polish

/polish - Codex Review + Cleanup

Codex reviews the diff, then critic-audit puts the findings through an adversarial challenge. Fix the survivors, then clean up with /simplify and enhancer-code. Do not stop at a findings report; apply the fixes directly. For an internal-reviewer audit, use /audit.

Input

$ARGUMENTS may contain the target scope. If empty, analyze git diff HEAD (staged and unstaged changes).

Phase 1: Codex review

Run an external Codex review to produce findings. If which codex fails, skip to Phase 3.

Run codex review "Review for logic, architecture, data flow, and code simplicity (flag over-complexity and unnecessary indirection)". In codex 0.141.0 the scope flags (--uncommitted / --base / --commit) are mutually exclusive with the PROMPT argument, so pass no scope flag when sending the simplicity-lens PROMPT. Codex reads git status itself and reviews the staged, unstaged, and untracked changes. Omitting the PROMPT falls back to Codex's default review and drops the simplicity lens, so always pass it.

Phase 2: Adversarial challenge + fix

critic-audit filters Codex false positives, then the surviving findings are fixed directly.

1. Map each finding to the critic-audit input schema (critic-audit § Input) and spawn Task(subagent_type: "critic-audit") once with the full list. Wait for the per-finding verdict
2. Triage each finding by its verdict, per the table below
3. From the survivors, keep Codex P1/P2. Drop Phase 3 territory (cleanup scope) and any fix touching files outside git diff scope
4. Fix the survivors high severity first, then detect the test command and validate. git stash any fix that breaks tests

Verdict	Action
confirmed	Keep as a fix candidate
disputed	Drop
downgraded	Apply the lowered severity
needs_context	Surface to the user with a one-line summary

Phase 3: Code cleanup

Two cleanup mutators run in sequence on the post-Phase-2 diff, then validate. Both apply fixes directly without a critic-audit challenge, since neither hunts bugs.

1. Run Skill("simplify") on the current diff for a cleanup-only pass (reuse, simplification, efficiency, altitude). If it rejects a no-arg invocation, pass the diff scope Phase 1 detected
2. Spawn enhancer-code via Task(subagent_type: "enhancer-code") to remove AI slop and apply simplification rules, then audit tests. Run it after /simplify so its preservation rule (when in doubt, keep) takes priority over /simplify's edits
3. Detect the test command and validate. git stash the cleanup edits on failure

Output

Report what each phase did. Codex findings (Phase 1), critic-audit verdict counts and fixes applied with skip reasons (Phase 2), and /simplify plus enhancer-code edits with file:line and the test validation result (Phase 3).

Error Handling

Error	Action
No changes in diff	Report that there is nothing to polish
codex review fails	Skip to Phase 3 with no findings
critic-audit fails or returns empty	Keep all findings as confirmed, proceed to fix
/simplify unavailable or fails	Skip to enhancer-code