Use when a plan with ordered tasks needs sequential execution. Triggers on "execute the plan", "implement this plan", "work through these tasks", "start the implementation".
Tasks execute sequentially. Each one must pass review before the next starts.
Read the plan. Create a task list with full task text, affected files, and dependencies. Note which tasks share files or modules — this informs the resume-vs-fresh decision in step 3.
Match each task to the appropriate builder skill:
Include the skill name in the builder's dispatch prompt.
Dispatch a builder agent with run_in_background: true. Include the full task text, affected files, and the skill name from step 2 in the prompt.
Resume or fresh? Does the next task list any of the same file paths or directories as the previous task? If yes, resume. If no, fresh.
| Signal | Resume | Fresh |
|---|---|---|
| Next task touches same files/module | Yes | -- |
| Next task is in a different package or directory tree | -- | Yes |
| Independent tasks that happen to be ordered | -- | Yes |
The builder from step 3 already completed work. Run evaluate-optimize-loop starting at step 2 (dispatch reviewer) with the builder's agent ID and the task requirements as review criteria. Mark the task complete on PASS.
Check remaining tasks. Use the resume-vs-fresh table from step 3 to decide whether to resume the existing builder or dispatch a fresh one. Repeat steps 2–5 until all tasks complete.
Run the verify-suite skill directly (not via builder dispatch). Individual tasks already verified locally — this catches cross-task regressions from interactions between tasks.