1 line
9.6 KiB
JSON
1 line
9.6 KiB
JSON
{"repo": "automataia/sir-bone-rs", "files": {"README.md": {"877169e84a4d48d9982d9d3e4196c4071244d395": {"t": 1785674438.0538652, "n": 9262, "h": "72c4cd6d5f2f84e8", "c": "<h1 align=\"center\">Sir Bone — Rust</h1>\n\n<img src=\"logos/sirbone.webp\" align=\"center\" />\n\n<p align=\"center\">\n <img src=\"https://img.shields.io/badge/license-MIT%2FApache--2.0-blue.svg\" alt=\"License: MIT OR Apache-2.0\">\n <img src=\"https://img.shields.io/badge/rust-1.88%2B-dea584.svg?logo=rust\" alt=\"Rust 1.88+\">\n <img src=\"https://img.shields.io/badge/Async-Tokio_1-000000.svg?logo=tokio&logoColor=white\" alt=\"Async Tokio\">\n <img src=\"https://img.shields.io/badge/TUI-ratatui_0.29-ffd33d.svg?logo=ratatui&logoColor=black\" alt=\"ratatui\">\n <img src=\"https://img.shields.io/badge/MCP-rmcp_1.7-6366f1.svg?logo=modelcontextprotocol&logoColor=white\" alt=\"MCP\">\n</p>\n\n---\n\n**Sir Bone** is a Rust coding-agent harness for developers who want the model to help, not take over:\nvisible tool calls, explicit permission gates, reversible workspace snapshots, token-aware context\nmanagement — in **one native binary**.\n\nIt started as a compact rewrite inspired by [Pi](https://github.com/earendil-works/pi), with one\nthesis: keep the developer in command while cutting runtime weight, wasted tokens, and accidental\ncomplexity.\n\n---\n\n## Why it's lighter\n\n| | TypeScript (Pi) | Rust (Sir Bone) | Saving |\n|---|---|---|---|\n| Source LOC (no tests) | ~81,222 | ~10,771 | **−87%** |\n| Equivalent scope (shared features only) | ~12,720 | ~4,090 | **−68%** |\n| Binary + runtime | ~246 MB (Node + node_modules) | 11.35 MB | **−95%** |\n\n<sub>The honest reading is **−68%** (features present in both). 224 tests (unit + integration + proptest), mutation-tested with a hard CI gate on changed lines.</sub>\n\n**Benchmark (exploratory):** end-to-end on **SWE-bench Lite** — clone → one-shot agent → `git diff` → official Docker harness — **37/50 resolved (74%)** on the first 50 instances with `glm-5.2` (z.ai). This measures *harness + that model*, single-seed (astropy + django): a smoke-level signal, not a leaderboard number.\n\n---\n\n## Install\n\n**Prebuilt binary** (no Rust toolchain) — built by [`dist`](https://opensource.axo.dev/cargo-dist/), SHA256-verified:\n\n```bash\n# Linux / macOS\ncurl --proto '=https' --tlsv1.2 -LsSf https://github.com/automataIA/sir-bone-rs/releases/latest/download/sir-bone-rs-installer.sh | sh\n```\n\n```powershell\n# Windows (PowerShell)\npowershell -ExecutionPolicy Bypass -c \"irm https://github.com/automataIA/sir-bone-rs/releases/latest/download/sir-bone-rs-installer.ps1 | iex\"\n```\n\n```bash\n# From crates.io (needs a Rust toolchain)\ncargo install sir-bone-rs --locked\n\n# Or the latest from source\ncargo install --git https://github.com/automataIA/sir-bone-rs --locked\n```\n\n`cargo binstall sir-bone-rs` also works once a release is published.\n\n---\n\n## Quick start\n\n```bash\nsirbone login # create a global ~/.sirbone/.env (chmod 600), then edit it\nsirbone doctor # check local readiness — offline, no model call\nsirbone # TUI (default)\nsirbone --repl \"in playground, inspect the calculator and suggest one safe improvement\"\nsirbone \"list the files in src/\" # one-shot, non-interactive\n```\n\n`sirbone login` seeds **one global** `~/.sirbone/.env` so you configure credentials **once** and run\nfrom any directory. A per-project `.env` (or a real environment variable) still wins where present.\n`login` and `doctor` run without a key, so you never have to enter the TUI/REPL to set up.\n\n### `.env` configuration\n\nEdit `~/.sirbone/.env` (created by `sirbone login`) and pick a provider — auto-detected\n(`ANTHROPIC_AUTH_TOKEN` → Anthropic, else OpenAI-compatible):\n\n```env\n# Anthropic\nANTHROPIC_AUTH_TOKEN=sk-ant-...\nSIRBONE_MODEL=claude-opus-4-7\n```\n\n```env\n# OpenAI / compatible (Ollama, Groq, LiteLLM, …)\nOPENAI_API_KEY=sk-...\nOPENAI_BASE_URL=https://api.openai.com/v1 # omit for OpenAI default\nSIRBONE_MODEL=gpt-4o\n```\n\n### Other commands\n\n```bash\nsirbone audit # export a local session audit (no model call)\nsirbone ground PLAN.md # verify a doc's claims about the code, deterministically — gates CI\nsirbone --continue \"follow-up\" # resume the most recent session\n```\n\nTry the scripted TUI without a key in the [WebAssembly demo](docs-site/book/src/demo.md).\n\n---\n\n## Why it's different\n\nMost coding-agent products compete on autonomy. Sir Bone's bet is narrower and testable:\n\n- **Supervision first** — every shell/file/MCP action goes through a permission pipeline, with\n interactive approval for ambiguous or risky work.\n- **Token discipline** — prompt caching, real context-window tracking, compaction, working notes,\n truncation, and an optional spend cap.