dsh-adaptive-reasoning

In the DeepSeek Harness Web GUI, reasoning effort has always been a single line of static text inside the model selector: two clicks, a dropdown, and when levels are sparse (off, high, max) or provider-specific identifiers are in play, the experience feels clunky. dsh-adaptive-reasoning turns that row into a true stepless slider. Drag anywhere and reasoningEffort is written live with a 16 ms throttle so requests never pile up mid-drag; release, blur, or finish keyboard navigation and the value snaps to the nearest declared level, issuing exactly one confirming write.

The panel leans into a pixel-art aesthetic: 4-px hard borders, segmented energy cells, Press Start 2P / Zpix fonts, and a track that sits as cool ash until the slider "ignites" it through a three-pass WebGL2 simulation (ignition → blur → composite). Each declared level burns its own colour from ash grey up to burning red. Browsers without WebGL2 or with reduced-motion preferences gracefully fall back to static energy cells without losing any functionality.

Perhaps the most practical touch is the one-click Auto-declare. When you open the panel on a model that has not yet declared reasoningEfforts, a hint and a button appear. Clicking it asks the Host to fill in a conservative declaration—first trying an exact models.dev match, then falling back to a per-family heuristic (deepseek/qwen/glm/claude/grok vs. gpt/gemini/mistral/…)—and writes the result into the llm-pi-ai settings. Models that already declare efforts are never touched. The plugin reads and writes exclusively through the normal session.selectModel RPC, leaving the official selector, /model command, and accessibility behaviour completely intact. Install it and you get a richer control; remove it and everything reverts.