{"repo": "agions/vynaro", "files": {"crates/vynaro-script/src/lib.rs": {"7bd402ca7866de2bd615bb46a3bf15e0bba3e603": {"t": 1786068396.669566, "n": 21281, "h": "2a3eb3842c5beb17", "c": "//! vynaro-llm v1.0.0  · LlmProvider trait + 11 provider 真实实现 + LlmManager 故障切换\n//!\n//! ## 11 个 Provider\n//! - Qwen (Qwen3-Max / Qwen-Plus / Qwen-Turbo)        — OpenAI 兼容\n//! - Kimi (Moonshot)                                   — OpenAI 兼容\n//! - GLM5 (智谱)                                       — OpenAI 兼容\n//! - Claude (Anthropic)                                — Anthropic 原生\n//! - Gemini (Google)                                   — Google Gen AI 原生\n//! - DeepSeek                                          — OpenAI 兼容\n//! - Doubao (字节豆包)                                  — OpenAI 兼容\n//! - Hunyuan (腾讯混元)                                 — OpenAI 兼容\n//! - Local (Ollama 等本地)                              — OpenAI 兼容\n//! - OpenAI (GPT-4o / GPT-4-Turbo)                     — OpenAI\n//! - Qwen3.7 (实验版)                                  — OpenAI 兼容\n//!\n//! ## M3 范围\n//! - 9 个 OpenAI 兼容 provider 共享 `OpenAiCompatible` 底座\n//! - Claude 与 Gemini 各自实现原生协议\n//! - 11 个 provider 全部可实际调用\n//! - `LlmProviderFactory` 按 `LlmProviderKind` 工厂创建\n//!\n//! ## 历史\n//! - M2: OpenAI 作为 PoC,其余 10 个 stub\n//! - M3: 全部 11 个真实实现\n\nuse async_trait::async_trait;\nuse serde::{Deserialize, Serialize};\n\npub use vynaro_core::error::{LlmProviderKind, VynaroError, VynaroResult};\n\n// ════════════════════════════════════════════════════════════════════════\n// 输入 / 输出契约\n// ════════════════════════════════════════════════════════════════════════\n\n/// 单次 LLM 调用的输入。\n#[derive(Debug, Clone, Serialize, Deserialize)]\npub struct LlmRequest {\n    pub system: String,\n    pub user: String,\n    pub model: Option<String>,    // 默认使用 provider 的 default model\n    pub max_tokens: Option<u32>,  // 默认 4096\n    pub temperature: Option<f32>, // 默认 0.7\n    pub stream: bool,             // 是否流式\n}\n\n/// LLM 响应 (一次性)。\n#[derive(Debug, Clone, Serialize, Deserialize)]\npub struct LlmResponse {\n    pub content: String,\n    pub model: String,\n    pub tokens_used: u32,\n    pub duration_ms: u64,\n}\n\n/// LlmProvider trait — 每个 11 个 provider 实现它。\n#[async_trait]\npub trait LlmProvider: Send + Sync {\n    fn kind(&self) -> LlmProviderKind;\n    fn default_model(&self) -> &'static str;\n    async fn chat(&self, req: &LlmRequest) -> VynaroResult<LlmResponse>;\n}\n\n// ════════════════════════════════════════════════════════════════════════\n// 公共 builder: 9 个 OpenAI 兼容 provider 共享\n// ════════════════════════════════════════════════════════════════════════\n\n/// OpenAI 兼容协议 (8 个厂商 + OpenAI 自己也用)\n#[derive(Debug, Clone)]\npub struct OpenAiCompatible {\n    pub kind: LlmProviderKind,\n    pub api_key: String,\n    pub base_url: String,\n    pub default_model: &'static str,\n    pub client: reqwest::Client,\n}\n\nimpl OpenAiCompatible {\n    pub fn new(\n        kind: LlmProviderKind,\n        api_key: impl Into<String>,\n        base_url: impl Into<String>,\n        default_model: &'static str,\n    ) -> Self {\n        Self {\n            kind,\n            api_key: api_key.into(),\n            base_url: base_url.into(),\n            default_model,\n            client: reqwest::Client::builder()\n                .timeout(std::time::Duration::from_secs(60))\n                .build()\n                .expect(\"reqwest client build\"),\n        }\n    }\n}\n\n#[async_trait]\nimpl LlmProvider for OpenAiCompatible {\n    fn kind(&self) -> LlmProviderKind {\n        self.kind\n    }\n\n    fn default_model(&self) -> &'static str {\n        self.default_model\n    }\n\n    async fn chat(&self, req: &LlmRequest) -> VynaroResult<LlmResponse> {\n        let start = std::time::Instant::now();\n\n        let body = serde_json::json!