\n            reason: \"当前 run 尚未完成 evidence/report，无法基于现有证据回答。\".to_string(),\n        },\n    }\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_agent_decision(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n    rationale: &str,\n) -> Result<()> {\n    let (input_hash, output_hash, _tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::AgentDecision {\n            actor: actor.to_string(),\n            input_hash,\n            output_hash,\n            rationale: rationale.to_string(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(trace: &TraceWriter, portfolio: &[QueryPortfolio]) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n    }\n    Ok(())\n}\n\nfn llm_config_can_call(llm_config: &LlmConfig) -> bool {\n    llm_config.enabled\n        && llm_config\n            .api_key\n            .as_deref()\n            .map(str::trim)\n            .is_some_and(|key| !key.is_empty())\n}\n\nfn fallback_chapter_plan(\n    topic: &str,\n    policy: &RunPolicy,\n    warning: Option<&str>,\n) -> planner::ChapterPlanOutput {\n    let default_plan = planner::default_plan(topic, policy.github_budget, policy.arxiv_budget);\n    let mut plan = planner::chapter_plan_from_plan_output(&default_plan, policy);\n    if let Some(warning) = warning {\n        plan.warnings.push(warning.to_string());\n    }\n    plan\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n\n#[cfg(test)]\nmod tests {\n    use std::time::{SystemTime, UNIX_EPOCH};\n\n    use super::create_run;\n    use crate::config::{AppConfig, LlmConfig};\n    use crate::run_policy::RunPolicy;\n    use crate::workflow_state::ResearchRunState;\n\n    #[tokio::test]\n    async fn create_run_falls_back_without_llm_key_and_traces_rule_nodes() {\n        let root = temp_root(\"stage3-create-run-fallback\");\n        let app_config = AppConfig {\n            github_token: None,\n            output: root.join(\"reports\"),\n            cache_dir: root.join(\"cache\"),\n            session_dir: root.join(\"sessions\"),\n            tags_file: None,\n            use_cache: false,\n            cache_ttl_hours: 24,\n            timeout_secs: 30,\n            enrich: false,\n        };\n        let llm_config = LlmConfig {\n            enabled: true,\n            api_key: None,\n            base_url: Some(\"https://api.deepseek.com\".to_string()),\n            main_model: \"deepseek-v4-pro\".to_string(),\n            side_model: Some(\"deepseek-v4-flash\".to_string()),\n            max_tokens: 4096,\n            timeout_secs: 30,\n        };\n\n        let run = create_run(\n            \"Rust agent framework\".to_string(),\n            app_config.clone(),\n            llm_config,\n            RunPolicy::default(),\n        )\n        .await\n        .expect(\"create_run should degrade to a deterministic plan\");\n\n        assert_eq!(run.state, ResearchRunState::PlanReady);\n        assert!(run\n            .plan_warnings\n            .iter()\n            .any(|warning| warning.contains(\"API key\")));\n        assert!(run\n            .warnings\n            .iter()\n            .any(|warning| warning.contains(\"0 real LLM request attempt\")));\n\n        let trace_path = app_config.session_dir.join(&run.run_id).join(\"trace.jsonl\");\n        let trace = std::fs::read_to_string(trace_path).expect(\"trace should be readable\");\n        assert!(trace.contains(\"\\\"event\\\":\\\"agent_decision\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"scoper\\\"\"));\n        assert!(trace.contains(\"\\\"actor\\\":\\\"plan_critic\\\"\"));\n        assert!(!trace.contains(\"\\\"event\\\":\\\"llm_request_started\\\"\"));\n    }\n\n    fn temp_root(name: &str) -> std::path::PathBuf {\n        let unique = SystemTime::now()\n            .duration_since(UNIX_EPOCH)\n            .unwrap()\n            .as_nanos();\n        std::env::temp_dir().join(format!(\n            \"litscout-rs-{name}-{}-{unique}\",\n            std::process::id()\n        ))\n    }\n}\n"}, "7c0c9f55db1e84d4039b524e6a5c6712e65c973e": {"t": 1785750668.8744934, "n": 22717, "h": "1e011587edacd5e6", "c": "use std::path::{Path, PathBuf};\n\nuse chrono::Utc;\nuse tokio::sync::mpsc;\nuse uuid::Uuid;\n\nuse crate::agent::{\n    arxiv_scout, citation_auditor, coverage_critic, evidence, followup_router, github_scout,\n    middleware::{\n        AgentContext, AgentMiddleware, CitationGuard, JsonSchemaGuard, TokenBudgetTracker,\n    },\n    plan_critic, planner, scoper, writer,\n};\nuse crate::checkpoint::{self, Checkpoint, CheckpointSnapshot};\nuse crate::config::{AppConfig, LlmConfig};\nuse crate::error::{AppError, Result};\nuse crate::model::{\n    ChapterNode, CoverageReport, EvidenceMemory, QueryAttempt, QueryPortfolio, SearchQuery,\n};\nuse crate::run_policy::RunPolicy;\nuse crate::trace::{stable_hash, token_estimate, TraceEvent, TraceWriter};\nuse crate::workflow_state::{ResearchRunRecord, ResearchRunState};\n\n#[derive(Debug, Clone, serde::Serialize)]\n#[serde(tag = \"event\", content = \"data\", rename_all = \"snake_case\")]\npub enum StatefulRunEvent {\n    StateChanged {\n        state: ResearchRunState,\n    },\n    CheckpointCreated {\n        checkpoint_id: String,\n        state: ResearchRunState,\n    },\n    EvidenceReady {\n        total: usize,\n    },\n    CoverageReady {\n        gaps: usize,\n    },\n    CitationAuditReady {\n        citation_coverage_ratio: f64,\n    },\n    Completed {\n        run_id: String,\n    },\n    Failed {\n        error: String,\n    },\n}\n\npub async fn create_run(\n    topic: String,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    policy: RunPolicy,\n) -> Result<ResearchRunRecord> {\n    let run_id = format!(\"run-{}\", Uuid::new_v4());\n    let run_dir = run_dir(&app_config, &run_id);\n    tokio::fs::create_dir_all(&run_dir).await?;\n    let trace = TraceWriter::new(&run_dir).await?;\n    let mut run = ResearchRunRecord::new(run_id, topic.clone(), policy.bounded());\n    save_run(&app_config, &run).await?;\n    let token_budget = TokenBudgetTracker::new();\n    let json_guard = JsonSchemaGuard;\n\n    let scoper_input = format!(\"topic={topic}; policy={:?