| name | clodex-adversarial-review |
| description | Runs an adversarial review cycle against a PR (Claude implements/fixes, Codex or a subagent stand-in reviews), up to 5 iterations. Read-only on the real PR — drafts fixes, never pushes them. |
clodex-adversarial-review procedure
The prompt carries intent; this skill carries the procedure — keep it a checklist,
not an essay. Adapted from Forward Future Loop Library #9 (Lukas Kucinski) to this
library's L1-first, report-only shape.
- Read — the PR diff and
clodex-adversarial-review-state.md's
iteration history for this PR.
- Review — run an adversarial review against the diff: the real
codex CLI if available, otherwise invoke the loop-verifier subagent
as a genuinely separate reviewer (never the same pass that wrote the
code).
- Fix findings above threshold — for each finding at or above the
severity threshold (medium by default): in a throwaway worktree, draft a
fix.
- Re-review — repeat steps 2–3 until the reviewer approves, only
accepted findings remain, no progress for 2 consecutive iterations, or 5
iterations are reached.
- Report honestly — write
clodex-review-report.md: the PR, checks,
verdict, and remaining findings. Never describe an errored or
exhausted run as approved — an iteration-cap exit with open findings
is exhausted, say so plainly.
- Record — update
clodex-adversarial-review-state.md's iteration
count and findings history (resumable), append one line to
loop-run-log.md, commit both. Discard the throwaway worktree.
- Take NO other action — never push a commit to the real PR branch.
This loop is L1 report-only until a human promotes it.