DAG runner idle timeout kills tasks during an in-progress SDK reconnect

Description

sdk/dag-task-runner (and the copied skill at .cursor/skills/dag-task-runner) wraps run.stream() with withTimeout(..., streamIdleTimeoutMs). The default idle window is 5 minutes. If no stream event arrives in that window, the task is marked ERROR as a stalled model.

Headless @cursor/sdk local runs keep enableAgentRetries: true by default. When the transport stalls, the SDK retries internally and emits nothing on the public stream. The runner cannot tell model silence from that reconnect, so it cancels a live retry and the canvas shows a dead task.

The same guard exists in both copies of run_dag.ts:

https://github.com/cursor/cookbook/blob/main/sdk/dag-task-runner/src/run_dag.ts#L363-L375

const timeoutForNext = Math.min(deadline - Date.now(), streamIdleTimeoutMs);
if (timeoutForNext <= 0) {
  throw new TimeoutError(`Task ${task.id} exceeded deadline of ${formatMs(taskTimeoutMs)}`);
}
const next = await withTimeout(
  iterator.next(),
  timeoutForNext,
  streamWaitTimeoutMessage({
    taskId: task.id,
    timeoutMs: timeoutForNext,
    streamIdleTimeoutMs,
  }),
);

Expected

While the SDK is reconnecting, the idle clock should pause. Idle should fire only on real model silence.

Suggested fix

Subscribe to a reconnect/retry signal (onConnectionStateChange / RETRYING) and skip the idle timer while reconnecting. Until @cursor/sdk forwards that hook, wire the runner to the helper in sdk/reconnect-signal (or an equivalent) so transport retries are not logged as task failures.

Related

• https://github.com/cursor/cookbook/issues/57

Compare the idle-timeout handling in sdk/dag-task-runner/src/run_dag.ts and .cursor/skills/dag-task-runner/run_dag.ts, then inspect sdk/reconnect-signal and the SDK retry signals described in the issue. Determine how the runner can distinguish reconnecting from model silence, and verify that both copies pause idle expiry during retries while still enforcing the overall task deadline.