}\", run.policy);\n    let scoper_ctx = agent_context(\"scoper\", scoper_input.clone(), &run.policy, Vec::new());\n    token_budget.before_llm_call(&scoper_ctx)?;\n    trace_llm_started(\n        &trace,\n        \"scoper\",\n        side_model_name(&llm_config),\n        &scoper_input,\n    )\n    .await?;\n    let brief = scoper::generate_research_brief(&topic, &run.policy, &llm_config).await?;\n    let brief_json = serde_json::to_string(&brief)?;\n    json_guard.after_llm_call(&scoper_ctx, &brief_json)?;\n    token_budget.after_llm_call(&scoper_ctx, &brief_json)?;\n    trace_llm_finished(&trace, \"scoper\", &scoper_input, &brief_json).await?;\n\n    let planner_input = serde_json::to_string(&brief)?;\n    let planner_ctx = agent_context(\"planner\", planner_input.clone(), &run.policy, Vec::new());\n    token_budget.before_llm_call(&planner_ctx)?;\n    trace_llm_started(\n        &trace,\n        \"planner\",\n        side_model_name(&llm_config),\n        &planner_input,\n    )\n    .await?;\n    let plan = planner::generate_chapter_plan(&brief, &run.policy, &llm_config).await?;\n    let plan_json = serde_json::to_string(&plan)?;\n    json_guard.after_llm_call(&planner_ctx, &plan_json)?;\n    token_budget.after_llm_call(&planner_ctx, &plan_json)?;\n    trace_llm_finished(&trace, \"planner\", &planner_input, &plan_json).await?;\n\n    let critic_input = serde_json::to_string(&(&brief, &plan.chapters, &plan.query_portfolio))?;\n    let critic_ctx = agent_context(\"plan_critic\", critic_input.clone(), &run.policy, Vec::new());\n    token_budget.before_llm_call(&critic_ctx)?;\n    trace_llm_started(\n        &trace,\n        \"plan_critic\",\n        side_model_name(&llm_config),\n        &critic_input,\n    )\n    .await?;\n    let critique =\n        plan_critic::critique_plan(&brief, &plan.chapters, &plan.query_portfolio, &run.policy);\n    let critique_json = serde_json::to_string(&critique)?;\n    json_guard.after_llm_call(&critic_ctx, &critique_json)?;\n    token_budget.after_llm_call(&critic_ctx, &critique_json)?;\n    trace_llm_finished(&trace, \"plan_critic\", &critic_input, &critique_json).await?;\n    let from = run.state.clone();\n    if !run.transition_to(ResearchRunState::PlanReady) {\n        return Err(invalid_transition(&from, &ResearchRunState::PlanReady));\n    }\n    run.brief = Some(brief);\n    run.chapters = plan.chapters;\n    run.query_portfolio = plan.query_portfolio;\n    run.plan_warnings = plan\n        .warnings\n        .into_iter()\n        .chain(critique.warnings)\n        .chain(critique.suggestions)\n        .collect();\n    run.warnings.push(format!(\n        \"Stage 3 agent planning used {} bounded LLM node call(s).\",\n        token_budget.calls()\n    ));\n    trace_transition(&trace, from, ResearchRunState::PlanReady).await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn continue_run(\n    run_id: &str,\n    app_config: AppConfig,\n    llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(&app_config, run_id).await?;\n    match run.state {\n        ResearchRunState::PlanReady => {\n            continue_from_plan_ready(&mut run, app_config, llm_config, tx).await\n        }\n        ResearchRunState::Completed => Ok(run),\n        _ => Err(AppError::Workflow(format!(\n            \"run `{run_id}` is in state {:?} and cannot continue from this endpoint\",\n            run.state\n        ))),\n    }\n}\n\nasync fn continue_from_plan_ready(\n    run: &mut ResearchRunRecord,\n    app_config: AppConfig,\n    _llm_config: LlmConfig,\n    tx: Option<mpsc::Sender<StatefulRunEvent>>,\n) -> Result<ResearchRunRecord> {\n    let run_dir = run_dir(&app_config, &run.run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Fetching,\n        tx.as_ref(),\n    )\n    .await?;\n    trace_planned_tool_calls(&trace, &run.query_portfolio).await?;\n\n    let (github_result, arxiv_result) = tokio::join!(\n        github_scout::scout_github(&run.query_portfolio, &app_config, 1),\n        arxiv_scout::scout_arxiv(&run.query_portfolio, &app_config, 1)\n    );\n    let (github_repos, github_attempts) = github_result?;\n    let (arxiv_papers, arxiv_attempts) = arxiv_result?;\n    let query_attempts = github_attempts\n        .into_iter()\n        .chain(arxiv_attempts.into_iter())\n        .collect::<Vec<QueryAttempt>>();\n    for attempt in &query_attempts {\n        trace\n            .append(&TraceEvent::ToolCallFinished {\n                tool: attempt.source.clone(),\n                result_count: attempt.result_count,\n                error: attempt.error.clone(),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    if github_repos.is_empty() && arxiv_papers.is_empty() {\n        let message = \"GitHub 与 arXiv 均未返回可用结果，无法生成 stateful report。\".to_string();\n        run.warnings.push(message.clone());\n        transition(\n            run,\n            &trace,\n            &app_config,\n            ResearchRunState::Failed,\n            tx.as_ref(),\n        )\n        .await?;\n        let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n        trace_checkpoint(&trace, &checkpoint).await?;\n        send_checkpoint(tx.as_ref(), &checkpoint).await;\n        if let Some(tx) = tx.as_ref() {\n            let _ = tx\n                .send(StatefulRunEvent::Failed {\n                    error: message.clone(),\n                })\n                .await;\n        }\n        return Err(AppError::Workflow(message));\n    }\n\n    let query = SearchQuery {\n        topic: run.topic.clone(),\n        github_limit: run.policy.github_budget,\n        arxiv_limit: run.policy.arxiv_budget,\n    };\n    let evidence = evidence::build_evidence_memory(\n        &query,\n        &app_config,\n        github_repos,\n        arxiv_papers,\n        query_attempts,\n    )?;\n    let ranked_total = evidence.ranked_items.len();\n    let group_total = evidence.groups.len();\n    let coverage = if run.policy.skip_coverage_critic {\n        CoverageReport::pass()\n    } else {\n        coverage_critic::evaluate_coverage(\n            &run.chapters,\n            &run.query_portfolio,\n            &evidence.memory,\n            &run.policy,\n        )\n    };\n    for gap in &coverage.gaps {\n        trace\n            .append(&TraceEvent::CoverageGapDetected {\n                chapter_id: gap.chapter_id.clone(),\n                gap_kind: format!(\"{:?}\", gap.gap_kind),\n                at: Utc::now(),\n            })\n            .await?;\n    }\n    run.evidence_memory = Some(evidence.memory.clone());\n    run.coverage_report = Some(coverage.clone());\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::EvidenceReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::EvidenceReady {\n                total: evidence.memory.items.len(),\n            })\n            .await;\n        let _ = tx\n            .send(StatefulRunEvent::CoverageReady {\n                gaps: coverage.gaps.len(),\n            })\n            .await;\n    }\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: format!