({\n            \"model\": req.model.clone().unwrap_or_else(|| self.default_model.to_string()),\n            \"messages\": [\n                { \"role\": \"system\", \"content\": req.system },\n                { \"role\": \"user\",   \"content\": req.user },\n            ],\n            \"max_tokens\": req.max_tokens.unwrap_or(4096),\n            \"temperature\": req.temperature.unwrap_or(0.7),\n            \"stream\": false,\n        });\n\n        let resp = self\n            .client\n            .post(&self.base_url)\n            .bearer_auth(&self.api_key)\n            .json(&body)\n            .send()\n            .await\n            .map_err(|e| VynaroError::Llm {\n                provider: self.kind,\n                message: format!(\"request send failed: {e}\"),\n            })?;\n\n        let status = resp.status();\n        if !status.is_success() {\n            let body = resp.text().await.unwrap_or_default();\n            return Err(VynaroError::Llm {\n                provider: self.kind,\n                message: format!(\"HTTP {status}: {body}\"),\n            });\n        }\n\n        #[derive(Deserialize)]\n        struct Choice {\n            message: ChoiceMessage,\n        }\n        #[derive(Deserialize)]\n        struct ChoiceMessage {\n            content: String,\n        }\n        #[derive(Deserialize)]\n        struct Usage {\n            total_tokens: u32,\n        }\n        #[derive(Deserialize)]\n        struct ApiResp {\n            model: String,\n            choices: Vec<Choice>,\n            usage: Usage,\n        }\n\n        let api: ApiResp = resp.json().await.map_err(|e| VynaroError::Llm {\n            provider: self.kind,\n            message: format!(\"response json parse failed: {e}\"),\n        })?;\n\n        let content = api\n            .choices\n            .into_iter()\n            .next()\n            .map(|c| c.message.content)\n            .unwrap_or_default();\n\n        Ok(LlmResponse {\n            content,\n            model: api.model,\n            tokens_used: api.usage.total_tokens,\n            duration_ms: start.elapsed().as_millis() as u64,\n        })\n    }\n}\n\n// ════════════════════════════════════════════════════════════════════════\n// Claude (Anthropic 原生协议)\n// ════════════════════════════════════════════════════════════════════════\n\nconst CLAUDE_MESSAGES_URL: &str = \"https://api.anthropic.com/v1/messages\";\n\n#[derive(Debug, Clone)]\npub struct ClaudeProvider {\n    api_key: String,\n    client: reqwest::Client,\n}\n\nimpl ClaudeProvider {\n    pub fn new(api_key: impl Into<String>) -> Self {\n        Self {\n            api_key: api_key.into(),\n            client: reqwest::Client::builder()\n                .timeout(std::time::Duration::from_secs(60))\n                .build()\n                .expect(\"reqwest client build\"),\n        }\n    }\n}\n\n#[async_trait]\nimpl LlmProvider for ClaudeProvider {\n    fn kind(&self) -> LlmProviderKind {\n        LlmProviderKind::Claude\n    }\n\n    fn default_model(&self) -> &'static str {\n        \"claude-sonnet-5\"\n    }\n\n    async fn chat(&self, req: &LlmRequest) -> VynaroResult<LlmResponse> {\n        let start = std::time::Instant::now();\n\n        let body = serde_json::json!