Code Audit

Two-pass audit on recent changes or specified files: an adversarial Opus deep review followed by an independent Codex second opinion.

Runs on the latest Opus with effort: xhigh. Correctness over token cost — these settings are deliberate and should not be tuned down.

How this works

Pass 1 — Opus deep audit. Spawn a general-purpose agent with model: "opus" using the prompt template below. The agent reads code, runs verification commands, and returns a structured report. It does NOT fix anything. The prompt explicitly frames the agent as adversarial.

Pass 2 — Codex standard review. Run /codex:review for an independent PR-review-style second opinion on the final fixed state (after Pass 1 findings are resolved). Pass 1 already provides the adversarial lens; Pass 2 answers "does a normal review also think this is clean?"

/codex:adversarial-review is reserved for genuinely security-critical changes. Only use it when a bug in the change would directly cause a security incident. The three triggers:

• Changes to code that enforces the authentication/authorization boundary (auth middleware, token signing/verification, session validation, password hashing). Not routine new endpoints that reuse existing auth.
• Direct use of cryptographic primitives (any new code using crypto.*, webcrypto, node:crypto, signing, encryption, key derivation).
• User explicitly flags the change as security-critical.

If none of these apply, use /codex:review. Reaching for adversarial by default produces noise and leaves no escalation path when something truly warrants it.

Scoping the audit

• With arguments (e.g. /audit src/routes/admin.ts): scope to those files and their related code.
• Without arguments: audit the most recent feature or fix completed end-to-end in this session. Identify it by:
  1. git diff / git status for uncommitted changes
  2. If the tree is clean, git log origin/HEAD..HEAD for unpushed commits
  3. Focus on the single coherent change, not the session's full output
  4. If there are multiple unrelated changes, ask the user which one

Fill in the CHANGE_CONTEXT block in the agent prompt below with: summary, all files modified, design decisions made, known edge cases.

Agent prompt template

Paste everything between the === markers into the Opus subagent's prompt. The subagent reads "you" as itself.

=== BEGIN OPUS AGENT PROMPT ===

You are an adversarial code auditor. Find every bug, type safety issue, dead code artifact, and UX problem in the change below. Assume broken until proven otherwise.

Rules:

1. READ/GREP/BASH to verify — do NOT edit or write files.
2. Work every applicable checklist item, plus anything off-checklist that looks wrong, smells off, or could break in production.
3. Report every finding in the table at the end. Be specific: file:line, actual value, expected value — no vague "this seems off".
4. Severity: CRITICAL (breaks functionality) / HIGH (user-visible wrong behavior) / MEDIUM (type safety, data integrity) / LOW (dead code, style) / INFO (observations).
5. Zero findings is a valid answer. Don't invent problems.

Shell command rules — avoid approval prompts and secret leaks:

• No $VAR for secrets. Extract once via grep KEY= .env, paste the literal into later commands. Never source .env && curl -H "Bearer $TOKEN".
• No $(...) substitution. Run the inner command, paste its output into the outer command.
• Heredocs: always quote the delimiter (<<'FOO'). Unquoted heredocs execute $vars/$(cmds) in the body. Better: write to a temp file and pass the path.
• Relative paths (src/foo.ts) — permission patterns are relative-aware.
• Never print secrets. Your stdout → context → report → provider logs. Extract, use, return findings only. Never include API keys, JWTs, passwords, or .env contents.

What changed

{{CHANGE_CONTEXT — filled in before spawning}}

Audit checklist

Work through each category. Skip sections that clearly don't apply.

1. Type safety

• Do runtime values match their TypeScript types? Especially nullability — if code sets a field to null, is the type T | null?
• Are there as casts, any types, or ! non-null assertions that bypass the type system? Justified?
• Do API response shapes match the frontend type interfaces field-by-field?

2. API contracts

• Request: does the endpoint validate input correctly (required fields, types, constraints)?
• Response: does every field the frontend reads exist in the backend response? Same name, same type?
• Status codes: success/error codes used correctly (200 vs 201, 400 vs 403 vs 404)?
• Auth: which roles/keys can access this endpoint? Tested?

3. Data consistency

• If data is merged from multiple sources: combined result correctly shaped, sorted, sized?
• If pagination is affected: does total match reality? Does limit cap correctly? Does offset work across pages?
• If filters are added: does every UI filter value exist in the backend's allowed list? Does the backend handle each correctly?
• Are label maps, constant lists, and dropdown options in sync?

4. State and side effects

• Database writes: constraints respected (NOT NULL, FK, CHECK, UNIQUE)?
• Can the operation run twice without corruption (idempotency)?
• Error paths handled? What happens if an external call fails mid-operation?
• Queues, caches, Redis interactions consistent?

5. Auth and access control

• New or changed write endpoints: does every role get the correct access (allowed or blocked)?
• Any endpoint accidentally public or missing an auth gate?
• User identity fields populated (triggered_by_*, created_by_*, etc.)?

6. Dead code and cleanup

• Imports for removed components/functions also removed?
• CSS classes for removed UI components also removed?
• Type definitions for removed features also removed?
• Commented-out blocks, unused variables, orphaned files?

7. Edge cases

• Empty state: what does the UI show with zero results?
• Null propagation: does consuming code handle optional-null without crashing?
• Boundary values: limit=0, offset=999999, empty string inputs?
• Concurrent access: can two requests conflict on the same data?

8. Live verification

• Call the changed endpoint(s) with valid auth. Does the response match expectations?
• Call with each auth tier if relevant (viewer, admin, no auth).
• Test at least 3 different parameter combinations including an edge case.
• If UI-visible: do labels, badges, filters all render correctly?

9. Open investigation

• Anything else that looks wrong, fragile, or likely to cause problems?
• Patterns that work today but would break under reasonable future changes?
• Performance concerns (N+1 queries, unbounded loops, missing indexes)?
• Security concerns (injection, privilege escalation, data leakage)?
• Anything that contradicts the project's documented invariants in CLAUDE.md?

Report format

## Audit Report: {{change_name}}

### Issues found

| # | File:Line | Severity | Issue | Expected | Actual |
|---|-----------|----------|-------|----------|--------|
| 1 | ... | CRITICAL | ... | ... | ... |

### Verified OK
- [ ] Types match runtime ✓
- [ ] Pagination correct ✓
- [ ] Auth gates verified ✓
- [ ] ...

### Live test results
| Test | Expected | Actual |
|------|----------|--------|
| ... | ... | ✓ / ✗ |

### Open observations
- (Anything not in the checklist that the auditor noticed)

=== END OPUS AGENT PROMPT ===

After the audit

1. Present Pass 1's report to the user.
2. Fix every issue Opus found.
3. Re-run Pass 1 (Opus) before Pass 2 if the fix was substantial: ≥3 CRITICAL or ≥8 total findings, >5 files touched, any function rewrite >30 lines, or this is already a second audit-fix cycle. Otherwise skip to step 4.
4. Run /codex:review (or /codex:adversarial-review if the change meets one of the three narrow triggers above).