Symptom

When the gateway restarts (via hermes gateway restart or launchd KeepAlive), any delegate_task subagents that are in-flight are not cancelled. They become orphaned, lose their API connections, and eventually time out after 12–16 minutes of no response. The user sees:

• Agent idle timeout (781s for main agent, 757s/765s for subagents)
• Subagent interrupted mid-API-call errors
• Lost work from the delegate tasks

Evidence

From gateway.log during a restart event (gateway received restart command at 06:56, old process lingered until 07:15):

No response from provider for 968s (model: deepseek-v4-pro, context: ~199,242 tokens)
No response from provider for 1097s (model: deepseek-v4-pro, context: ~199,242 tokens)
[subagent-0] No response from provider for 757s (model: kimi-k2.6, context: ~3,080 tokens)
[subagent-1] No response from provider for 765s (model: kimi-k2.6, context: ~5,538 tokens)
[subagent-1] Interrupted during API call.
[subagent-0] Interrupted during API call.

All 56 timeout/disconnect/interrupt events in the gateway log across multiple sessions occurred during gateway restart transition periods.

Root Cause

cancel_session_processing in the gateway shutdown path does not recursively cancel in-flight delegate subagents. When the gateway is shutting down:

1. The main agent session gets a shutdown signal
2. But delegate_task subagents continue running independently
3. The gateway process loses its API connections during restart
4. Subagents wait indefinitely for responses (757s, 765s observed)
5. Eventually they hit connection-level timeouts and are interrupted

Proposed Fix

In the gateway shutdown sequence, before disconnecting platform adapters:

1. Collect all active sessions with in-flight delegate tasks
2. Cancel each subagent (set interrupt flag, cancel pending API calls)
3. Wait for subagents to acknowledge cancellation (with a short grace period, e.g. 5s)
4. Only then proceed with transport disconnect and gateway shutdown

Related

• [Bug]: sessions still leak open rows and ghost stubs across gateway/cli/cron after completion or exception paths #12029 — session leak across restart/exception paths
• hermes gateway restart: CLI deadline equals gateway drain budget, false-fires 'still running after 60s' warning and force-kills mid-cleanup #25966 — gateway restart force-kill races cleanup tail