Bug type

Regression (worked before, now fails)

Summary

OpenClaw hallucinates fake inbound "User Messages" (with full System: [...] headers) that predict/simulate user intent, contaminating the chat stream.

Steps to reproduce

1. Start a conversation with OpenClaw in Feishu (Lark) DM.
2. Engage in a multi-turn task (e.g., asking the agent to analyze data or write a script).
3. Wait for the agent to propose a next step (e.g., "Should I run the analysis?").
4. Do not reply immediately. Observe the chat stream.
5. Watch for a new message block appearing in the conversation history that looks exactly like a user reply (starting with System: [Timestamp] Feishu[...] DM from <user_id>:), but was not sent by the user.

Expected behavior

• The agent should wait indefinitely for actual user input.
• The conversation history should only contain messages genuinely sent by the user or the agent.
• No predictive or simulated user turns should ever be output to the chat stream or treated as real context.

Actual behavior

• The agent outputs a message block that mimics a user's affirmative reply (e.g., "Yes, go ahead and analyze it") before the user has typed anything.
• This "phantom message" includes authoritative system headers (System: [Timestamp]...), making it look like a real inbound event.
• The agent sometimes proceeds to execute the task based on this hallucinated permission, effectively bypassing user confirmation.

OpenClaw version

2026.2.15 (3fe22ea)

Operating system

macOS 26.3

Install method

npm global

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response