This patch improves DeepSeek integration with new settings, removed limits, and critical tool-call fixes, while ensuring compatibility with OpenRouter and MistralAI.
Detailed list of changes
- Added
ReasoningEffortandTopPsettings for DeepSeek provider, enabling control over thinking mode and nucleus sampling. - Removed the artificial
MaxTokenscap of8192for DeepSeek, allowing values up to the API output limit. - Fixed multi-step tool call requests in DeepSeek by properly structuring assistant and tool messages to avoid
invalid_request_error. - Removed optional
namefield fromrole=toolresult messages in DeepSeek and OpenRouter providers for compatibility.
Auto-generated release notes
What's Changed
- feat(deepseek): release 1.4.4 with DeepSeek settings, model catalog updates, and tool-call fixes by @marc-romu in #671
Full Changelog: 1.4.3...1.4.4