openclaw - 💡(How to fix) Fix Possible regression: less explicit prompt/tooling guidance may increase wrong media-send tool choice in v2026.4.5 [1 participants]

Using MiniMax M2.7, an agent that should have returned media through normal reply-time media delivery instead chose an exec/CLI-style path.

TL;DR

The issue can be addressed by revising the prompt assembly around reply generation to make the distinction between returning media in the current conversation and invoking a tool/command more explicit.

Guidance

• Review the changes made to the prompt assembly around reply generation in the newer versions to identify areas where the distinction between media return and tool invocation can be clarified.
• Analyze the structured tool definitions and inference mechanisms to ensure they are correctly handling the media delivery paths.
• Consider adding more explicit human-readable tool-purpose guidance to the newer versions to mitigate the issue of weaker prompt redundancy.
• Test the media delivery functionality with different prompts and tool choices to verify the effectiveness of the changes.

Example

No code snippet is provided as the issue does not contain specific code references.

Notes

The issue seems to be related to the changes in the prompt assembly and tool definitions, and addressing these changes may require a thorough review of the code and testing.

Recommendation

Apply workaround: revise the prompt assembly and add more explicit guidance to mitigate the issue, as the root cause is related to the changes in the newer versions and not a direct breakage of the MEDIA: functionality.