DeepInfra provider for DeepSeek Harness

Use DeepInfra models from DeepSeek Harness (dsh). The plugin registers a deepinfra provider route and streams over DeepInfra's OpenAI-compatible chat-completions API.

• Live model catalog. Models come from GET https://api.deepinfra.com/v1/openai/models, so new DeepInfra models appear without a plugin update. Only chat models are listed.
• Real model facts. Each model reports its own context window, image input (vision), and reasoning-effort levels. Compaction timing and the model picker use these values, so nothing has to be typed into settings.
• Tool calls, reasoning, and images. Parallel tool calls, reasoning_content (also sent back on later turns so thinking models can continue a tool loop), and image input on vision models.
• Prompt caching. The DSH session id goes out as prompt_cache_key, so repeated turns reuse their cached prefix.
• No conflict with your own routes. It adds a route beside llm-pi-ai and does not modify it.

Install

Needs Node ≥ 22.15 and a DeepInfra API token (https://deepinfra.com/dash/api_keys).

dsh plugin --profile web add github:deepinfra/dsh-provider-deepinfra
export DEEPINFRA_API_KEY=...   # or store it from the web Models page
dsh web

Pick a DeepInfra model in the model picker, or make one the default:

# $DSH_HOME/profiles/<profile>/cordis.patch.yml
- id: agent-default-model
  config:
    provider: deepinfra
    model: deepseek-ai/DeepSeek-V4-Flash

Configuration

Field	Default	Meaning
baseURL	https://api.deepinfra.com/v1/openai	OpenAI-compatible API root
apiKeyEnv	DEEPINFRA_API_KEY	Credential reference resolved for each request through the DSH credential service, or from the environment when the composition has none

- id: provider-deepinfra
  config:
    apiKeyEnv: MY_DEEPINFRA_TOKEN

Behavior

• The catalog is cached for 10 minutes. If the listing is unreachable, the last good listing is served, or a small built-in list if there was never one, so the route never comes up empty. A model id missing from the listing (for example, a private deployment) can still be requested; its context window is reported as unknown rather than guessed.
• Without an API key the picker lists no DeepInfra models, because every request would fail.
• Reasoning: models DeepInfra tags reasoning_effort offer Off, Low, Medium, High; models tagged only reasoning offer Off/On (none/high on the wire). The choice goes out as reasoning_effort, and leaving it unset keeps the model's own default. Some models, such as DeepSeek-V4.1-Flash, do not think unless you turn it on.
• HTTP failures map to DSH failure codes: 401 → INVALID_CREDENTIAL, 402 or quota text → QUOTA, 429 → RATE_LIMIT (with Retry-After), context-overflow 400 → CONTEXT_WINDOW_EXCEEDED, 5xx → SERVER. dsh-llm-retry retries the transient ones.
• Images go inline as base64 data URLs, scaled to ~4 MP and 1 MiB each. When one request's images exceed 20 MiB, DSH is asked to offload the oldest ones.

Development

npm install
npm test

The tests include a contract check: every GenerateOptions field declared by the installed @deepseek-ai/dsh-llm must have an entry in REQUEST_FIELD_DISPOSITION (lib/request.js). When DSH adds a request field, the test fails until the field is handled or deliberately ignored.

License

MIT