\n- **Mechanical reliability over clever prompting** — stale-edit guards, same-file mutation\n serialization, retry classification, secret redaction, lifecycle hooks, snapshots, rollback.\n- **Measured honestly** — experimental features (e.g. `architect`) stay opt-in/default-off when\n ablations don't show a net win.\n\n---\n\n## Features at a glance\n\n- **~15 native tools** by default — `bash` (+ background jobs), `read`/`write`/`edit`/`undo`,\n `grep`/`glob`, `web_fetch`/`web_search`, `load_skill`, `note`, `verify`, `code_map`, `historia`.\n Plus opt-in `architect` and feature-gated `doc_search` (`rag`).\n- **Permission pipeline** — `allow`/`soft_deny` globs + destructive detection + a command-injection\n guard (every chained segment must pass on its own) + an optional LLM classifier for ambiguous bash.\n- **MCP** — generic stdio client; servers declared in `~/.sirbone/mcp.json`, enabled per project;\n each remote tool registered as `mcp__<server>__<tool>`.\n- **Two providers, one trait** — Anthropic (SSE + prompt caching + extended thinking) and\n OpenAI-compatible (`async-openai`), both with classified retry/backoff and secret redaction.\n- **Real context compaction** at 87.5% — LLM summary of old turns, keeps the last 6, persisted as a\n session checkpoint so resume rebuilds the compacted transcript.\n- **Workspace snapshots + `/rollback`** — a shadow git repo commits the work-tree once per run before\n the first mutation; your project's own `.git` is never touched. Non-git dirs work too.\n- **Edit safety** — fuzzy multi-pass matching + a staleness guard that rejects edits to files changed\n since the last `read`, forcing a re-read instead of a lost-update overwrite.\n- **Background jobs** survive restarts — detached `bash`, live `⚙` gauge with ETA, `/jobs` report,\n completion bell.\n- **Lifecycle hooks** (no LLM round-trip) — `pre_tool_use` exit-code gate, `post_tool_use` lint-on-edit,\n `stop` re-loop gate.\n- **Project memory** (`HISTORIA.md`) — a per-project changelog the agent keeps and reads back across\n sessions.\n- **SSRF-guarded `web_fetch`**, **secret redaction in logs**, **token spend cap**, **tracing spans**.\n\n> Full detail per subsystem lives in the **[mdBook](docs-site/book/src/)**:\n> [tools](docs-site/book/src/tools.md) ·\n> [permissions](docs-site/book/src/permissions.md) ·\n> [MCP](docs-site/book/src/mcp.md) ·\n> [providers](docs-site/book/src/providers.md) ·\n> [sessions & snapshots](docs-site/book/src/sessions.md) ·\n> [configuration](docs-site/book/src/configuration.md)\n\n---\n\n## TUI\n\n<p align=\"center\">\n <img src=\"logos/mock-tui.webp\" alt=\"Scripted Sir Bone TUI mock demo\" width=\"960\">\n</p>\n\nSplit-screen ratatui: output panel + input. Renders markdown, diffs, tables, mermaid, code blocks;\nlive tool boxes (spinner → ✓/✗ + output preview, click to expand); y/n confirm dialogs for\ndestructive commands; background-job gauge; 6 palettes (`Alt+P`); info bar with provider, model,\ncontext %, and prompt-cache hit share (`⚡N%`).\n\n| Key | Action |\n|---|---|\n| `Enter` | Send prompt (or queue if the agent is busy) |\n| `Tab` / `Shift+Tab` | Move focus chat ↔ input |\n| `↑` `↓` | Scroll output (chat) · input history (input) |\n| `PgUp` / `PgDn` | Fast scroll |\n| `G` | Go to bottom (auto-scroll) |\n| `Alt+B` / `Alt+P` | Toggle boar panel · cycle palette |\n| `Alt+A` / `Alt+S` | About · Settings (localize, architect, thinking budget; skills / MCP pickers) |\n| `y` / `n` | Allow / deny a destructive command |\n| `Esc` | Drop queued message · cancel turn · close popup |\n| `Ctrl+C` | Cancel agent · twice within 2s to quit |\n\nMouse text selection needs a modifier (`Shift`/`Alt`/`Ctrl` depending on terminal) — details in\n[tui.md](docs-site/book/src/tui.md).\n\n---\n\n## Build from source\n\n```bash\ncargo build\ncargo clippy -- -D warnings\ncargo test\ncargo run # TUI\ncargo run --example mock_tui # TUI sandbox, no API key\n```\n\nThe [`playground/`](playground/) is a Rust project with intentional bugs and an automated task\nrunner (`bash playground/tasks.sh`) that runs the agent then verifies with `cargo test` — an\nend-to-end check of agent behavior.\n\n---\n\n## Stack\n\n`tokio` · `reqwest` + SSE (Anthropic) · `async-openai 0.40` (OpenAI) · `ratatui 0.29` +\n`ratatui-markdown` · `rmcp 1.7` (MCP) · `clap 4` · `schemars 1` · `serde` · `crossterm 0.29`.\nCrate-wide `unsafe_code = \"forbid\"`.\n\n---\n\n## License\n\nDual-licensed under [MIT](LICENSE-MIT) or [Apache-2.0](LICENSE-APACHE), at your option. See\n[NOTICE](NOTICE) — Sir Bone is derived in part from [Pi](https://github.com/earendil-works/pi)\nby Mario Zechner (MIT).\n\nUnless you state otherwise, any contribution you submit for inclusion shall be dual-licensed as\nabove, without additional terms.\n"}}}} |