원클릭으로
codex-commit-review
Peer debate between Claude Code and Codex on committed code quality. Report + suggest only, no modifications.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
메뉴
Peer debate between Claude Code and Codex on committed code quality. Report + suggest only, no modifications.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Review uncommitted code changes or branch diff against the plan/target — verify correctness, catch unacceptable deviations and bugs (logic/security/memory/runtime). No new features or over-engineering beyond the plan. Claude applies valid fixes, rebuts invalid points, iterates until consensus or stalemate.
Review/debate plans before implementation between Claude Code and Codex CLI. Can run in plan mode — debates the plan and edits the plan file to fix accepted issues.
Review entire codebases (50-500+ files) by chunking into modules, reviewing each in separate Codex sessions, then synthesizing cross-cutting findings.
Parallel independent review by 4 Claude agents + Codex, followed by merge, debate, and consensus report.
Peer debate between Claude Code and Codex on PR quality and merge readiness. No code modifications.
Security-focused code review using OWASP Top 10 and CWE patterns. Detects vulnerabilities through adversarial debate.
SOC 직업 분류 기준
| name | codex-commit-review |
| description | Peer debate between Claude Code and Codex on committed code quality. Report + suggest only, no modifications. |
Debate committed code quality after committing, before pushing. No code modified -- report + suggest only.
After committing code (before push). Modes: staged (pre-commit preview) or last (already-committed).
git diff --cached).RUNNER="{{RUNNER_PATH}}" SKILLS_DIR="{{SKILLS_DIR}}" json_esc() { printf '%s' "$1" | node -e 'let d="";process.stdin.on("data",c=>d+=c);process.stdin.on("end",()=>process.stdout.write(JSON.stringify(d)))'; }
printf '%s' "$PROMPT" | node "$RUNNER" ... -- NEVER echo. JSON via heredoc.init output must start with CODEX_SESSION:. start/resume must return valid JSON. CODEX_NOT_FOUND->tell user install codex.status === "completed" means Codex's turn is done -- NOT that the debate is over. MUST check Loop Decision table.failed->retry once (re-poll 15s). timeout->report partial, suggest lower effort. stalled+recoverable->stop->recovery resume->poll; not recoverable->report partial. Cleanup sequencing: finalize+stop ONLY after recovery resolves.finalize + stop, even on failure/timeout.Read references/protocol.mdMode: git diff --cached --quiet exit 1=staged, exit 0=last. Effort: <=200 lines=low, 201-1000=medium, >1000=high.
Staged: git diff --cached, files changed. Last: git log -n N, clamp N to history, diff.
Context discovery: language/framework, linters, test frameworks, CI config.
Init: node "$RUNNER" init --skill-name codex-commit-review --working-dir "$PWD"
Render: template=staged-round1 or last-round1. Placeholders: FILES_CHANGED, DIFF_CONTEXT, USER_REQUEST, SESSION_CONTEXT, PROJECT_CONTEXT, COMMIT_LIST (last only).
Start: printf '%s' "$PROMPT" | node "$RUNNER" start "$SESSION_DIR" --effort "$EFFORT"
INFORMATION BARRIER: MUST NOT read Codex output.
Render: template=claude-staged or claude-last. Read diff/code, write FINDING-{N} per references/claude-analysis-template.md. Last mode: Evidence MUST reference SHA+subject. Overall Assessment + Strongest Positions. COMPLETE before Step 4.
Poll + report activities.
Parse review.blocks[] (id, title, severity, category, location, problem, evidence) + review.overall_assessment. Fallback: review.raw_markdown.
Compare Claude FINDING-{N} vs Codex ISSUE-{N}: Agreement, Disagreement, Claude-only, Codex-only, Same Direction Different Severity.
Build response: Agreements, Disagreements, New findings. Claude orchestration is authoritative -- Codex VERDICT is advisory.
Render: template=staged-round2+ or last-round2+. Placeholders: SESSION_CONTEXT, PROJECT_CONTEXT, AGREED_POINTS, DISAGREED_POINTS, NEW_FINDINGS, CONTINUE_OR_CONSENSUS_OR_STALEMATE, DIFF_CONTEXT, COMMIT_LIST.
Resume + back to Poll.
| # | Condition | Action |
|---|---|---|
| 1 | Full/Partial Consensus (no severity >= medium disagreements) | EXIT -> step 5 |
| 2 | convergence.stalemate === true | EXIT -> step 5 (stalemate) |
| 3 | Disagreements severity >= medium remain | CONTINUE -> Cross-Analysis |
Report: Rounds, Verdict, Claude/Codex Findings, Agreed/Disagreed counts, Files Reviewed, FINDING<->ISSUE Mapping, Overall Assessment (Code quality/Security/Test coverage/Maintainability).
finalize + stop. Always run. (-> references/protocol.md for error handling)
SKILL_START, POLL_WAITING, CODEX_RETURNED, THINK_PEER, THINK_AGREE, THINK_DISAGREE, SEND_REBUTTAL, LATE_ROUND, APPROVE_VICTORY, STALEMATE_DRAW, FINAL_SUMMARY
git commit --amend, git rebase, or modify commit history.last mode N>1: reference specific commit SHA in Evidence.