More model providers and brand icons, pinned models in the picker, mini tool timeout control, and sweeping improvements to context compaction and permissions.
New model providers: added NVIDIA, MiniMax International, Moonshot AI International, Qwen Token Plan International, and completed the full Qwen CN model list.
Model brand icons: added brand icons for Doubao, Hunyuan, NVIDIA, Meta, Mistral, StepFun and more, so you can spot models at a glance.
Pin frequently used models: the model menu now supports pinning models to the top with a slot limit, keeping your go-to models within reach.
Markdown local path links: local file paths and links in messages are now clickable to preview or open; ⌘/Ctrl click reveals the file in your file manager.
Mini tool timeout control: mini tools can declare a timeoutMs so tool execution timeouts are no longer affected by provider timeouts.
Improvements
Smarter context compaction: auto-compaction now works mid-task on reload and resumes automatically; the compaction threshold self-corrects when a model's context window is mislabeled.
Clearer permission prompts: danger notices now state consequences and point to the specific action; parallel commands in Chinese are joined with enumeration commas; long command lines in permission cards prioritize the area around the dangerous command.
Mini tool risk grading: mini tool high-risk actions now use orange; red is reserved only for high-risk Shell commands; MCP is downgraded to medium risk.
Longer login timeout: login timeout extended to 120 seconds.
Rendering performance: reduced timeline rendering overhead and lowered the draw frequency of the running-text shimmer effect.
Mini tool ID strategy change: the mini tool ID strategy has been updated; some already-added mini tools need to be re-logged-in to obtain fresh credentials.
Fixes
Pinned model slot leak: fixed an issue where closing a model still occupied a pinned-model slot.
Auto-compaction failure: fixed auto-compaction no longer working after the context filled up; improved detection wording for context limits so they're no longer misidentified as output-length limits.
Local image display: fixed local images under /tmp and Windows temp directories not displaying.
Model detection false positives: custom model detection now trusts successful model-list endpoints, fixing false positives for large-directory providers like NVIDIA.
OpenAI probe request: fixed the OpenAI Responses API probe request sending max_output_tokens below the minimum.
ToolSearch query loss: fixed the query term being ignored when searching a specific MCP server via ToolSearch.