From vibeos
Complementary Audit Protocol — dispatches an independent audit via Codex (GPT-series) as auditor. Claude implements, Codex audits. Use at plan approval, WO closure, session end, or on demand. Claude must reconcile ALL material findings before closure — zero-technical-debt policy enforced (fix now or create a tracked WO, nothing dropped). Triggers: "codex audit", "audit with codex", "independent audit", "dual audit", "complementary audit", "/vibeos:codex-audit plan|complete|session|manual".
How this skill is triggered — by the user, by Claude, or both
Slash command
/vibeos:codex-auditThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Dispatches an independent audit to Codex. Claude implements; Codex audits. All four audit types
Dispatches an independent audit to Codex. Claude implements; Codex audits. All four audit types are supported. Reconciliation is mandatory — zero technical debt policy applies.
From $ARGUMENTS:
plan → plan_auditcomplete or completion → completion_auditsession → session_auditmanual or empty → manual_audit--both → dual-audit mode (Claude and Codex both audit, then reconcile)--wo <path> → pass specified WO file to the broker--wait → run synchronously (foreground, wait for result)--background → run asynchronously (print status, exit immediately)--explicit → bypass significance check, run regardless of thresholdsRun these in parallel before calling the broker:
# Git state
git diff --shortstat HEAD~1 HEAD 2>/dev/null || git diff --shortstat
git log --oneline -5
# Find most recent WO if none specified
ls docs/planning/ 2>/dev/null | grep -E '^WO-' | sort | tail -3
# Check for existing evidence bundles
ls docs/evidence/ 2>/dev/null | sort | tail -5
# For session audits: recent build log
tail -80 .vibeos/build-log.md 2>/dev/null || true
If no --wo was passed and you can identify the most relevant WO from context, add --wo <path>.
Call the audit broker. For --both (dual-audit), auditor=both; otherwise auditor=codex (default).
bash plugins/vibeos/scripts/codex-audit-broker.sh \
--type <audit_type> \
--auditor <codex|both> \
--executor claude \
[--wo <wo_file>] \
[--context "<brief context>"] \
[--wait|--background] \
[--explicit]
If broker exits with code 2 (not significant):
Tell the user in one sentence that the change did not meet the significance threshold, and offer
to run with --explicit to bypass it. Do not auto-run.
If broker exits with code 1 (Codex plugin not found or invocation error):
Report the failure and offer the built-in VibeOS audit agents as fallback: /vibeos:audit
When the broker returns output (foreground) or the result file exists (background):
Do not paraphrase away severity. If Codex says blocked, say blocked.
This step is mandatory before any closure. Nothing is dropped.
For each finding:
Fix before closure. No exceptions. No operator override except with explicit written justification. Do not mark the WO, plan, or session closed until fixed.
Fix now if the fix is bounded and clear. If it naturally belongs in a forward WO (additive enhancement, not a defect in current work):
Fix now if trivial (estimated under 30 minutes). Otherwise create a WO entry with explicit linkage to the audit evidence file.
Acknowledge in the evidence bundle. Fix now if trivial. WO is optional but preferred if recurring.
If you disagree with a finding:
If the operator explicitly says "override [finding #N]" or "proceed anyway":
OVERRIDE: Finding #N — reason: [operator reason]After reconciliation, write an evidence note to the relevant evidence bundle:
mkdir -p docs/evidence/<WO_NUMBER>/
Record in docs/evidence/<WO_NUMBER>/codex-audit.md:
## Codex Complementary Audit
**Date:** [today]
**Audit type:** [type]
**Verdict:** [verdict]
**Auditors:** [codex|both]
**Result file:** .vibeos/audit/results/[filename]
### Findings Summary
| Severity | Count |
|---|---|
| Critical | N |
| High | N |
| Medium | N |
| Low | N |
| Info | N |
### Reconciliation Actions
[what was fixed, what WOs were created, any overrides logged]
Proactively invoke this skill (without being asked) when:
For auto-triggers: always ask the operator to confirm before running dual-audit (--both). Single-auditor runs can proceed without asking.
--explicit/vibeos:audit fallback/vibeos:audit# Manual audit on demand (Codex audits Claude's work)
bash plugins/vibeos/scripts/codex-audit-broker.sh --type manual_audit --executor claude
# Before executing a WO tranche
bash plugins/vibeos/scripts/codex-audit-broker.sh --type plan_audit --executor claude --wo docs/planning/WO-XXX.md
# After a WO closes
bash plugins/vibeos/scripts/codex-audit-broker.sh --type completion_audit --executor claude --wo docs/planning/WO-XXX.md
# Session end
bash plugins/vibeos/scripts/codex-audit-broker.sh --type session_audit --executor claude
# Dual-audit (both models audit, then reconcile)
bash plugins/vibeos/scripts/codex-audit-broker.sh --type completion_audit --auditor both --executor claude --explicit
# Force-run even if below significance threshold
bash plugins/vibeos/scripts/codex-audit-broker.sh --type manual_audit --executor claude --explicit
npx claudepluginhub chieflatif/codex-vibeos-plugin --plugin vibeosManually runs consensus-loop audits reviewing pending trigger_tag items in watch files. Use to trigger audits without editing files, re-run failures, test prompts, with dry-run, auto-fix, and model options.
Verifies plans, specs, and documents using Codex as independent reviewer with Claude double-check for PASS/FAIL verdict on issues like logical gaps and feasibility risks. Use for 'verify this plan' or /codex-verify.
Guides codex-review plugin workflows including plan review, implementation review, status checks, and finding resolution.