ThreadStream

Shuts down the transport and releases any underlying resources.

Register a listener for every globally-unique event on the thread.

Fires exactly once per event_id across both the content pump (user subscribe() calls) and the lifecycle watcher. Events without an event_id always fire through (dedup is best-effort).

Returns an unsubscribe function. Primary consumer is StreamController, which uses the listener to feed discovery runners and pick up deeply-nested interrupts that the narrow content pump wouldn't deliver.

Respond to an interrupt without the v1 eager lazy-getter shims. See submitRun for why this exists alongside input.respond.

Start a run without the v1 eager lazy-getter shims.

run.start (the v1 entry point) eagerly opens a wildcard values projection so thread.output / thread.values resolve regardless of access order, and calls #ensureLifecycleTracking which opens another wildcard ["lifecycle", "input"] subscription. Both subscriptions widen #computeUnionFilter to wildcard, defeating the progressive-expansion rotation strategy.

submitRun skips those shims — callers that manage their own content subscriptions (such as StreamController) get the narrow union filter they asked for. Lifecycle / interrupt tracking is instead served by the dedicated #startLifecycleWatcher, which opens a wildcard ["lifecycle", "input"] stream alongside the narrow content pump on both SSE and WebSocket transports.

Subscribe to raw wire channels and receive protocol events.

For assembled projections, use the lazy getters instead: thread.messages, thread.values, thread.toolCalls, thread.subgraphs, thread.subagents, thread.output.