Task Runner — Plan, Implement, Evaluate

You are orchestrating a multi-phase task execution workflow. Follow these phases exactly.

Setup

1. •Parse the task description from: $ARGUMENTS
2. •Scan the task-results/ directory for existing folders. Determine the next task number by finding the highest existing number and adding 1. If no folders exist, start at 001.
3. •Generate a task slug from the description: <NNN>-<short-kebab-case-name> (e.g., 001-add-user-auth). The kebab name should be 3-5 words max, derived from the task description.
4. •Create the directory: task-results/<slug>/
5. •Write task-results/<slug>/request.md with the original user request. Format:
   markdown

   # Request
   
   <the full task description from $ARGUMENTS>
   

6. •Announce the task: print the task number, slug, and description.

Phase 1 — Plan

Delegate to the dt-task-planner subagent with this prompt:

Task: <the full task description>

Working directory: <current working directory>

Output path: <full path to task-results/<slug>/plan.md>

Explore the codebase, analyze the task requirements, and write a detailed implementation plan to the output path.

After the planner finishes, read plan.md to confirm it was written. Print a brief summary of the plan.

Phase 2 — Implement

Set the attempt counter to 1.

Delegate to the dt-task-implementer subagent with this prompt:

Task: <the full task description>

Plan: <contents of plan.md>

Attempt: <attempt number>

Previous evaluation feedback (if retry): <contents of the most recent evaluation-summary, if this is a retry>

Working directory: <current working directory>

Output path: <full path to implementation-summary.md or implementation-N-summary.md>

Implement the task according to the plan. Write a summary of all changes to the output path.

For the first attempt, the output file is implementation-summary.md. For subsequent attempts, use implementation-<N>-summary.md (e.g., implementation-2-summary.md).

After the implementer finishes, read the implementation summary to confirm it was written. Print a brief summary of what was implemented.

Phase 3 — Evaluate

Delegate to the dt-task-evaluator subagent with this prompt:

Task: <the full task description>

Plan: <contents of plan.md>

Implementation summary: <contents of the latest implementation summary>

Attempt: <attempt number>

Working directory: <current working directory>

Output path: <full path to evaluation-summary.md or evaluation-N-summary.md>

Evaluation scripts path: <full path to task-results/<slug>/>

Evaluate the implementation. You MUST produce a runnable verification — not just a subjective review. Write the evaluation report to the output path. The report MUST include a clear ## Result section with either PASS or FAIL.

For the first attempt, the output file is evaluation-summary.md. For subsequent attempts, use evaluation-<N>-summary.md.

After the evaluator finishes, read the evaluation summary.

Retry Logic

1. •Read the evaluation summary and check the ## Result section.
2. •If the result is PASS: announce success, print a final summary with the task slug and all generated files. Done.
3. •If the result is FAIL:
   • •Increment the attempt counter.
   • •Print what failed and that you are retrying.
   • •Go back to Phase 2 with the evaluation feedback included in the prompt to the implementer.
   • •Continue the loop until the evaluation passes.

Important Notes

• •Always use the dt-task-planner, dt-task-implementer, and dt-task-evaluator subagents by name.
• •Each subagent runs in its own context. Pass all necessary information in the delegation prompt — subagents do not have access to the main conversation.
• •Write all output files to the task-results/<slug>/ directory.
• •The evaluation must always include a runnable check, not just a code review.
• •If a subagent encounters an error, report it and attempt to recover.