session-task-planninglisted

# Session Task Planning Convert an accepted work item's plan into session-scoped task records with explicit parallelization opportunities. Open with one sentence saying what you are about to do and what it will produce. ## Non-Negotiables 1. **Every task is a record under its work item.** One file per task at `{work}/seq-NNN/tasks/<ID>.md`, written through the runtime. Never a phase file in `todo/`, never a task inside `plans/`, and never all tasks in one document. 2. **Every task carries allowed paths, instructions, criteria, and a test.** No exceptions. A task missing any of them is not valid and must be corrected before it is captured. 3. **Criteria must be observable.** "Code is cleaner" is not a criterion. `pytest tests/foo.py::test_bar passes` is. 4. **Action verbs only.** "Create", "Add", "Move", "Extract" — not "Consider", "Look into", "Think about". 5. **Every parallelism claim is tested against the file-modification graph.** Two tasks whose `allowed_paths` overlap cannot run in parallel — and dispatch will stop them, so a wrong claim here costs a run. 6. **Exact paths, not "relevant files".** `src/auth/login.py:45-120` — not "auth files". 7. **Breaking work down grants no permission to do it.** Task records are captured, not accepted. A file existing under `tasks/` is never eligibility. ## Core Principle Each task must be completable in **one Claude Code session** (~30 min focused work). Tasks too large get split. Tasks too small get merged. **Predecessor:**