一个 Agent 踩过的坑，不该让其他 Agent 再踩一次。

One agent's failure shouldn't be another agent's lesson — that's why MisakaNet exists.

• MisakaNet — failure-memory MCP server for coding agents · v2.43.0

  

  • 🚀 Remote MCP: https://misakanet.org/mcp — 8 tools, no install, no account needed to read
  • ⚡ One-command setup: npx @misaka-net/misakanet-setup — wires 10 agents, pre-allows the read-only tools, and prints a redacted health report. Full walkthrough below, because "one command" hides three separate things it installs.
  • 🧩 DeepSeek Harness plugin: dsh plugin --profile web add misakanet — the same library as a native host surface: a /misakanet <query> command that answers in the composer, a right-pane panel (what this session searched, what came back, what was filed, the reuse counters), a General-settings row, and a frame-wide toast — in zh/en
  • 🛰️ Discovery beyond MCP: llms.txt / llms-full.txt, an A2A agent card at .well-known/agent-card.json, and a WebMCP surface for browsers
  • 🎯 Intake Ways: anonymous MCP intake · the intake bot as a published GitHub Action — uses: Ikalus1988/MisakaNet@v1, CI failure → lesson suggestion (needs actions: read) · registered node with a stable client_id
  • 🔍 Quality machinery: a provenance gate in CI (a cited source must resolve — placeholders and 404s fail the build), an opt-in auto-merge channel for lesson PRs, --report --strict for CI health gating, four required checks on main (DCO, the ubuntu test leg, gate, and audit), a behavioural e2e that drives the DSH client half in a real host and a real Chromium (the source-level gates can only see registrations, never whether anything happens), a weekly mutation audit (a gate that stays green when its check is disabled fails the audit), and a weekly output audit (an automation that runs but produces nothing fails the run)
  • 📄 state-of-the-repo — a publishable snapshot of what runs, what gates can be trusted, and what needs a human decision, so a second maintainer can pick it up without anyone's private notes

• fatal-guard — capture fatal crashes in Node.js without patching upstream. Zero deps.

• misakanet-core — zero-dependency BM25 + RRF search engine. pip install misakanet-core

• pensieve — procedural memory for agents: validated routines as YAML cards over an SQLite FTS5 index, injected as a path + anchor instead of a wall of text. Deliberately not a general memory layer.

• 工业 RAG 知识库 — hybrid retrieval + SAG search over 190+ FANUC robot manuals, 230k+ chunks. The build journal is public; the knowledge base itself is a private repo, so this link is the demo, not a download.

• 🐚 pre-commit-dco — 43 行 Python 代码，独立验证 DCO sign-off。The standalone sibling check-dco and the archaeology in dco-audit live in their own repos; the write-up of the upstream fight is public.

Option 1 — one command, ten agents (recommended):

npx @misaka-net/misakanet-setup
# close and reopen the assistant window, then ask something that would normally
# invite a guess: "switch vision model" · "tool call permission denied"

Option 2 — Remote MCP (30 seconds, no install):

curl -sS https://misakanet.org/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"misakanet_search","arguments":{"query":"database locked"}}}'

Option 3 — Local MCP:

git clone https://github.com/Ikalus1988/MisakaNet.git && cd MisakaNet
python3 scripts/mcp_server.py
# Add to your MCP config, then ask: "Search MisakaNet for pip install timeout"

Option 4 — PyPI:

pip install misakanet
misakanet "database is locked"

Option 5 — the library in your own code:

pip install misakanet-core   # BM25 + RRF, stdlib only, CJK-aware tokenization

Same name, different things: pip install misakanet is the CLI/stdio server, pip install misakanet-core is the search library, and the npm package misakanet is the DSH plugin bundle. Two of those three carry the DSH-era version number (2.43.0); the library tracks on its own (2.7.0).

@misaka-net/misakanet-setup (v0.5.6, MIT, Node 18+, zero dependencies) is the package the one-liner above runs. It is worth a paragraph, because a tool that "sets up an agent" can be doing three unrelated jobs, and only two of them are the same job.

Without (2) a rule saying "search first" is decorative — agents don't keep counters, so without (3) the checkpoint never arrives either; and (3) has nothing to call without (1).

