| name | audit-harness |
| description | Audit the harness itself for drift, weak spots, overgrown instructions, missing deterministic checks, or language-pack gaps. Invoke automatically when harness-level changes accumulate and need consistency review. |
| allowed-tools | Read, Grep, Glob, Bash, Write |
Audit the harness, not the product code alone.
Inspect
AGENTS.md and CLAUDE.md.claude/rules/.claude/skills/ (including /test skill).claude/hooks/.codex/config.toml, .codex/AGENTS.override.md, .codex/README.md, .codex/hooks/.agents/skills/ (Codex skill mirrors)scripts/run-verify.sh, scripts/run-static-verify.sh, scripts/run-test.sh, scripts/check-skill-sync.shpacks/languages/- CI and report templates
AGENTS.md size budget
Codex caps the merged AGENTS.md prompt at project_doc_max_bytes
(default 32 KiB, see config-reference).
Anything beyond that is silently truncated, which strips the bottom of the
file with no warning. Run during the audit:
size="$(wc -c < AGENTS.md)"
if [ "$size" -gt 24576 ]; then
echo "WARN: AGENTS.md=${size} bytes (>24 KiB)"
fi
if [ "$size" -gt 32768 ]; then
echo "FAIL: AGENTS.md=${size} bytes (>32 KiB Codex cap)"
fi
Promote anything that crosses the warning threshold into .claude/rules/,
.claude/skills/, or .codex/AGENTS.override.md instead of bloating the
shared map. Record the size and remediation in the audit memo.
Questions
- Is always-on context too large?
- Are there rules that should become scripts, tests, or hooks?
- Are there repeated review comments that justify automation?
- Are there missing language packs or pack-specific verifiers?
- Are reports and plans actually helping, or only adding ceremony?
- Is the harness complexity still justified by the current task and model quality?
- Do
/self-review, /verify, and /test have clear non-overlapping responsibilities?
Quality gate alignment
Check whether docs/quality/ still matches reality:
- Does
docs/quality/definition-of-done.md reflect the actual completion workflow? Compare against /work, /self-review, /verify, /test, and /pr skill steps.
- Does
docs/quality/quality-gates.md list the verifiers and CI checks that actually exist in scripts/ and .github/workflows/?
- Are there new verification tools, linters, or test frameworks in use that are not mentioned in the quality gates?
- Are there gates listed that no longer apply or have been removed?
If drift is found, update the quality docs or flag them in the audit memo.
Output
Write a short audit memo with:
- strengths
- pain points
- missing guardrails
- proposed promotions from prose to code
- simplifications worth trying
CLI execution modes
This skill runs under both Claude Code and Codex. The execution mode follows
the conventions in AGENTS.md and .codex/AGENTS.override.md.
| Aspect | Claude Code | Codex |
|---|
| Skill invocation | /skill-name slash command | $skill-name mention or the /skills menu (avoid the /skill-name form — it collides with built-ins) |
| Skill body path | .claude/skills/<name>/SKILL.md | .agents/skills/<name>/SKILL.md |
| Subagent mechanism | Task(subagent_type=...) when a policy delegates | .codex/agents/ custom agents when a policy delegates |
| Structured prompts | AskUserQuestion | Numbered options printed to stdout, awaiting a digit reply |
| Artifacts | docs/reports/, docs/plans/, docs/specs/ (shared) | Same (CLI-agnostic) |
The drift check (./scripts/check-skill-sync.sh) cross-checks both bodies and
invocation metadata — editing only one side will fail CI.