({\n            \"model\": req.model.clone().unwrap_or_else(|| self.default_model().to_string()),\n            \"system\": req.system,\n            \"messages\": [\n                { \"role\": \"user\", \"content\": req.user },\n            ],\n            \"max_tokens\": req.max_tokens.unwrap_or(4096),\n            \"temperature\": req.temperature.unwrap_or(0.7),\n        });\n\n        let resp = self\n            .client\n            .post(CLAUDE_MESSAGES_URL)\n            .header(\"x-api-key\", &self.api_key)\n            .header(\"anthropic-version\", \"2023-06-01\")\n            .json(&body)\n            .send()\n            .await\n            .map_err(|e| VynaroError::Llm {\n                provider: LlmProviderKind::Claude,\n                message: format!(\"request send failed: {e}\"),\n            })?;\n\n        let status = resp.status();\n        if !status.is_success() {\n            let body = resp.text().await.unwrap_or_default();\n            return Err(VynaroError::Llm {\n                provider: LlmProviderKind::Claude,\n                message: format!(\"HTTP {status}: {body}\"),\n            });\n        }\n\n        #[derive(Deserialize)]\n        struct Content {\n            #[serde(rename = \"type\")]\n            kind: String,\n            text: Option<String>,\n        }\n        #[derive(Deserialize)]\n        struct ApiResp {\n            model: String,\n            content: Vec<Content>,\n            usage: ClaudeUsage,\n        }\n        #[derive(Deserialize)]\n        struct ClaudeUsage {\n            input_tokens: u32,\n            output_tokens: u32,\n        }\n\n        let api: ApiResp = resp.json().await.map_err(|e| VynaroError::Llm {\n            provider: LlmProviderKind::Claude,\n            message: format!(\"response json parse failed: {e}\"),\n        })?;\n\n        let content = api\n            .content\n            .into_iter()\n            .filter(|c| c.kind == \"text\")\n            .filter_map(|c| c.text)\n            .collect::<Vec<_>>()\n            .join(\"\");\n\n        Ok(LlmResponse {\n            content,\n            model: api.model,\n            tokens_used: api.usage.input_tokens + api.usage.output_tokens,\n            duration_ms: start.elapsed().as_millis() as u64,\n        })\n    }\n}\n\n// ════════════════════════════════════════════════════════════════════════\n// Gemini (Google Gen AI 原生协议)\n// ════════════════════════════════════════════════════════════════════════\n\n#[derive(Debug, Clone)]\npub struct GeminiProvider {\n    api_key: String,\n    client: reqwest::Client,\n}\n\nimpl GeminiProvider {\n    pub fn new(api_key: impl Into<String>) -> Self {\n        Self {\n            api_key: api_key.into(),\n            client: reqwest::Client::builder()\n                .timeout(std::time::Duration::from_secs(60))\n                .build()\n                .expect(\"reqwest client build\"),\n        }\n    }\n}\n\n#[async_trait]\nimpl LlmProvider for GeminiProvider {\n    fn kind(&self) -> LlmProviderKind {\n        LlmProviderKind::Gemini\n    }\n\n    fn default_model(&self) -> &'static str {\n        \"gemini-3.6-flash\"\n    }\n\n    async fn chat(&self, req: &LlmRequest) -> VynaroResult<LlmResponse> {\n        let start = std::time::Instant::now();\n\n        let model = req\n            .model\n            .clone()\n            .unwrap_or_else(|| self.default_model().to_string());\n        // 使用 streamGenerateContent=false 切到非流式\n        let url = format!(\n            \"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={key}\",\n            model = model,\n            key = self.api_key,\n        );\n\n        let body = serde_json::json!