sum-rundown

Run `./bin/sumctl inbox --live`, which takes one bounded agent list snapshot per Herdr session and reads every active task's state from it. Twelve workers cost one observation call instead of twelve sequential waits.

Record the answer with `./bin/sumctl answer TASK_ID QUESTION_ID --text 'The authorized answer'`. The answer stays visible until the worker marks it applied, and unanswered decisions surface first in the rundown.

No, the rundown is not a background monitor or daemon. It reconciles missed events once from one snapshot per session and runs only when invoked, so it does not promise instant unattended detection.

A `cleanup: pending` row means the exact PR was observed merged but cleanup remains unfinished; offer `sumctl cleanup TASK_ID`. A `blocked` row names blocker codes that should be relayed, never cleared by force.

Run `./bin/sumctl init` in the new pane; only after confirming the old coordinator is gone, use `init --role coordinator --reclaim`, which proceeds only when Herdr reports the old pane as `pane_not_found`. Then bind existing tasks with `sumctl bind TASK_ID --parent-only`.

It reads saved records and Herdr only, never GitHub, so CI rows reflect the last observation rather than the current state. It has no automatic retry, process fencing, or merge polling, and idle workers are not assumed done without inspection.