Description

When a subagent encounters a prompt-error (abort) during execution, the parent session becomes stuck. Incoming messages to the parent channel are queued with "Queued announce messages while agent was busy" but not processed until the session eventually recovers after a significant delay (~10+ minutes) or manual restart.

Reproduction

1. Parent session (Discord channel, requireMention: false) has an active subagent running
2. Subagent runs multiple tool calls (e.g. read, edit)
3. Subagent receives openclaw:prompt-error with "error": "aborted" (e.g. due to provider timeout or context limit)
4. Subagent status becomes failed
5. User sends new messages to the parent channel — they are queued but the parent session does not process them
6. Parent session only resumes after a long delay or manual intervention

Evidence from logs

Subagent transcript ends with:

{
  "type": "custom",
  "customType": "openclaw:prompt-error",
  "data": {
    "error": "aborted"
  }
}

Parent session transcript shows queued messages:

[Queued announce messages while agent was busy]

These queued messages are eventually delivered but with significant delay.

Expected behavior

When a subagent fails/aborts, the parent session should:

1. Immediately detect the subagent failure and release any locks/blocks
2. Drain the message queue and resume processing new incoming messages
3. Optionally notify the user that the subagent task failed

Environment

• OpenClaw 2026.3.23-2
• Model: Bedrock Claude Opus 4.6 (custom provider)
• Channel: Discord with requireMention: false
• Subagent spawned via sessions_spawn with runtime: "subagent"