devflow-executing

DevFlow — Execution

You are in in_progress. The plan has been approved. Your job: implement it task by task, tracking progress via task_list / task_update, then submit for review.

Iron Law — Execution stays in the flow

• Work the plan, not the inbox. The implementation plan is the single source of work order. No mid-flow scope creep — file new flows for out-of-scope discoveries.
• No external agent-delegation. Don't farm tasks out to other tool-skill systems; run the loop below, with task_list / task_update as the progress marker.
• TDD per task. Failing test first, then minimal code, then verify — captured in the devflow-tdd discipline-skill (token-emitted when each cycle has its evidence).
• One commit per task when gitEnabled: true — keeps the diff auditable + lets the flow attach commits via flow_update({commits}).

Step 0 — Prepare Tasks (run before the Execution Loop)

Tasks are the unit the tasksAllCompleted-gate (DF-416) checks at in_progress → review. If the flow has zero tasks, the gate may already block at the planning-side via tasksRequired (project-config task_enforcement = 'gate'). Don't enter the loop blind:

1. Call task_list({flowId}).
2. If the list is empty, derive tasks from the approved plan and create them now:
   • Preferred: copy the entries from the plan's ## Tasks (will be created in 'in_progress') section.
   • Fallback: derive one task per Acceptance Criterion (the AC numbering matches the plan).
3. For each derived item: task_create({flowId, summary, description?, acceptanceCriteria?}).
4. Only then proceed to the Execution Loop below.

Iron Law: No tasks → no observable implementation loop. The gate has nothing to clear at review time, and the next session has no resume-point. If the plan really has no work-units (rare — usually a docs-only flow), say so explicitly in agentSummary so the reviewer knows the zero-task state is intentional.

Execution Loop

1. Call task_list to get the current task state
2. For each open task: a. Mark in_progress via task_update b. Write failing test c. Implement minimal code d. Run test, verify pass e. If gitEnabled: true → commit with Gitmoji prefix. If gitEnabled: false → skip the commit, just save your edits. f. Mark completed via task_update
3. After all tasks done: verify build/tests pass globally
4. Submit for review:

   flow_update({
     flowId: <current>,
     currentState: 'review',
     agentSummary: <what was implemented>,
     testingInstructions: <how to verify>
   })