audit-worker-output

Audit Worker Output

Use independent audit workers when a stage requires evidence stronger than the implementing worker's self-check.

When Required

• Security, privacy, credential, migration, release, or destructive-operation risk.
• Stage-unblocking acceptance claims.
• Material conflict between worker output, inbox, recall, or repo evidence.
• Reader-facing claims that need independent source checking.
• Broad refactors or integration work where a narrow test is insufficient.

Audit Brief

Audit Task: <audit_id>
Subject: <worker/report/diff/claim>
Stage Contract: <path/ref>
Evidence To Inspect: <reports, diffs, tests, transcript refs, docs>
Questions: <specific yes/no or findings requested>
Output: Audit Report with pass/fail/concerns, evidence, and blocking status
Authority: audit can block acceptance, but cannot rewrite orchestration state directly

Rules

• Self-audit is useful but cannot unblock a stage that requires independent review.
• Audit workers run in visible worker rooms when their output is an independent artifact.
• Auditors inspect evidence; they do not merge, archive, mutate coordination state, or broaden scope.
• Blocking findings require Orchestrator decision: retry, fix, dispatch another worker, recall Guiding Principal, or abandon.

Audit Report Format

When available, use prompts/ccm/auditor.md and docs/orchestration/_templates/audit-report.md so audits are portable across Claude Code and Codex.

Audit Report: <audit_id>
Verdict: pass | pass-with-concerns | block
Evidence Reviewed: <paths/commands/transcript refs>
Findings: <numbered claims>
Blocking Issues: <required fixes or none>
Recommended Orchestrator Action: <accept | retry | fix | recall | abandon>