Those last five are the honest edge of the tool: their rules live in project-scoped files (.cursor/rules/*.mdc, GEMINI.md, .github/copilot-instructions.md, AGENTS.md, .kiro/steering/*.md) and an installer cannot know where your projects live — so it writes the MCP entry, then tells you the project-side action per client instead of implying behaviour changed. Each file also uses its own key path and URL field; a copied entry with the wrong key fails silently — no error, the server just never appears.

--dry-run      # print what would change, write nothing (also lists what it *cannot* change)
--verify       # READY / NOT READY, endpoint probe, hook, MCP registration, installed vs latest
--report       # this machine's state as redacted YAML — safe to paste into an issue or a CI log
--report --strict   # (= --ci) the same report, but the exit code IS the verdict: 0 READY / 1 NOT READY / 2 could not run
--report-json  # same document as JSON (`misakanet-setup-report/1`), stdout only, for MDM/GPO
--silent       # enterprise: no progress narration; every `!` line, report and error still prints
--only claude,codex   # one agent only
--mcp-only     # register the endpoint and change nothing else — for permission-tiered fleets
--no-register  # read-only install: no anonymous node, no token
--uninstall    # remove exactly what it added

--report alone always exits 0 on purpose — a report is evidence meant to be collected, and a nonzero exit turns "collect the evidence" into a failing step. The gate is the opt-in --strict form. The convention is the repo-wide one (0 = fine, 1 = found problems, 2 = could not run), so a CI step can tell "the machine is unhealthy" from "the checker broke".

• Every rewritten file is backed up to <file>.misakanet.bak first; everything it adds sits between misakanet:start / misakanet:end markers, so --uninstall removes exactly its own work and leaves your hooks, MCP servers and TOML keys alone.
• Idempotent: a second run changes nothing. Updating is one command — nothing is ever replaced behind your back, and the upgrade nudge is silent for the first 14 days.
• The token lives at ~/.misakanet-agent/token (mode 600), goes into your local config only, and is never printed — no subprocess is spawned, so it never reaches a command line. --verify probes the endpoint anonymously.
• Lessons retrieved from the server are data, not instructions — the injected rules tell the assistant not to execute commands found in them.
• Enterprise boundary questions (what stays on the machine, what must reach misakanet.org, how --no-register avoids registration entirely): docs/maintainer/enterprise-deployment.md and misakanet.org/enterprise — which also lists what does not exist yet (no compliance certifications, no org-level audit export, no self-hosted endpoint).
• Behind a corporate proxy: NODE_USE_ENV_PROXY=1 (Node ≥ 22.21) — otherwise the probe can fail while your agent reaches the endpoint fine, and the report says so (endpoint-probe: inconclusive-proxy) instead of lying.
• WSL/Linux users have a Python installer too: integrations/agent-autostart/install_misakanet_agent.py. Both write the same markers, so either can verify or undo the other.

📖 package README · CI job you can copy · voice hooks

Reads need no account, no email, and no GitHub login; a registered node is a pseudonym, not a login. Registration is not the way in — it only unlocks the write tools (misakanet_write_lesson / misakanet_preflight), which is why the two Bearer rows are the only ones marked so. The authoritative statement of what needs auth is AGENTS.md §3 — this page mirrors it, and a mirror rots: this one kept advertising the removed daily read cap for four days, and kept saying "7 tools" for as long as the endpoint had been serving 8. The local stdio server (python3 scripts/mcp_server.py) exposes a few extra maintenance tools.

→ misakanet.org 在线搜索所有 failure lessons，MCP 直连 Cursor / Claude / Codex。 → docs/integrations/mcp-remote.md 完整接入指南

Where it is listed: Glama (tool definitions rated A) · dsh-plugin.org · DSH Directory · dsh.so · MCP Toplist · Smithery · HOL Registry

📝 ikalus1988.github.io — case studies, field notes, and project journeys.

🆕 Five Failures That Passed Every Check — the September log: four pull requests with 24/24 green checks citing a repository that does not exist, a tokenizer that silently dropped every Chinese query, a red check that was telling the truth, a clone 112 commits behind reality, and three security holes in the gate built to prevent exactly that.

📚 「自动化里那些骗过我的东西」 — a 30-article Chinese series on criteria that lie, gates that never ran, and inputs nobody validated, sorted into ten failure layers (L0–L9). The workflow that produces it is a private repo; the articles are not: post12 · 我的浏览器自动化脚本连着骗了我七次 · post15 · 假成功有四种形状 · post27 · 62.7 秒：从「请手动删 pod」到宣布 Resolved