---

name: worktree-executor description: SUB-AGENT ONLY: This skill should be used when spawned in isolated worktree - executes PLAN.md and syncs status to task.toon. version: 0.1.0

Worktree Executor

Execute implementation plans in isolated worktrees with mandatory status synchronization.

Announce at start: "I'm executing the plan in this worktree."

---

Important: Sub-Agent Context

This is a sub-agent started by the main agent via spawn-session.sh, running in an isolated worktree.

Do NOT call any polydev scripts. The responsibilities are:

1. Read PLAN.md
2. Execute tasks step by step
3. Update task.toon status
4. Commit code changes

The main agent monitors task.toon status via poll.sh.

---

Iron Law - Violation = Failure

Every status change MUST be immediately written to task.toon:

---

Execution Flow

Step 1: Initialize

1. Update task.toon -> overall_status: in_progress, clear blocking_reason
2. Read PLAN.md from worktree root
3. Review critically - identify any questions or concerns
4. If concerns: Update task.toon -> hil, set blocking_reason, STOP
5. If no concerns: Create TodoWrite and proceed

Step 2: Execute Batch

Default batch size: 3 tasks

For each task in batch:

1. Mark task as in_progress in TodoWrite
2. Follow each step exactly as written in plan
3. Run verification commands as specified
4. If step fails:
   • Analyze the failure
   • If fixable: fix and retry
   • If blocked: go to Step 4
5. Mark task as completed in TodoWrite
6. Commit changes (if plan specifies)

Step 3: Report & Checkpoint

When batch complete:

CopyBatch [N] complete:
- Task 1: [description] done
- Task 2: [description] done
- Task 3: [description] done

Verification output:
[test/build output]

Update task.toon -> hil, blocking_reason: Awaiting code review

STOP and wait for feedback.

Step 4: Handle Blockers

blocked vs hil - Critical Distinction:

blocked flow:

1. Determine main agent might be able to solve
2. Update task.toon:

   Copyoverall_status: blocked
   blocking_reason: Needs UserService from feature/auth branch, not yet complete
   

3. STOP - do not retry, wait for main agent

hil flow:

1. Determine human intervention required
2. Update task.toon:

   Copyoverall_status: hil
   blocking_reason: Unclear if user wants OAuth or JWT
   

3. STOP - wait for human decision

Step 5: Continue After Feedback

Based on feedback:

1. Update task.toon -> in_progress, clear blocking_reason
2. Apply requested changes if any
3. Execute next batch
4. Repeat until complete

Step 6: Complete

When all tasks done:

1. Run final verification (all tests, build, lint)
2. Commit all changes (MANDATORY):

   Copygit add -A
   git diff --cached --quiet || git commit -m "feat(<branch-name>): implementation complete"
   

   Never skip this step - uncommitted changes will be lost when worktree is removed.
3. Update task.toon -> overall_status: completed
4. Output completion summary

---

When to Stop and Ask for Help

STOP executing immediately when:

• Hit a blocker mid-batch (missing dependency, test fails repeatedly)
• Plan has critical gaps preventing progress
• You don't understand an instruction
• Verification fails and you can't fix it

Ask for clarification rather than guessing.

---

Critical Reminders

• If you pause but don't update task.toon, the workflow deadlocks!
• ALWAYS set blocking_reason - main agent needs it to decide how to handle
• The main agent monitors you only through task.toon
• When blocked: STOP immediately, don't retry endlessly
• Follow plan steps exactly - don't improvise
• Don't skip verifications

---

Rule Reflection (Check Before Completion)

Trigger conditions (all must be met):

1. Encountered environment/compatibility/parameter usage issue
2. Issue will recur when new Agent executes
3. Issue has been solved with a clear solution

Action: Write file to .agent-memory/proposed-rules/<issue-summary>.md

Format:

Copy# <Issue Summary>

## Problem
<Describe the issue and trigger conditions>

## Solution
<Specific solution>

## Example
\`\`\`bash
# Wrong approach
...

# Correct approach
...
\`\`\`

Do NOT trigger for:

• Business logic issues
• One-time issues
• Uncertain if generalizable