dsh-llm-ai

Configuration reference

Cordis composition entry:

- id: llm-ai
  name: '@deepseek-ai/dsh-llm-ai'
  # Omitting config mounts dormant (zero routes); a settings section can
  # activate routes at any time.
  config:
    catalogUrl: https://models.dev/api.json      # optional: self-hosted mirror
    catalogCachePath: /path/to/cache.json        # optional: cache location
    providers:
      openai:                    # Catalog route: endpoint, protocol, and models
        apiKeyEnv: OPENAI_API_KEY    # all inherited from models.dev.
      zai-coding-cn:             # Hand-declared route: api + baseURL + a
        apiKeyEnv: ZAI_CODING_CN_API_KEY   # non-empty models list are required.
        api: openai-completions
        baseURL: https://open.bigmodel.cn/api/coding/paas/v4
        models:
          - { id: glm-5.3, contextWindow: 1000000, maxTokens: 131072 }

Provider profile fields: apiKeyEnv (credential reference), displayName, api (openai-completions only in v1), baseURL, models (replaces the route's catalog; unset fields default from the registry entry of the same id), modelOverrides (reshapes individual models while the rest of the catalog keeps serving), compat (the three switches), reasoning (the deployment default level), retryPolicy (omission = normal mode with five retries), headers, defaultContextWindow / defaultMaxTokens / defaultInput (fallbacks for configured entries that state no capacity), streamIdleTimeoutMs (five-minute default), maxRequestImageBytes (20 MiB default).

Catalog resolution notes: a models.dev model with no context window is refused rather than guessed; a modelOverrides key naming a model the catalog does not describe is refused; timeoutMs is gone (it named pi-ai runtime behavior) and configuring it fails with migration directions.