openclaw - 💡(How to fix) Fix video_generate model override inconsistently ignored — falls back to MiniMax default [1 participants]

• No error or warning is surfaced to indicate the override was ignored

TL;DR

The issue can be mitigated by ensuring that the video_generate call is made with a fresh session or by investigating the session management in the OpenClaw version 2026.4.20 to prevent silent fallbacks to the default model.

Guidance

• Investigate the session management in OpenClaw to understand why subsequent requests in the same session silently fall back to the default model fal/fal-ai/minimax/video-01-live.
• Verify that the model parameter is correctly set for each video_generate call and that there are no caching or session-related issues that could cause the model to be overridden.
• Consider setting up logging or debugging to capture any potential errors or warnings that might indicate why the specified model is being ignored in favor of the default.
• Test the behavior with different models and sessions to see if the issue is specific to the fal/bytedance/seedance-2.0/fast/text-to-video model or if it's a more general problem.

Example

No specific code example can be provided without more context on the OpenClaw API and its usage, but ensuring that each video_generate call is made with the correct model parameter and monitoring the response for any signs of model override is crucial.

Notes

The issue seems to be related to session management and model selection in OpenClaw. The fact that the first request in a session works correctly but subsequent ones do not suggests a potential issue with how sessions are handled or how model selections are cached.

Recommendation

Apply a workaround by ensuring each video_generate call is made with a fresh session or by implementing a check to verify that the correct model is being used before proceeding with the video generation. This approach is chosen because the root cause of the silent fallback is not explicitly stated, and directly addressing the session management or model selection logic might require more in-depth knowledge of the OpenClaw internals.