codex-review（Codex 审查）

Codex Review — Three-Tier Code Quality Defense

Unified orchestration layer: picks audit depth based on trigger phrases. bug-audit is invoked as an independent skill — never modified.

Security & Privacy

• Read-only by default: This skill only reads your project files for analysis. It does NOT modify, delete, or upload your code anywhere.
• Optional external model: L1/L3 can use an external code-review API (OpenAI-compatible) for a second opinion. This is opt-in — if no API key is configured, the skill works fine with agent-only review.
• Credentials via environment variables only: API keys are loaded from CODEX_REVIEW_API_KEY env var. Never hardcoded, never logged, never stored.
• Local-only artifacts: Hotspot files are written to system temp directory and auto-cleaned. No network transmission of analysis results.
• No data exfiltration: Code snippets sent to the external API are limited to the files being reviewed. No telemetry, no analytics, no third-party data sharing beyond the configured review model.

Prerequisites

• External model API (optional, for L1 Round 1 and L3): Any OpenAI-compatible endpoint.

- Set env vars: CODEX_REVIEW_API_BASE (default: https://api.openai.com/v1), CODEX_REVIEW_API_KEY, CODEX_REVIEW_MODEL (default: gpt-4o) - Works without this — falls back to agent-only audit

• bug-audit skill (optional): Required for L2/L3. Without it, L2 uses a built-in fallback.
• curl: For API calls (standard on macOS/Linux)

Trigger Mapping

Level 1: Quick Scan (core of codex-review)

Flow

1. Gather code — local read, git clone <url>, server scp, user-pasted snippet, or PR diff
2. Exclude — node_modules/, .git/, package-lock.json, dist/, *.db, __pycache__/, vendor/
3. Round 1 — send to external model API for automated scan (skipped if no API key)
4. Round 2 — current agent does deep supplementary pass
5. Merge & dedup — output severity-graded report
6. Write hotspot file (for L1→L2 handoff)

External Model API Call

curl -s "${CODEX_REVIEW_API_BASE:-https://api.openai.com/v1}/chat/completions" \
  -H "Authorization: Bearer ${CODEX_REVIEW_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "${CODEX_REVIEW_MODEL:-gpt-4o}",
    "messages": [
      {"role": "system", "content": "<REVIEW_SYSTEM_PROMPT>"},
      {"role": "user", "content": "<code content>"}
    ],
    "temperature": 0.2,
    "max_tokens": 6000
  }'

Fallback: If API call fails or times out (120s), skip Round 1 and complete with agent-only audit.

System Prompt (L1 External Scan)

You are an expert code reviewer. Find ALL bugs and security issues:
1. CRITICAL — Security vulnerabilities (XSS, injection, auth bypass), crash bugs
2. HIGH — Logic errors, race conditions, unhandled exceptions
3. MEDIUM — Missing validation, edge cases, performance issues
4. LOW — Code style, dead code, minor improvements

For each: Severity, File+line, Issue, Fix with code snippet.
Focus on real bugs, not style opinions. Output language: match the user's language.

Agent Round 2 — Universal Checklist

• [ ] Cross-file logic consistency (imports, exports, shared state)
• [ ] Authentication & authorization bypass
• [ ] Race conditions (concurrent requests, DB write conflicts)
• [ ] Unhandled exceptions / missing error boundaries
• [ ] Input validation & sanitization (SQL injection, XSS, path traversal)
• [ ] Memory/resource leaks (unclosed connections, event listener buildup)
• [ ] Sensitive data exposure (keys in code, logs, error messages)
• [ ] Timezone handling (UTC vs local)
• [ ] Dependency vulnerabilities (outdated packages, known CVEs)

Agent Round 2 — Tech-Stack Specific (auto-detect & apply)

Node.js/Express:

• [ ] SQLite pitfalls (DEFAULT doesn't support functions, double-quote = column name)
• [ ] Middleware ordering (auth before route handlers)
• [ ] pm2/cluster mode compatibility

Python/Django/Flask:

• [ ] ORM N+1 queries
• [ ] CSRF protection enabled
• [ ] Debug mode in production

Frontend (React/Vue/vanilla):

• [ ] innerHTML / dangerouslySetInnerHTML without sanitization
• [ ] WebView compatibility (WeChat, in-app browsers)
• [ ] Nginx sub-path / base URL issues

Other stacks: adapt checklist to detected technology.

