openclaw - 💡(How to fix) Fix Bug: Empty user message gets hallucinated content from DeepSeek, causing unauthorized config mutations [2 comments, 2 participants]

What happened

During a WebChat session, an empty user message (content: "") was presented to the assistant. The assistant (DeepSeek V4 Pro) then acted on specific text ("putz den kleinen ram weg und aktivier tts") that never existed in the actual transcript. This caused the assistant to execute openclaw config set and openclaw sessions cleanup without real user intent.

Evidence from session transcript

Raw JSONL session file shows the message between assistant question and action:

{
  "type": "message",
  "id": "9b277e78",
  "parentId": "5a180f11",
  "timestamp": "2026-04-27T20:07:06.967Z",
  "message": {
    "role": "user",
    "content": [{"type": "text", "text": ""}],
    "timestamp": 1777320426952
  }
}

The text field is empty. Yet the model's thinking traces reference and act on: "putz den kleinen ram weg und aktivier tts" (German: "clean up the small RAM and activate TTS").

Impact

The assistant executed these actions based on hallucinated user input:

1. openclaw sessions cleanup --store ... --enforce --fix-missing — benign in this case
2. openclaw config set messages.tts.enabled true — unauthorized config mutation

The user never saw this "message" in their WebChat UI and confirmed via screenshot that no such message was ever sent.

Suspected root cause

When OpenClaw assembles the conversation context for the DeepSeek provider, an empty user message either:

• a) Gets filled/misassembled during prompt building
• b) Causes the DeepSeek model to hallucinate plausible continuation text in the empty slot

Either way, the model receives non-existent user intent and acts on it — a dangerous failure mode because it leads to real mutations without real user input.

Environment

• OpenClaw: v2026.4.25 (aa36ee6)
• Provider: DeepSeek (deepseek/deepseek-v4-pro)
• Channel: WebChat (Control UI)
• Session ID: 1ac4a847-2dbd-4cdd-b068-86dd4838ef12
• OS: Linux 6.8.0-110-generic (x64), Node v22.22.2

Steps to reproduce (suspected)

1. Use DeepSeek as the model provider in a WebChat session
2. Have an assistant message that ends with a clear yes/no question
3. Somehow trigger an empty user message (possible WebChat edge case?)
4. Observe model filling the empty slot with contextually plausible text and acting on it

Mitigation

Config was reverted immediately. Session transcript preserved as evidence.