implement

name	implement
description	Analyzes ROADMAP and implements tasks in parallel with automatic orchestration. Use when starting implementation work.

Parallel Implementation

spec/tasks/ROADMAP.md を確認し、並列実装可能なタスクを分析して実装を進めます。

CRITICAL: Main Agent Role

メインエージェントは管理・指揮のみを行い、直接作業は一切行いません。

以下は全てサブエージェント（ワーカー）に委譲すること：

• 実装: コードの作成・編集
• テスト: テストの実行・確認
• レビュー: PRのレビュー
• 調査: コードベースの調査・分析

メインエージェントが行うのは：

• タスク分析と優先順位付け
• ワーカーのスポーン・監視
• オーケストレーション管理
• レビュー結果の報告・ユーザー判断の仲介
• マージとROADMAP更新

Planned Tasks

!grep "Planned\|Pending" spec/tasks/ROADMAP.md

Active Worktrees

!git worktree list

Steps

1. Task Analysis

• Check spec/tasks/ROADMAP.md for "Planned"/"Pending" tasks
• Identify tasks with satisfied dependencies (parallel candidates)
• Select tasks to implement
  • Multiple parallelizable tasks → use spawn-worker.sh

2. Spawn Workers

Worker limit: max 4 workers (plus the main agent). Before spawning, check current agent count:

./scripts/monitor-agents.sh  # workers (non-main rows) must be < 4

If 5+ tasks exist, spawn sequentially — wait for one to complete before spawning the next.

# Spawn worker for each task
./scripts/spawn-worker.sh <branch-name> <task-keyword>

spawn-worker.sh automatically:

1. Creates worktree at ~/.herdr/worktrees/reference-manager/<branch-name>
2. Runs npm install
3. Opens a herdr workspace and starts Claude with the task (--permission-mode auto)

3. Start Orchestration

./scripts/orchestrate.sh --background

Orchestrator (detect + notify model):

• Detects agent state changes (idle, error, etc.)
• Writes event files to /tmp/claude-orchestrator/events/
• Sends 1-line notification to main agent pane
• Does NOT auto-execute actions — main agent decides

4. Monitor Implementation

Periodically check worker progress:

herdr agent read <pane-id> --lines 30

Context exhaustion handling: If agent output shows signs of context exhaustion (repetitive output, no progress):

1. Send commit instruction: ./scripts/send-to-agent.sh <pane-id> "現状をcommitしてpushしてください"
2. Wait for commit, then kill: ./scripts/kill-agent.sh <pane-id>
3. Re-spawn: ./scripts/spawn-worker.sh <branch-name> <task-keyword>

Blocked (permission) handling: If check-agent-state.sh reports permission, or the orchestrator sends an agent-blocked event:

1. Inspect the pane: herdr agent read <pane-id> --lines 30
2. Startup dialog (MCP/trust confirmation) → accept it: herdr pane send-keys <pane-id> Enter
3. Auto-mode fallback to manual prompts, or an unexpected interactive prompt → kill and re-spawn: ./scripts/kill-agent.sh <pane-id> then ./scripts/spawn-worker.sh <branch-name> <task-keyword>

5. Implementation Complete → Kill → Spawn Reviewer

When worker finishes (PR created, visible in pane output):

1. Confirm PR exists: gh pr list --head <branch>
2. Kill implementation agent: ./scripts/kill-agent.sh <pane-id>
3. Spawn reviewer: ./scripts/spawn-reviewer.sh --pr <pr-number>

6. Review Complete → Report to User

When reviewer finishes (review comment posted, visible in pane output):

1. Confirm review posted: gh pr view <pr-number> --json reviews
2. Kill reviewer: ./scripts/kill-agent.sh <pane-id>
3. Report ALL findings to user (minor含む全件報告):
   • Critical/Major issues
   • Minor issues and suggestions
   • Test/CI status
4. Ask user for decision — merge or fix

7. Fix (if user requests)

When user requests fixes for review findings:

1. Launch fix agent with specific instructions:

   ./scripts/launch-agent.sh <worktree-dir> "<fix instructions>"
   

2. Monitor until complete
3. Kill fix agent: ./scripts/kill-agent.sh <pane-id>
4. Report to user and ask for next action (re-review or merge)

8. Merge (after user approval)

./scripts/merge-pr.sh <pr-number>

9. Post-Merge (main branch)

• Update ROADMAP.md status to "Done"
• Move task file to spec/tasks/completed/
• Clean up worktree if needed: git worktree remove <path>

Notes

• メインエージェントは直接コードを書かない・テストを実行しない・レビューしない・調査しない
• All agents run in herdr workspaces (one workspace per worktree)
• Use git worktree list to see all worktrees
• Use ./scripts/monitor-agents.sh to see agent states
• Orchestrator commands: --status, --stop, --clean (clear persisted states)
• Be aware of dependency conflicts during parallel work