Misakanet

English | 日本語 | 简体中文

MisakaNet

mcp-name: io.github.Ikalus1988/misakanet

The same failure, paid for twice.

A pip install times out behind a corporate proxy. An agent finds the fix. Days later the same

timeout appears on another machine, under another model — and is debugged from scratch, because the

fix was never written down anywhere the next agent looks.

MisakaNet is that place: a git-backed failure memory. Every lesson is a Markdown file in this

repository — graded by evidence, reviewed like code, retrieved by an agent instead of guessed at.

You fail once; the next agent does not.

Agent-native interfaces: MCP server (8 tools), WebMCP, llms.txt, A2A discovery via .well-known/agent-card.json. The local stdio server (pip install misakanet) ships 11 — the hosted endpoint is a subset; docs/mcp.md has both rows.

Core

Install

Ecosystem

---

Install (30 seconds)

The point of an install is a first successful search, not the install. None of these needs an account or a token.

The installer-managed row is what npx @misaka-net/misakanet-setup writes into each host's config. For DeepSeek Harness that is the rules half only ($DSH_HOME/AGENTS.md): its tools come from the plugin in the row above, which the installer prints as the one manual step.

Updates: dsh plugin --profile web update misakanet@latest. Installer updates: npx @misaka-net/misakanet-setup@latest. Full prerequisites and the two-package trap: How to use it.

Your first five minutes

1. Ask it something you have actually broken. No client installed? This curl is the whole interface:

curl -sS https://misakanet.org/mcp \

-H 'Content-Type: application/json' -H 'Accept: application/json' \ -H 'MCP-Protocol-Version: 2025-06-18' -H 'Origin: https://misakanet.org' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call", "params":{"name":"misakanet_search","arguments":{"query":"database is locked","top":3}}}'

2. Read what came back. A hit returns the lesson, its evidence_level, and its path — open the file and read the fix before running any command in it. A miss returns no_match plus a ready-to-call intake: the corpus saying it does not know this one yet.

3. Check the trust signal, not the tone. E0–E3 say how far a lesson was verified; E4 is the one nobody can write or vote into.

What is MisakaNet?

Git-backed failure memory for AI coding agents. An error shows up → the agent searches → it applies a fix somebody already verified → if nothing matches, the miss becomes an intake, and the intake becomes a lesson for the next agent.

Every lesson is a Markdown file: reviewed like code, graded by evidence, retrieved with BM25 over the Python standard library. The hosted MCP endpoint adds an optional semantic fallback (bge-m3 + Vectorize) only when BM25 returns no match; MISAKANET_SEMANTIC=0 disables it. No vector database is required locally, and no server unless you want one.

What it answers — and what it does not

MisakaNet answers for the failures it has indexed, and only those. When nothing matches the reply is no_match plus a ready-to-call intake — the corpus saying "not this one yet", not a dead end.

Lesson vs Skill

A skill teaches an agent how to do something; a lesson records what went wrong before and how not to fail again. MisakaNet is only the second thing. → FAQ

Where it is today

• The numbers that move are badges, on purpose: the Lessons badge renders data/badges/lessons.json, regenerated by update-badges.yml, so it can lag a run but is never a hand-typed total.
• The top of the evidence ladder is empty, and this page says so. E4 needs an observation — two distinct clients actually retrieving a lesson — and the corpus has none yet (distribution, regenerated and checked by CI).
• Search is BM25 first; the hosted endpoint can fall back to embeddings. Local/offline remains stdlib BM25; the fallback runs only when BM25 returns no match, and MISAKANET_SEMANTIC=0 turns it off. A corpus gap is still a miss: the reply is no_match plus an intake, not an empty page.
• Reported gaps arrive faster than they are triaged. The intake queue is the honest bottleneck.
• The corpus is downloadable and citable. data.misakanet.org publishes a versioned bundle — the full corpus, the benchmark query set, and every run artifact — with a per-file SHA-256 and an explicit license per file (DATASET.md). The counts live in each version's manifest.json, not in prose.

Retrieval benchmark: 267 queries, BM25 vs semantic fallback

On 2026-10-27 the hosted endpoint's retrieval path was evaluated on 267 queries, one per lesson generated by Workers AI 70B. Both arms used the same eval worker and corpus; only MISAKANET_SEMANTIC changed.

The fallback improves R@1 by 23.6 points and cuts no_match by 25.9 points. It does not meet the R@1 half of the target (0.446 < 0.46); no_match is within the target (0.453 ≤ 0.50). A 113-query subset reaches 0.460/0.460, while 250 reaches 0.404/0.504: the sample is large enough to show the fallback works, not large enough to call R@1 solved.

This is 267 queries from the 300-entry pool, not the full 650; 33 pool entries yielded no query, 2 semantic requests errored, and the eval corpus was not refreshed. → full report

What changed since: the gate, not the model

The table above measures a configuration production no longer uses. Until 2026-10-28 the semantic path ran only when BM25 returned nothing, so a confident-but-wrong lexical hit suppressed it — 13 of the 20 recorded misses were exactly that. For queries containing CJK characters the two ranked lists are now fused (reciprocal-rank) instead of one gating the other.

Measured on the live endpoint, 64 Chinese queries, 0 transport failures:

For scale: bge-m3 pure cosine already reached 0.636 on the cross-lingual bucket, and a gemini embedding 0.727 — so most of the gap was the decision to consult the vector path, not the model. English queries are untouched (the fuse is gated on CJK characters) and sit at 15/20 against a floor of 16, which is where they were before this change. → full report

Recitation benchmark: what it measured, and why it is paused

One narrow question: when you hand a model a lesson, how much of that lesson comes back? The metric is lesson_hit_rate — the share of the injected lesson's commands reproduced in the answer. Read the setup before the number: the scenario in this benchmark is each lesson's own title, the injected lesson is *that same lesson*, no retrieval call is made, and correctness is not checked. It measures the recitation half of RAG: necessary for MisakaNet to help, and not evidence that search finds anything.

The run is paused. Two independent runs of ≈500 scenarios each on 2026-09-22: 46.1% with the lesson pasted in, 23.3% without. What remains is t