Bug type

Behavior bug (incorrect output/state without crash)

Summary

OpenClaw may leak internal tool-call/commentary payload into user-visible Feishu DM messages

• OpenClaw version: 2026.3.8
• Channel: Feishu
• Surface: DM
• Streaming: enabled
• Feishu config included blockStreaming: true

Steps to reproduce

1. Use OpenClaw in a Feishu DM conversation
2. Ask the assistant to perform a multi-step task involving multiple tool lookups/writes
3. In this case, the task was to record an expense and set its reimbursement status to "unreimbursed"
4. The assistant performed several Feishu Bitable / Wiki lookup calls before writing the final record
5. Observe the DM output

Expected behavior

Only user-facing natural-language messages should appear in the DM.

Internal orchestration data such as tool payloads, tool routing metadata, or commentary-only structures should never be visible.

Actual behavior

The user received raw internal JSON blobs in the Feishu DM, for example:

{"tool_uses":[{"recipient_name":"functions.feishu_bitable_app","parameters":{"action":"list","page_size":50}},{"recipient_name":"functions.feishu_search_doc_wiki","parameters":{"action":"search","query":"XXXX","filter":{"doc_types":["BITABLE"],"sort_type":"EDIT_TIME"},"page_size":10}}]}

and similar payloads containing:

• tool_uses
• recipient_name
• functions.*
• raw tool parameters

OpenClaw version

2026.3.8

Operating system

Ubuntu 22.04

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

What I found

I checked the session transcript and the leaked payload was already recorded as assistant text content with phase: "commentary".
That suggests the problem is not Feishu rendering a real tool call incorrectly.
Instead, it looks like:

1. internal orchestration payload was emitted as assistant-visible commentary text
2. outbound message delivery did not suppress/filter it
3. the text was then sent to Feishu as a normal visible message
   So this appears to be an OpenClaw-side bug in the agent-output / messaging path.

Impact

• exposes internal execution metadata to end users
• creates confusing / noisy chat output
• makes the assistant feel unstable or "leaky"
• especially bad on messaging surfaces where every emitted text becomes part of the permanent conversation

Suggested fix

1. Prevent internal multi-tool payloads from being emitted as assistant-visible commentary text
2. Add an outbound safeguard to suppress messages that look like internal orchestration payloads
3. Restrict user-visible commentary to natural-language progress updates only

Classification

OpenClaw bug, not a Feishu platform bug.