Agent Skills: iterative-converging-audit — Audit to Convergence, Not to Exhaustion-of-Patience

iterative-converging-audit — Audit to Convergence, Not to Exhaustion-of-Patience

A single audit pass is never thorough. The first sweep has blind spots; the fixes themselves reveal new instances or introduce new ones; and "I looked once and stopped finding things" usually means "I got tired," not "there is nothing left." The only honest definition of a complete audit is a converged one: a re-run that finds zero new discoveries.

This skill turns that into a structurally enforced loop. It applies to ANY "find all instances of X" task — security audits, safety audits, code reviews, research sweeps, compliance checks, dependency audits, dead-code hunts, "are there other places we do this wrong" investigations.

This skill enforces the Iterative Audit to Convergence constitution standard (docs/STANDARDS-REGISTRY.md). You do not declare an audit complete on round 1. You track rounds. You re-audit after every fix batch.

When to Activate This Skill

• Any request to "find all", "audit", "sweep", "review everything", "make sure we got everything".
• After fixing a bug, when the same class of bug likely exists elsewhere ("where else do we do this?").
• A security or safety audit where a missed instance is dangerous.
• A research/review question where one source or one angle is not enough.
• Whenever you catch yourself about to say "I checked, looks clean" after a SINGLE pass.

The Loop (do not skip steps)

Step 0 — Frame the audit (write this down first)

Before sweeping, define — explicitly, in the audit ledger:

1. Target pattern — what exactly are you finding? Be precise. "Silent LLM fallbacks in gating paths", not "bad error handling".
2. Search surface — where could instances live? List the angles/locations. The first list is always incomplete; you will add to it.
3. Classification rule — for each finding, what are the buckets? (e.g. DANGEROUS / MITIGATED / ADVISORY-OK). A finding is not "done" until it is classified.
4. Fix policy — what does "fixed" mean per bucket? (remediate / annotate-as-accepted-with-reason / escalate).
5. Convergence criterion — usually: a full re-sweep surfaces zero findings not already in the ledger. State it concretely.

Step 1 — Audit pass (round N)

Sweep the search surface for the target pattern. Record EVERY finding in the ledger with: location, current behavior, bucket. Cast wide — false positives are cheap to classify out; missed instances are the failure mode this skill exists to prevent. Use multiple search angles (by-name, by-content, by-structure); one angle is blind to what the others catch.

Step 2 — Fix pass

For each finding, apply the fix policy: remediate it, OR classify it as accepted with an explicit written reason (an accepted finding is a DECISION, not a TODO). Capture the diff. Fixing changes the code, which is exactly why you must re-audit.

Step 3 — RE-audit (round N+1)

Sweep again — the FULL surface, not just what you touched. Two things happen on a real re-audit:

• Your search surface grew (round N taught you new places/angles to look). Add them.
• The fixes may have moved, masked, or created instances. Catch them.

Record any NEW findings. If there are new findings → back to Step 2. If a re-sweep finds nothing not already in the ledger → converged.

Step 4 — Declare convergence (honestly, and MACHINE-EARNED where possible)

Only now may you say the audit is complete — and you state it as: "Converged after K rounds; round K surfaced no new findings. Ledger: X total, Y fixed, Z accepted-advisory (each with a reason)." If you stopped for any reason OTHER than a clean re-sweep (ran out of time, budget, patience), say so explicitly — that is an INCOMPLETE audit, not a converged one. Never dress up an exhausted audit as a thorough one.

In a repo that carries the validator (the instar source tree, or any repo vendoring scripts/write-audit-convergence.mjs), the converged claim is EARNED, not asserted: write the canonical report at docs/audits/<slug>.md (see The Ledger below) and run node scripts/write-audit-convergence.mjs --audit docs/audits/<slug>.md. The validator refuses the stamp unless the ledger genuinely shows ≥2 rounds ending in a zeroed final round, every finding closed, and a standing guard (or a closed-enum exemption). A hand-typed converged: is rejected at commit and re-checked in CI. In a repo WITHOUT the validator, still write the canonical report and self-check against the shape — but never fabricate a machine stamp you cannot earn.

Step 5 — Leave a standing guard (so it stays converged)

A converged audit decays the moment new code lands. Where possible, encode the target pattern as a ratchet (a test/lint that fails when a new instance appears) so the audit does not silently un-converge. The ledger of accepted findings becomes the ratchet's allowlist. (Example: the no-silent-llm-fallback test is the standing guard for the LLM-fallback audit.) Name the ratchet in the report's standing-guard: field — or, if the pattern genuinely is not CI-expressible, record a closed-enum exemption: (non-ci-expressible | external-system | one-time-human-review) with a real rationale.

The Ledger (the durable artifact) — canonical at docs/audits/<slug>.md

The audit ledger IS a canonical report at docs/audits/<slug>.md with frontmatter (audit, target-pattern, search-surface, converged [stamped by the validator only], standing-guard XOR exemption) and one ## Round N section per pass. Each round records the search angles run, the surface delta since the prior round, a findings table (location | behavior | bucket | disposition), and a New findings this round: <count> line. Dispositions are closed: fixed:<commit/PR>, accepted:<reason>, or deferred:<tracking-ref>. Reference each finding by path+line — NEVER paste the offending secret/credential material into the ledger (the commit gate scans audit reports for credentials and will block). The report is what makes "converged after K rounds" a machine-verifiable claim instead of a feeling; the new-findings-per-round count falling to zero is the convergence signal.

Anti-Patterns (this skill exists to forbid these)

| Anti-pattern | Why it fails | The fix | |---|---|---| | "I checked, looks clean" (one pass) | Round 1 always has blind spots | Re-audit at least once; converge | | "Fixed the 3 I found, done" | Fixes reveal/create new instances | Re-sweep AFTER fixing | | Re-auditing only what you touched | New instances hide in untouched code | Re-sweep the FULL surface every round | | Treating an accepted finding as a TODO | It silently rots; later reader can't tell reviewed-OK from missed | Every accepted finding carries a written reason | | Declaring "thorough" when you stopped for time/budget | Dishonest; the gap reads as "covered" | Say "incomplete — stopped at round K with N open", never "thorough" | | One search angle | Each angle is blind to what others catch | Sweep by-name AND by-content AND by-structure | | No standing guard | The audit un-converges on the next commit | Leave a ratchet (test/lint) keyed on the pattern |

Relationship to other tools

• For a fan-out across many files where you only need the conclusion, delegate the per-angle sweeps to subagents and converge their findings — but the convergence loop (re-audit until clean) is still yours to run.
• The Coherence Gate and DegradationReporter are about runtime behavior; this skill is about pre-merge thoroughness.
• When the audit's pattern is enforceable in CI, pair this skill with a no-* ratchet test (the standing guard from Step 5).

The principle: thoroughness is not how hard you looked once — it is whether a fresh look finds anything new. Audit until the fresh look comes back empty.