@kshlm/dsh-neuralwatt-provider

Neuralwatt LLM provider plugin for DeepSeek Harness.

It dynamically populates models from Neuralwatt's GET /v1/models endpoint and sets each model's capabilities from the endpoint metadata: input modalities (text or image), context window, max output tokens, and reasoning efforts. It speaks Neuralwatt's OpenAI-compatible /chat/completions API.

Install

dsh plugin --profile <profile> add @kshlm/dsh-neuralwatt-provider

The package declares a dsh.bundle layer (cordis.patch.yml) that inserts the llm-neuralwatt plugin into the profile automatically, so no manual patch row is needed. To pass configuration, restate the row with a config block in your profile's cordis.patch.yml:

- id: llm-neuralwatt
  name: "@kshlm/dsh-neuralwatt-provider"
  config:
    baseURL: https://api.neuralwatt.com/v1

Then remove any hand-written neuralwatt block under llm-pi-ai.providers in your profile's cordis.patch.yml (the two routes would otherwise collide).

Configuration

Configure under the plugin (llm-neuralwatt) in cordis.patch.yml:

- id: llm-neuralwatt
  name: "@kshlm/dsh-neuralwatt-provider"
  config:
    baseURL: https://api.neuralwatt.com/v1   # optional; override here
    apiKeyEnv: NEURALWATT_API_KEY            # optional; env var *name*
    route: neuralwatt                        # optional
    displayName: Neuralwatt                  # optional

Field	Default	Env fallback	Meaning
route	neuralwatt	None	Provider route key / selector id.
displayName	Neuralwatt	None	Provider display name.
baseURL	https://api.neuralwatt.com/v1	NEURALWATT_BASE_URL	Endpoint root (overridable).
apiKeyEnv	NEURALWATT_API_KEY	None	Env-var name holding the API key.

Precedence: config > env > built-in default. Set the API key by exporting NEURALWATT_API_KEY (or whatever apiKeyEnv names) or storing it via the credentials service.

Behavior

• Models are fetched once at load (boot snapshot) from GET {baseURL}/models.
• All models are exposed; only entries marked deprecated are skipped.
• Vision models advertise image input; text-only models advertise text.
• Reasoning efforts are mapped from Neuralwatt's accepted_efforts, folding effort_aliases onto canonical ids and normalizing none to off.
• Context and max-output capacities come from the endpoint; when it omits them the defaults are 262144 context and 32768 output tokens.
• Every request carries attributionHeaders() and forwards the caller's abort signal.

License

MIT