openclaw - 💡(How to fix) Fix [Bug]: Ollama thinking modes intermittently fall back to hardcoded defaults [1 comments, 2 participants]

TL;DR

The issue can be mitigated by ensuring that the per-model thinking mode data is properly cached and not invalidated by subsequent requests, such as /status calls.

Guidance

• Verify that the caching mechanism for per-model thinking modes is correctly implemented and not being bypassed or invalidated by certain code paths, such as model re-resolve or /status calls.
• Investigate the specific conditions under which the regression occurs, focusing on the timing and sequence of requests, to identify potential triggers for the cache invalidation.
• Consider adding logging or debugging statements to track the cache state and per-model thinking mode data throughout the session to gain more insight into the issue.
• Review the code handling /status calls and model re-resolve paths to ensure they do not inadvertently cause the cache to fall back to default values.

Example

No specific code snippet can be provided without more context, but ensuring that cache invalidation is properly handled might involve checking cache update logic, especially around points where the default values are used.

Notes

The solution may involve adjusting the caching strategy or ensuring that all relevant code paths respect and update the cache correctly. The issue's intermittent nature suggests that timing and the sequence of operations play a significant role.

Recommendation

Apply a workaround to ensure that the per-model thinking mode data is persisted and not overwritten by default values, possibly by reviewing and adjusting the caching logic around /status calls and model re-resolve paths. This is recommended because the root cause seems related to cache handling rather than a version-specific issue that could be resolved by an upgrade.