implement

Implement

Execute a plan. Each step gets test → code → commit, with the plan doc updated as you go.

Pre-conditions

• Plan exists at docs/plans/YYYY-MM-DD-<topic>.md with Status: approved
• Working tree is clean (or the user has explicitly authorized building on dirty state)
• Branch is the right one (verify with the user if unclear)

If pre-conditions fail, return status=blocked with the gap.

Loop

For each plan step in declared order:

1. Read the step — confirm files, action, test, dependencies
2. Invoke tdd skill — write failing test → minimal code → refactor (test stays green)
3. Run the step's verification test — must pass
4. Run the suite — full suite must remain green
5. Commit — message format: <step-N>: <action summary> (plan: <plan-path>)
6. Update plan — mark step status. Note any deviations.
7. Stop and report on failure — do not skip a failing step

Failure handling

Inside a step:

• Test won't pass after reasonable iterations → stop, report, return to orchestrator
• Step turns out to depend on something not in the spec → stop, surface to user, possibly return to design

Between steps:

• Plan revealed wrong → stop, return to orchestrator with revision request
• Earlier step's choice blocks later step → return to plan with the conflict

Refusal conditions

Refuse to proceed if:

• The user asks you to skip a step's test
• The user asks you to commit a failing test as "todo: fix later"
• The plan calls for an action that contradicts the spec (re-design needed)

In each case, state the refusal reason and return to the appropriate prior stage.

Output

{
  "completed_steps": <int>,
  "failed_step": null | { "step": <int>, "reason": "<>" },
  "commits": [<sha>, ...],
  "plan_path": "<path>"
}

Handoff

On full completion, return to the orchestrator for stage 4 (verify).