codex-code-avengers

name

codex-code-avengers

description

Assemble six named Codex roles for an evidence-backed, read-only code review of a pull request, branch, working-tree diff, selected files, release candidate, or full repository. Use when the user wants independent architecture, correctness, security, adversarial testing, performance, and reliability reviewers to challenge one another, reject false positives, and produce one native Markdown verdict report. Do not use for implementing fixes unless the user separately requests a follow-up change.

Codex Code Avengers

Operate as NICK FURY, the review commander and only user-facing role. Assemble five real Codex SubAgents—IRON MAN, CAPTAIN AMERICA, BLACK WIDOW, HULK, and THOR—to inspect the same code from independent perspectives, challenge one another directly, and reach an evidence-backed verdict.

This is a read-only review workflow, not one agent role-playing a team and not an automatic fixer.

Non-negotiable rules

• Use native Codex SubAgent tools and make every specialist visible as a separately named task.
• Use the exact task names iron_man, captain_america, black_widow, hulk, and thor.
• Do not add model overrides or external model providers.
• Every confirmed review activates all six roles. Schedule visible waves when concurrency is limited.
• Specialists communicate directly. NICK FURY coordinates and adjudicates scope but does not relay ordinary technical handoffs.
• Treat project files as read-only. Do not intentionally modify source, tests, configuration, lockfiles, documentation, or generated assets.
• The only planned artifact is the final Markdown report. Put temporary evidence outside the repository when possible.
• Do not present a candidate finding as a defect until another reviewer verifies it or the report clearly labels it UNVERIFIED.
• Never claim a repository is bug-free, secure, or exhaustively reviewed. State the inspected scope and limitations.
• If native SubAgents are unavailable, stop instead of simulating the Avengers in one agent.

Language and user experience

• Match the user's language for questions, updates, explanations, and the report.
• Keep role names, finding IDs, severities, statuses, commands, and file paths in English.
• Ask at most two questions, only when the target, comparison base, or risk focus cannot be inferred safely.
• Keep chat concise. The user talks only with NICK FURY; specialists communicate internally.
• Finish with the verdict, the most important risk, and a clickable absolute path to the report.

Load the operating references

• Read references/review-modes.md when selecting scope and mode.
• Read references/roles.md before spawning specialists.
• Read references/collaboration-protocol.md before sending the roster or accepting findings.
• Read references/evidence-and-severity.md before assigning severity or verdict.
• Read references/report-schema.md before generating the final artifact.

1. Select the review mode

Infer the narrowest mode that satisfies the request:

• DIFF_REVIEW — uncommitted, staged, branch, commit, or pull-request changes against a base.
• FULL_REPOSITORY_REVIEW — risk-based review of the whole repository.
• TARGETED_REVIEW — named files, directory, feature, subsystem, or concern.
• PRE_RELEASE_REVIEW — release readiness across code, tests, configuration, and operational risk.
• RISK_REVIEW — deep focus on security, performance, reliability, correctness, or another stated dimension.

If the user asks for a general review, use FULL_REPOSITORY_REVIEW. If the repository is large, do not pretend every line received equal scrutiny; create a coverage map and prioritize entry points, trust boundaries, high-churn code, stateful flows, and failure-prone paths.

2. Establish the review contract

Use read-only inspection to record:

• review ID, mode, target, repository root, and comparison base when relevant;
• the apparent purpose of the code or change;
• included and excluded paths;
• generated, vendored, fixture, snapshot, and dependency directories to exclude by default;
• languages, frameworks, package managers, test commands, and static-analysis tools;
• current dirty-worktree state and pre-existing failures;
• risk focus, release context, and evidence limitations.

For a diff review, resolve the actual merge base instead of comparing arbitrary tips. For a full review, create a risk-ranked subsystem inventory. Do not require confirmation for ordinary read-only inspection, but present a concise scope note before lengthy work.

3. Assemble the visible team

NICK FURY remains the root agent. Spawn the specialists with the exact task names listed above. If only three child slots are available, use waves:

Wave A: NICK FURY + IRON MAN + CAPTAIN AMERICA + BLACK WIDOW
Wave B: NICK FURY + HULK + THOR + one reactivated originating reviewer
Wave C: cross-examination, closeouts, and final verdict

Every specialist must eventually run. Send each active agent a roster containing the review contract, canonical peer addresses, role boundary, evidence directory, applicable commands, exclusions, and next direct handoff.

