Agent Skills: Dev Step All

Dev Step All

stepファイルを全て読み込み、ユーザー確認なしで連続実装する。エラー・テスト失敗時は停止して報告。

Workflow Schema

See workflow.yaml for YAML structure.

Procedure

1. Validation

1. Get branch name from git: git branch --show-current

   • Extract number: feat/#1473 → 1473

2. Check if docs/{branch_name}/dev-workflow.yaml exists

   • If not exists → Ask user:

     「dev-workflow.yaml が見つかりません。
     先に /dev-plan を実行して計画を作成しますか？」
     

   • Stop and wait for user response

3. Read YAML and validate state:

   • If plan.status != Done →

     「計画が未完了です（status: {current_status}）。
     先に /dev-plan を完了してください。」
     

   • Stop

4. Check execute.status:

   • If Done → "全stepが完了しています。再実行しますか？"
   • If Pending → Update to In progress, proceed
   • If In progress → Resume from current_step

2. Load Step File

1. Determine step file path:

   • Project: Extract from pwd (e.g., mu-muc-app-service2 → mu-muc-app-service)
   • Path: ~/uuta/Projects/{project}/{branch_name}/*-step-{current_step:02d}.md

2. Read step file using Glob + Read

   • If not found → Error with available files list

3. Display step info:

   ## Step {current}/{total}: {step_title}
   実装中...
   

3. Execute Step

1. Use Claude Code Agent Teams (TeamCreate + Task tool) as much as possible to parallelize work within a step:

   • Identify independent sub-tasks (e.g., writing tests for different modules, implementing unrelated files, running checks)
   • Spawn teammates to handle sub-tasks concurrently
   • Coordinate results before marking the step complete
   • Fall back to sequential execution only when tasks have strict dependencies

2. Implement according to step file instructions

3. Follow TDD approach if applicable:

   • Red: Write failing tests first
   • Green: Implement minimum code to pass
   • Refactor: Clean up

4. Run verification checks defined in step file

   • On test failure or error → STOP immediately and report the issue. Do not continue to the next step.

4. Complete Step

1. Update step file status to Done (in Obsidian)

2. Update YAML:

   execute:
     current_step: {current + 1}
   

3. Check if all steps complete:

   • If current_step > total_steps:

     execute:
       status: Done
     

     Report: "全てのstepが完了しました！" → STOP

5. Auto-Continue

• If steps remain → Loop back to Phase 2 immediately (no user confirmation)
• Only stop when:
  • All steps are Done
  • An error or test failure occurs (report the issue)