Work the plan one thin slice at a time. A slice is the smallest piece that changes observable behavior. When a spec exists at docs/specs/<slug>.md, its acceptance criteria are the plan: work criterion by criterion, and let the spec's non-goals fence every diff. A spec that still carries an ## Open decisions section is a draft: stop and route it back to drill.

Open a todo list (a checklist in the reply when the harness has no todo tool) before the first test: one item per acceptance criterion when a spec exists, one per slice otherwise, then verify, then rubric when the work touches a quality surface, as rubric defines it. A list of fewer than three items is not opened. Mark each item done as it lands; a skipped item stays as skip: <reason>, and a criterion is never skipped, only left open.
Pick the smallest unfinished slice of the plan. When the plan has three slices or more, the harness dispatches subagents, and the user's instructions ask for them, run the slice as subagent slices lays out; otherwise work steps 3 to 5 in this session.
Write the test that fails for it. Test external behavior through the interface, never implementation details. If no failing test can be written, the seam is wrong: stop and fix the plan, not the test; a change that preserves behavior is the exception.
Write the minimum code that makes it pass.
Refactor only with tests green. Match the surrounding style, including naming and layout.
Gate the slice through verify and commit on its record.
Repeat from step 2 until the plan has no unfinished slices.
When no slice is left, run verify on the whole work, then the rubric skill when the work touches a quality surface, as rubric defines it, even when every claim is a behavior. Then offer the user land, which ends the work, and /recall on the code the agent wrote.

Rules:

implement