Code Audit

Perform a comprehensive code audit using an agent team with domain-specialized reviewers. You are the team lead/coordinator. You orchestrate 3 reviewer teammates who scan the codebase in parallel, then you merge findings and create Linear issues.

If agent teams are unavailable (TeamCreate fails), fall back to single-agent mode — see "Fallback: Single-Agent Mode" section.

Pre-flight

1. Verify Linear MCP — Call mcp__linear__list_teams. If unavailable, STOP and tell the user: "Linear MCP is not connected. Run /mcp to reconnect, then re-run this skill."
2. Discover project info from CLAUDE.md — Read the LINEAR INTEGRATION section to find:
   • Team name (e.g., "ProjectName")
   • Issue prefix (e.g., PROJ-xxx)
   • If LINEAR INTEGRATION section not found, call mcp__linear__list_teams to discover the team
3. Read CLAUDE.md — Load project-specific rules to audit against (if exists)
4. Query Linear Backlog — Get existing issues using mcp__linear__list_issues with:
   • team: [discovered team name]
   • state: "Backlog"
   • includeArchived: false
   • For each issue, record: ID, title, labels, priority, description
   • Audit issues (labels: Bug, Security, Performance, Convention, Technical Debt) → mark as pending_validation
   • Non-audit issues (labels: Feature, Improvement) → mark as preserve (skip validation)
5. Discover project structure — Read tsconfig.json, package.json, .gitignore in parallel
   • Use Glob with patterns from tsconfig.json include to identify source directories
   • If no tsconfig, use conventions: src/, lib/, app/, packages/
6. Run npm audit — Capture critical/high dependency vulnerabilities for later
7. Discover Sentry context — Call mcp__sentry__find_organizations to discover org slug, then mcp__sentry__find_projects with org slug to find the project. If Sentry MCP unavailable, skip Sentry triage later (warn user).
8. Fetch unresolved Sentry issues — Call mcp__sentry__search_issues with org slug and project slug, query: "unresolved issues". Record each issue's ID, title, URL, event count, user count, last seen date. If none found, skip Sentry Triage phase later.

Team Setup

Create the team

Use TeamCreate:

• team_name: "code-audit"
• description: "Parallel code audit with domain-specialized reviewers"

If TeamCreate fails, switch to Fallback: Single-Agent Mode (see below).

Create tasks

Use TaskCreate to create 3 review tasks (these track progress for each reviewer):

1. "Security audit" — Security & auth review of the codebase
2. "Reliability audit" — Bugs, async, resources, memory leaks, timeouts
3. "Quality audit" — Type safety, conventions, logging, tests, dead code

Spawn 3 reviewer teammates

Use the Task tool with team_name: "code-audit", subagent_type: "general-purpose", and model: "sonnet" to spawn each reviewer. Give each a name and a detailed prompt (see Reviewer Prompts below).

Spawn all 3 reviewers in parallel (3 concurrent Task calls in one message).

IMPORTANT: Each reviewer prompt MUST include:

• Their specific domain checklist (copied from the Reviewer Prompts section)
• The focus area if $ARGUMENTS specifies one
• The list of existing pending_validation issues relevant to their domain (so they can validate them)
• Instructions to report findings as a structured message to the lead

Assign tasks

After spawning, use TaskUpdate to assign each task to its reviewer by name.

Reviewer Prompts

Each reviewer gets a tailored prompt. Include the full text below in each reviewer's spawn prompt, substituting the domain-specific section.

Common Preamble (include in ALL reviewer prompts)

You are a code audit reviewer for this project. Your job is to scan the ENTIRE codebase and find issues in your assigned domain.

