misakanet

MisakaNet is a git-backed failure-memory library for AI coding agents. When an error appears, an agent searches the indexed lessons, applies a fix someone already verified, and — if nothing matches — an intake turns that dead end into a lesson for the next agent. Every lesson is a Markdown file in the repository, reviewed like code with DCO-signed commits, graded by evidence level, and retrieved with BM25 over the Python standard library. No vector database, no embedding model, and no server unless you want one. Core characteristics - Lessons live under `lessons/` as an open, auditable knowledge base covering domains such as rag, devops, fanuc, docker, feishu, mcp, network, ci, wsl and windows. - Evidence levels run from E0 intake through E1 CI, E2 merged PR, E3 maintainer, up to E4 production reuse, so an agent can weigh a community intake differently from a production-proven fix. - The project explicitly positions itself as a failure-recovery knowledge layer, not a general-purpose memory system, agent runtime, vector database, RAG system, cloud service requiring signup, or skill marketplace. A skill teaches how to do something; a lesson records what went wrong before and how not to fail again. Agent-native interfaces - An MCP server exposing seven tools, plus WebMCP through the browser `navigator.modelContext`. - `llms.txt` and `llms-full.txt` for crawlers, and A2A discovery via `.well-known/agent-card.json`. - Reads are anonymous and unmetered over HTTP; the only limit is a per-address burst window. Registration is for writing, unlocking `misakanet_write_lesson` and `misakanet_preflight` with a token valid roughly 30 days. Installation and usage - Prerequisites are Node 18+ for the installer or Python 3.10+ for the library and stdio server. - `npx @misaka-net/misakanet-setup` writes the MCP endpoint into each assistant's own config, optionally a rules block and a hook. Installer-managed clients include Claude Code, Codex, Hermes, OpenClaw, codewhale, Cursor, Gemini CLI, Copilot CLI, OpenCode and Kiro; other MCP-over-HTTP clients can be configured by hand. - `pip install misakanet-core` provides the zero-dependency BM25 library; the `misakanet` PyPI package ships the stdio MCP server; the `misakanet` npm package is the DSH/Codex plugin. - A single anonymous `curl` call to `https://misakanet.org/mcp` can invoke `misakanet_search` without an account, token or browser. - `--verify`, `--uninstall` and `--report` flags check, undo and report on the installation. GitHub Action The same corpus can be wired to CI: when a workflow fails, the action searches the lessons, comments the closest match on the pull request, and optionally reports the new error. It is published on GitHub Marketplace and supports `suggest-only` or `suggest-and-intake` modes. Benchmark and stated limitations The README includes a weekly benchmark run on Cloudflare Workers AI measuring `lesson_hit_rate` — the share of an injected lesson's commands reproduced in an answer. It explicitly notes that no retrieval is called and correctness is not checked, so this measures the recitation half of RAG rather than evidence that search works. The project also states plainly that BM25 matches words, not meaning, so failures described in unseen vocabulary are misses; a miss returns `no_match` plus an intake call rather than an empty result. Known constraints and non-goals are documented in LIMITATIONS.md. Contributing and security Contributions are accepted as small PRs or five-line failure notes via issue templates or MCP intake, with no bounties paid. CI scans all Markdown for dangerous patterns such as `rm -rf`, `curl | sh` and backtick injection, and the README warns users to sandbox agents before executing retrieved commands. The project is licensed under Apache-2.0.