long-task-relaylisted

# Long-Task Relay Use a mechanical watcher for waiting and an agent only for judgment. The agent arms the watcher, verifies it, and ends the current turn. The watcher polls local state without calling a model, then delivers one compact event to the same conversation when attention is required. Do not keep an agent alive to run `sleep`/status loops. Do not let the watcher interpret results, edit files, or choose a new task. ## Goal mode and relay are mutually exclusive per objective Treat control ownership per agent conversation and unresolved objective, not globally across a distributed job. An unfinished Goal and a relay must not own the same control loop: the Goal keeps that agent active, while a relay requires an idle conversation that can be resumed only when judgment is needed. - If the user explicitly selects Goal mode for the unfinished long-running objective, do not arm a relay for that objective and do not inject relay events into that active Goal session. - If relay mode is selected, use ordinary mode: arm and verify the watcher, then end the agent turn. The watcher owns waiting until it delivers one deduplicated event. - A bounded Goal may repair one already-delivered incident and complete before the relay is re-armed. It must not remain active while the relay watches the same unresolved objective. - Different agents may use different modes for distinct role-owned objectives. In particular, a coordinator Goal may own bounded cluster recovery thro