(\n                \"EvidenceBuilder ranked {ranked_total} item(s) into {group_total} group(s).\"\n            ),\n            at: Utc::now(),\n        })\n        .await?;\n\n    let draft = writer::draft_report(&run.topic, &run.chapters, &evidence.memory);\n    let audit = if run.policy.require_citation_audit {\n        citation_auditor::audit_citations(&draft, &evidence.citations, &evidence.memory)\n    } else {\n        crate::model::CitationAuditReport::pass()\n    };\n    let markdown = writer::render_report_markdown(&draft, &evidence.citations);\n    let citation_ctx = agent_context(\n        \"writer\",\n        \"render_report_markdown\".to_string(),\n        &run.policy,\n        evidence\n            .citations\n            .citations\n            .iter()\n            .map(|citation| citation.url.clone())\n            .collect(),\n    );\n    CitationGuard.after_llm_call(&citation_ctx, &markdown)?;\n    run.report_draft = Some(draft);\n    run.citation_audit = Some(audit.clone());\n    run.report_markdown = Some(markdown.clone());\n    if !audit.url_whitelist_passed {\n        run.warnings\n            .push(\"Citation audit failed URL whitelist check.\".to_string());\n    }\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::SynthesisReady,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx.as_ref() {\n        let _ = tx\n            .send(StatefulRunEvent::CitationAuditReady {\n                citation_coverage_ratio: audit.citation_coverage_ratio,\n            })\n            .await;\n    }\n\n    write_report_file(&app_config, run, &markdown).await?;\n    transition(\n        run,\n        &trace,\n        &app_config,\n        ResearchRunState::Completed,\n        tx.as_ref(),\n    )\n    .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    send_checkpoint(tx.as_ref(), &checkpoint).await;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::Completed {\n                run_id: run.run_id.clone(),\n            })\n            .await;\n    }\n    Ok(run.clone())\n}\n\npub async fn load_run(app_config: &AppConfig, run_id: &str) -> Result<ResearchRunRecord> {\n    let body = tokio::fs::read_to_string(run_summary_path(app_config, run_id)).await?;\n    Ok(serde_json::from_str(&body)?)\n}\n\npub async fn save_run(app_config: &AppConfig, run: &ResearchRunRecord) -> Result<()> {\n    tokio::fs::create_dir_all(&app_config.session_dir).await?;\n    let body = serde_json::to_string_pretty(run)?;\n    tokio::fs::write(run_summary_path(app_config, &run.run_id), body).await?;\n    Ok(())\n}\n\npub async fn evidence_for_run(app_config: &AppConfig, run_id: &str) -> Result<EvidenceMemory> {\n    load_run(app_config, run_id)\n        .await?\n        .evidence_memory\n        .ok_or_else(|| AppError::Workflow(\"run has no EvidenceMemory yet\".to_string()))\n}\n\npub async fn coverage_for_run(app_config: &AppConfig, run_id: &str) -> Result<CoverageReport> {\n    load_run(app_config, run_id)\n        .await?\n        .coverage_report\n        .ok_or_else(|| AppError::Workflow(\"run has no CoverageReport yet\".to_string()))\n}\n\npub async fn revise_plan(\n    app_config: &AppConfig,\n    run_id: &str,\n    chapters: Option<Vec<ChapterNode>>,\n    query_portfolio: Option<Vec<QueryPortfolio>>,\n    user_feedback: Option<String>,\n) -> Result<ResearchRunRecord> {\n    let mut run = load_run(app_config, run_id).await?;\n    if !matches!(run.state, ResearchRunState::PlanReady) {\n        return Err(AppError::Workflow(\n            \"only PlanReady runs can revise the Stage 3 plan\".to_string(),\n        ));\n    }\n    if let Some(chapters) = chapters {\n        if chapters.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised chapter plan cannot be empty\".to_string(),\n            ));\n        }\n        run.chapters = chapters;\n    }\n    if let Some(query_portfolio) = query_portfolio {\n        if query_portfolio.is_empty() {\n            return Err(AppError::Workflow(\n                \"revised query portfolio cannot be empty\".to_string(),\n            ));\n        }\n        run.query_portfolio = query_portfolio;\n    }\n    if let Some(feedback) = user_feedback.map(|value| value.trim().to_string()) {\n        if !feedback.is_empty() {\n            run.plan_warnings\n                .push(format!(\"用户修订计划备注：{feedback}\"));\n        }\n    }\n    run.updated_at = Utc::now();\n    let run_dir = run_dir(app_config, run_id);\n    let trace = TraceWriter::new(&run_dir).await?;\n    trace\n        .append(&TraceEvent::QualityWarning {\n            message: \"用户已修订 PlanReady 搜索计划，并创建新的 checkpoint。\".to_string(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&run_dir, &run).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(app_config, &run).await?;\n    Ok(run)\n}\n\npub async fn branch_from_plan_ready(\n    app_config: AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<ResearchRunRecord> {\n    let snapshot = load_checkpoint_snapshot_by_id(&app_config, run_id, checkpoint_id).await?;\n    if !snapshot.checkpoint.rollback_allowed\n        || !matches!(snapshot.checkpoint.state, ResearchRunState::PlanReady)\n    {\n        return Err(AppError::Workflow(\n            \"Stage 3 currently supports rollback branches only from PlanReady checkpoints\"\n                .to_string(),\n        ));\n    }\n    let mut branch = snapshot.run;\n    branch.run_id = format!(\"run-{}\", Uuid::new_v4());\n    branch.created_at = Utc::now();\n    branch.updated_at = Utc::now();\n    branch.state = ResearchRunState::PlanReady;\n    branch.origin_run_id = Some(run_id.to_string());\n    branch.origin_checkpoint_id = Some(checkpoint_id.to_string());\n    branch.evidence_memory = None;\n    branch.coverage_report = None;\n    branch.report_draft = None;\n    branch.citation_audit = None;\n    branch.report_markdown = None;\n    branch.output_report = None;\n    let branch_dir = run_dir(&app_config, &branch.run_id);\n    tokio::fs::create_dir_all(&branch_dir).await?;\n    let trace = TraceWriter::new(&branch_dir).await?;\n    trace\n        .append(&TraceEvent::RollbackBranchCreated {\n            origin_run_id: run_id.to_string(),\n            new_run_id: branch.run_id.clone(),\n            at: Utc::now(),\n        })\n        .await?;\n    let checkpoint = checkpoint::write_checkpoint(&branch_dir, &branch).await?;\n    trace_checkpoint(&trace, &checkpoint).await?;\n    save_run(&app_config, &branch).await?;\n    Ok(branch)\n}\n\nasync fn load_checkpoint_snapshot_by_id(\n    app_config: &AppConfig,\n    run_id: &str,\n    checkpoint_id: &str,\n) -> Result<CheckpointSnapshot> {\n    let run_dir = run_dir(app_config, run_id);\n    let checkpoint_dir = run_dir.join(\"checkpoints\");\n    let mut entries = tokio::fs::read_dir(checkpoint_dir).await?;\n    while let Some(entry) = entries.next_entry().await? {\n        let body = tokio::fs::read_to_string(entry.path()).await?;\n        let snapshot: CheckpointSnapshot = serde_json::from_str(&body)?;\n        if snapshot.checkpoint.checkpoint_id == checkpoint_id {\n            return checkpoint::read_checkpoint(&run_dir, &snapshot.checkpoint).await;\n        }\n    }\n    Err(AppError::Workflow(format!