Skip to content

dsh-model-info

Verified

@mzwing/dsh-model-info Β· v0.2.1 Β· MIT

Lists third-party models in DeepSeek Harness from their endpoints and completes their metadata from the pi.dev and models.dev catalogs

Install

dsh plugin add @mzwing/dsh-model-info

Confirm the layer applied with dsh --profile default --dump-config β€” see the install guide.

Source

Tags

Creators

Readme

@mzwing/dsh-model-info

npm

A DeepSeek Harness plugin that lists the models of your llm-pi-ai routes from their endpoints and completes their context window, input modalities and reasoning efforts from the pi.dev and models.dev catalogs, where DSH would otherwise fall back to the route's defaults. It fills the model entries in memory each time llm-pi-ai resolves its config, so nothing is written to your files.

Install

dsh plugin --profile web add @mzwing/dsh-model-info

Usage

Nothing happens until a route is opted in, in ~/.dsh/cordis.patch.yml or the profile's own cordis.patch.yml:

- id: model-info
  config:
    providers:
      my-relay: {}

Only routes that list their models are completed. A model is matched by its alias, then its exact id, then its id with leading / segments dropped one at a time, then its id with at most one prefix and one suffix rule removed. An ambiguous match changes nothing, and fields you wrote on a model entry always win. maxTokens is never taken from a catalog, because llm-pi-ai reserves a configured maxTokens on every request.

A models entry whose id contains * stands for every id the route's endpoint lists that it matches, with * matching any run of characters, so this route serves the endpoint's whole list:

- id: llm-pi-ai
  config:
    providers:
      my-relay:
        models:
          - id: '*'

Each model a pattern brings in carries the fields written on the pattern, and entries you write out keep their place and fields. The list comes from llm-pi-ai's own model discovery, so the route's API key and headers apply, and it is cached and fetched again in the background. Until the first list arrives, llm-pi-ai serves the pattern entry as written.

/model-info summarises what was listed and completed, /model-info <route>/<model> shows where each field came from, and /model-info refresh re-checks the catalogs and the endpoints.

Configuration

Per route, under providers:

Key Default Description
catalogProvider β€” Scope lookups to one catalog provider, e.g. openrouter
contextWindowPolicy catalog min never raises the route's defaultContextWindow, keep leaves it alone
capabilityPolicy catalog widen only adds capabilities, keep leaves them alone
useCatalogName false Rename models to their catalog names
mapThinkingLevels false Derive reasoning efforts from models.dev
models β€” Per-model alias, override, skip, prefixes and suffixes

An override takes the same fields as an llm-pi-ai model entry: name, contextWindow, maxTokens, input and reasoningEfforts. A models key containing * covers every model it matches, so one entry reaches a model under all of a relay's routing prefixes; an exact key wins, then the longest matching pattern.

Rules strip an affix before matching. -free and :free are built in:

rules:
  - id: thinking
    kind: suffix
    value: -thinking
    override:
      contextWindow: 250000

See the example config and src/config.ts for every option.

License

MIT