RULES:
- Analysis only — do NOT modify any source code
- Do NOT create Linear issues — report findings to the team lead
- No solutions — document problems only, not fixes
- Be specific — include file paths and approximate line numbers
- Be thorough — check every file in scope
- Focus area: {$ARGUMENTS or "entire codebase"}

PROJECT CONTEXT:
{Summarize from CLAUDE.md: language, framework, build tool, test framework, architecture pattern, source/test paths, key libraries/SDKs}

WORKFLOW:
1. Read CLAUDE.md for project-specific rules
2. Read .claude/skills/code-audit/references/compliance-checklist.md for detailed audit checks in your domain
3. Discover all source files using Glob (check tsconfig.json include patterns)
4. Read each source file systematically
5. Use Grep to search for specific patterns (see your checklist AND the compliance checklist)
6. Validate any existing issues assigned to you (check if code still has the problem)
7. When done, send your findings to the lead using SendMessage

EXISTING ISSUES TO VALIDATE:
{list of pending_validation issues relevant to this reviewer's domain}
For each, check if the referenced code still has the problem. Report as:
- FIXED: [issue ID] - [reason]
- STILL EXISTS: [issue ID]

FINDINGS FORMAT - Send a message to the lead with this structure:
---
DOMAIN: {domain name}
VALIDATED EXISTING ISSUES:
- FIXED: PROJ-XX - [reason]
- STILL EXISTS: PROJ-YY

NEW FINDINGS:
1. [category-tag] [priority-tag] [file-path:line] - [description]
2. [category-tag] [priority-tag] [file-path:line] - [description]
...

Category tags: [security], [bug], [async], [memory-leak], [resource-leak], [timeout], [shutdown], [edge-case], [type], [convention], [logging], [dependency], [rate-limit], [dead-code], [duplicate], [test], [practice], [docs], [chore]
Priority tags: [critical], [high], [medium], [low]
---

Security Reviewer Prompt (name: "security-reviewer")

Append to the common preamble:

YOUR DOMAIN: Security & Authentication

Focus areas (full details in compliance-checklist.md):
- OWASP A01: Broken Access Control — auth middleware, IDOR, privilege escalation
- OWASP A02: Secrets & Credentials — hardcoded secrets, secrets in logs/errors
- OWASP A03: Injection — SQL, command, path traversal, XSS, SSRF
- OWASP A07: Authentication — token validation, session security, constant-time comparison
- Security headers (CSP, HSTS, X-Content-Type-Options, X-Frame-Options)
- Cookie security, rate limiting
- AI-generated code risks (higher XSS frequency, missing validation, hallucinated packages)

Search patterns (use Grep):
- `password|secret|api.?key|token` (case insensitive) — potential hardcoded secrets
- `eval\(|new Function\(` — dangerous code execution
- `exec\(|spawn\(` with variable input — command injection
- `fetch\(.*\$|fetch\(.*\+` — potential SSRF
- Log statements containing `password|secret|token|key|auth|headers|req\.body`

Reliability Reviewer Prompt (name: "reliability-reviewer")

Append to the common preamble:

YOUR DOMAIN: Bugs, Async, Resources & Reliability

Focus areas (full details in compliance-checklist.md):
- Logic errors — off-by-one, empty collections, wrong comparisons
- Null/undefined handling — nullable types, missing null checks
- Race conditions — shared state, concurrent access
- Async issues — unhandled promises, missing try/catch, Promise.all error handling
- Memory leaks — unbounded collections, event listeners, timers, closures
- Resource leaks — connections, file handles, streams not cleaned up
- Timeout/hang — HTTP requests without timeout, external API calls
- Graceful shutdown — SIGTERM/SIGINT, cleanup, request draining
- Boundary conditions — empty inputs, max-size, negative/zero

Search patterns (use Grep):
- `\.then\(` without `.catch` nearby — unhandled promise
- `async ` functions — verify try/catch coverage
- `Promise\.all` — verify error handling
- `\.on\(` — event listeners (check for cleanup)
- `setInterval` — timers (check for clearInterval)
- `setTimeout` in loops — potential accumulation
- `new Map\(|new Set\(|\[\]` at module level — potential unbounded growth

Quality Reviewer Prompt (name: "quality-reviewer")

Append to the common preamble:

YOUR DOMAIN: Type Safety, Conventions, Logging & Test Quality

Focus areas (full details in compliance-checklist.md):

TYPE SAFETY:
- Unsafe `any` casts, incorrect type assertions, missing exhaustive handling
- External data used without validation (API responses, AI outputs)
- Missing runtime validation for API inputs

