. =%%*******##%@@+.
:%*++#*+%%%##%%*+:: .*@%%=
+#%*%#*#%%%##+#**#%%%%#+%%- #*%+
=%*%%*#%%********%**%%%%%%#+###:#**.
. .: #@%%***#*#%##%%%%#%%%*%@%%%###+**#* =
.-=-%%%%%%%*#%%%%%%%@@%#@@%%**%%####+***#*
:=+@#%@@@@%*%%@@@%@@@%%%#%%%%*+*#+*#*****#*=
-%*%%@%%%%*%@%%%@@%%%%#-#%%#@+#=#%%#-#******
%#*%#%%%%%%#%####%@##*#**##%%%#%+-*#--#-==+:*
=##*%#*#***##*%%###%@%%%*=#####*##=-+**-*+--+=
*%#%%%#%%%%%#=#%%##%%%%#*#*+####***-=++-=:.*#=*#-
%****##*##%###=*######%#**#****+-*+---=++*..*=***#
+**#=*=*=*=*##-*=**%##*#.#%#*=*+ +--- .++#.:.#.++*
*#+##.+-:*+.=-+=:%#=++=--##%#:=- -*- --:+* --=+===
#:-*-*#*:...=.**+####+*:*-% =:*--== .--.: :-=*===
#--%**#%==#**+###%%%*##%%*####%+ +*+. :-:.:-==-=+-
*%-.=+#**+%+#*#%%%%%%%%%%%##%%%#*=++:-:--:*=:*=-:-
+:: +-*##%*%%%%%%%%@%%%%%%%%%%*++*- .=*:#:-.:.*.*+
-----.-*+-##+=##%%%%%%%%%%%%%#:= :.=----=:-+* *.
=-+---.-:*=#-:=%%%%%%%%%%%%%%- .---+- -. -
:=*+==- :**+#*#%%%%%%%%%%%. + .:::=-# =
: =----..:- : .-*%%%%%%#+ ..-:.-=.-:-.+
#+--- :+:-:-+++**+++## : + := .
:*#++++*#*# .
:**#%+*##**:
.-#**##:=##**##..
.::.#%%%%%. *#%###-....-:
.:........:.:-#%#*....+%%%#:..:.::.....:..
. ........-#**## .#*+## ...........::...
.....+#%%.:..=%%%-.:........ ..:....
. ..%%#:....### .... ..
. :@#. ..%@: .
## :
.
AI Agent Infrastructure
I build auditable knowledge systems for AI agents.
MisakaNet turns real engineering failures into searchable lessons for Cursor, Claude, Codex, and other coding agents. One agent's failure shouldn't be another agent's lesson.
一个 Agent 踩过的坑，不该让其他 Agent 再踩一次。MisakaNet 将真实工程失败经验转化为编程 Agent 可搜索的课程。
Automation pollution refers to the progressive degradation of knowledge quality when automated systems (bots, scrapers, LLM-generated content) write back into a knowledge base without human audit.
当机器人、爬虫、LLM 生成内容等自动化系统未经人工审计就回写知识库时，知识质量会逐步退化；相关工程实践的目标就是阻止这种退化进入长期知识资产。
02 / Projects
03 / Artifacts
Primary artifact and narrative center: a Git-backed swarm knowledge system for verified lessons, autonomous nodes, audit loops, and public construction.
Snapshot updated: 2026-07-29
Feishu-native RAG system for FANUC robot manuals, built around hybrid retrieval, entity indexes, RRF fusion, brand-pollution filtering, and a four-layer quality loop.
Snapshot updated: 2026-07-29
04 / Case Studies
How we rejected pygrep, built a standalone check-dco in 43 lines, survived upstream rejection, and achieved full supply chain autonomy.
How we shipped a fatal-error hook to OpenClaw through 3 architecture audits, shell injection fixes, secret isolation, and ClawSweeper's gold shrimp rating — without ever building the runtime.
How Git commits, lessons, nodes, and audit loops turn public construction into a durable decentralized knowledge state.
How 190+ FANUC PDFs became a Feishu-native RAG loop with hybrid retrieval, entity search, and bad-case learning.
Zero-dependency Node.js runtime guard — capture crashes via wrapper mode (fatal-guard -- <cmd>) or preload (node -r), route 4-field payload to any external handler.
The shared pattern behind MisakaNet and Industrial RAG: resist pollution, preserve evidence, and keep knowledge auditable.
05 / Field Notes
A colleague didn't know what a "skill" was. They just dumped all of Claude's memory. From that chaos, we distilled kcantrans VR-variable access — a fix that was already working before it became a lesson. MisakaNet turns messy agent memory dumps into reusable failure lessons.
14 days: 6,074 clones, 738 unique cloners. Helpful votes: zero. The API worked. Nobody clicked. This revealed that GitHub traffic ≠ lesson consumption — and launched the Usefulness Proof Sprint.
Skills teach agents how to do things. Lessons teach agents what went wrong before and how not to fail again. MisakaNet is not another skill marketplace — it is a shared failure-memory layer.
7 contributors reached. Most came via bounty labels. Post-PR interaction: zero. This shifted strategy from "chase PR counts" to "chase usefulness proof."
Non-technical users don't need lessons. They need 3-step rescue cards: "Move file to desktop. Use 7-Zip. Re-download." Screenshot + one sentence → structured knowledge.
glama.json ownership. awesome-mcp-servers PR #10044. 16 Glama Chat QA queries covering DCO, pip, GitHub API, secret scan, agent setup, and negative tests.