openclaw - 💡(How to fix) Fix Stuck session not auto-killed — API call hung for 49 minutes blocking Telegram session [1 participants]

When an LLM API call hangs (no response, no error), OpenClaw's diagnostic heartbeat correctly detects the stuck session every 30 seconds but never kills it. In our case, a gpt-5.4 API call hung for ~49 minutes (2,928,499 ms), completely blocking the Telegram session for that entire duration. All subsequent messages queued behind the stuck run. 3. Triggers failover to the next model in the fallback chain (if available) or surfaces an error to the user.

TL;DR

Implement a configurable sessionMaxRunTimeMs to auto-kill stuck API calls and trigger failover or error handling.

Guidance

• Introduce a sessionMaxRunTimeMs setting to track elapsed time since run_started and abort the run if exceeded.
• Set a default value (e.g., 300s) and allow overrides per agent or model for flexibility.
• Upon exceeding the threshold, forcibly abort the run and trigger failover to the next model in the fallback chain or surface an error to the user.
• Review the diagnostic heartbeat logs to ensure correct detection of stuck sessions and verify the new setting's effectiveness.

Example

No code snippet is provided as the issue does not contain specific code references, but the suggested fix implies modifying the OpenClaw configuration or implementation to include the sessionMaxRunTimeMs setting.

Notes

The suggested fix assumes that the OpenClaw implementation allows for such a configuration change. It is essential to review the OpenClaw documentation and codebase to ensure that this change is feasible and compatible with the existing architecture.

Recommendation

Apply the suggested workaround by implementing the sessionMaxRunTimeMs setting to prevent indefinite blocking of sessions due to stuck API calls, as this addresses the root cause of the issue and provides a configurable solution for handling such scenarios.