Agent Failure Loop
v1.0.1An end-to-end self-improvement loop that automatically detects agent failures, classifies them, tracks recurrence, auto-generates rules, and promotes them to...
⭐ 0· 74·0 current·0 all-time
by@reikys
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
The name/description (auto-detect, classify, track, promote failures) aligns with the provided artifacts: two Python scripts and a SKILL.md describing reading failure records, analyzing patterns, and promoting rules into AGENTS.md/CLAUDE.md/.cursorrules. There are no unexpected external dependencies or credentials required.
Instruction Scope
SKILL.md and the scripts instruct the agent to read local failure records (default: memory/failures or failures directory), write structured analysis to .learnings/, and insert rules into target files (default AGENTS.md). This is consistent with the purpose, but it does grant the skill the ability to modify repository/local documentation files (AGENTS.md, CLAUDE.md, etc.). The scripts use only local file I/O; no network endpoints are present.
Install Mechanism
There is no install spec and the Python scripts use only the standard library. No external downloads, package installs, or extracts are performed by the skill bundle itself.
Credentials
The skill declares no required environment variables or credentials. The scripts optionally read environment overrides (AFL_LEARNINGS_DIR, AFL_TARGET_FILE, AFL_FORMAT, AFL_MIN_COUNT) which are reasonable convenience hooks and do not grant extra privileges.
Persistence & Privilege
The skill is not always-included and does not request elevated agent privileges, but it does modify local files (creates/updates .learnings/*, promotable.json, and can insert lines into AGENTS.md/CLAUDE.md or create new files in plain mode). This is expected for its purpose but you should be aware it can autonomously change local rule files when run (there is a --dry-run option).
Assessment
This skill appears internally consistent: it reads local failure logs, analyzes them, and can write promoted rules back into local rule files. Before running or installing: 1) backup AGENTS.md/CLAUDE.md (or run in a branch) so automatic inserts can be reviewed; 2) run the scripts with --dry-run first to inspect what would be changed; 3) verify the failures/ or memory/failures directory contains only the entries you expect (the tool groups by a hash of the 'cause'); 4) if you intend to run automatically (cron), ensure you trust the promotion logic and review promotable.json periodically. No network exfiltration or credential access was found in the provided files.Like a lobster shell, security has layers — review code before you run it.
latestvk9776ffq89b11z9p0xg0efz39h83nm58
License
MIT-0
Free to use, modify, and redistribute. No attribution required.