investigate

<user_request>$ARGUMENTS</user_request>

<critical_override> THIS COMMAND OVERRIDES THE CLAUDE.md TASK ROUTING TABLE. WHY: This is a READ-ONLY orchestrator. It must NEVER self-handle investigation work. RULE: ALL investigation must be delegated via Agent tool to the resolved agent. NEVER: Read files, trace code, or analyze architecture inline. NOTE: No autonomy question — investigation is inherently read-only and non-destructive. </critical_override>

  question: "What do you want to investigate?"
  header: "Investigation Scope"
  options:
    - label: "Architecture"
      description: "How system components connect and why — map the big picture"
    - label: "Implementation"
      description: "How a specific feature or flow works — trace the code path"
    - label: "Bug hunt"
      description: "Trace error origins and find where failures occur"
    - label: "Test coverage"
      description: "What is and isn't tested — find coverage gaps"
    - label: "Comprehensive"
      description: "Full multi-perspective analysis — all of the above"

  Inference rules (skip AskUserQuestion if match is confident):
  - "how does X work", "explain", "understand", "trace" → implementation
  - "architecture", "design", "components", "structure", "system" → architecture
  - "error", "crash", "why does", "broken", "bug", "fail" → bug-hunt
  - "test", "coverage", "tested", "untested", "gaps" → test-audit
  - "everything", "full", "comprehensive", "deep", "complete" → comprehensive
</step>

<step number="2" name="Plugin Check">
  Check if the code-analysis plugin is available:

  ```bash
  ls "${HOME}/.claude/plugins/cache/" 2>/dev/null | grep -q "code-analysis"
  ```

  If code-analysis is installed: proceed to step 3a.
  If absent: proceed to step 3b.
</step>

<step number="3a" name="Delegate (code-analysis present)">
  Map scope to investigation mode:
    architecture  → mode: architecture, skill: code-analysis:investigate
    implementation → mode: implementation, skill: code-analysis:investigate
    bug-hunt      → mode: bug, skill: code-analysis:investigate
    test-audit    → mode: test, skill: code-analysis:investigate
    comprehensive → skill: code-analysis:deep-analysis (no mode parameter — it has none)

  Launch the agent:

  Agent(
    run_in_background: false,
    description: """
      Investigate: {$ARGUMENTS}

      Investigation mode: {mode}
      Load skill: {skill} and use mode={mode}.

      RULES:
      - This is READ-ONLY analysis. Do not modify any files.
      - Use the code-analysis MCP tools (mcp__plugin_code-analysis_ca__*) for semantic code navigation.
      - Produce a clear investigation report with findings and evidence.

      FOCUS BY MODE:
      - architecture: system boundaries, component relationships, data flows,
        architectural patterns, PageRank analysis of core abstractions.
      - implementation: entry points, call chains, data transformations,
        exact file/line evidence for each claim.
      - bug: error origins, failure conditions, blast radius,
        reproduction steps, root cause identification.
      - test: what is tested, what is not, coverage gaps,
        risky untested paths, test quality assessment.
      - (comprehensive scope routes to code-analysis:deep-analysis, which takes no
        mode) all 4 modes above, synthesized into unified report
        with cross-cutting observations.
    """,
    subagent_type: "code-analysis:detective"
  )
</step>