openclaw - 💡(How to fix) Fix [Bug]: Thinking stream not forwarded via WebSocket agent events during streaming [1 comments, 2 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Summary

When extended thinking is enabled (Think: high, Reasoning: stream, model reasoning: true), the gateway stores thinking content in the transcript (visible via chat.history) but does not emit stream: "thinking" agent events via WebSocket during streaming, making real-time thinking display impossible for WebChat frontends.

Steps to reproduce

1. Configure a custom Anthropic provider with reasoning: true on the model definition
2. Set session-level thinking via /thinking high and /reasoning stream
3. Confirm session status shows Think: high · Reasoning: stream
4. Send a message via WebChat (Control UI)
5. Monitor WebSocket frames for agent events during the response

Expected behavior

During streaming, the gateway should emit agent events with stream: "thinking" containing thinking content chunks, in addition to stream: "assistant" for text content. The chat delta/final events should report hasThinking: true and/or hasThinkingDelta: true when thinking content is present.

Actual behavior

• All agent events during streaming have stream: "assistant" only — no stream: "thinking" events are emitted
• All chat events (both delta and final state) report hasThinking: false and hasThinkingDelta: false
• After the run completes, calling chat.history returns the message with thinking content included
• This confirms the gateway receives and stores thinking content but does not forward it via WebSocket during streaming

OpenClaw version

2026.3.13 (commit 61d171a)

Operating system

Linux 6.18.7-arch1-1 (x64), Node v24.14.0

Install method

npm global

Model

custom-anthropic/claude-opus-4-6

Provider / routing chain

openclaw -> custom-anthropic (gateway.lazycat.ai, api: anthropic-messages) -> anthropic

Additional provider/model setup details

{
  "models": {
    "mode": "merge",
    "providers": {
      "custom-anthropic": {
        "baseUrl": "xxxx",
        "api": "anthropic-messages",
        "models": [
          {
            "id": "claude-opus-4-6",
            "name": "claude-opus-4-6 (Custom Provider)",
            "reasoning": true
          }
        ]
      }
    }
  }
}

Session-level: /thinking high and /reasoning stream both enabled. Session status confirms Think: high · Reasoning: stream.

Logs, screenshots, and evidence

### WebSocket agent events (excerpt — all stream values during a full response)


stream: "lifecycle"  (phase: startedAt)
stream: "assistant"  (dataKeys: [text, delta])  ← repeated for all content chunks
stream: "assistant"  (dataKeys: [text, delta])
stream: "assistant"  (dataKeys: [text, delta])
...
stream: "lifecycle"  (phase: endedAt)


No `stream: "thinking"` event observed across the entire response.

### Chat events (all delta + final)


{
  "event": "chat",
  "state": "delta",
  "hasText": true,
  "hasTextDelta": false,
  "hasThinking": false,
  "hasThinkingDelta": false
}

{
  "event": "chat",
  "state": "final",
  "hasText": true,
  "hasTextDelta": false,
  "hasThinking": false,
  "hasThinkingDelta": false
}

### Evidence that thinking IS stored

After the response completes, `chat.history` returns the message with thinking content present. Historical messages in the WebChat UI display a "thinking" section. This proves the gateway receives thinking from the provider and persists it, but does not stream it in real-time via WebSocket events.

Impact and severity

• Affected: All WebChat / Control UI users who want to display real-time thinking during streaming
• Severity: Medium — does not block core functionality, but prevents a key UX feature (real-time thinking display)
• Frequency: 100% reproducible — observed across multiple messages with thinking enabled
• Consequence: Custom WebChat frontends cannot implement real-time "thinking process" display; thinking content is only accessible after the full response completes via chat.history, defeating the purpose of streaming thinking

Additional information

• The Reasoning: stream session setting implies thinking should be streamed, but the WebSocket events do not reflect this
• The reasoning: true model flag was initially false (auto-defaulted); after manually setting it to true and restarting, thinking content began appearing in chat.history but still not in real-time WebSocket events
• No stream: "thinking" agent event type was observed in any test run