AgentSkillsCN

maestro-critic

Maestro Critic——运用代理团队进行跨功能冲突解决与全面质量巡检（第二阶段）

SKILL.md
--- frontmatter
name: maestro-critic
description: "Maestro Critic — cross-feature conflict resolution + full quality sweep using an agent team (Phase 2)"
user-invocable: true

User Input

text
$ARGUMENTS

You MUST consider the user input before proceeding (if not empty).

Outline

Run the reviewer: a comprehensive pass across the entire project — you (the lead) run tests, then spawn an agent team with 2 teammates working in parallel. This is designed to run in an interactive Claude Code session after all features have been implemented by Maestro.

Steps

  1. •

    Read CLAUDE.md for project standards, build/test commands, quality requirements, and any available MCP servers, plugins, or skills. Read queue/masterplan.md and queue/*.md feature files (excluding masterplan.md) to understand what each feature is supposed to do — pass this context to teammates so they can distinguish bugs from intentional design choices. Then run all tests. Before spawning the team, run the full test suite. If tests fail, fix the implementation (never modify tests) and re-run until they pass. This gives the teammates a clean baseline.

  2. •

    Create an agent team named reviewer with 2 teammates:

    • •

      conflict-checker: Find and fix cross-feature conflicts across the codebase:

      • •Duplicate routes or API endpoints
      • •Naming collisions (functions, variables, CSS classes, IDs)
      • •Import errors and circular dependencies
      • •Conflicting global state or shared resource access
      • •Inconsistent data models or schemas across features
      • •Report all conflicts found and how they were resolved.
    • •

      quality-sweep: Run a sequential deep scan across the entire codebase covering:

      1. •Security — OWASP Top 10: injection, auth gaps, sensitive data exposure, CSRF, cross-feature data leakage
      2. •Code quality — bugs, dead code, unused imports, magic numbers, code smells, inconsistent error handling
      3. •Performance — N+1 queries, unnecessary re-renders, missing indexes, unoptimized loops, large bundle imports
      • •Fix all issues found in each category before moving to the next. Report a summary per category.
  3. •

    Wait for both teammates to finish. Synthesize their results into a single reviewer report.

  4. •

    Validate: Run the full test suite one final time to ensure no fixes broke anything.

  5. •

    Commit: If all tests pass, stage and commit all changes:

    code
    maestro-critic: resolve cross-feature conflicts and quality issues

<Brief summary of what was fixed by each teammate.>

    This gives traceability in git history and crash safety for ralph-loop restarts.

  6. •

    Result:

    • •If all tests pass and all checks are resolved: output <promise>ALL_TESTS_PASS</promise>
    • •If issues remain: output TESTS_FAILED with a summary of remaining issues
  7. •

    Clean up the team when done.

Fresh-Context Retries

Run the critic inside a Ralph Loop for automatic fresh-context retries:

bash
claude --dangerously-skip-permissions
/ralph-loop "/maestro-critic" --completion-promise "ALL_TESTS_PASS"

Each iteration gets a fresh context window. If validation fails, exit with TESTS_FAILED — the Ralph Loop will restart with clean context, re-read the codebase (including any fixes from the previous iteration), and try again.

Notes

  • •The conflict-checker is the highest-value teammate — cross-feature conflicts are invisible to per-feature testing
  • •The quality-sweep handles security, quality, and perf sequentially in one context window — no file conflicts, no overlap
  • •You can message either teammate directly using Shift+Up/Down to redirect their approach