(\n        \"checkpoint `{checkpoint_id}` was not found for run `{run_id}`\"\n    )))\n}\n\npub fn route_followup_for_run(\n    question: &str,\n    memory: Option<&EvidenceMemory>,\n    report_markdown: Option<&str>,\n) -> followup_router::FollowupRoute {\n    match (memory, report_markdown) {\n        (Some(memory), Some(markdown)) => {\n            followup_router::route_followup(question, memory, markdown)\n        }\n        _ => followup_router::FollowupRoute::IncrementalResearchRequired {\n            reason: \"当前 run 尚未完成 evidence/report，无法基于现有证据回答。\".to_string(),\n        },\n    }\n}\n\nasync fn transition(\n    run: &mut ResearchRunRecord,\n    trace: &TraceWriter,\n    app_config: &AppConfig,\n    next: ResearchRunState,\n    tx: Option<&mpsc::Sender<StatefulRunEvent>>,\n) -> Result<()> {\n    let from = run.state.clone();\n    if !run.transition_to(next.clone()) {\n        return Err(invalid_transition(&from, &next));\n    }\n    trace_transition(trace, from, next.clone()).await?;\n    save_run(app_config, run).await?;\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::StateChanged { state: next })\n            .await;\n    }\n    Ok(())\n}\n\nasync fn trace_transition(\n    trace: &TraceWriter,\n    from: ResearchRunState,\n    to: ResearchRunState,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::StateTransition {\n            from,\n            to,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_started(\n    trace: &TraceWriter,\n    actor: &str,\n    model: String,\n    input: &str,\n) -> Result<()> {\n    trace\n        .append(&TraceEvent::LlmRequestStarted {\n            actor: actor.to_string(),\n            model,\n            input_hash: stable_hash(input),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_llm_finished(\n    trace: &TraceWriter,\n    actor: &str,\n    input: &str,\n    output: &str,\n) -> Result<()> {\n    let (_input_hash, output_hash, tokens) = decision_hash(input, output);\n    trace\n        .append(&TraceEvent::LlmRequestFinished {\n            actor: actor.to_string(),\n            output_hash,\n            token_estimate: tokens,\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn trace_checkpoint(trace: &TraceWriter, checkpoint: &Checkpoint) -> Result<()> {\n    trace\n        .append(&TraceEvent::CheckpointCreated {\n            checkpoint_id: checkpoint.checkpoint_id.clone(),\n            state: checkpoint.state.clone(),\n            at: Utc::now(),\n        })\n        .await\n}\n\nasync fn send_checkpoint(tx: Option<&mpsc::Sender<StatefulRunEvent>>, checkpoint: &Checkpoint) {\n    if let Some(tx) = tx {\n        let _ = tx\n            .send(StatefulRunEvent::CheckpointCreated {\n                checkpoint_id: checkpoint.checkpoint_id.clone(),\n                state: checkpoint.state.clone(),\n            })\n            .await;\n    }\n}\n\nasync fn trace_planned_tool_calls(trace: &TraceWriter, portfolio: &[QueryPortfolio]) -> Result<()> {\n    for item in portfolio {\n        for query in &item.github_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"github\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n        for query in &item.arxiv_queries {\n            trace\n                .append(&TraceEvent::ToolCallStarted {\n                    tool: \"arxiv\".to_string(),\n                    query: query.clone(),\n                    at: Utc::now(),\n                })\n                .await?;\n        }\n    }\n    Ok(())\n}\n\nfn agent_context(\n    actor: &str,\n    input: String,\n    policy: &RunPolicy,\n    allowed_urls: Vec<String>,\n) -> AgentContext {\n    AgentContext {\n        actor: actor.to_string(),\n        input,\n        allowed_urls,\n        max_llm_calls: policy.max_llm_calls_per_run,\n    }\n}\n\nfn side_model_name(llm_config: &LlmConfig) -> String {\n    llm_config\n        .side_model\n        .clone()\n        .unwrap_or_else(|| llm_config.main_model.clone())\n}\n\nfn invalid_transition(from: &ResearchRunState, to: &ResearchRunState) -> AppError {\n    AppError::Workflow(format!(\"invalid state transition from {from:?} to {to:?}\"))\n}\n\nfn run_summary_path(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(format!(\"{run_id}.json\"))\n}\n\nfn run_dir(app_config: &AppConfig, run_id: &str) -> PathBuf {\n    app_config.session_dir.join(run_id)\n}\n\nasync fn write_report_file(\n    app_config: &AppConfig,\n    run: &mut ResearchRunRecord,\n    markdown: &str,\n) -> Result<()> {\n    let path = resolve_report_path(&app_config.output, &run.topic);\n    if let Some(parent) = path.parent() {\n        tokio::fs::create_dir_all(parent).await?;\n    }\n    tokio::fs::write(&path, markdown).await?;\n    run.output_report = Some(path.display().to_string());\n    Ok(())\n}\n\nfn resolve_report_path(output: &Path, topic: &str) -> PathBuf {\n    if output.extension().is_some() {\n        return output.to_path_buf();\n    }\n    output.join(format!(\n        \"{}-stage3-{}.md\",\n        slugify(topic),\n        Utc::now().format(\"%Y%m%d-%H%M%S\")\n    ))\n}\n\nfn slugify(topic: &str) -> String {\n    let mut slug = String::new();\n    let mut last_dash = false;\n    for ch in topic.chars().flat_map(char::to_lowercase) {\n        if ch.is_ascii_alphanumeric() {\n            slug.push(ch);\n            last_dash = false;\n        } else if !last_dash {\n            slug.push('-');\n            last_dash = true;\n        }\n    }\n    let slug = slug.trim_matches('-');\n    if slug.is_empty() {\n        \"report\".to_string()\n    } else {\n        slug.to_string()\n    }\n}\n\npub fn decision_hash(input: &str, output: &str) -> (String, String, usize) {\n    (\n        stable_hash(input),\n        stable_hash(output),\n        token_estimate(input) + token_estimate(output),\n    )\n}\n"}}, "README.zh-CN.md": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1785747287.9926121, "n": 7525, "h": "9d217ce1685c1a7c", "c": "# LitScout-RS\n\n<p align=\"center\">\n  <a href=\"README.md\">English</a> | 简体中文\n</p>\n\n<p align=\"center\">\n  <img src=\"icon.png\" alt=\"LitScout-RS logo\" width=\"132\" />\n</p>\n\n<p align=\"center\">\n  <strong>用 Rust 构建的证据约束型研究侦察 Agent。