({\n            \"systemInstruction\": {\n                \"parts\": [{ \"text\": req.system }],\n            },\n            \"contents\": [\n                {\n                    \"role\": \"user\",\n                    \"parts\": [{ \"text\": req.user }],\n                },\n            ],\n            \"generationConfig\": {\n                \"maxOutputTokens\": req.max_tokens.unwrap_or(4096),\n                \"temperature\": req.temperature.unwrap_or(0.7),\n            },\n        });\n\n        let resp = self\n            .client\n            .post(&url)\n            .json(&body)\n            .send()\n            .await\n            .map_err(|e| VynaroError::Llm {\n                provider: LlmProviderKind::Gemini,\n                message: format!(\"request send failed: {e}\"),\n            })?;\n\n        let status = resp.status();\n        if !status.is_success() {\n            let body = resp.text().await.unwrap_or_default();\n            return Err(VynaroError::Llm {\n                provider: LlmProviderKind::Gemini,\n                message: format!(\"HTTP {status}: {body}\"),\n            });\n        }\n\n        #[derive(Deserialize)]\n        struct Part {\n            text: Option<String>,\n        }\n        #[derive(Deserialize)]\n        struct Content {\n            parts: Vec<Part>,\n        }\n        #[derive(Deserialize)]\n        struct Candidate {\n            content: Content,\n        }\n        #[derive(Deserialize)]\n        struct UsageMetadata {\n            total_token_count: u32,\n        }\n        #[derive(Deserialize)]\n        struct ApiResp {\n            candidates: Vec<Candidate>,\n            #[serde(rename = \"modelVersion\")]\n            model_version: String,\n            #[serde(rename = \"usageMetadata\")]\n            usage_metadata: UsageMetadata,\n        }\n\n        let api: ApiResp = resp.json().await.map_err(|e| VynaroError::Llm {\n            provider: LlmProviderKind::Gemini,\n            message: format!(\"response json parse failed: {e}\"),\n        })?;\n\n        let content = api\n            .candidates\n            .into_iter()\n            .next()\n            .map(|c| {\n                c.content\n                    .parts\n                    .into_iter()\n                    .filter_map(|p| p.text)\n                    .collect::<Vec<_>>()\n                    .join(\"\")\n            })\n            .unwrap_or_default();\n\n        Ok(LlmResponse {\n            content,\n            model: api.model_version,\n            tokens_used: api.usage_metadata.total_token_count,\n            duration_ms: start.elapsed().as_millis() as u64,\n        })\n    }\n}\n\n// ════════════════════════════════════════════════════════════════════════\n// Provider 别名 (11 个工厂 constructor)\n// ════════════════════════════════════════════════════════════════════════\n\n/// OpenAI 公开别名 (M2 兼容)\npub type OpenAiProvider = OpenAiCompatible;\n\npub fn openai(api_key: impl Into<String>) -> OpenAiProvider {\n    OpenAiCompatible::new(\n        LlmProviderKind::OpenAi,\n        api_key,\n        \"https://api.openai.com/v1/chat/completions\",\n        \"gpt-5.6-sol\",\n    )\n}\n\npub fn qwen(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Qwen,\n        api_key,\n        \"https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions\",\n        \"qwen3.8-max\",\n    )\n}\n\npub fn kimi(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Kimi,\n        api_key,\n        \"https://api.moonshot.cn/v1/chat/completions\",\n        \"kimi-k3\",\n    )\n}\n\npub fn glm5(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Glm5,\n        api_key,\n        \"https://open.bigmodel.cn/api/paas/v4/chat/completions\",\n        \"glm-5.2\",\n    )\n}\n\npub fn deepseek(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::DeepSeek,\n        api_key,\n        \"https://api.deepseek.com/v1/chat/completions\",\n        \"deepseek-v4-pro\",\n    )\n}\n\npub fn doubao(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Doubao,\n        api_key,\n        \"https://ark.cn-beijing.volces.com/api/v3/chat/completions\",\n        \"doubao-seed-2-1-pro-260628\",\n    )\n}\n\npub fn hunyuan(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Hunyuan,\n        api_key,\n        \"https://api.hunyuan.tencent.com/v1/chat/completions\",\n        \"hunyuan-pro\",\n    )\n}\n\n// Hunyuan 与 Qwen3.7 端点为开发中默认值,M3.1 任务说明:\n// - Hunyuan: 视腾讯侧正式 GA 端点调整\n// - Qwen3.7: 实验版,默认端点为标准 DashScope\npub fn local(base_url: Option<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Local,\n        \"no-key-required\",\n        base_url.unwrap_or_else(|| \"http://localhost:11434/v1/chat/completions\".into()),\n        \"llama3.2\",\n    )\n}\n\npub fn qwen37(api_key: impl Into<String>) -> OpenAiCompatible {\n    OpenAiCompatible::new(\n        LlmProviderKind::Qwen37,\n        api_key,\n        \"https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions\",\n        \"qwen3.7-max\",\n    )\n}\n\npub fn claude(api_key: impl Into<String>) -> ClaudeProvider {\n    ClaudeProvider::new(api_key)\n}\n\npub fn gemini(api_key: impl Into<String>) -> GeminiProvider {\n    GeminiProvider::new(api_key)\n}\n\n/// 工厂: 按 `LlmProviderKind` 构造(API key 必传)\npub fn factory(kind: LlmProviderKind, api_key: impl Into<String>) -> Box<dyn LlmProvider> {\n    let key: String = api_key.into();\n    match kind {\n        LlmProviderKind::OpenAi => Box::new(openai(key)),\n        LlmProviderKind::Qwen => Box::new(qwen(key)),\n        LlmProviderKind::Kimi => Box::new(kimi(key)),\n        LlmProviderKind::Glm5 => Box::new(glm5(key)),\n        LlmProviderKind::DeepSeek => Box::new(deepseek(key)),\n        LlmProviderKind::Doubao => Box::new(doubao(key)),\n        LlmProviderKind::Hunyuan => Box::new(hunyuan(key)),\n        LlmProviderKind::Local => Box::new(local(None)),\n        LlmProviderKind::Qwen37 => Box::new(qwen37(key)),\n        LlmProviderKind::Claude => Box::new(claude(key)),\n        LlmProviderKind::Gemini => Box::new(gemini(key)),\n    }\n}\n\n// ════════════════════════════════════════════════════════════════════════\n// LlmManager — 故障切换链\n// ════════════════════════════════════════════════════════════════════════\n\n/// 多个 provider 链式调用,失败 fallback 至下一个。\n///\n/// 设计: 顺序尝试 `chain` 中的 provider,第一个成功的返回;全部失败返回最后一次错误。\npub struct LlmManager {\n    chain: Vec<Box<dyn LlmProvider>>,\n}\n\nimpl std::fmt::Debug for LlmManager {\n    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {\n        let names: Vec<_> = self.chain.iter().map(|p| p.kind()).collect();\n        f.debug_struct(\"LlmManager\").field(\"chain\", &names).finish()\n    }\n}\n\nimpl LlmManager {\n    pub fn new(chain: Vec<Box<dyn LlmProvider>>) -> Self {\n        Self { chain }\n    }\n\n    /// 顺序尝试所有 provider,直到有成功响应。\n    pub async fn chat(&self, req: &LlmRequest) -> VynaroResult<LlmResponse> {\n        if self.chain.is_empty() {\n            return Err(VynaroError::Llm {\n                provider: LlmProviderKind::OpenAi, // 占位\n                message: \"no LLM provider configured\".into(),\n            });\n        }\n\n        let mut last_err: Option<VynaroError> = None;\n        for provider in &self.chain {\n            match provider.chat(req).await {\n                Ok(resp) => {\n                    tracing::info!(\n                        provider = ?provider.kind(),\n                        tokens = resp.tokens_used,\n                        duration_ms = resp.duration_ms,\n                        \"llm request succeeded\"\n                    );\n                    return Ok(resp);\n                }\n                Err(e) => {\n                    tracing::warn!