ADR 0017: operator control of new task runs

Status: implemented; qualification in progress.

Decision

Store a repository-wide task scheduling decision in repository metadata: paused, revision, reason, operator identity and update time. Readers can inspect it; changing it requires agent.operate (delegator or administrator) and the expected revision. The decision, audit entry and outbox event commit atomically.

The agent engine checks the decision inside the transaction that claims the task and creates its run. A pause therefore cannot be bypassed by an old scheduler, another replica, or a direct task-bound StartRun request. A run that committed before the pause keeps running. Pausing is not process cancellation. Untasked/manual sessions are outside this task-launch gate; the control is deliberately labelled accordingly.

The orchestrator reads the decision before scheduling, exposes the pause in status, and defers ready tasks with operator_pause. It continues to reconcile existing work and cancel sessions whose tasks were ended. Failure to read an implemented gate stops new scheduling for that pass. A 404 from an older server is treated as lack of the additive feature; its behavior remains unchanged.

The Web UI shows the actual stored decision, its reason and author. Resume explicitly makes all ready tasks eligible for configured executors. The page does not claim that an executor is installed or connected merely because task launches are allowed.

Compatibility and recovery

Metadata schema 5 initializes the decision. Earlier binaries reject schema 5, instead of silently ignoring a persisted pause. Back up before upgrading; downgrade requires the pre-upgrade database. Both SQLite backups and PostgreSQL dumps include the gate. Protocol additions are optional and advertised as task.scheduling; older Web clients continue to work but cannot change the gate.

Limits

This is operator scheduling control, not a browser-based executor installer. Configuration of executable commands, model credentials and the orchestrator process is still an infrastructure step. Per-task executor selection and browser-managed runner setup remain open. No external coding-model run is claimed by the command-executor checks.