dsh-llm-autorouter
适用于 DeepSeek Harness 的 AutoRouter（模型路由网关 / 中继）LLM 适配器
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:autorouter0-ai/dsh-llm-autorouter#d5c8bbb733d5ae2a66a5564bc17459771b5a6049Llm Autorouter 以 dsh-llm-autorouter 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
DeepSeek Harness LLM adapter for AutoRouter — a model routing gateway that unifies access to multiple upstream channels. AutoRouter supports mainstream API formats (OpenAI-compatible, Anthropic, Gemini, and others) and handles routing, channel failover, billing, and observability inside the gateway.
This plugin mounts on the Harness so every model call goes through your AutoRouter deployment instead of hitting providers directly. It registers one provider route (autorouter by default) on ctx.llm; first fetch and import the models exposed by the gateway, then use those imported model ids in your agent config.
Tested with DeepSeek Harness 0.1.0-rc.5. Harness is still Developer Preview — later versions may break compatibility.
Choose the installation method that matches how you run DeepSeek Harness.
If you use DeepSeek Harness directly from npm:
npx @deepseek-ai/dsh plugin --profile web add dsh-llm-autorouter
If you want to install the plugin from a git repository:
npx @deepseek-ai/dsh plugin --profile web add github:autorouter0-ai/dsh-llm-autorouter
A git install fetches sources and runs the prepare build. If the first attempt fails, follow the allowBuilds instructions printed by the command.
If you are developing this plugin locally, build the package first and then install it into a Web profile:
pnpm install
pnpm build
pnpm pack # produces dsh-llm-autorouter-0.1.0.tgz
npx @deepseek-ai/dsh plugin --profile web add ./dsh-llm-autorouter-0.1.0.tgz
The bundle has two entries with one installation: the Host entry supplies the LLM adapter, settings namespace, and model discovery; the browser entry supplies the AutoRouter card on the Web Plugins page.
After install, open the AutoRouter card under Plugins. Enter the gateway URL and API key. The key is stored only through the credentials service (AUTOROUTER_API_KEY by default) — never in settings.yaml or the browser.
Use Fetch models and Import selected to populate the models allowlist that chat selectors read from.
Headless profiles load only the Host entry. Set AUTOROUTER_API_KEY in the launching environment, or store the key through the credentials service.
Defaults apply when the bundle is installed with no extra config. Override any field in your profile's cordis.patch.yml or a --patch overlay:
| Field | Default | Meaning |
|---|---|---|
baseURL | $AUTOROUTER_BASE_URL → https://api.autorouter.top | Gateway origin, without a trailing slash. |
apiKeyEnv | AUTOROUTER_API_KEY | Credential reference resolved per request through the credentials service. |
providers | ['autorouter'] | Provider routes this adapter owns. |
thinking | (neutral) | 'enabled' / 'disabled'. Neutral: nothing is sent; upstream keeps its own default. |
reasoningEffort | (neutral) | 'off' / 'low' / 'high' / 'max'. When set, materialized into requests that omit one. |
defaultContextWindow | 128000 | Context capacity when the selected model has no exact value. |
models | [] | Saved chat-model allowlist ({ id, name?, description?, contextWindow?, maxTokens?, inputModalities? }). |
streamIdleTimeoutMs | 300000 | Max idle time while one stream read is outstanding. |
discoveryTimeoutMs | 10000 | Network timeout for model listing. |
retryPolicy | normal defaults | Provider-owned request retry policy. |
Example cordis.patch.yml:
- insert:
- id: llm-autorouter
name: dsh-llm-autorouter
config:
baseURL: https://gateway.example.com
apiKeyEnv: AUTOROUTER_API_KEY
thinking: enabled
reasoningEffort: high
models:
- id: deepseek-v4-flash
contextWindow: 1000000
Environment overrides for local overlays:
AUTOROUTER_BASE_URL — gateway originAUTOROUTER_API_KEY_ENV — credential env var namePoint an agent at the autorouter provider:
- id: agent-loop
name: '@deepseek-ai/dsh-agent-loop'
config:
agents:
- id: main
provider: autorouter
model: deepseek-v4-flash # any model id your gateway routes
Neutral by default: with no thinking / reasoningEffort configured, neither field is sent on the wire and each upstream channel keeps its own default. AutoRouter relays the request body per channel (including DeepSeek-V4 thinking inferred from model-name suffixes).
reasoningEffort: high (etc.) to send thinking + reasoning effort on every request.thinking: disabled to force thinking off on every request.reasoning_content from DeepSeek-family channels is streamed back as reasoning blocks.Chat selectors list only the saved models allowlist — they never probe the gateway on their own. The Web card's Fetch models action queries the gateway with your token and returns up to 1000 candidates grouped by capability. Import selected saves the chosen ids and their advertised modalities / token limits into the allowlist. Re-import after a gateway upgrade to pick up new fields. A failed discovery leaves the existing catalog unchanged.
pnpm install
pnpm typecheck
pnpm build
For a local Host overlay without installing the bundle:
pnpm build
pnpm dev:overlay # writes dev-cordis.local.yml
pnpm dsh web --patch ./dev-cordis.local.yml
Peer packages (@deepseek-ai/*) must be linked from a deepseek-harness checkout:
DSH_HARNESS_DIR=/path/to/deepseek-harness pnpm dev:link-peers
pnpm build && pnpm smoke
MIT. Portions adapted from @deepseek-ai/dsh-llm-deepseek (MIT, deepseek-harness).