loop-audit
Skill1 starsChanged 5 months agoMITView on GitHub
Health and security audit of running loops, automations, and scheduled agents — quiet failures, gate rot, self-grading, permission scope creep, secrets in logs, unread diffs, cost per accepted change. Use monthly, or when asked whether existing loops are still safe and earning their keep.
3 other projects use a version of this.
Adopted by
Most recently taken up by
- HTMLGuyLLC/ClaudeSkillstheir own version, first seen Nov 1, 2026
- Moriarty33/jb-kaizentheir own version, first seen Oct 20, 2026
- patr-coder/robobuilder-protheir own version, first seen Jul 3, 2026
Other versions3 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 3
- Moriarty33/jb-kaizen★ 1same name
- HTMLGuyLLC/ClaudeSkills★ 0same name
- patr-coder/robobuilder-pro★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository13 other files it gives its agents.
- adversarial-verifySkill
- closed-loopSkill
- cluster-feedbackSkill
- doubt-gateSkill
- fleet-loopSkill
- hill-climbSkill
- keep-greenSkill
- loop-readinessSkill
- loop-stateSkill
- prod-watchSkill
- scaffold-planningSkill
- ship-preflightSkill
- source-groundedSkill
The file
Files
Loop Audit
A loop running unattended is also an attack surface and a cost center running unattended. Audit every active loop monthly. Verdict per loop: healthy, fix, or kill.
The metric
Cost per accepted change — not tokens spent, not tasks attempted. If the accepted-change rate is below 50%, you're doing the review work the loop was supposed to save, and the loop is losing. Measure: changes merged ÷ changes proposed since last audit.
Failure modes to check
| Mode | Symptom | Fix |
| Quiet exit (Ralph Wiggum) | Loop declares done on a half-finished job; no objective signal failed it | Replace soft conditions with a gate that returns pass/fail: test, build, lint exit code |
| Self-grading | The agent that wrote the output also verified it | Separate verifier context with no exposure to the maker's reasoning (see /adversarial-verify) |
| Goal drift | Constraints from the original brief no longer enforced after long sessions | Standing spec file reread each run (see /loop-state) |
| Agentic laziness | "Done enough" at partial completion | Objective stop condition checked by a fresh context, not the maker |
| No hard stop | Loop runs until a rate limit or invoice kills it | Token budget, iteration cap, or time limit — pick one, enforce it |
| Gate rot | The test that approves the loop's PRs no longer catches the failure mode it was built for | Spot-check: pick 2–3 recent loop PRs, verify the approving gate actually fails when the bug is reintroduced |
Security checks
- Unreviewed merges — does the gate include security checks (SAST, dependency audit, secret scanning), or just functional tests? A loop opening PRs faster than humans read them merges insecure code automatically.
- Skill injection — any skill the loop loads was source-audited before install. Never auto-install community skills; audited samples show a meaningful fraction leak credentials or carry prompt injection in descriptions.
- Credentials in logs — verbose/debug logging off in production loops; spot-check recent logs for secrets.
- Permission scope creep — re-audit permissions every 30 days. A loop tested read-only that gained "just one" write permission for convenience is the canonical hole.
Comprehension debt checks
- Diffs are being read. Unread loop output is comprehension debt at compound interest — the bill is the day you debug a system no one has read.
- The loop is blocked from judgment-call work. Small, machine-checkable changes only. Architecture, auth, payments stay human.
Report format
Loop: <name> | Last audit: <date>
Accepted-change rate: <n merged>/<n proposed> (<pct>) — <above|below> 50%
Failure modes: <none | mode: evidence>
Security: <clean | finding: detail>
Comprehension: <diffs read | n unread PRs>
Verdict: HEALTHY | FIX (<what>) | KILL (<why>)Loop usage
/schedule monthly: run /loop-audit across all active loops and routines.Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.