dsh-adaptive-reasoning

dsh-adaptive-reasoning

English · 中文

A capability-aware reasoning-effort slider for the DeepSeek Harness Web GUI.

The plugin keeps DSH's official model selector and /model command intact. Clicking the selector's existing Reasoning effort / 推理等级 row opens an accessible slider panel. The panel reads the selected model's exact reasoning metadata from the Host and writes selections back through the normal session.selectModel RPC.

Install (two ways)

From npm:

dsh plugin --profile web add dsh-adaptive-reasoning

From GitHub (pin to a released tag for reproducibility):

dsh plugin --profile web add github:zhuifengqug/dsh-adaptive-reasoning#v2.0.0

Restart dsh web afterwards. Details, removal, and the optional models.dev enrichment are covered below.

What It Does

• Uses only the current model's advertised reasoning.efforts; it does not invent a global effort list.
• Supports sparse and provider-specific levels such as off, high, max, or an adapter-defined identifier.
• Stepless drag: continuous 0–100 dragging writes reasoningEffort live with a 16ms throttle, so requests never pile up during a drag.
• Snap on release: releasing, blurring, or ending keyboard navigation snaps to the nearest declared level and issues one confirming write.
• WebGL fire: a three-pass WebGL2 simulation (ignition → blur → composite) whose leading edge follows the slider; the track sits as cool ash until a level is dragged over and "ignites".
• Per-level ember palette: each effort level burns its own colour from ash grey up to burning red.
• Pixel-art UI: 4px-pixelated fire, jagged hard borders, segmented energy cells, and Press Start 2P / Zpix pixel fonts (with system-font fallback).
• Falls back to static energy cells when WebGL2 is unavailable or reduced motion is preferred.
• Uses DSH semantic design tokens and respects prefers-reduced-motion.
• Leaves official model selection, load failures, accessibility behavior, and the /model command unchanged.

One-Click Auto-Declare

When you open the effort panel on a model that declares no reasoningEfforts, the panel shows a hint and an 自动补全 (Auto-declare) button. Clicking it asks the Host to fill in a conservative declaration for every model of the current provider that lacks one, then reloads the panel so the slider appears.

Source priority follows mode C: a confident models.dev match wins when the catalog is reachable; otherwise a per-family heuristic decides (deepseek/qwen/glm/claude/grok → off/low/high/max; gpt/gemini/minimax/step/kimi/moonshot/mimo/mistral/… → off/low/medium/high; image models are skipped). Models that already declare efforts are never touched.

The button is an explicit per-click user action, so it is not gated behind a config flag; what it writes is the same reasoningEfforts field in llm-pi-ai settings (persisted to settings.yaml) that you would type by hand. It is dispatched as the host command /adaptive-reasoning-declare <provider>, which you can also run directly to backfill a whole provider.

Optional models.dev Enrichment

Some manually configured llm-pi-ai models do not declare reasoningEfforts, so DSH correctly exposes no effort UI. This plugin can optionally enrich those missing declarations from models.dev at startup.

This behavior is disabled by default because a catalog record cannot prove that a private gateway accepts the same reasoning protocol.

When enabled, the Host half:

• fetches https://models.dev/api.json and caches it in the system temporary directory;
• only updates models with no existing reasoningEfforts field;
• accepts only exact or uniquely normalized model-id matches;
• never supplies a family-heuristic fallback for unknown or ambiguous models (that fallback exists only behind the explicit Auto-declare button above);
• writes inferred entries to the llm-pi-ai settings section, where DSH validates them before use.

Enable it in the profile's cordis.patch.yml by overriding this entry:

- replace:
    - id: adaptive-reasoning
      name: dsh-adaptive-reasoning
      config:
        enrichFromModelsDev: true
        cacheHours: 24

cacheHours must be a positive number. An unavailable catalog only disables enrichment for that run; it does not prevent DSH from starting.

Install

The package is plain JavaScript with no prepare script, so a git install needs no pnpm build permission. Install it straight from GitHub:

dsh plugin --profile web add github:zhuifengqug/dsh-adaptive-reasoning

# Pin to a released version for reproducibility (recommended):
dsh plugin --profile web add github:zhuifengqug/dsh-adaptive-reasoning#v2.0.0

# Verify that the bundle patch participates in the Web profile.
dsh --profile web --dump-config

You can also install from a local checkout instead — run dsh plugin --profile web add ./dsh-adaptive-reasoning from the directory that contains it.

Restart the existing dsh web process after installation. A client-plugin update hot-reloads only while the DSH checkout's pnpm run dev:web watcher is already rebuilding its bundles; this standalone bundle otherwise loads at the next Web process start.

To remove it:

dsh plugin --profile web remove dsh-adaptive-reasoning

Design Notes

The Host is authoritative. The UI does not interpret an effort's wire spelling, map off to another value, or silently clamp an unsupported selection. The only values shown are those published by session.models for the exact selected route, and the Host decides whether a submitted value becomes the session's next request configuration.

The browser integration deliberately uses the official selector's effort row as its entry point rather than taking the single conversation.input.model seat. This avoids replacing DSH's official selector and preserves its model catalog, retry, locale, and error surfaces.

Troubleshooting

Custom relays reject requests with role: developer

pi-ai sends the system prompt with the OpenAI developer role whenever a model has reasoning enabled and the endpoint is judged to support that role. For unknown/custom relay URLs that auto-detection defaults to supported, and many OpenAI-compatible relays only accept system — the request then fails with something like developer is not one of ['system', 'assistant', 'user', 'tool', 'function'].

The llm-pi-ai configuration seam does not expose pi-ai's supportsDeveloperRole switch (as of 0.1.0-rc.7), so it must be patched into the installed bundle. If your deployment hits this, apply the same surgical patch to <dsh-install>/node_modules/@deepseek-ai/dsh-llm-pi-ai/lib/index.js (back up the file first, and re-apply after any dsh upgrade), then set per route:

compat:
  supportsDeveloperRole: false

A relay that does accept the developer role can leave the flag at its default or set it to true.

Files

dsh-adaptive-reasoning/
├── index.js            # Host half and optional models.dev enrichment
├── lib/client.js       # Browser half and slider panel
├── cordis.patch.yml    # Web profile bundle patch
├── package.json
└── README.md

License

MIT