</strong>\n</p>\n\n<p align=\"center\">\n  <a href=\"https://www.rust-lang.org/\"><img alt=\"Rust\" src=\"https://img.shields.io/badge/Rust-2021-b7410e?logo=rust&logoColor=white\"></a>\n  <img alt=\"Status\" src=\"https://img.shields.io/badge/status-active-brightgreen\">\n  <img alt=\"LLM\" src=\"https://img.shields.io/badge/LLM-DeepSeek-4b6bfb\">\n  <img alt=\"Sources\" src=\"https://img.shields.io/badge/sources-GitHub%20%7C%20arXiv%20%7C%20Academic%20Indexes-0f766e\">\n  <img alt=\"License\" src=\"https://img.shields.io/badge/license-not%20set-lightgrey\">\n</p>\n\nLitScout-RS 是一个基于 Rust 的研究侦察 Agent，可以把一个研究主题转化为可审查、带引用的调研报告，并继续进入论文阅读笔记与单篇追问流程。\n\n项目面向课程调研、早期文献综述、开源技术选型和研究型开发。它不把自己设计成无限制浏览器 Agent，而是从受控来源收集证据，记录每条证据的查询来源，筛掉噪声较高的学术候选，审计引用边界，并允许用户从生成报告继续进入论文级阅读笔记和问答。\n\n## 项目亮点\n\n- **完整研究流程，而不只是搜索列表**：生成研究简介、章节计划、查询组合、证据池、覆盖度报告、引用审计和最终 Markdown 报告。\n- **受控多源收集**：默认使用 GitHub 和 arXiv，可显式启用 Semantic Scholar、DBLP、OpenAlex、Crossref 组成的 Stage A 学术扩展源。\n- **证据优先的报告生成**：写作模块只能引用登记在 `CitationLedger` 中的 URL，来源 lineage 和 query attempt 均可检查。\n- **准确率优先的扩源策略**：学术候选需要经过 canonical merge、ranking、classification 和 `EvidenceQualityGate`，才能进入报告证据池。\n- **有状态的 Agent 控制环**：状态迁移、trace 事件、checkpoint 和 PlanReady 分支让每次运行都可复盘，而不是只看到一个黑盒结果。\n- **内置论文阅读库**：证据池中的 arXiv 论文可以加入阅读库，继续做全文或近全文抓取、结构化笔记和单篇论文追问。\n- **Rust 后端和类型化边界**：API 响应、session、checkpoint、证据模型和 LLM 输出都用 Rust 类型与 serde 兼容 JSON 表达。\n\n## 系统工作流\n\n<p align=\"center\">\n  <img src=\"workflow.png\" alt=\"LitScout-RS workflow\" width=\"860\" />\n</p>\n\n```text\nUser research topic\n  -> Scoper creates ResearchBrief\n  -> Planner creates ChapterPlan and QueryPortfolio\n  -> PlanCritic reviews plan quality\n  -> User approves or revises the plan\n  -> GitHubScout / ArxivScout collect default evidence\n  -> Optional AcademicScout collects Semantic Scholar / DBLP / OpenAlex / Crossref\n  -> EvidenceBuilder performs canonical merge, ranking, classification, and EvidenceQualityGate\n  -> EvidenceMemory and CitationLedger are built with query lineage\n  -> CoverageCritic checks chapter-level evidence gaps\n  -> Writer generates a cited Markdown report\n  -> CitationAuditor checks URL whitelist, citation coverage, and source diversity\n  -> Reading Library supports arXiv paper notes and single-paper Q&A\n```\n\nLLM 不会被允许自由浏览网页或编造来源。它只能接收 LitScout-RS 已收集并登记的证据，最终报告中的引用也会被 `CitationLedger` 约束和检查。\n\n## 架构概览\n\n```text\nsrc/\n  main.rs                 # 程序入口和 Web server 启动\n  cli.rs                  # 命令行参数\n  config.rs               # AppConfig 和环境变量处理\n  model.rs                # SourceItem、EvidenceMemory、CitationLedger、run model\n  sources/                # GitHub、arXiv、Semantic Scholar、DBLP、OpenAlex、Crossref\n  agent/                  # scoper、planner、scout、evidence builder、writer、auditor\n  reading/                # arXiv 阅读库、正文抓取、笔记、单篇论文问答\n  server/                 # axum 路由和 SSE endpoint\n  ranking.rs              # 排序信号\n  dedup.rs                # canonical work merge\n  classify.rs             # 证据分类\n  checkpoint.rs           # 不持久化密钥的 checkpoint\n  trace.rs                # jsonl run trace\nweb/src/\n  components/             # React 工作台视图\n  api/                    # 类型化前端 API client\n```\n\n## 环境要求\n\n- Rust stable toolchain 和 Cargo\n- Node.js 与 npm，用于 Web 工作台\n- DeepSeek 兼容的 chat completion endpoint\n- 可选 GitHub token，用于提高 GitHub API rate limit\n\n## 配置\n\nLitScout-RS 从 CLI 参数和环境变量读取运行配置。参考配置写在 `config.example.toml` 中。\n\n```bash\nexport DEEPSEEK_API_KEY=...\nexport DEEPSEEK_BASE_URL=https://api.deepseek.com\nexport DEEPSEEK_MODEL=deepseek-v4-pro\nexport DEEPSEEK_SIDE_MODEL=deepseek-v4-flash\nexport DEEPSEEK_MAX_TOKENS=4096\nexport DEEPSEEK_TIMEOUT_SECS=30\n\n# 可选，建议用于提高 GitHub API rate limit。\nexport GITHUB_TOKEN=...\n\n# 可选学术扩展源。\nexport SEMANTIC_SCHOLAR_API_KEY=...\nexport OPENALEX_API_KEY=...\nexport CROSSREF_MAILTO=you@example.com\n```\n\n不要提交真实 API key。checkpoint 和 trace 的设计目标是不持久化密钥，但环境变量和本地 shell history 仍需要正常保护。\n\n## 快速开始：Web 工作台\n\n构建前端并启动 Rust server：\n\n```bash\ncd web\nnpm install\nnpm run build\ncd ..\n\ncargo run -- --serve --port 3000\n```\n\n打开：\n\n```text\nhttp://127.0.0.1:3000\n```\n\n推荐的 Web 使用流程：\n\n1. 配置 DeepSeek 和可选 GitHub token。\n2. 输入中文或英文主题创建调研任务。\n3. 在正式收集前审查生成的计划。\n4. 批准计划，观察运行进度、事件和 checkpoint。\n5. 检查证据、query lineage、覆盖度和引用审计。\n6. 阅读生成的 Markdown 报告。\n7. 将有价值的 arXiv 论文加入阅读库，继续生成笔记和单篇论文问答。\n\n学术扩展源需要显式启用。只有当你希望 Semantic Scholar、DBLP、OpenAlex 和 Crossref 参与证据池时，才在工作台 policy 中打开对应能力。\n\n## CLI 使用\n\n当前主流程需要 LLM mode：\n\n```bash\ncargo run -- \"rust agent framework\" --llm\ncargo run -- \"llm tool calling\" --llm --github-limit 10 --arxiv-limit 10\ncargo run -- \"llm agent benchmark\" --llm --academic-extra --academic-limit 10\n```\n\n需要时也可以通过命令行传入 DeepSeek 配置：\n\n```bash\ncargo run -- \"rust agent framework\" --llm \\\n  --deepseek-api-key \"$DEEPSEEK_API_KEY\" \\\n  --deepseek-base-url https://api.deepseek.com \\\n  --deepseek-model deepseek-v4-pro \\\n  --deepseek-side-model deepseek-v4-flash \\\n  --llm-timeout 45 \\\n  --llm-max-tokens 4096\n```\n\n报告会写入 `reports/<topic>-<timestamp>.md`。session summary、trace 和 checkpoint 会写入 `sessions/`，其中不会保存 API key。\n\n## Web 工作台能力\n\n- 证据收集前的计划审查和修订。\n- 基于 SSE 的运行进度展示，覆盖 fetching、evidence building、synthesis 和 audit 阶段。\n- 证据池展示 source link、query attempt、lineage 和 selection summary。\n- 覆盖度矩阵展示章节级 `QueryGap` 与 `SourceGap` 诊断。\n- 引用审计检查 URL 白名单、引用覆盖度和来源多样性。\n- checkpoint 列表和 PlanReady 分支创建。\n- Markdown 报告预览。\n- 阅读库支持 arXiv 论文入库、正文抓取诊断、结构化笔记和单篇论文流式问答。\n\n## API 概览\n\n核心 endpoints：\n\n- `GET /api/health`\n- `POST /api/plan`\n- `POST /api/plan/revise`\n- `POST /api/run`\n- `POST /api/run/stream`\n- `POST /api/report/translate`\n\n有状态运行 endpoints：\n\n- `POST /api/runs`\n- `GET /api/runs/:run_id`\n- `GET /api/runs/:run_id/events`\n- `POST /api/runs/:run_id/approve-plan`\n- `POST /api/runs/:run_id/continue`\n- `POST /api/runs/:run_id/revise-plan`\n- `GET /api/runs/:run_id/evidence`\n- `GET /api/runs/:run_id/coverage`\n- `GET /api/runs/:run_id/citation-audit`\n- `GET /api/runs/:run_id/checkpoints`\n- `POST /api/runs/:run_id/branch-from-checkpoint`\n\n阅读库 endpoints：\n\n- `GET /api/library`\n- `POST /api/library/items`\n- `GET /api/library/items/:paper_key`\n- `DELETE /api/library/items/:paper_key`\n- `POST /api/library/items/:paper_key/generate-note`\n- `POST /api/library/items/:paper_key/chat/stream`\n\n报告级追问 endpoints 已移除。