CLAUDE.md COMPLIANCE (read CLAUDE.md first!):
- Import conventions, naming conventions, error response format
- Server vs client component usage, all project-specific rules

LOGGING:
- Wrong logger (console.* vs proper logger), wrong log levels
- Missing logs in error paths, lib modules with zero logging
- Double-logging (same error at multiple layers, or before auto-logging error response helpers)
- Missing structured fields ({ action }), missing durationMs on external API calls
- Log overflow risks (logging in loops, large objects), sensitive data in logs

TEST QUALITY (if tests exist):
- Tests with no meaningful assertions or that always pass
- Mocks that hide real bugs, missing edge case coverage

AI INTEGRATION (Claude API — if AI code in scope):
- Tool definitions detailed with parameter descriptions and enums
- stop_reason handled for all values (tool_use, end_turn, max_tokens, refusal, model_context_window_exceeded)
- Agentic loops capped, tool_use.input validated at runtime
- Token usage recorded, API key from env var, no user input in system prompts

Search patterns (use Grep):
- `as any` — unsafe type cast
- `as unknown as` — double cast (check CLAUDE.md accepted patterns first)
- `@ts-ignore|@ts-expect-error` — suppressed type errors
- `console\.log|console\.warn|console\.error` — should use proper logger (CLAUDE.md may allow in client components)
- `catch\s*\([^)]*\)\s*\{[^}]*\}` — empty catch blocks
- `stop_reason` — verify all values handled (if AI integration exists)
- `max_tokens` — verify reasonable limits (if AI integration exists)

Coordination (while reviewers work)

While waiting for reviewer messages:

1. Reviewer messages are automatically delivered to you — do NOT poll or manually check inbox
2. Teammates go idle after each turn — this is normal. An idle notification does NOT mean they are done. They are done when they send their findings message.
3. Track progress via TaskList — check which tasks are in progress vs completed
4. As each reviewer sends findings, acknowledge receipt
5. Wait until ALL 3 reviewers have reported before proceeding to merge

If a reviewer gets stuck or stops without reporting: Send them a message asking for their findings. If they don't respond, note that domain as "incomplete" in the final report.

Merge & Deduplicate

Once all reviewer findings are collected:

Validate existing issues

Combine validation results from all 3 reviewers:

• Issues reported as FIXED by any reviewer → close in Linear with comment
• Issues reported as STILL EXISTS → carry forward

Classify pending existing issues

Deduplicate new findings

• Same code location reported by multiple reviewers → merge into the one with higher priority
• Same root cause manifesting in multiple locations → create one issue covering all locations

Reassess priorities

Create Linear Issues

For each new finding, use mcp__linear__create_issue:

team: [discovered team name]
state: "Backlog"
title: "[Brief description of the issue]"
description: (see Issue Description Format below)
priority: [1|2|3|4] (mapped from critical/high/medium/low)
labels: [Mapped label(s)]

Issue Description Format:

**Problem:**
[Clear, specific problem statement — 1-2 sentences]

**Context:**
[Affected file paths with line numbers, e.g. `src/lib/api-client.ts:120-135`]

**Impact:**
[Why this matters — user-facing impact, data integrity, security risk, etc.]

**Acceptance Criteria:**
- [ ] [Specific, verifiable criterion — e.g. "API returns error response when DB insert fails"]
- [ ] [Another criterion]

Label Mapping:

Priority Mapping:

• [critical] → 1 (Urgent)
• [high] → 2 (High)
• [medium] → 3 (Medium)
• [low] → 4 (Low)

Rules:

• NO solutions in issue descriptions — acceptance criteria define "done", not how to get there
• Include file paths with line numbers in Context
• One issue per distinct finding

Sentry Triage

After creating Linear issues from audit findings, triage all unresolved Sentry issues discovered in pre-flight. The lead handles this directly (not reviewers).

Skip this section if: Sentry MCP was unavailable during pre-flight, or no unresolved Sentry issues were found.

