When to use
- Starting or resuming a task that has prior state, rules, or lessons.
- Long / multi-session / multi-stage work where chat context is unreliable.
Procedure
- Read
control/rules.md(read-only authority — you must obey, cannot edit). - Read
control/approved_lessons/index.yaml; pull the lessons relevant to this task by their summaries (load full lesson files only as needed — progressive disclosure). - Read the task's
state/tasks/<task_id>/run_state.yaml(working memory — what was believed/done so far; not ground truth) and itscorrection_journal.mdif present. - Produce an active_constraints list (from rules + relevant approved lessons + any journal entries from earlier in this task) and a short relevant_lessons summary.
- If
state/lessonbook/inbox.md($LESSONBOOK_INBOX) hasUNREVIEWEDplaceholders left by aPreCompacthook, surface them — they may hold a correction from before a compaction. Hand them tocorrection-captureto triage. - Any unmet precondition or open question → record it in
state/tasks/<task_id>/run_state.yaml(do not silently proceed past it).
Outputs
- Updated
run_state.yaml(resume marker, active_constraints). - A short "active constraints + relevant lessons" summary for this session.
Do not
- Do not treat
run_state.yamlas proof a condition is met — it records belief, not verified state. - Do not edit
control/. If a rule seems wrong, file acandidate_lessonsproposal; do not change authority yourself.