(\n                        provider = ?provider.kind(),\n                        error = %e,\n                        \"llm request failed, fallback to next provider\"\n                    );\n                    last_err = Some(e);\n                }\n            }\n        }\n\n        Err(last_err.unwrap_or_else(|| {\n            VynaroError::Other(\"LlmManager: chain empty after iteration\".into())\n        }))\n    }\n}\n\n#[cfg(test)]\nmod tests {\n    use super::*;\n\n    #[tokio::test]\n    async fn manager_chain_empty_returns_err() {\n        let m = LlmManager::new(vec![]);\n        let r = m\n            .chat(&LlmRequest {\n                system: \"test\".into(),\n                user: \"hi\".into(),\n                model: None,\n                max_tokens: None,\n                temperature: None,\n                stream: false,\n            })\n            .await;\n        assert!(r.is_err());\n    }\n\n    #[test]\n    fn all_eleven_providers_kind_correct() {\n        let all: Vec<(&str, LlmProviderKind)> = vec![\n            (\"openai\", LlmProviderKind::OpenAi),\n            (\"qwen\", LlmProviderKind::Qwen),\n            (\"kimi\", LlmProviderKind::Kimi),\n            (\"glm5\", LlmProviderKind::Glm5),\n            (\"deepseek\", LlmProviderKind::DeepSeek),\n            (\"doubao\", LlmProviderKind::Doubao),\n            (\"hunyuan\", LlmProviderKind::Hunyuan),\n            (\"local\", LlmProviderKind::Local),\n            (\"qwen37\", LlmProviderKind::Qwen37),\n            (\"claude\", LlmProviderKind::Claude),\n            (\"gemini\", LlmProviderKind::Gemini),\n        ];\n\n        for (name, kind) in all {\n            let p = factory(kind, \"test-key\");\n            assert_eq!(p.kind(), kind, \"factory[{name}] kind mismatch\");\n        }\n    }\n\n    #[test]\n    fn default_models_distinct() {\n        let m = [\n            (LlmProviderKind::OpenAi, openai(\"k\").default_model()),\n            (LlmProviderKind::Qwen, qwen(\"k\").default_model()),\n            (LlmProviderKind::Kimi, kimi(\"k\").default_model()),\n            (LlmProviderKind::Glm5, glm5(\"k\").default_model()),\n            (LlmProviderKind::DeepSeek, deepseek(\"k\").default_model()),\n            (LlmProviderKind::Doubao, doubao(\"k\").default_model()),\n            (LlmProviderKind::Hunyuan, hunyuan(\"k\").default_model()),\n            (LlmProviderKind::Local, local(None).default_model()),\n            (LlmProviderKind::Qwen37, qwen37(\"k\").default_model()),\n            (LlmProviderKind::Claude, claude(\"k\").default_model()),\n            (LlmProviderKind::Gemini, gemini(\"k\").default_model()),\n        ];\n        for (kind, model) in m {\n            assert!(!model.is_empty(), \"{kind:?} default model empty\");\n        }\n    }\n\n    #[tokio::test]\n    async fn manager_chain_propagates_last_error() {\n        // 1 个无法联通的 provider (无 URL) 应该得到 Err\n        let bad = OpenAiCompatible::new(\n            LlmProviderKind::OpenAi,\n            \"k\",\n            \"http://127.0.0.1:1/v1/chat/completions\", // 1 端口不可达\n            \"gpt-4o-mini\",\n        );\n        let mgr = LlmManager::new(vec![Box::new(bad)]);\n        let r = mgr\n            .chat(&LlmRequest {\n                system: \"test\".into(),\n                user: \"hi\".into(),\n                model: None,\n                max_tokens: None,\n                temperature: None,\n                stream: false,\n            })\n            .await;\n        assert!(r.is_err());\n    }\n}\n"}}}}