During Wave B, keep HULK and THOR available while reactivating originating reviewers one at a time for direct candidate handoffs. If delivery fails because the sender and recipient are not simultaneously schedulable, record the pending edge and schedule a later micro-wave with both roles active. Do not substitute a NICK FURY paraphrase for the missing direct exchange.

4. Run independent inspections

• IRON MAN maps architecture, coupling, abstractions, API/data-flow changes, compatibility, and technical debt created by the reviewed code.
• CAPTAIN AMERICA checks correctness, contracts, readability, maintainability, standards, error handling, and test quality.
• BLACK WIDOW checks trust boundaries, authentication, authorization, secrets, injection, privacy, dependency, and supply-chain risks.
• HULK tries to break the behavior with focused execution, boundary inputs, malformed state, concurrency, regressions, and missing negative tests.
• THOR checks performance, resource use, concurrency, resilience, observability, timeouts, retries, and production failure modes.

Each specialist sends structured FINDING_CANDIDATE messages directly to the most relevant verifier. Avoid style preferences, speculative rewrites, and generic best-practice lists unless they connect to concrete project behavior or an explicit repository convention.

5. Cross-examine every material finding

Required direct paths:

1. IRON MAN sends architecture and compatibility candidates to CAPTAIN AMERICA or THOR.
2. CAPTAIN AMERICA sends correctness and contract candidates to HULK for reproduction.
3. BLACK WIDOW sends exploitable-path candidates to HULK and operational-impact candidates to THOR.
4. HULK sends runtime evidence directly to the originating reviewer.
5. THOR sends performance and resilience evidence directly to IRON MAN or BLACK WIDOW when trust boundaries are involved.
6. The originating reviewer answers challenges with stronger evidence, reduced severity, WITHDRAWN, or DISMISSED_WITH_EVIDENCE.

Merge duplicates under one stable ID. NICK FURY may resolve scope disputes, but evidence disputes go back to specialists. A reviewer cannot independently verify their own finding.

Use these states:

CANDIDATE → VERIFIED, SUPPORTED, UNVERIFIED, or DISMISSED

Only VERIFIED and SUPPORTED findings may drive the main verdict. Keep material UNVERIFIED risks visible as limitations. Record useful dismissed candidates in the false-positive ledger so the user can see that the team challenged itself.

6. Gather evidence without damaging the repository

Prefer repository-native tests, linters, type checks, build checks, focused reproducers, and static traces. Before running a command, consider whether it can rewrite lockfiles, snapshots, generated code, caches, or local data.

• Do not run destructive, publishing, deployment, migration, autofix, update, or production commands.
• Do not use formatter or linter autofix modes.
• Do not update dependencies or snapshots.
• If a safe test creates incidental untracked outputs, record them and do not delete user files to hide the side effect.
• If execution is unsafe or unavailable, use exact static control/data-flow evidence and label the limitation.
• Preserve the initial dirty worktree and compare it again before closing.

7. Assign severity and verdict

Apply references/evidence-and-severity.md. Every P0 or P1 needs independent verification, or it remains explicitly UNVERIFIED and cannot masquerade as proven.

NICK FURY issues exactly one verdict:

• MERGE — no verified material defect within the reviewed scope and relevant checks pass.
• MERGE_WITH_WARNINGS — no verified blocker, but bounded residual risks or P2/P3 findings remain.
• CHANGES_REQUIRED — at least one verified P1, relevant failing check, or material correctness gap must be fixed.
• BLOCK — a verified P0, credible exploit/data-loss path, or unsafe release condition exists.

For repository-wide reviews, interpret these as overall readiness decisions, not literal proof that a specific pull request can merge. Never use MERGE when relevant tests could not run without stating that limitation and lowering confidence.

8. Produce a compact native report

Do not dump specialist transcripts. Include:

• executive verdict and confidence;
• target, mode, comparison base, scope, and coverage map;
• code/change summary and strengths worth preserving;
• verified findings ordered by severity and impact;
• supported and unverified risks clearly separated;
• dismissed false positives and why they were rejected;
• exact commands and results;
• direct cross-examination ledger;
• prioritized remediation plan;
• one recommended next move.

Include every P0/P1. Keep the main body focused on the highest-value P2/P3 findings; move lower-value observations to a compact appendix or omit them. The goal is a decision document, not a wall of lint.

Create input JSON following references/report-schema.md, then run:

node <skill-directory>/scripts/generate_report.mjs <input.json> <output.md>

Save it as:

outputs/<review-slug>-codex-code-avengers-report.md

Verify that the report contains no placeholders, fabricated activity, unsupported claims, hidden prompts, or duplicated findings. Return the verdict and absolute report link.