追问现在属于阅读库，语境限定在单篇论文内部。\n\n## 开发检查\n\nRust：\n\n```bash\ncargo fmt\ncargo check\ncargo test\n```\n\n前端：\n\n```bash\ncd web\nnpm run build\n```\n\nStage 3 fixture bench：\n\n```bash\nnode scripts/stage3_eval.mjs\n```\n\n这个 mini bench 用 fixture/mock data 验证控制环结构，不能替代 GitHub、arXiv、学术扩展源和 LLM 的真实集成检查。\n\n## 当前边界\n\n- 默认来源是 GitHub 和 arXiv。\n- Stage A 学术扩展源需要通过 `--academic-extra` 或 Web run policy 显式启用。\n- 开放 Web 搜索、浏览器自动化、无限制 ReAct、微信公众号抓取和通用新闻抓取都不在当前范围内。\n- Academic index 和 bibliography 候选必须通过 `EvidenceQualityGate`，才能进入 `EvidenceMemory` 和 `CitationLedger`。\n- 阅读库当前聚焦 arXiv 论文。它可以通过 Jina Reader 和本地 PDF extraction 抓取正文，但不是通用 PDF/文档分析系统。\n- `CoverageCritic` 负责报告缺口和建议，不会启动开放式自主爬取循环。\n- 分支能力目前支持从 PlanReady checkpoint 创建新 run。\n- 论文问答使用 SSE 接口。当前后端可能先缓冲模型输出，再发送 Markdown chunk；后续可以补 token-native streaming。\n\n## 仓库维护建议\n\n推荐的 GitHub project settings：\n\n- 添加简短仓库描述，比如：`Rust research scouting agent with evidence-gated multi-source search, cited reports, and arXiv reading notes.`\n- 添加 topics，例如 `rust`、`llm`、`research-agent`、`literature-review`、`arxiv`、`github-api`、`evidence`、`citation-audit`。\n- 如果仓库准备公开发布或复用，请添加 license file。\n- 不要提交 `.env`、API key、生成的 sessions、本地 cache 和私有 reports。\n- 如果希望 README 在 GitHub 上展示截图或架构图，建议把图片放在可追踪的 `docs/assets/` 目录下。\n\n## License / 许可证\n"}, "42bc98a2548dae3e116a8e54ae1fb9bc03b4ffd2": {"t": 1785750660.999696, "n": 7582, "h": "91139754b5cf1f10", "c": "# LitScout-RS\n\n<p align=\"center\">\n  <a href=\"README.md\">English</a> | 简体中文\n</p>\n\n<p align=\"center\">\n  <img src=\"icon.png\" alt=\"LitScout-RS logo\" width=\"132\" />\n</p>\n\n<p align=\"center\">\n  <strong>用 Rust 构建的证据约束型研究侦察 Agent。</strong>\n</p>\n\n<p align=\"center\">\n  <a href=\"https://www.rust-lang.org/\"><img alt=\"Rust\" src=\"https://img.shields.io/badge/Rust-2021-b7410e?logo=rust&logoColor=white\"></a>\n  <img alt=\"Status\" src=\"https://img.shields.io/badge/status-active-brightgreen\">\n  <img alt=\"LLM\" src=\"https://img.shields.io/badge/LLM-DeepSeek-4b6bfb\">\n  <img alt=\"Sources\" src=\"https://img.shields.io/badge/sources-GitHub%20%7C%20arXiv%20%7C%20Academic%20Indexes-0f766e\">\n  <img alt=\"License\" src=\"https://img.shields.io/badge/license-not%20set-lightgrey\">\n</p>\n\nLitScout-RS 是一个基于 Rust 的研究侦察 Agent，可以把一个研究主题转化为可审查、带引用的调研报告，并继续进入论文阅读笔记与单篇追问流程。\n\n项目面向课程调研、早期文献综述、开源技术选型和研究型开发。它不把自己设计成无限制浏览器 Agent，而是从受控来源收集证据，记录每条证据的查询来源，筛掉噪声较高的学术候选，审计引用边界，并允许用户从生成报告继续进入论文级阅读笔记和问答。\n\n## 项目亮点\n\n- **完整研究流程，而不只是搜索列表**：生成研究简介、章节计划、查询组合、证据池、覆盖度报告、引用审计和最终 Markdown 报告。\n- **受控多源收集**：默认使用 GitHub 和 arXiv，可显式启用 Semantic Scholar、DBLP、OpenAlex、Crossref 组成的 Stage A 学术扩展源。\n- **证据优先的报告生成**：写作模块只能引用登记在 `CitationLedger` 中的 URL，来源 lineage 和 query attempt 均可检查。\n- **准确率优先的扩源策略**：学术候选需要经过 canonical merge、ranking、classification 和 `EvidenceQualityGate`，才能进入报告证据池。\n- **有状态的 Agent 控制环**：状态迁移、trace 事件、checkpoint 和 PlanReady 分支让每次运行都可复盘，而不是只看到一个黑盒结果。\n- **内置论文阅读库**：证据池中的 arXiv 论文可以加入阅读库，继续做全文或近全文抓取、结构化笔记和单篇论文追问。\n- **Rust 后端和类型化边界**：API 响应、session、checkpoint、证据模型和 LLM 输出都用 Rust 类型与 serde 兼容 JSON 表达。\n\n## 系统工作流\n\n<p align=\"center\">\n  <img src=\"workflow.png\" alt=\"LitScout-RS workflow\" width=\"860\" />\n</p>\n\n```text\nUser research topic\n  -> Scoper creates ResearchBrief\n  -> Planner creates ChapterPlan and QueryPortfolio\n  -> PlanCritic reviews plan quality\n  -> User approves or revises the plan\n  -> GitHubScout / ArxivScout collect default evidence\n  -> Optional AcademicScout collects Semantic Scholar / DBLP / OpenAlex / Crossref\n  -> EvidenceBuilder performs canonical merge, ranking, classification, and EvidenceQualityGate\n  -> EvidenceMemory and CitationLedger are built with query lineage\n  -> CoverageCritic checks chapter-level evidence gaps\n  -> Writer generates a cited Markdown report\n  -> CitationAuditor checks URL whitelist, citation coverage, and source diversity\n  -> Reading Library supports arXiv paper notes and single-paper Q&A\n```\n\nLLM 不会被允许自由浏览网页或编造来源。它只能接收 LitScout-RS 已收集并登记的证据，最终报告中的引用也会被 `CitationLedger` 约束和检查。\n\n## 架构概览\n\n```text\nsrc/\n  main.rs                 # 程序入口和 Web server 启动\n  cli.rs                  # 命令行参数\n  config.rs               # AppConfig 和环境变量处理\n  model.rs                # SourceItem、EvidenceMemory、CitationLedger、run model\n  sources/                # GitHub、arXiv、Semantic Scholar、DBLP、OpenAlex、Crossref\n  agent/                  # scoper、planner、scout、evidence builder、writer、auditor\n  reading/                # arXiv 阅读库、正文抓取、笔记、单篇论文问答\n  server/                 # axum 路由和 SSE endpoint\n  ranking.rs              # 排序信号\n  dedup.rs                # canonical work merge\n  classify.rs             # 证据分类\n  checkpoint.rs           # 不持久化密钥的 checkpoint\n  trace.rs                # jsonl run trace\nweb/src/\n  components/             # React 工作台视图\n  api/                    # 类型化前端 API client\n```\n\n## 环境要求\n\n- Rust stable toolchain 和 Cargo\n- Node.js 与 npm，用于 Web 工作台\n- DeepSeek 兼容的 chat completion endpoint\n- 可选 GitHub token，用于提高 GitHub API rate limit\n\n## 配置\n\nLitScout-RS 从 CLI 参数和环境变量读取运行配置。参考配置写在 `config.example.toml` 中。\n\n```bash\nexport DEEPSEEK_API_KEY=...\nexport DEEPSEEK_BASE_URL=https://api.deepseek.com\nexport DEEPSEEK_MODEL=deepseek-v4-pro\nexport DEEPSEEK_SIDE_MODEL=deepseek-v4-flash\nexport DEEPSEEK_MAX_TOKENS=4096\nexport DEEPSEEK_TIMEOUT_SECS=30\n\n# 可选，建议用于提高 GitHub API rate limit。\nexport GITHUB_TOKEN=...\n\n# 可选学术扩展源。\nexport SEMANTIC_SCHOLAR_API_KEY=...\nexport OPENALEX_API_KEY=...\nexport CROSSREF_MAILTO=you@example.com\n```\n\n不要提交真实 API key。checkpoint 和 trace 的设计目标是不持久化密钥，但环境变量和本地 shell history 仍需要正常保护。\n\n## 快速开始：Web 工作台\n\n构建前端并启动 Rust server：\n\n```bash\ncd web\nnpm install\nnpm run build\ncd ..\n\ncargo run -- --serve --port 3000\n```\n\n打开：\n\n```text\nhttp://127.0.0.1:3000\n```\n\n推荐的 Web 使用流程：\n\n1. 配置 DeepSeek 和可选 GitHub token。\n2. 输入中文或英文主题创建调研任务。\n3. 在正式收集前审查生成的计划。\n4. 批准计划，观察运行进度、事件和 checkpoint。\n5. 检查证据、query lineage、覆盖度和引用审计。\n6. 阅读生成的 Markdown 报告。\n7. 将有价值的 arXiv 论文加入阅读库，继续生成笔记和单篇论文问答。\n\n学术扩展源需要显式启用。只有当你希望 Semantic Scholar、DBLP、OpenAlex 和 Crossref 参与证据池时，才在工作台 policy 中打开对应能力。\n\n## CLI 使用\n\n当前主流程需要 LLM mode：\n\n```bash\ncargo run -- \"rust agent framework\" --llm\ncargo run -- \"llm tool calling\" --llm --github-limit 10 --arxiv-limit 10\ncargo run -- \"llm agent benchmark\" --llm --academic-extra --academic-limit 10\n```\n\n需要时也可以通过命令行传入 DeepSeek 配置：\n\n```bash\ncargo run -- \"rust agent framework\" --llm \\\n  --deepseek-api-key \"$DEEPSEEK_API_KEY\" \\\n  --deepseek-base-url https://api.deepseek.com \\\n  --deepseek-model deepseek-v4-pro \\\n  --deepseek-side-model deepseek-v4-flash \\\n  --llm-timeout 45 \\\n  --llm-max-tokens 4096\n```\n\n报告会写入 `reports/<topic>-<timestamp>.md`。