watchdogs

Watchdogs and recovery

Every wakeup runs the required coordinator cycle to actually ask pending questions and fill free lanes. Reading or reporting a sweep alone is not an adequate wakeup result. Recheck immediately after each completion; do not wait for a timer.

Use the available harness's completion events plus an authorized fallback wake-up for in-flight work (typically 20–30 minutes). Record its actual id and scope in the canonical state. A missing scheduler is a disclosed limitation, never a reason to claim a wake-up is armed. Do not create duplicate monitors or revive a schedule the founder disabled.

Sweep the recorded queue

On wake/resume, run the bundled coord.py --root WORKSPACE sweep alongside the normal state read. It reports releases/user blockers without a linked question, an idle question slot with queued work, outdated questions, and quiet in-progress/verifying tasks. It never sends, changes status, dispatches or cancels. Use --available-slots N only after a fresh native fleet/capacity check; its dispatch candidates still require scoped authorization and lane/resource checks. check_capacity requires that live check, prepare_task requires intake, and review_dispatch_scope requires checking a hold against the named task rather than freezing unaffected work. Finish the cycle with --check-idle; exit 2 means outstanding action/blocker accounting, not permission to go silently idle. Unknown capacity is not zero agents, and missing recent DB notes is not proof an agent died. queue_conflict and clear_closed_queue expose stale ordering; correct canonical records after checking their intent. Conflicting positions withhold those tasks from dispatch candidates until reconciled; unrelated eligible work can proceed. review_lane flags running work without a lane or active lanes whose task is missing/closed. Inspect native ownership and remaining processes before repairing records; do not launch a replacement or release a checkout merely to make the sweep green.