audit-code

name	audit-code
description	Review and fix any scope, round after round until a dry round
disable-model-invocation	true

audit-code

Role: the audit loop. Reviews and fixes existing code at any scope, running review rounds until a round finds nothing new, then offers the stack to land.

Prerequisites

~/.agentic-sdk/<project>/project.edn exists with :lanes, :dimensions-active, and :spine. At least one module has landed; there is code to audit.

Procedure

The per-round procedure is run-review-round; this skill is the loop around it. The topology and conflict rules are in skills/shared/references/worktree-model.md.

1. Initialize. Start the audit change as the next commit on the current tip, never on the main bookmark: jj new, then jj describe -m "Refactor: Audit <scope>". Hold the scope (a module, a list of modules, or a theme) in context; the rest of the run cites it. A long audit may keep a minimal, gitignored resume checkpoint under ~/.agentic-sdk/<project>/runs/<slug>/ recording what is done (rounds completed, findings still open); never committed, never the hand-off medium.
2. Shard the scope by module ownership: one shard is one module directory plus its tests. For a theme scope, grep the theme to a file list first, then group by module. Dispatch gather-module-context once per shard and reuse the briefs across all dimensions in the round.
3. Rounds while findings warrant. Dispatch review-round-runner with the audit change id, the shards, and the round number. It runs run-review-round: fans out the active check-* dimensions over each shard, runs the spine triage task to dedupe and order findings, dispatches editor waves by level via apply-findings, and re-verifies. It returns one summary line; its escalations and out-of-scope items ride back in that line and you hold them. A full audit is not bound by the per-phase two-round cap implement-change uses: finding everything is the point. Keep going while a round still surfaces high- or medium-severity findings. A round that finds only low-severity or style findings is the last one: record those and stop rather than spend another full round. Cap the loop at a round budget stated up front (default four) to bound cost; say so if hit. Resume after a crash per the worktree-model resume law.
4. Report. The audit change is already a linear stack on the tip; nothing to restack. Report from context: the rounds table (the one-line summaries), what was fixed by level, the escalations the rounds returned, and the merged changelog set (the editors' CHANGELOG: lines from each round summary). Offer the stack to land; advancing main is the maintainer's call, never autonomous.

Boundaries

Reads only the one-line returns of gather-module-context and each review-round-runner dispatch, never their findings dumps, diffs, or reasoning. The per-round detail lives in the round-runner's context, thrown away when it returns. Holds the scope, the shard briefs, and the round summary lines. An audit fixes what reviews find inside the scope; out-of-scope findings ride back in the round summaries, reported not chased, input to a future audit. The final land waits for the maintainer. Atoms dispatched: gather-module-context (once per shard), review-round-runner (run-review-round), write-commit and write-changelog at the land.

Return

A rounds table (one line per round), the fix counts by level, the escalations, and the stack head ready for the maintainer to land.