AgentSkillsCN

refine-loop

持续运行评审 + 修复循环，直至 ISSUE_COUNT = 0，或达到迭代式代码质量改进的最大循环次数。

SKILL.md
--- frontmatter
name: refine-loop
description: "Runs a review + fix loop that continues until ISSUE_COUNT = 0 or max iterations for iterative code quality improvements."
user-invocable: true
allowed-tools: Bash, Write, Read, AskUserQuestion, Skill, Task
hooks:
  Stop:
    - hooks:
        - type: command
          command: "bash -c 'node ~/.claude/scripts/refine-loop-hook.js'"

/refine-loop

Start a review + fix loop that continues until ISSUE_COUNT = 0 or max iterations.

Progress Checklist

code
- [ ] Step 1: Parse args and set variables
- [ ] Step 2: Collect type-specific context
- [ ] Step 3: Determine type directory name
- [ ] Step 4: Save context
- [ ] Step 5: Create state file
- [ ] Step 6: Execute the loop

Steps

  1. •

    Parse args and set variables:

    • •type=code → set type to code
    • •type=tests → set type to tests
    • •type=code-tests → set type to code-tests
    • •type=ui-tests → set type to ui-tests
    • •type=docs → set type to docs
    • •type=ac → set type to ac
    • •iterations=N → MAX_ITERATIONS = N (default: 10)
    • •scope → use provided value, skip corresponding question in Step 2
    • •ticket → use provided value, skip ticket selection in Step 2
    • •file → use provided value, skip file path question in Step 2 (docs mode)

    If type not specified, ask user to select type via AskUserQuestion:

    OptionAgentDescription
    codecode-reviewer/fixer + security-reviewer/fixerコード + セキュリティレビュー（並列）
    teststests-reviewer/fixerテストコードレビュー + 修正
    docsdocument-reviewer/fixerドキュメントレビュー + 修正
    acac-reviewerACレビュー（レポートのみ、修正なし）
  2. •

    Collect type-specific context:

    If code selected:

    • •Resolve ticket:
      • •If ticket arg provided → TICKET_PATH = ticket value
      • •Else → invoke ticket-reader list, ask user to select via AskUserQuestion → TICKET_PATH
    • •"スコープは？" → Uncommitted / Branch
      • •Uncommitted → SCOPE = "uncommitted"
      • •Branch → SCOPE = "branch"

    If tests selected:

    • •Resolve ticket:
      • •If ticket arg provided → TICKET_PATH = ticket value
      • •Else → invoke ticket-reader list, ask user to select via AskUserQuestion → TICKET_PATH
    • •"スコープは？" → Uncommitted / Branch
      • •Uncommitted → SCOPE = "uncommitted"
      • •Branch → SCOPE = "branch"
    • •If $TEST_FILTER is empty or "<test-filter>": Error "Run /init-project first to configure test filter"

    If code-tests type (from args):

    • •If ticket arg provided → TICKET_PATH = ticket value
    • •Else → invoke ticket-reader list, ask user to select via AskUserQuestion → TICKET_PATH
    • •"スコープは？" → Uncommitted / Branch
      • •Uncommitted → SCOPE = "uncommitted"
      • •Branch → SCOPE = "branch"
    • •If $TEST_FILTER is empty or "<test-filter>": Error "Run /init-project first to configure test filter"

    If ui-tests type (from args):

    • •If ticket arg provided → TICKET_PATH = ticket value
    • •Else → invoke ticket-reader list, ask user to select via AskUserQuestion → TICKET_PATH

    If docs selected:

    • •If file arg provided → FILE_PATH = file value
    • •Else → "どのドキュメントをチェックしますか？" → File path input via AskUserQuestion → FILE_PATH
    • •Read document file → CONTENT
    • •Invoke knowledge-reader agent: OPERATION=resolve WORKFLOW=/refine-loop → KNOWLEDGE

    If ac selected:

    • •Resolve ticket:
      • •If ticket arg provided → TICKET_PATH = ticket value
      • •Else → invoke ticket-reader list, ask user to select via AskUserQuestion → TICKET_PATH
    • •"スコープは？" → Uncommitted / Branch
      • •Uncommitted → SCOPE = "uncommitted"
      • •Branch → SCOPE = "branch"
  3. •

    Determine type directory name:

    • •code → code
    • •tests → tests
    • •code-tests → code and tests (creates both)
    • •ui-tests → ui-tests
    • •docs → docs
    • •ac → ac
  4. •

    Save context to .claude/refine-loop/{type}/context.json:

    • •code: {"TICKET_PATH": "...", "SCOPE": "..."}
    • •tests/code-tests: {"TICKET_PATH": "...", "SCOPE": "..."}
    • •ui-tests: {"TICKET_PATH": "..."}
    • •docs: {"FILE_PATH": "...", "CONTENT": "...", "KNOWLEDGE": [...]}
    • •ac: {"TICKET_PATH": "...", "SCOPE": "..."}
  5. •

    Create state file .claude/refine-loop/{type}/state.local.md with checklist:

markdown
---
type: {TYPE}
max_iterations: {MAX_ITERATIONS}
iteration: 1
---

# Refine Loop

- [ ] Iteration 1: Review
- [ ] Iteration 1: Fix
- [ ] Iteration 2: Review
- [ ] Iteration 2: Fix
... (max_iterations)
- [ ] False Positive Check
- [ ] Report
- [ ] Cleanup
  1. •Execute the loop following the Protocol below

Protocol

Async Handling: Task tool may switch to async mode automatically (due to context pressure).