Code Volume Control

• Single API request: backend core files only (server + routes + db + config)
• Send frontend as a second batch if needed
• Very large projects (>50 files): summarize file tree first, then scan in priority order

Hotspot File (L1→L2 handoff)

After L1, write issue summary to ${TMPDIR:-/tmp}/codex-review-hotspots.json:

{
  "project": "my-project",
  "timestamp": "2026-03-05T22:00:00",
  "hotspots": [
    {"file": "routes/admin.js", "severity": "CRITICAL", "brief": "Admin auth bypass via localhost"},
    {"file": "routes/game.js", "severity": "CRITICAL", "brief": "Score submission no server validation"}
  ]
}

This file is only used internally for L1→L2 handoff. bug-audit is unaware of it.

Level 2: Deep Audit

Flow (bug-audit available)

1. Read bug-audit's SKILL.md and execute its full flow (6 Phases)
2. bug-audit itself is never modified
3. Agent strictly follows bug-audit's specification

Flow (bug-audit NOT available — built-in fallback)

1. Phase 1: Project Dissection — read all source files, build dependency graph
2. Phase 2: Build Check Matrix — generate project-specific checklist from actual code patterns
3. Phase 3: Exhaustive Verification — verify every checklist item against real code
4. Phase 4: Reproduce — for each finding, trace the exact execution path
5. Phase 5: Report — output full severity-graded report
6. Phase 6: Fix Suggestions — provide concrete code patches

Level 1→2 Cascade: Pre-Deploy Check

Flow

1. Execute L1 quick scan
2. Write hotspot file
3. Execute L2 (bug-audit or fallback)
4. After L2, agent does hotspot gap analysis:

- Read hotspot file - Check if L2 report covers each L1 hotspot - Uncovered hotspots → targeted deep analysis, add to report - L1 vs L2 conclusions conflict → flag for manual review

1. Output final merged report

Level 3: Cross-Validation (highest level)

Flow

Step 1: External model independent audit
  → Full code to external API with detailed system prompt
  → Output: Report A

Step 2: Agent independent audit (bug-audit or fallback)
  → Full bug-audit flow (or built-in fallback)
  → Output: Report B

Step 3: Cross-compare
  → Both found       → 🔴 Confirmed high-risk (high confidence)
  → Only external    → 🟡 Agent verifies (possible false positive)
  → Only agent       → 🟡 External verifies (possible deep logic bug)
  → Contradictory    → ⚠️ Deep analysis, provide judgment

Step 4: Adversarial testing
  → Ask external model to bypass discovered fixes
  → Validate fix robustness

Adversarial Test Prompt

You are a security researcher. The following security fixes were applied to a project.
For each fix, analyze:
1. Can the fix be bypassed? How?
2. Does the fix introduce new vulnerabilities?
3. Are there edge cases the fix doesn't cover?
Be adversarial and thorough. Output language: match the user's language.

Report Format (all levels)

# 🔍 Code Audit Report — [Project Name]
## Audit Level: L1 / L2 / L3
## 📊 Overview
- Files scanned: X
- Issues found: X (🔴 Critical X | 🟠 High X | 🟡 Medium X | 🔵 Low X)
- [L3 only] Cross-validation: Both agreed X | External only X | Agent only X | Conflict X

## 🔴 Critical Issues
### 1. [Issue Title]
- **File**: `path/to/file.js:42-55`
- **Found by**: External model / Agent / Both
- **Description**: ...
- **Fix**:
(code snippet)

## ✅ Highlights
- [What's done well]

User Options

Users can customize behavior by saying:

• "only scan backend" / "只扫后端" → skip frontend files
• "ignore LOW" / "忽略低级别" → filter out LOW severity
• "output in English/Chinese" → control report language
• "scan this PR" / "审这个PR" → fetch PR diff instead of full codebase
• "skip external model" / "不用外部模型" → agent-only audit

Notes

1. External API timeout: 120 seconds. On failure, skip that round — agent completes independently
2. Large projects: split into batches (backend → frontend → config)
3. Long reports: split across multiple messages, adapted to current channel
4. L2/L3 bug-audit execution strictly follows its own SKILL.md — no modifications or shortcuts
5. Hotspot file is ephemeral — overwritten each L1 run, not persisted
6. All secrets/keys must come from env vars or user config — never hardcoded in this skill