openclaw - 💡(How to fix) Fix Feishu: Agent-generated replies silently dropped (replies=0, queuedFinal=false) after WebSocket reconnect

Description

After a Feishu WebSocket reconnection, agent-generated text replies are silently dropped — the gateway logs dispatch complete (queuedFinal=false, replies=0) even though the agent session transcript shows a reply was successfully generated. System commands (/tasks, /stop, /model) and delivery-mirror messages still deliver correctly.

This issue started on 2026-05-27 at ~10:03 GMT+8 after a Feishu WebSocket disconnect/reconnect event and persists across gateway restarts and version upgrades.

Environment

• OpenClaw version: 2026.5.26
• Feishu plugin version: @openclaw/[email protected]
• OS: macOS (Darwin 25.5.0, arm64)
• Node: v22.22.1
• Affected agents: doctorc, unitzero (all feishu accounts)
• Config: renderMode=auto, blockStreaming=false (previously raw/true, changed during debugging)

Steps to Reproduce

1. Send a message to a Feishu group that triggers a long-running agent task (>30s, involving tool calls)
2. Observe the gateway log

Expected Behavior

Agent reply is delivered to the Feishu group.

Actual Behavior

• Gateway log shows dispatch complete (queuedFinal=false, replies=0)
• The agent session transcript confirms a reply was generated with stopReason=stop
• The reply is never delivered to Feishu

Key Log Evidence

Case 1: Long task reply dropped (May 28, 08:34)

08:34:17 feishu[unitzero]: message in group: 开始麒龙城市花园审计工作
08:34:18 feishu[unitzero]: dispatching to agent
08:36:24 feishu[unitzero]: dispatch complete (queuedFinal=false, replies=0)  ← DROPPED

Session transcript confirms reply was generated:

role=assistant stop=stop model=mimo-v2.5-pro provider=xiaomi-coding
text=杨老师，麒龙城市花园审计报告及全套底稿已生成...

Case 2: Short command succeeds immediately after (May 28, 08:36)

08:36:57 feishu[unitzero]: message in group: /tasks
08:36:58 feishu[unitzero]: dispatch complete (queuedFinal=true, replies=1)  ← DELIVERED

Case 3: Blocked message delivered after timeout eviction (May 28, 08:37-08:42)

08:37:15 feishu[unitzero]: message in group: 麒龙城市花园审计工作完成了吗
08:42:15 feishu[unitzero]: per-chat task exceeded 300000ms cap; evicting from queue
08:42:35 feishu[unitzero]: dispatch complete (queuedFinal=true, replies=1)  ← DELIVERED after eviction

Pattern

• Short/fast responses → queuedFinal=true, replies=1 → delivered
• Long-running tasks (>2min with tool calls) → queuedFinal=false, replies=0 → dropped
• Messages that get queued behind a long task and hit the 300s timeout eviction → queuedFinal=true, replies=1 → delivered

Trigger Event (May 27, 10:03)

10:03:12 [error]: code: 230013, msg: 'Bot has NO availability to this user.'
10:03:21 [info]: [ 'client ready' ]  ← Feishu WebSocket reconnect
10:03:22 [info]: [ 'client ready' ]
10:03:56 dispatch complete (queuedFinal=false, replies=0)  ← First occurrence

Additional Context

• /tasks, /stop, /model commands always deliver (queuedFinal=true, replies=1)
• Gateway restart does not fix the issue
• delivery-mirror system messages (e.g., "新会话" notifications) always deliver
• Manual sends via message(action=send) always work
• The bug affects both renderMode=raw + blockStreaming=true and renderMode=auto + blockStreaming=false

Suspected Root Cause

The reply delivery pipeline appears to treat long-running agent responses differently from short/fast responses. After a Feishu WebSocket reconnection, the mechanism that queues agent-generated text replies into the final delivery path breaks — queuedFinal stays false and replies stays 0 even though the reply was generated. The "visible replies" split (system commands vs agent text) may be involved.