  • •If Task returns inline result → use directly as REVIEW_RESULT/FIX_RESULT
  • •If Task returns "Async agent launched" with agentId → use TaskOutput(task_id=agentId, block=true) to wait and retrieve result
  • •If TaskOutput returns empty → retry with longer timeout or re-invoke Task

Single type (code): Review Phase:

  1. •Review: invoke code-reviewer AND security-reviewer agents in parallel
    • •code-reviewer: SCOPE={scope}
    • •security-reviewer: SCOPE={scope}
  2. •Capture outputs as CODE_REVIEW_RESULT and SECURITY_REVIEW_RESULT
  3. •Extract ISSUE_COUNT from each (Critical/High + fixable only)
    • •CODE_ISSUES from CODE_REVIEW_RESULT
    • •SECURITY_ISSUES from SECURITY_REVIEW_RESULT
    • •TOTAL_ISSUES = CODE_ISSUES + SECURITY_ISSUES
  4. •If TOTAL_ISSUES = 0: skip to Loop End

Fix Phase: 5. Fix: invoke code-fixer AND security-fixer agents in parallel

  • •code-fixer: SCOPE={scope} REVIEW_RESULT={code_review_result}
  • •security-fixer: SCOPE={scope} REVIEW_RESULT={security_review_result}
  1. •Continue to next iteration

Single type (tests):

  1. •Review: invoke tests-reviewer with SCOPE={scope} → capture output as REVIEW_RESULT
  2. •Extract ISSUE_COUNT from REVIEW_RESULT (Critical/High + fixable only)
  3. •If ISSUE_COUNT = 0: skip to Loop End
  4. •Fix: invoke tests-fixer with SCOPE={scope} REVIEW_RESULT={review_result}
  5. •Continue to next iteration

Single type (ui-tests):

  1. •Review: invoke ui-tests-reviewer with TICKET_PATH={ticket_path} → capture output as REVIEW_RESULT
  2. •Extract ISSUE_COUNT from REVIEW_RESULT (Critical/High + fixable only)
  3. •If ISSUE_COUNT = 0: skip to Loop End
  4. •Fix: invoke ui-tests-fixer with TICKET_PATH={ticket_path} REVIEW_RESULT={review_result}
  5. •Continue to next iteration

code-tests type: This mode invokes code-reviewer/fixer and tests-reviewer/fixer agents in parallel.

Review Phase:

  1. •Review: invoke code-reviewer AND tests-reviewer agents in parallel
    • •code-reviewer: SCOPE={scope}
    • •tests-reviewer: SCOPE={scope}
  2. •Capture outputs as CODE_REVIEW_RESULT and TESTS_REVIEW_RESULT
  3. •Extract ISSUE_COUNT from each (Critical/High + fixable only) → TOTAL_ISSUES = CODE_ISSUES + TEST_ISSUES
  4. •If TOTAL_ISSUES = 0: skip to Loop End

Fix Phase: 5. Fix: invoke code-fixer AND tests-fixer agents in parallel

  • •code-fixer: SCOPE={scope} REVIEW_RESULT={code_review_result}
  • •tests-fixer: SCOPE={scope} REVIEW_RESULT={tests_review_result}
  1. •Continue to next iteration

Single type (docs):

  1. •Review: invoke document-reviewer with CONTENT={content} KNOWLEDGE={knowledge} → capture output as REVIEW_RESULT
  2. •Extract ISSUE_COUNT from REVIEW_RESULT (all fixable issues)
  3. •If ISSUE_COUNT = 0: skip to Loop End
  4. •Fix: invoke document-fixer with CONTENT={content} KNOWLEDGE={knowledge} REVIEW_RESULT={review_result}
  5. •Capture FIXED_CONTENT from output
  6. •Write FIXED_CONTENT to FILE_PATH
  7. •Update CONTENT = FIXED_CONTENT for next iteration
  8. •Continue to next iteration

Single type (ac):

  1. •Review: invoke ac-reviewer with TICKET_PATH={ticket_path} SCOPE={scope} → capture output as REVIEW_RESULT
  2. •Extract ISSUE_COUNT from REVIEW_RESULT (Critical/High, regardless of fixable flag)
  3. •Loop End (no fix phase)

Loop End (after ISSUE_COUNT = 0 or max iterations):

  1. •

    False Positive Check:

    • •Review the final iteration's reported issues
    • •Signs of false positive: intentional design, context misunderstanding, valid pattern flagged as violation
    • •If false positive found → revert the fix, exclude from FIXED_COUNT
    • •If ALL issues were false positives → treat as successful completion
  2. •

    Report:

    • •Read TICKET_PATH from .claude/refine-loop/{type}/context.json
    • •If TICKET_PATH exists: Write <TICKET_PATH>/refine-{type}-report.md with:
      • •Summary: total iterations, total fixed issues
      • •Per-iteration: REVIEW_RESULT table and FIXED_COUNT
    • •If TICKET_PATH not found: skip report output
  3. •

    Cleanup: Delete .claude/refine-loop/{type}/ directory only

    • •IMPORTANT: Other types may be running in parallel, only delete the current type's directory