Planner to executor handoff

Goal: produce a plan the executor tier can execute without ever talking to the planner. The plan is the entire interface. Phase doc: docs/workflows/02-planning.md; tier policy: docs/models/routing.md.

The handoff contract

A handoff is complete when the executor needs nothing but the plan and the repository: no verbal context, no "as discussed", no tribal memory.

Procedure (planner)

1. Read the research artifact and the relevant code paths yourself; do not plan from summaries alone.
2. Write docs/plans/<slug>.md:

# <plan title>
Work item: <tracker id>
## Goal and non-goals
## Context          (links to research, incidents, prior plans)
## Tasks            (each: acceptance criteria + how it will be tested + assigned tier)
## Test plan        (what the test gate will run)
## Risks            (each with its mitigation and owner)
## Rollback         (what makes this change safe to revert)

3. Slice tasks to finishable-in-one-session size. A task an executor cannot complete green in a session is two tasks.
4. Assign each task a tier. Most are executor; anything re-deciding architecture is strategist.
5. Get the plan approved (human or second strategist session) before any code exists. Approval is recorded on the work item.

Procedure (executor)

1. Read the plan fully before the first command.
2. Execute tasks in order, TDD where applicable, committing per docs/standards/git.md.
3. Deviations: small ones proceed and get logged in the handoff note; scope changes stop and go back to the planner as a work-item comment.
4. Hand back: diff, green gates, deltas from plan, what the reviewer should probe.

Rules

• The planner does not implement; the executor does not re-plan. Crossing that line is how responsibility blurs and defects hide.
• If the executor is stuck twice on one task, it goes back to the planner with both attempts written down. That is a planning defect until proven otherwise.