| name | adversarial-plan-review-codex |
| description | Use when a plan needs hostile review before execution, especially high-risk coding, business deliverables, migrations, no-git changes, weak validation, stale assumptions, rollback gaps, or plans that must be safe for another agent to execute. |
Codex Adversarial Plan Review
Overview
Red-team a plan before execution. The goal is to find blocking flaws while changes are still cheap: stale paths, unsupported assumptions, missing validation, rollback gaps, stakeholder gaps, and contradictions with prior evidence.
Workflow
- Read the plan, success criteria, failure criteria, out-of-scope boundaries, repo map, evidence catalog, assumption ledger, dead ends, probes, and verification plan.
- Select review lenses: software, business, or mixed.
- Check whether every plan step is evidence-backed, ordered safely, and verifiable.
- Classify findings as
BLOCKING, IMPORTANT, or NOTE.
- If blockers exist, revise the source plan or tell the user exactly what decision/evidence is missing.
- Return the review in conversation. Write a review artifact only when the user explicitly requests one, and place it in the active session folder or another user-approved destination.
Native Plan Mode
- Treat the current proposed plan and inspected evidence as the source inputs; do not require a repository planning workfolder.
- Return
PASS only when there are no BLOCKING findings. Use FAIL when inspected evidence proves a blocking flaw. Use PARTIAL when required evidence cannot be inspected well enough to determine whether the plan is safe.
- Do not add a second approval gate. Report the verdict and required corrections so the active Plan Mode can incorporate them.
- Do not write files while the current collaboration mode prohibits mutations.
References
Read as needed:
references/software-review.md for code, tests, migrations, dependencies, and rollback.
references/business-review.md for stakeholders, source data, approval, privacy, and deliverable risks.
references/mixed-review.md for combined business and coding work.
PASS Criteria
Return PASS only when:
- critical paths and source materials exist;
- assumptions are either verified or explicitly accepted risks;
- validation proves the stated success criteria;
- rollback or recovery is defined;
- sequencing does not depend on impossible or circular steps;
- no blocked stakeholder, data, security, or permission issue remains.
Output Format
Use this structure:
# Adversarial Plan Review
## Verdict
PASS | FAIL | PARTIAL
## Findings
- **Severity**: BLOCKING | IMPORTANT | NOTE
- **Issue**: ...
- **Evidence**: ...
- **Required fix**: ...
## Re-review
[What changed, or why no re-review was needed]