full-review（全面审查）

Full Review: Quad Audit Orchestrator

Run four independent audits in parallel and present each report separately. One command that catches bugs, slop, security issues, and stale docs across the entire codebase without invoking each skill manually.

The four audits:

1. Code Review (code-review skill) - bugs, logic errors, edge cases, race conditions, resource leaks, convention violations. Uses confidence-based filtering (>= 80%), adversarial self-check, and evidence-based verification.
2. Slop Check (anti-slop skill) - machine-generated patterns, over-abstraction, verbose code, stale idioms
3. Security Audit (security-audit skill) - vulnerabilities, secrets, dependency risks, OWASP mapping
4. Docs Sweep (update-docs skill) - stale docs, bloated instruction files, missing gotchas, broken links, companion-file drift

Each audit runs in its own parallel agent/subprocess with a fresh context window, so they don't compete for tokens or bias each other's findings.

When to use

• Running a repo-wide quality gate before merge, release, or handoff
• Auditing an unfamiliar codebase across correctness, security, slop, and docs in one pass
• Getting a broad review when the user explicitly wants multiple audit lenses at once

When NOT to use

• A targeted correctness review on specific files - use code-review
• Style/slop cleanup without the other audit passes - use anti-slop
• A dedicated security review only - use security-audit
• A documentation-only maintenance sweep - use update-docs
• A comprehensive audit across all applicable skills (up to 21) - use deep-audit
• Auditing the skill collection for consistency or quality - use skill-creator

AI Self-Check

Run this checklist after all agents return but before presenting the combined report to the user. Do not present results until every item passes.

Verify:

• All 4 agents dispatched as general-purpose type (NOT feature-dev:*, code-simplifier:*, or other restricted types)
• Each agent invoked its assigned custom skill (code-review, anti-slop, security-audit, update-docs) via the Skill tool
• Each report presented under its own header, unedited
• No cross-report merging or editorializing (findings from different audits stay separate)
• SECURITY-AUDIT.md gitignore reminder included
• Failed agents noted with reason (don't silently drop a missing audit)
• Preflight context block was passed to all agents
• When user specified a scope, the Scope: line in every agent's context block reflects that scope (not "full codebase review")
• Scope held in output: each agent's findings reference only files/modules within the requested scope. If any agent's output references out-of-scope paths, flag it in that report's header (see Step 3 scope verification)

Workflow

Step 0: Preflight

Gather context before dispatching agents. Run these in parallel (guard each with ; true so one failure doesn't cancel siblings):

1. Repo state: git rev-parse --show-toplevel; true and git rev-parse --short HEAD; true
2. Branch: git branch --show-current; true
3. Language detection: check for manifest files (package.json, requirements.txt, go.mod, Cargo.toml, pyproject.toml, composer.json, Gemfile, *.tf, helmfile.yaml)
4. Repo size estimate: git ls-files | wc -l; true

If not a git repo (step 1 fails): stop and tell the user. The audits rely on git context (history, blame, diff). Running without it produces low-quality results.

Record preflight values - each subagent prompt uses them. Substitute {placeholders} in the agent prompts below with the actual values from preflight (e.g., replace {repo_root} with the output of git rev-parse --show-toplevel). Default {scope} to "full codebase review - scan everything"; override in Step 1 if the user specifies a narrower target.

Step 1: Determine Scope

Default is full codebase since the user is running this as a quality gate. Adapt if context suggests otherwise:

• Uncommitted changes present -> mention this, but still audit the full repo.
• Detached HEAD / bare repo -> warn the user, proceed with what's available.
• User specified a narrower scope (specific files, directory, module) -> pass that scope constraint to all four agents. Each agent only audits within the specified scope. This is the key to scoped reviews: narrowing the target, not the audit dimensions. Set {scope} in the context block to the user's scope (e.g., "src/auth/ directory only") instead of the default "full codebase review - scan everything".

Step 2: Dispatch Four Parallel Agents

Spawn all four agents concurrently. Each agent invokes one of the four custom skills and runs a full codebase audit.

Agent type selection (critical): Each agent MUST be dispatched as a general-purpose agent (or equivalent full-access agent type). Do NOT use specialized agent types like feature-dev:code-reviewer, feature-dev:code-explorer, code-simplifier:*, or any other restricted-toolset agent - these lack access to the Skill tool and cannot invoke custom skills. The agent type name should reflect its capabilities (full tool access), not the audit it performs.

Skill invocation: Each general-purpose agent MUST invoke the named custom skill via the Skill tool (or equivalent skill-loading mechanism) as its first action. Custom skills from the user's installed collection take priority over built-in reviewers or platform-provided audit modes. Specifically:

• Agent 1 invokes code-review via Skill tool, not a built-in code-review mode
• Agent 2 invokes anti-slop via Skill tool, not a built-in code simplifier
• Agent 3 invokes security-audit via Skill tool, not a built-in security scanner
• Agent 4 invokes update-docs via Skill tool, not a built-in documentation reviewer

Fallback: If a custom skill is not available (skill lookup/load returns "not found" or similar), THEN fall back to the best available alternative (manual review following the skill's principles) and note which skill was unavailable in the output header.

If parallel execution is unavailable (restricted sandbox, no subagent support): run sequentially in this order: Security Audit, Code Review, Slop Check, Docs Sweep. Security first because those findings are most time-sensitive. If any agent exceeds 5 minutes wall-clock, note the timeout in the output header and continue with the remaining agents.

