openclaw - 💡(How to fix) Fix Session lane not released after LLM provider timeout, blocking all subsequent messages [1 comments, 2 participants]

Description

When an LLM provider (OpenAI Codex / gpt-5.5) returns consecutive server_error timeouts during a group chat agent run, the session lane is not properly released. All subsequent messages to the same session are queued indefinitely behind the "ghost run", effectively freezing the session until a full gateway restart.

Environment

• OpenClaw version: 2026.5.7
• OS: Linux 6.17.0-20-generic (x64), Ubuntu
• Node: v24.14.1
• Channel: Feishu (group chat, 4 bot accounts: sanqi/taiyi/shengongbao/nezha)
• Model: openai-codex/gpt-5.5
• Session type: agent:main:feishu:group:<chat_id>

Steps to Reproduce

1. Configure a Feishu group chat with requireMention: true
2. Have the agent perform a multi-step task in the group chat (involving tool calls like exec, edit)
3. While the agent is executing, the LLM provider (Codex) returns consecutive server_error / timeout errors
4. The agent run ends with isError: true
5. Send a new message to the group chat (with @mention)

Expected Behavior

The session lane should be released after the failed agent run, allowing new messages to trigger a fresh agent run.

Actual Behavior

• The session lane remains locked
• New messages are dispatched but queued (queueAhead=1)
• Diagnostic log shows: lane wait exceeded: lane=session:agent:main:feishu:group:... waitedMs=403570 queueAhead=1
• dispatch complete (queuedFinal=false, replies=0) — messages are accepted but never processed
• Only a full openclaw gateway restart recovers the session

Relevant Logs

15:21:37 WARN embedded_run_agent_end: isError=true, error="LLM error server_error", failoverReason="timeout", model="gpt-5.5", provider="openai-codex"
15:22:34 WARN embedded_run_agent_end: isError=true, error="LLM error server_error" (second timeout)
15:19:08 WARN lane wait exceeded: lane=session:agent:main:feishu:group:oc_d497bd2b11243f8c001af50d983504ba waitedMs=403570 queueAhead=1
15:22:46 INFO feishu[sanqi]: dispatching to agent (session=...)
15:22:46 INFO feishu[sanqi]: dispatch complete (queuedFinal=false, replies=0)

Impact

• Group chat sessions become completely unresponsive after LLM provider errors
• No CLI command or API endpoint exists to abort a stuck session lane (openclaw sessions only has cleanup, no abort)
• Users must restart the entire gateway to recover, which affects all sessions

Suggested Fix

1. Error recovery: When an agent run ends with isError: true due to provider timeout, ensure the session lane is properly released
2. Lane timeout: Add a configurable lane lock timeout (e.g., 5 minutes) — if a run doesn't complete within the timeout, force-release the lane
3. CLI command: Add openclaw session abort <session-key> to allow manual recovery without full gateway restart
4. Diagnostic: Log a warning when a lane has been held for an unusually long time

Workaround

Restart the gateway: openclaw gateway restart or systemctl --user restart openclaw-gateway.service