Overnight: Two Desktops

STOP unless the human explicitly selected two desktop agents for this run. Otherwise use overnight-one-desktop. Independent review alone does not require two desktops. bb is an optional named peer, not an inferred third topology.

Roles: one runner/gate-captain/human relay, one cross-provider independent gatekeeper/co-overseer. Both need qualified native task re-entry and explicit native task/inbox identities. The gatekeeper does not implement. Additional workers need separate authorization and single-writer leases. Read desktop-live-comms, desktop-multi-agent and mesh-comms-core. Both overseers must ACK final closeout.

Contract and Preflight

Read the canonical Overnight Cookbook, especially chapters 2, 4-12, 16, 18-20 and Appendices A, I and J. These are the maintained procedures; historical Part A/B/C and M-number references are not separate authorities.

Confirm back the resolved run contract before starting: exact project/branch, human goal, allowed writes, forbidden effects, named writer and independent reviewer, model and harness routes, budget/time limits, first useful delivery, expected unattended period, alert owner/SLA, and local versus external acceptance. A skill is not a launch or spending authorization. Preserve dirty/staged work and identify single-writer leases.

Prove a real first round trip with each critical-path peer and one harmless tool action. Render, task re-entry, nonce ACK and completed work are separate evidence. Qualify a next-turn mechanism after arming; a process or scheduled-task ID is not proof it wakes the model. Keep a lifecycle sentinel responsible until joint closeout. Reconcile the whole current outstanding request set before moving a cursor; a shallow ledger tail cannot prove no work is pending.

Before unattended reliance, exercise the fixture fault matrix in the liveness runbook. Use structured, generation-fenced heartbeat records; distinguish WAKE_SEEN from HEALTH_VALIDATED. Test credentials and the actual human wake ingress before absence. Supervision must survive the agent's credential failure. Declare shared-host failure domains: use an independent observer or record the human's explicit risk acceptance. A substitute verifier does not cancel the duty to alert the human to a dead gatekeeper. If a required mechanism is unavailable, report UNPROVEN/BLOCKED; do not invent it.

Operating Loop

Read own inbox, reconcile scoped outstanding nonces and artifacts, then classify progress, liveness and health separately. A missed ACK is an investigation trigger, not death or permission to replace a peer. Mirror material direct sends to the ledger and require receiver ACK; uncertainty requires inspection before retry. See desktop-live-comms for desktop lanes and mesh-comms-core for CLI lanes.

Deliver useful increments during the first hour. Use bounded review/repair rounds with clear acceptance and escalation; do not let process work substitute for product work. Only release the next specifically authorized milestone after an independent reviewer reruns relevant proof. The runner recommends; it cannot mint its own gate approval. Relayed human instructions need quoted/file-backed evidence and must not conflict with the latest direct request. Maintain checkpoints, decision/change records, hashes, and the automation registry. No secret values in evidence or comms.

For a Claude Code wedge, use the cookbook's qualified signature and recovery ladder (Appendix I.4), not spinner appearance alone. Preserve CPU/children/RSS, checkpoint, staged state and open nonces. At most one controlled Escape after the full signature, then verify postconditions and record RESUME. Diagnose recurrence; no blind restart. Historical watchdog environment knobs require runtime support checks, not assumption. Run heavy verification as bounded commands with retained output. Track stall/detection latency and useful progress, not just wake counts.

Acceptance and Closeout

Stage 1 is local implementation plus independent review/tests and integrity checks. Stage 2 is any required credentialed, external, web, or human acceptance. Never claim Stage 1 proves Stage 2. Complete the final handoff, required peer closeout ACKs and human report before retiring the task-owned supervision; record what remains active. No push, merge, deployment, production mutation or new agent without the specific authority in the contract. Never use Ctrl-C as routine Codex clearing.

Report: topology; exact roles/task IDs; artifact hashes; proof commands/results; Stage-1/Stage-2 status; liveness evidence; unresolved blockers; next human decision.