dsh-llm-aimux

 

DeepSeek Harness (dsh) LLM adapter backed by aimux — one Rust engine exposing 325+ AI providers as dsh model routes.

> 中文简介：本插件把 aimux（Rust 统一 LLM 访问层，325+ 提供商）接入 DeepSeek Harness——每个 aimux provider 都可以配置成一条 dsh 模型路由。

Install

dsh plugin --profile web add @arcships/dsh-llm-aimux

Then store a credential (or export the environment variable) and pick the model:

dsh --profile web
# Models page → provider "deepseek" → paste your DEEPSEEK_API_KEY

Headless without the credentials store works too — aimux reads the provider's conventional env var (DEEPSEEK_API_KEY, GROQ_API_KEY, …) directly.

Configuration

Routes are keyed by dsh provider id; each maps to one aimux provider. The composition entry below is only the base layer — users add or edit routes through their settings (web Models page → provider card, or the llm-aimux: section in settings.yaml) and changes apply without a restart:

- insert:
    - id: llm-aimux
      name: '@arcships/dsh-llm-aimux'
      config:
        providers:
          deepseek:                    # dsh route id
            provider: deepseek         # aimux provider name (default: the route id)
            apiKeyEnv: DEEPSEEK_API_KEY # credential ref (default: <PROVIDER>_API_KEY)
          kimi:
            provider: moonshot         # credential ref derives as MOONSHOT_API_KEY
          groq:
            displayName: Groq

The same shape works in the user settings layer:

llm-aimux:
  streamIdleTimeoutMs: 300000     # chunk-idle watchdog relayed to aimux
  providers:
    groq: {}
    relay:                        # endpoint override through a gateway/relay
      provider: deepseek
      baseUrl: https://relay.example/v1
      headers:
        X-Org: acme

• provider — an aimux provider name. aimux ships 325+ (251 of them

OpenAI-compatible registry entries); see the aimux provider docs.

• Every route appears as a configurable provider on the Models page

(ctx.llm.registerConfigurableProviders, addressed at llm-aimux.providers.<route>); added or removed routes swap atomically.

• Credentials resolve through the harness ctx.credentials seam when mounted;

otherwise aimux falls back to the provider's own env var. Resolution is per-request, so a rotated key reaches the next call.

Status & scope

This is a young experiment tracking a developer-preview harness — expect breaking changes on both sides (dsh rc bumps, aimux pre-1.0).

Currently supported:

• Streaming chat with tool calls (raw JSON tool arguments, per the adapter contract)
• Reasoning blocks in and out
• Disjoint token usage accounting (cache read/write split out)
• Attribution headers on every outbound provider request
• Settings integration: llm-aimux user-settings section, per-route Models

page cards, and hot route swaps (registration.replace) on section changes

• Model metadata merging (aimux RFC-0027 host role): runtime /models

discovery + the community catalogue (getModelSpecs), feeding Models-page cards, resolveModel context windows / output caps / reasoning efforts, and endpoint interrogation for stored routes

• Per-route endpoint overrides (baseUrl, headers) for relays and gateways
• Streaming chunk-idle watchdog (streamIdleTimeoutMs, default 300 s)

Retry policy: aimux always sends exactly one attempt (max_retries: 0) — the dsh agent layer owns the retry budget, matching llm-pi-ai's treatment.

Not yet supported:

• Image input (dsh attachment references are not resolved to bytes yet) — fails

closed with UNSUPPORTED

• Subscription/OAuth channels (aimux RFC-0018 keeps those integrator-side)
• Endpoint interrogation for draft (not-yet-stored) routes — arbitrary

baseURL drafts cannot go through the aimux registry and answer with nothing

Route conflicts: routes owned here cannot overlap with the built-in llm-pi-ai routes (dsh rejects duplicate adapters). If you configure both, give the aimux entry a distinct route id.

Development

npm install
npm run typecheck
npm test

The adapter is a thin shell over src/translate.ts (pure, synchronous, fully unit-tested without the native binding). @arcships/aimux ships prebuilt platform binaries via npm optional dependencies — no build scripts, so dsh plugin add works without pnpm allowBuilds flags.

License

[MIT](LICENSE)