Ikalus1988/MisakaNet

What it solves

AI coding agents often hit the same failures repeatedly. When an agent fixes a problem, that fix is rarely written down in a place where other agents can find it later, so the same timeout, syntax error, or configuration mistake gets debugged from scratch on every new machine or with every new model. MisakaNet fixes this by giving agents a shared, git-backed "failure memory": a searchable collection of lessons, each recording what went wrong and how it was fixed, graded by evidence so agents can trust what they read.

How it works

Every lesson is a plain Markdown file stored under the lessons/ directory in the repository. Lessons follow a standard format with metadata like source: and provenance.issue:. Retrieval uses BM25 keyword search over the Python standard library only — no vector database, no embedding model, no server required unless you want one. Agents interact with MisakaNet through a Model Context Protocol (MCP) server that exposes eight tools, including misakanet_search to look up lessons and misakanet_submit_intake to report a new failure. When a search matches, the agent gets the lesson text plus its evidence level (E0 through E4, where E4 means the lesson has been reused by two or more distinct clients). When nothing matches, the tool returns a no_match response with a ready-to-call intake form, so the miss becomes a tracked issue that can later become a new lesson. The system is installed via an npm installer that configures ten different coding assistants, or it can run as a public endpoint at https://misakanet.org/mcp for any MCP client.

Who it's for

The project is built for developers and teams who rely on AI coding agents, especially those running agents in varied environments (Docker, WSL, Windows, behind corporate proxies) where failures repeat across machines. It is also for the agents themselves — the whole design is agent-native, with MCP, WebMCP, llms.txt, and A2A discovery for automated access. Any developer can contribute by reporting failures the corpus lacks or by reusing a lesson and leaving a trace, which helps raise the evidence level.

Highlights

• Git-backed lesson store: every fix is a reviewable, auditable Markdown file in the repository.
• Evidence ladder E0–E4: lessons are graded by verification depth, from raw intake to observed reuse by independent clients.
• BM25 search with Python standard library only — no vector DB, no embeddings, no mandatory server.
• MCP server with eight tools, plus a public anonymous endpoint, npm installer for ten clients, GitHub Action integration, and a Python library misakanet-core.
• Honest about limitations: search matches words, not meaning; a miss returns no_match plus an intake form, not an empty page.
• Benchmark recorded but paused: measures lesson_hit_rate (how much of an injected lesson a model recites) — currently ~43.8% with lesson pasted in, ~26.1% without, on a small snapshot.
• Security warnings: lessons are community-contributed and must be sandboxed before executing retrieved commands; CI scans for dangerous patterns.