LLM Layer
Domain: Models & Providers
Motivation
@deepseek-ai/dsh-llm is the provider-neutral model-call service an agent loop, session-title generator, or compaction summarizer streams through. It keeps one vocabulary across the loop, the session log, and every consumer, while each provider adapter owns its own wire format. The service itself has no configuration and no provider code — you mount it with at least one adapter.
Concrete Example
Mount the service plus an adapter, then name the route in each request:
- name: '@deepseek-ai/dsh-llm'
- name: '@deepseek-ai/dsh-llm-deepseek-api-key'
config: { apiKeyEnv: DEEPSEEK_API_KEY }A request streams with ctx.llm.stream({ provider: 'deepseek-official', model: 'deepseek-flash', messages }) and always ends in one terminal finish chunk.
Analogy
A universal power strip — one socket, any adapter plugs in and you get the same electricity.