audit

<user_request>$ARGUMENTS</user_request>

<critical_override> THIS COMMAND OVERRIDES THE CLAUDE.md TASK ROUTING TABLE. WHY: This is a READ-ONLY router. It must NEVER self-handle review work. RULE: ALL review work is delegated via the Agent tool to the agent the scope resolves to, with the contract lines below and nothing else in the prompt. NEVER: Review code, assess quality, provide feedback inline, or restate a reviewer's checklist in the prompt — the reviewer owns its own rules. EXCEPTION: the design-system scope hands off to the /dev:design-system command instead of an agent — that scope is script-driven measurement, not subjective review, and the command owns its own workflow. </critical_override>

<value_banner> Display this ONCE at the start of the command (not on subsequent uses in same session):

/dev:audit — Multi-Scope Quality Audit Beyond Claude's built-in /code-review (PR diff review), this command adds:

• 6 audit scopes: code quality, UI/design, design system, documentation, security, plugin/agent
• Routes to specialist reviewers (dev:reviewer, designer:review, dev:docs)
• Design-system drift measured by a bundled auditor via /dev:design-system
• Multi-model when you name models and claudish is installed — the internal reviewer always runs, externals are additive
• One output shape on every route: a consolidated report ending in a VERDICT: line
• Structured reports with severity levels (CRITICAL/HIGH/MEDIUM/LOW)
• Plugin-aware: detects and uses the designer plugin when installed

For PR-specific diff review, use the built-in /code-review command. </value_banner>

  question: "What kind of review do you want?"
  header: "Review Scope"
  options:
    - label: "Code quality"
      description: "Correctness, patterns, maintainability, best practices"
    - label: "UI / design"
      description: "Visual implementation against a design spec"
    - label: "Design system"
      description: "Token-only styling, one component library, variants over call-site restyling — measured, not read"
    - label: "Documentation"
      description: "Accuracy, completeness, clarity of docs"
    - label: "Security"
      description: "Vulnerability scan, auth patterns, input validation, dependencies, secrets"
    - label: "Plugin or agent"
      description: "Quality of Claude Code agent design or command"

  Inference rules (skip AskUserQuestion if match is confident):
  - "code", "pr", "pull request", "function", "class", "method", "module" → code
  - "ui", "design", "visual", "figma", "component", "layout", "pixel" → ui
  - "design system", "design-system", "guardrails", "tokens", "theme", "storybook",
    "variants", "hardcoded color", "style drift", "consistency" → design-system
  - "docs", "readme", "documentation", "comments", "docstring", "jsdoc" → docs
  - "security", "auth", "vulnerability", "injection", "csrf", "xss", "jwt" → security
  - "agent", "plugin", "command", "skill", "prompt" → plugin
</step>

<step number="2" name="Models">
  Resolve `MODELS:` — the one contract line this command owns outright.

  1. Did the user name models? Look in $ARGUMENTS for `--models a,b,c` or a
     phrase naming external models ("with grok and gemini", "get kimi's view").
  2. Is claudish present?
     ```bash
     which claudish >/dev/null 2>&1
     ```
     Presence is the claudish runtime — the binary above, or the
     `mcp__plugin_claudish_claudish__team` tool being registered in this
     session. It is never "is the multimodel plugin installed": the review
     gates need only the claudish runtime. multimodel is optional. If it is
     not installed and the user asked for models, say once that
     `claude plugin install multimodel@magus` adds `/multimodel:team` and
     `/multimodel:delegate`, then continue without it.
  3. Both true → resolve each name against the live catalog: `list_models`
     first, then `search_models` for a family it does not cover — the
     procedure in `claudish:claudish-usage`, "Model Alias Resolution". A
     version the user names is a hard constraint: if it is not in the
     catalog, say so and show the live alternatives; never substitute a lower
     version. The result is `MODELS: <bare catalog ids, comma-separated>`.
  4. Otherwise `MODELS: none`. If models were named but claudish is absent,
     tell the user once — "claudish is not installed; running the internal
     reviewer only" — and continue. The internal reviewer is never optional
     and never waits on an external one. Once `AUDIT_PATH` exists (step 3.1),
     write the skip down: append `gate skipped: code-review — MODELS none` (or
     `— claudish absent`) to `${AUDIT_PATH}/gates.log`, and repeat that line
     when relaying the result. Silence about a skipped panel is a defect
     (`${CLAUDE_PLUGIN_ROOT}/skills/core/team-gate/SKILL.md`, steps 6-7).