better-model-provider

Better custom-provider model onboarding for DeepSeek Harness (dsh), with per-model declaration of native vision input, reasoning-effort levels, and token capacities (context window / max output).

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). This page edits those two plus the per-model token capacities (contextWindow / maxTokens) — on the models you declare — so one row fully configures one model. Until the capability fields are declared:

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

Install

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

# Local checkout:
dsh plugin --profile web add link:<absolute path to this repo>

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; fill the context window / max output tokens when the provider default is wrong. Provider and model lifecycle stays on the official Models page.
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 Off should be available.

Development

See CONTRIBUTING.md for the gate list (build / test / coverage / lint / typecheck / verify:pack) and the two real-harness lanes; ci runs the hermetic gates per push and live runs nightly.

Compatibility

Verified against a DeepSeek Harness checkout at master commit 47f943859b (2026-08-13; the published contract line is 0.1.0-rc.6, whose real npm ancestry is rc.2 → rc.3 → rc.6 — there never was an rc.5). Minimum-known-good: the contract that commit serves — settings.describe/mutate, llm.providers, the { rpcId, result } envelope, and the model row schema. Internal-only surfaces (settings.section registration shape) are treated as experimental compatibility: a newer harness that doesn't serve them degrades silently instead of breaking the page.

License

MIT