If agent dispatch is unavailable (non-Claude harness, no subagent API): run each audit sequentially in separate CLI sessions, invoking each skill manually in its own conversation.

Pass this context block to every agent, substituting the {placeholders} from preflight:

Context:
- Repo: {repo_root}
- Commit: {short_sha}
- Branch: {branch}
- Languages: {detected_languages}
- File count: {file_count}
- Scope: {scope}

Each agent receives the context block above plus a task prompt. Use these templates:

Agent 1: Code Review

{context_block}

Invoke the `code-review` skill via the Skill tool, then run a full code review on the codebase.
Scope: {scope}. Return the complete report.

Agent 2: Slop Check

{context_block}

Invoke the `anti-slop` skill via the Skill tool, then audit the codebase for machine-generated
patterns, over-abstraction, and code quality issues. Scope: {scope}. Return the complete report.

Agent 3: Security Audit

{context_block}

Invoke the `security-audit` skill via the Skill tool, then run a security audit on the codebase.
Scope: {scope}. Return the complete report including SECURITY-AUDIT.md content.

Agent 4: Docs Sweep

{context_block}

Invoke the `update-docs` skill via the Skill tool as a read-only audit. Scope: {scope}.
Focus on: stale docs, instruction-file bloat (40,000 char limit), companion-file drift, broken
links, orphaned gotchas, missing docs on recent changes. Do NOT make changes or commit anything.
Return the complete report.

Step 3: Present Results

After all four agents return, present each report under its own header. Do not merge, summarize, or editorialize across reports - each stands alone. The user reads the skill's native output, not a reinterpretation.

Scoped reviews: when the user specified a narrower scope, each report focuses on that scope. Use this routing table to emphasize domain-relevant checks:

For scopes not in the table, apply each skill's standard checklist narrowed to the specified files/module. Do not skip an audit just because the scope seems domain-specific - every skill may surface relevant findings on arbitrary code.

Scope verification before presenting: when a scope was specified, confirm each agent's output before including it in the report. If an agent's findings reference files or modules outside the requested scope, that agent ignored the scope constraint - note the discrepancy in its report header and, if possible, filter out-of-scope findings. If an agent returned zero findings, confirm it actually ran against the scoped target (not an empty or wrong path) before reporting "no issues found."

User requests synthesis: if the user asks for a combined summary after seeing the reports, prioritize: security fixes > correctness bugs > slop cleanup > doc updates. Keep synthesis brief - the individual reports are the source of truth.

After presenting results, remind the user: "Check that SECURITY-AUDIT.md is in .gitignore - it contains vulnerability details that shouldn't be committed."

Use this structure:

# Full Review: {repo_name} @ {short_sha}

Languages: {detected_languages} | Files: {file_count} | Branch: {branch}
Scope: {scope}

---

## 1. Code Review

{agent 1 output verbatim}

---

## 2. Slop Check

{agent 2 output verbatim}

---

## 3. Security Audit

{agent 3 output verbatim}

---

## 4. Docs Sweep

{agent 4 output verbatim}

---

Step 4: Handle Failures

If an agent fails or times out:

• Note which audit failed and why (timeout, skill not found, tool permission denied)
• Present whatever completed successfully
• Do not re-run failed agents unless the user asks

If a skill is not available, perform a manual review in the same general-purpose agent. Note the substitution in the output header so the user knows a fallback was used. Partial results are still useful.

Related Skills

• code-review - one of the four parallel audits. Finds bugs, logic errors, correctness issues.
• anti-slop - one of the four parallel audits. Finds quality/style issues and AI-generated patterns.
• security-audit - one of the four parallel audits. Finds vulnerabilities, secrets, dependency risks.
• update-docs - one of the four parallel audits. Finds stale docs, bloated instruction files, and missing gotchas.
• skill-creator - audits the skill collection itself. Full-review audits application code.

Rules

• General-purpose agents only. Every subagent MUST be a general-purpose (full-access) agent type. Never use feature-dev:*, code-simplifier:*, or other restricted agent types - they cannot invoke custom skills. The agent type controls tool access, not the audit topic.
• Custom skills first. Each agent invokes its assigned custom skill (code-review, anti-slop, security-audit, update-docs) via the Skill tool as its first action. Fall back to manual review only if the skill is not installed.
• Parallel dispatch is strongly preferred. Run all four agents concurrently when the environment supports it. If parallel execution is unavailable, run sequentially (security first - see Step 2).
• Don't editorialize. Present each report as the skill produced it. No unsolicited synthesis across reports.
• Respect each skill's output format. The anti-slop skill has its own format. The security audit writes SECURITY-AUDIT.md. The code reviewer and docs sweep have their formats. Don't normalize them into a single style.
• Don't duplicate work. If a finding appears in multiple reports (e.g., dead code in both slop check and code review), that's fine - independent auditors catching the same thing is signal, not noise.
• Preflight is fast. The parallel git commands in Step 0 should take under 2 seconds. Don't skip them - the agent prompts are much better with context.
• Large repos. If file count exceeds 1000, mention to the user that this will take a while. Don't reduce scope unless asked.
• SECURITY-AUDIT.md gitignore. The security audit writes a report file containing vulnerability details to the repo root. After presenting results, remind the user to check that SECURITY-AUDIT.md is in .gitignore - the sub-skill warns too, but it's easy to miss buried in output.
• Docs sweep is read-only. The update-docs agent must not make changes or commit anything during a full review. It reports what needs updating; the user decides when to act on it.