audit-loop

Audit Loop Skill

Trigger

"Run an audit loop" / "Drain findings" / "Audit my app" / "Find bugs in the running system" / "Recursive audit"

Steps

1. Start the Drain Loop

Call the forge_tempering_drain MCP tool to run a multi-round tempering drain. The drain discovers findings from the running system, applies the classifier, and repeats until convergence (zero real findings) or the round cap is hit.

Pass through any user-supplied options:

• maxRounds — cap on iterations (default 5)
• env — target environment (default dev)
• dryRun — if true, discover findings but skip triage side effects

Review the drain result for:

• Number of rounds executed
• Per-round finding counts (the drain curve)
• Whether the loop converged or hit the round cap
• The final set of unresolved findings

2. Triage Each Finding

For every finding returned by the drain, call the forge_triage_route MCP tool to classify it into one of three lanes:

Lane	Meaning	Next Action
bug	Real product defect	File via forge_bug_register
spec	Feature or spec gap	Route to plan pipeline
classifier	Noise / false positive	Propose classifier PR

Collect triage results and group by lane.

3. Report the Drain Curve

Present a summary showing the drain trajectory and triage breakdown:

Audit Loop Results:
  Rounds:      N executed (converged: yes/no)
  Drain curve: R1: 12 → R2: 5 → R3: 1 → R4: 0

  Triage:
    🐛 Bug:        N findings → filed / ready to file
    📋 Spec:       N findings → routed to plan pipeline
    🔇 Classifier: N findings → classifier PR proposed

  Overall: CONVERGED at round N / HIT CAP at round N

Safety Rules

• Default environment is dev — never target production unless explicitly requested
• --dry-run skips all side effects (bug filing, spec routing, classifier PRs)
• Respect audit.mode in .forge.json — if set to off, warn the user and stop
• This skill orchestrates existing tools; it does NOT crawl or classify directly