better-model-provider

A DeepSeek Harness settings plugin that declares, per model, your OpenAI-compatible provider's reasoning-effort levels (with wire spellings) and request modalities (text / image) — no hand-editing settings.yaml.

English | 中文

Why you may need it

The official Models page manages providers and model rows, but two per-model capability fields stayed YAML-only: reasoningEfforts (which reasoning levels the model accepts, plus each level's wire spelling) and input (which request modalities the model admits). Until they are declared:

• the composer's model picker shows no reasoning-effort control for your model;
• the host refuses image sessions for it (Model ... does not accept image input), because a hand-declared model defaults to text-only.

There are a few cousin plugins for the reasoning half (dsh-effort-tweak, dsh-model-thinking, dsh-effort-config) — all separate pages with a hardcoded 7-level list. This one puts capability editing inside each model row, reads the vocabulary from the adapter's own schema, and writes catalog models through modelOverrides so the built-in model list is never touched. For the vision half (input modalities) this is currently the only plugin that exists.

What it does

• Lists every configured llm-pi-ai provider (hand-declared or catalog route) and its models.
• Per model row:
  • Reasoning effort — provider default, non-reasoning (false), or a custom level list with per-level wire spellings. Checking a level seeds the spelling with the level id; a blank off spelling stores null ("supported, send nothing").
  • Input modalities — provider default or a custom modality list. Checking image is literally what makes a vision model usable: image sessions stop refusing to switch to it, and images can be sent.
• Catalog routes write modelOverrides. — the built-in catalog stays intact, and clearing the capabilities lifts the override whole.
• "Add provider" enables a catalog route (empty profile) in one click; removing a provider is a two-step confirm.
• Writes are minimal settings.mutate path ops with a compare-and-set revision; an invalid declaration fails on its own row before anything is written.

Install

# From GitHub:
dsh plugin --profile web add github:sanshanya/better-model-provider

# Local checkout:
dsh plugin --profile web add link:

Restart dsh web, hard-refresh the browser, and the settings sidebar gains Model capabilities. Uninstall:

dsh plugin --profile web rm better-model-provider

Use

1. Configure the provider on the official Models page first (that's where the API key lives; this page never touches credentials).
2. Open Settings → Model capabilities.
3. Expand a model row: pick Custom under reasoning effort and check the levels; for vision models check image under input modalities.
4. Apply. The picker immediately offers exactly the declared levels, and the host's image admission admits the model.

Small note: declaring { high } alone gives the picker only High; also check off (blank wire → null) if you want Off available.

How it is built (and why it stays maintainable)

Decision	Reason
Zero runtime harness deps	The bundle takes React from the shell and inlines everything else. The dsh family isn't fully on npm yet (@deepseek-ai/dsh-type-meta is missing), so a runtime dep would break installs today
Vocabulary read from schema	Levels and modalities come from the serialized llm-pi-ai schema — an adapter upgrade changes the page with no plugin release
Catalog routes write modelOverrides	The smallest possible declaration for a catalog model; nothing in its catalog gets rewritten
Structurally open edits	Hand-written YAML keys the editor doesn't show survive every write

The stable boundary is the host's wire contract (settings.describe/mutate, llm.providers/models, plus the settings/document-updated / llm/adapters-updated events). The plugin is browser-side only.

Development

pnpm install
pnpm build           # lib/index.js + lib/client.js (module-loader bundle)
pnpm test            # unit + interaction flows + bundle smoke
pnpm test:coverage   # v8, 100% per-file gates
pnpm lint; pnpm typecheck

License

MIT