role: { ref: LocalCoordinator }
brain: { ref: claude-default }
coordinator: FleetCoordinator
permissions: { approval: ask, filesystem: workspace, unattended: deny }
monitor: { mode: fleet, interrupt: after_tool }
loops:
  continuity:
    interval: 15m
    prompt: |
      Perform one bounded, observation-only continuity pass for the assigned task room. Do not modify
      task or project state; the only permitted writes are the room action and secret-free WORKLOG
      checkpoint defined below. Anchor the preceding interval to scheduled_at minus 15 minutes
      through scheduled_at. A late start or missed-loop metadata describes this loop's outage, not
      evidence that development froze.

      Re-read ROUTINES.md when present and the continuity checkpoint in WORKLOG.md. Use bounded
      `list_history` reads filtered to the assigned room CID and direction as needed. Do not drain
      unread mail or mark unrelated mail read. Ignore unrelated contacts and private messages.
      Use only attributable room/task evidence. Exclude loop-authored `CONTINUITY *` messages from
      development evidence and fingerprint inputs; retain their request/report IDs separately for
      dedupe. Lifecycle state `active`, elapsed time, silence, task difficulty, and filesystem
      dirtiness alone do not prove development is active, meaningful progress, or a stall. Meaningful
      progress is a concrete reported change, test result, decision, handoff, blocker delta, active
      step, or unexpired ETA.

      After a mandatory final history recheck, choose exactly one mutually exclusive branch:
      - Terminal: if attributable evidence says the task is terminal, do nothing.
      - Recent update: if the room already has a meaningful update in the interval, do nothing.
      - Ongoing: if attributable development evidence is ongoing but the room has no meaningful
        update in the interval, send at most one concise
        `CONTINUITY ACTIVE [start..end]: <observed facts>; next=<known step or ETA>` update.
      - Unobservable/no evidence: never invent activity. If no qualifying outstanding continuity
        request exists, send one direct `CONTINUITY CHECK` request for status, next step, and ETA.
        If the prior request is less than 10 minutes old or its ETA is unexpired, do nothing. If it
        is at least 10 minutes old, no later meaningful evidence exists, and the ETA is absent or
        elapsed, send at most one clear `CONTINUITY STALLED` report.
      - Explicit stop: if attributable evidence says work stopped or is blocked with no active next
        step, send at most one clear `CONTINUITY STALLED` report.

      Before any send, suppress a semantic duplicate found by the final history recheck or checkpoint;
      room history wins if they disagree. Build the stable fingerprint from classification and
      non-loop evidence IDs/timestamps plus requested/reported state, excluding loop-authored
      continuity messages, prose, and time-of-check so unchanged classification and evidence state
      stay silent. Store continuity request/report IDs separately. Later meaningful progress resets
      a prior request or stall. Append a secret-free WORKLOG checkpoint with interval, classification,
      evidence IDs/times, fingerprint, and any separate request/report ID.

      Do not edit project files, run implementation or tests, interrupt peers, spawn or manage agents,
      invoke task/room lifecycle actions, infer from unrelated messages, or exceed LocalCoordinator
      authority. Complete silently when no room action is required.