session summary、trace 和 checkpoint 会写入 `sessions/`，其中不会保存 API key。\n\n## Web 工作台能力\n\n- 证据收集前的计划审查和修订。\n- 基于 SSE 的运行进度展示，覆盖 fetching、evidence building、synthesis 和 audit 阶段。\n- 证据池展示 source link、query attempt、lineage 和 selection summary。\n- 覆盖度矩阵展示章节级 `QueryGap` 与 `SourceGap` 诊断。\n- 引用审计检查 URL 白名单、引用覆盖度和来源多样性。\n- checkpoint 列表和 PlanReady 分支创建。\n- Markdown 报告预览。\n- 阅读库支持 arXiv 论文入库、正文抓取诊断、结构化笔记和单篇论文流式问答。\n\n## API 概览\n\n核心 endpoints：\n\n- `GET /api/health`\n- `POST /api/plan`\n- `POST /api/plan/revise`\n- `POST /api/run`\n- `POST /api/run/stream`\n- `POST /api/report/translate`\n\n有状态运行 endpoints：\n\n- `POST /api/runs`\n- `GET /api/runs/:run_id`\n- `GET /api/runs/:run_id/events`\n- `POST /api/runs/:run_id/approve-plan`\n- `POST /api/runs/:run_id/continue`\n- `POST /api/runs/:run_id/revise-plan`\n- `GET /api/runs/:run_id/evidence`\n- `GET /api/runs/:run_id/coverage`\n- `GET /api/runs/:run_id/citation-audit`\n- `GET /api/runs/:run_id/checkpoints`\n- `POST /api/runs/:run_id/branch-from-checkpoint`\n\n阅读库 endpoints：\n\n- `GET /api/library`\n- `POST /api/library/items`\n- `GET /api/library/items/:paper_key`\n- `DELETE /api/library/items/:paper_key`\n- `POST /api/library/items/:paper_key/generate-note`\n- `POST /api/library/items/:paper_key/chat/stream`\n\n报告级追问 endpoints 已移除。追问现在属于阅读库，语境限定在单篇论文内部。\n\n## 开发检查\n\nRust：\n\n```bash\ncargo fmt\ncargo check\ncargo test\n```\n\n前端：\n\n```bash\ncd web\nnpm run build\n```\n\nStage 3 fixture bench：\n\n```bash\nnode scripts/stage3_eval.mjs\n```\n\n这个 mini bench 用 fixture/mock data 验证控制环结构，不能替代 GitHub、arXiv、学术扩展源和 LLM 的真实集成检查。\n\n## 当前边界\n\n- 默认来源是 GitHub 和 arXiv。\n- Stage A 学术扩展源需要通过 `--academic-extra` 或 Web run policy 显式启用。\n- 开放 Web 搜索、浏览器自动化、无限制 ReAct、微信公众号抓取和通用新闻抓取都不在当前范围内。\n- Academic index 和 bibliography 候选必须通过 `EvidenceQualityGate`，才能进入 `EvidenceMemory` 和 `CitationLedger`。\n- 阅读库当前聚焦 arXiv 论文。它可以通过 Jina Reader 和本地 PDF extraction 抓取正文，但不是通用 PDF/文档分析系统。\n- `CoverageCritic` 负责报告缺口和建议，不会启动开放式自主爬取循环。\n- 分支能力目前支持从 PlanReady checkpoint 创建新 run。\n- 论文问答使用 SSE 接口。当前后端可能先缓冲模型输出，再发送 Markdown chunk；后续可以补 token-native streaming。\n\n## 仓库维护建议\n\n推荐的 GitHub project settings：\n\n- 添加简短仓库描述，比如：`Rust research scouting agent with evidence-gated multi-source search, cited reports, and arXiv reading notes.`\n- 添加 topics，例如 `rust`、`llm`、`research-agent`、`literature-review`、`arxiv`、`github-api`、`evidence`、`citation-audit`。\n- 如果仓库准备公开发布或复用，请添加 license file。\n- 不要提交 `.env`、API key、生成的 sessions、本地 cache 和私有 reports。\n- 如果希望 README 在 GitHub 上展示截图或架构图，建议把图片放在可追踪的 `docs/assets/` 目录下。\n\n## License / 许可证\n\n当前仓库没有包含 license file。如果项目准备发布或复用，建议在 release 前添加明确许可证。\n"}}, "src/agent/planner.rs": {"df730e4a09eb4f1c5830a9f3deacc3ae94c1aa19": {"t": 1786068337.9169388, "n": 9831, "h": "a47ef55c1564492d", "c": "use serde::{Deserialize, Serialize};\nuse uuid::Uuid;\n\nuse crate::config::LlmConfig;\nuse crate::error::{AppError, Result};\nuse crate::llm::deepseek::{\n    ChatCompletionsRequest, ChatMessage, DeepSeekClient, DeepSeekConfig, ResponseFormat,\n};\nuse crate::model::{ChapterNode, QueryPortfolio, ResearchBrief};\nuse crate::run_policy::RunPolicy;\n\n#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]\npub struct ChapterPlanOutput {\n    pub chapters: Vec<ChapterNode>,\n    pub query_portfolio: Vec<QueryPortfolio>,\n    pub warnings: Vec<String>,\n}\n\npub async fn generate_chapter_plan(\n    brief: &ResearchBrief,\n    policy: &RunPolicy,\n    llm_config: &LlmConfig,\n) -> Result<ChapterPlanOutput> {\n    let plan = generate_chinese_plan(\n        &brief.topic,\n        policy.github_budget,\n        policy.arxiv_budget,\n        llm_config,\n    )\n    .await?;\n    Ok(chapter_plan_from_plan_output(&plan, policy))\n}\n\npub fn chapter_plan_from_plan_output(plan: &PlanOutput, policy: &RunPolicy) -> ChapterPlanOutput {\n    let aspects = plan\n        .aspects\n        .iter()\n        .take(policy.max_aspects_per_round.max(1))\n        .collect::<Vec<_>>();\n    let mut warnings = plan.warnings.clone();\n    if aspects.is_empty() {\n        warnings.push(\"搜索计划为空，使用默认章节。\".to_string());\n    }\n\n    let fallback_aspect = AspectOutput {\n        name_zh: \"默认研究方向\".to_string(),\n        rationale_zh: \"围绕原始主题进行 GitHub 与 arXiv 检索。\".to_string(),\n        github_query: plan.original_topic.clone(),\n        arxiv_query: plan.original_topic.clone(),\n        github_limit: policy.github_budget,\n        arxiv_limit: policy.arxiv_budget,\n    };\n    let effective_aspects = if aspects.is_empty() {\n        vec![&fallback_aspect]\n    } else {\n        aspects\n    };\n\n    let chapters = effective_aspects\n        .iter()\n        .enumerate()\n        .map(|(index, aspect)| ChapterNode {\n            id: format!(\"ch-{}\", index + 1),\n            parent_id: None,\n            title_zh: aspect.name_zh.clone(),\n            research_question: aspect.rationale_zh.clone(),\n            required_evidence_kinds: vec![\"github\".to_string(), \"arxiv\".to_string()],\n            evidence_quota: (aspect.github_limit + aspect.arxiv_limit).max(1),\n            sort_order: index + 1,\n        })\n        .collect::<Vec<_>>();\n\n    let query_portfolio = effective_aspects\n        .iter()\n        .enumerate()\n        .map(|(index, aspect)| QueryPortfolio {\n            chapter_id: format!(\"ch-{}\", index + 1),\n            github_queries: vec![aspect.github_query.clone()],\n            arxiv_queries: vec![aspect.arxiv_query.clone()],\n            rationale: aspect.rationale_zh.clone(),\n            budget: (aspect.github_limit + aspect.arxiv_limit).max(1),\n        })\n        .collect();\n\n    ChapterPlanOutput {\n        chapters,\n        query_portfolio,\n        warnings,\n    }\n}\n\npub async fn generate_chinese_plan(\n    topic: &str,\n    github_limit: usize,\n    arxiv_limit: usize,\n    llm_config: &LlmConfig,\n) -> Result<PlanOutput> {\n    if !llm_config.enabled {\n        return Ok(default_plan(topic, github_limit, arxiv_limit));\n    }\n\n    let config = DeepSeekConfig::from_llm_config(llm_config).ok_or_else(|| {\n        AppError::InvalidConfig(\n            \"`--llm` requires a DeepSeek API key. Set DEEPSEEK_API_KEY or pass --deepseek-api-key.\"\n                .to_string(),\n        )\n    })?;\n    let client = DeepSeekClient::new(config)?;\n    let request = build_plan_request(\n        client.side_model(),\n        format!(\n            \"请为以下技术主题生成搜索计划。