openclaw - 💡(How to fix) Fix Gateway should enforce runTimeoutSeconds and emit terminal child.timeout event

Problem

When sessions_spawn is called with runTimeoutSeconds, the value is accepted but does not appear to be enforced as a hard release signal. Today (2026-05-27) two subagent runs in the same workspace stalled past their declared timeouts:

• TSK-20260527-0006 tester subagent: 1800s timeout, still showed status='active (waiting on 1 child)' over an hour past the deadline.
• TSK-20260527-0010 v1 coder subagent: 1800s timeout, no fs activity for 90 minutes, parent orchestrator never received any completion event.

The parent was parked via sessions_yield expecting a push-based completion event. None arrived. Recovery required a full gateway restart.

Expected behavior

At exactly runTimeoutSeconds after spawn, the gateway should:

1. Mark the child run as failed (reason=timeout).
2. Emit a synthetic completion event to the parent session so sessions_yield unparks.
3. Free the slot in the active subagents list.

This should happen regardless of whether the underlying agent process is still alive — runTimeoutSeconds is a contract with the parent, not a hint to the child.

Repro sketch

1. Spawn a child via sessions_spawn with runTimeoutSeconds: 60.
2. Have the child intentionally hang (e.g. infinite sleep loop on a tool call) or simulate a model-side drop.
3. Parent calls sessions_yield.
4. Observe: parent remains parked indefinitely; subagents action=list still reports the child as active well past 60s.

Workaround

Today the only recovery is gateway restart. A subagents action='kill' would be the manual escape hatch (filed separately).

Impact

Any multi-lane orchestration is fragile under transient failures. Parent agents cannot make progress decisions because the contract they were given (timeout) is not honored.