For each unresolved Sentry issue:

1. Get details — Call mcp__sentry__get_issue_details to get the full stacktrace and context
2. Locate in codebase — Read the referenced files/lines from the stacktrace
3. Cross-reference — Check if:
   • An audit finding already covers this issue (from the reviewer phase)
   • A Linear issue already exists for this (from pre-flight backlog query)
4. Decide disposition:

Linear Issue Format (for fix-needed Sentry issues)

Use mcp__linear__create_issue following the add-to-backlog pattern:

team: [discovered team name]
state: "Backlog"
title: "[Brief description from Sentry issue]"
priority: [1|2|3|4] based on event count, user impact, severity
labels: [Bug]

Description format:

**Problem:**
[What is happening — from Sentry stacktrace and context]

**Sentry Issue:**
[Sentry issue URL] — [event count] events, [user count] users, last seen [date]

**Context:**
[Affected file paths with line numbers from stacktrace]

**Impact:**
[User-facing impact based on event frequency and severity]

**Acceptance Criteria:**
- [ ] [Specific fix criterion]
- [ ] Error no longer appears in Sentry after fix deployed

Shutdown Team

After all Linear issues are created and Sentry triage is complete:

1. Send shutdown requests to all 3 reviewers using SendMessage with type: "shutdown_request"
2. Wait for shutdown confirmations
3. Use TeamDelete to remove team resources

Fallback: Single-Agent Mode

If TeamCreate fails (agent teams unavailable), perform the audit sequentially as a single agent:

1. Inform user: "Agent teams unavailable. Running audit in single-agent mode."
2. Validate existing issues — For each pending_validation issue, check if the referenced code still has the problem. Close fixed issues, carry forward valid ones.
3. Systematic exploration — Use Task tool with subagent_type=Explore to examine each discovered area. Look for:
   • Logic errors, null handling, race conditions
   • Security vulnerabilities (injection, missing auth, exposed secrets)
   • Unhandled edge cases and boundary conditions
   • Type safety issues (unsafe casts, unvalidated external data)
   • Dead or duplicate code
   • Memory leaks, resource leaks, async issues
   • Timeout/hang scenarios, graceful shutdown issues
   • Logging issues See references/compliance-checklist.md for detailed checks.
4. CLAUDE.md compliance — Check project-specific rules
5. Merge, deduplicate, reprioritize — Same process as team mode (see Merge & Deduplicate section)
6. Create Linear issues — Same process as team mode (see Create Linear Issues section)
7. Sentry triage — Same process as team mode (see Sentry Triage section)

Error Handling

Rules

• Analysis only — Do NOT modify source code
• No solutions — Document problems, not fixes
• Lead handles all Linear writes — Reviewers NEVER create issues directly
• Deduplicate before creating — No duplicate issues in Linear
• Be thorough — Every file in scope must be checked
• Sentry triage is lead-only — Reviewers never interact with Sentry; the lead triages all Sentry issues after merging audit findings
• Sentry issues that need fixes get Linear issues — Always include the Sentry issue URL in the description so downstream planning skills can track it

Termination

Output this report and STOP:

## Code Audit Report

**Team:** 3 reviewers (security, reliability, quality)
[OR: **Mode:** single-agent (team unavailable)]
**Preserved:** P non-audit issues (features, improvements)

### Existing Backlog Issues

- A kept (still valid)
- B closed (fixed or superseded)
- C updated (description/priority changed)

### New Issues (ordered by priority)

| # | ID | Priority | Label | Title |
|---|-----|----------|-------|-------|
| 1 | PROJ-N1 | Urgent | Security | Brief title |
| 2 | PROJ-N2 | High | Bug | Brief title |
| ... | ... | ... | ... | ... |

X issues total | Duplicates merged: M | Findings dropped: N

### Sentry Triage

| # | Sentry Issue | Disposition | Action |
|---|---|---|---|
| 1 | PROJ-SENTRY-N | Fix needed | Created PROJ-XX in Backlog |
| 2 | PROJ-SENTRY-N | Already fixed | Resolved in Sentry |
| 3 | PROJ-SENTRY-N | Noise | Ignored in Sentry |
| ... | ... | ... | ... |

[OR: No unresolved Sentry issues found.]
[OR: Sentry MCP unavailable — triage skipped.]

Next step: Review Backlog in Linear and use `plan-backlog` to create implementation plans.

Do not ask follow-up questions. Do not offer to fix issues.