\\n主题：{topic}\\nGitHub limit: {github_limit}\\narXiv limit: {arxiv_limit}\"\n        ),\n    );\n    let response = client.chat_completions_with_retry(request).await?;\n    let content = response.first_content()?;\n    parse_plan_json(content, topic, github_limit, arxiv_limit)\n}\n\npub fn default_plan(topic: &str, github_limit: usize, arxiv_limit: usize) -> PlanOutput {\n    PlanOutput {\n        plan_id: Uuid::new_v4().to_string(),\n        original_topic: topic.to_string(),\n        aspects: vec![AspectOutput {\n            name_zh: \"默认搜索\".to_string(),\n            rationale_zh: \"未启用 LLM，使用原始主题作为搜索词。\".to_string(),\n            github_query: topic.to_string(),\n            arxiv_query: topic.to_string(),\n            github_limit,\n            arxiv_limit,\n        }],\n        llm_generated: false,\n        warnings: vec![\n            \"LLM 未启用，使用默认搜索计划；如需中文智能规划，请以 --llm 启动并配置 DeepSeek API key。\"\n                .to_string(),\n        ],\n    }\n}\n\nfn build_plan_request(model: String, content: String) -> ChatCompletionsRequest {\n    ChatCompletionsRequest {\n        model,\n        messages: vec![\n            ChatMessage {\n                role: \"system\".to_string(),\n                content: SEARCH_PLAN_SYSTEM_PROMPT.to_string(),\n            },\n            ChatMessage {\n                role: \"user\".to_string(),\n                content,\n            },\n        ],\n        temperature: Some(0.2),\n        max_tokens: Some(1024),\n        stream: Some(false),\n        response_format: Some(ResponseFormat {\n            r#type: \"json_object\".to_string(),\n        }),\n    }\n}\n\npub(crate) fn parse_plan_json(\n    content: &str,\n    topic: &str,\n    github_limit: usize,\n    arxiv_limit: usize,\n) -> Result<PlanOutput> {\n    let json = crate::llm::deepseek::strip_json_fence(content);\n    let llm_response: LlmPlanResponse = serde_json::from_str(json)\n        .map_err(|err| AppError::Llm(format!(\"章节计划 JSON 解析失败: {err}\")))?;\n\n    let raw_aspects = llm_response\n        .aspects\n        .into_iter()\n        .take(3)\n        .filter(|aspect| {\n            !aspect.github_query.trim().is_empty() && !aspect.arxiv_query.trim().is_empty()\n        })\n        .collect::<Vec<_>>();\n\n    if raw_aspects.is_empty() {\n        return Err(AppError::Llm(\"章节计划未包含可用搜索方向\".to_string()));\n    }\n\n    let github_per_aspect = split_limit(github_limit, raw_aspects.len());\n    let arxiv_per_aspect = split_limit(arxiv_limit, raw_aspects.len());\n    let aspects = raw_aspects\n        .into_iter()\n        .map(|aspect| AspectOutput {\n            name_zh: non_empty_or(aspect.name_zh, \"研究方向\"),\n            rationale_zh: non_empty_or(aspect.rationale_zh, \"基于主题关键词的搜索方向。\"),\n            github_query: aspect.github_query.trim().to_string(),\n            arxiv_query: aspect.arxiv_query.trim().to_string(),\n            github_limit: github_per_aspect,\n            arxiv_limit: arxiv_per_aspect,\n        })\n        .collect();\n\n    Ok(PlanOutput {\n        plan_id: Uuid::new_v4().to_string(),\n        original_topic: topic.to_string(),\n        aspects,\n        llm_generated: true,\n        warnings: Vec::new(),\n    })\n}\n\nfn split_limit(limit: usize, parts: usize) -> usize {\n    if parts == 0 {\n        return limit.max(1);\n    }\n    limit.div_ceil(parts).max(1)\n}\n\nfn non_empty_or(value: String, default: &str) -> String {\n    let trimmed = value.trim();\n    if trimmed.is_empty() {\n        default.to_string()\n    } else {\n        trimmed.to_string()\n    }\n}\n\n#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]\npub struct PlanOutput {\n    pub plan_id: String,\n    pub original_topic: String,\n    pub aspects: Vec<AspectOutput>,\n    pub llm_generated: bool,\n    pub warnings: Vec<String>,\n}\n\n#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]\npub struct AspectOutput {\n    pub name_zh: String,\n    pub rationale_zh: String,\n    pub github_query: String,\n    pub arxiv_query: String,\n    pub github_limit: usize,\n    pub arxiv_limit: usize,\n}\n\n#[derive(Debug, Deserialize)]\nstruct LlmPlanResponse {\n    aspects: Vec<LlmAspect>,\n}\n\n#[derive(Debug, Deserialize)]\nstruct LlmAspect {\n    name_zh: String,\n    rationale_zh: String,\n    github_query: String,\n    arxiv_query: String,\n}\n\nconst SEARCH_PLAN_SYSTEM_PROMPT: &str = \"\\\n你是 LitScout-RS 的研究侦察规划器。用户用中文描述技术主题，你生成受控搜索计划。\n\n规则：\n1. 生成 1-3 个搜索角度。\n2. name_zh 和 rationale_zh 必须使用中文。\n3. github_query 和 arxiv_query 必须使用英文技术术语或中英混合术语，以保证 GitHub/arXiv 检索效果。\n4. 不要使用任意网页、浏览器、PDF 或外部搜索源。\n5. 不要使用 AND/OR/NOT 等复杂布尔表达式，使用自然关键词即可。\n6. 只返回 JSON，不要任何额外文本。\n\n返回格式：\n{\n  \\\"aspects\\\": [\n    {\n      \\\"name_zh\\\": \\\"核心框架方向\\\",\n      \\\"rationale_zh\\\": \\\"优先定位可直接复用的实现框架。\\\",\n      \\\"github_query\\\": \\\"llm agent framework tool calling\\\",\n      \\\"arxiv_query\\\": \\\"large language model tool calling agent\\\"\n    }\n  ]\n}\";\n\n#[cfg(test)]\nmod tests {\n    use super::{default_plan, parse_plan_json};\n\n    #[test]\n    fn default_plan_is_valid_without_llm() {\n        let plan = default_plan(\"Rust Agent 框架\", 8, 6);\n\n        assert!(!plan.llm_generated);\n        assert_eq!(plan.original_topic, \"Rust Agent 框架\");\n        assert_eq!(plan.aspects.len(), 1);\n        assert_eq!(plan.aspects[0].github_limit, 8);\n        assert_eq!(plan.aspects[0].arxiv_limit, 6);\n        assert!(!plan.plan_id.is_empty());\n    }\n\n    #[test]\n    fn parse_chinese_plan_json_from_fixture() {\n        let content = r#\"{\n            \"aspects\": [\n                {\n                    \"name_zh\": \"核心框架方向\",\n                    \"rationale_zh\": \"寻找可复用 Rust agent 框架。\",\n                    \"github_query\": \"rust agent framework tool calling\",\n                    \"arxiv_query\": \"large language model agent framework\"\n                },\n                {\n                    \"name_zh\": \"评测基准方向\",\n                    \"rationale_zh\": \"补充 agent benchmark 相关论文和项目。\",\n                    \"github_query\": \"agent benchmark rust\",\n                    \"arxiv_query\": \"agent benchmark tool use\"\n                }\n            ]\n        }\"#;\n\n        let plan = parse_plan_json(content, \"Rust Agent 框架\", 9, 7).expect(\"plan should parse\");\n\n        assert!(plan.llm_generated);\n        assert_eq!(plan.aspects.len(), 2);\n        assert_eq!(plan.aspects[0].name_zh, \"核心框架方向\");\n        assert_eq!(plan.aspects[0].github_limit, 5);\n        assert_eq!(plan.aspects[0].arxiv_limit, 4);\n        assert_eq!(plan.aspects[1].arxiv_query, \"agent benchmark tool use\");\n    }\n}\n"}}}}