harness-implement

Harness Implement

Drive one feature through the harness lifecycle, honoring minimal-context reading and durable state.

Boot

1. Read ai-harness/START-HERE.md once (if not already read this session).
2. For the chosen feature, read ai-harness/specs/<feature>/state.md: phase, next action, read budget, tasks.
3. Open only what the read budget lists (usually one section of spec.md).
4. Adopt the role for the current phase from ai-harness/roles.md (read only that role's section).

Loop (stop conditions are mandatory)

Run task by task. For an implementation task:

1. Set the task doing in state.md.
2. TDD where practical (red -> green -> refactor) per ai-harness/context/testing.md; if deferred, follow the DV plan and add a follow-up task.
3. Keep the domain pure; reach outside only through ports/adapters per ai-harness/context/architecture.md.
4. Move the task to review, then done once its validation passes.
5. Update state.md after every task.

STOP and hand back when ANY of:

• a blocker appears (set the task blocked, record blocker + what would unblock, set overall blocked);
• a decision genuinely needs the human (surface it clearly);
• the spec is done (all tasks done/cancelled, requirements satisfied, validation run or scheduled).

Before stopping

Update state.md (phase, last completed step, next action, task states). Report the stop reason and the single next action. The file is the truth — never rely on conversation memory.