dsh-infron-provider

Infron models in DeepSeek Harness.

Adds one infron provider route to DeepSeek Harness: the gateway's OpenAI-compatible streaming API, its live model catalog, and a Models-page card for the API key and the models you want to expose.

Install

dsh plugin --profile web add github:gintasz/dsh-infron-provider

No npm package is needed. The plugin declares no runtime dependencies: the harness packages it extends resolve from the running installation. An npm copy, if you ever want the in-app Plugin Hub listing, installs with dsh plugin --profile web add dsh-infron-provider.

Restart the app, then open Settings → Models → Infron, paste the key, and press Save. Leave the model list on All models from the gateway, or select a subset. The composer then offers an Infron group.

Configure

Every field is live; edits apply to the next request without a restart.

Field	Default	Meaning
apiKeyEnv	INFRON_API_KEY	Credential reference resolved through the credential seam
baseURL	https://llm.onerouter.pro/v1	OpenAI-compatible root; INFRON_BASE_URL overrides it per launch
models	[]	Empty serves the live catalog; a list serves exactly those entries
reasoningEffort	medium	Route default; off sends no reasoning hint
catalogRefreshMinutes	360	Background catalog refresh; 0 disables it
streamIdleTimeoutMs	300000	Longest tolerated gap between stream events
maxTokensField	max_tokens	Switch to max_completion_tokens for strict gateways
retryPolicy	normal, 5 retries	Passed to the harness retry executor

Models that report no capacity fall back to defaultContextWindow (131072) and defaultMaxTokens (32768).

The card writes through the settings service, so the profile patch stays the source of truth. The key is separate: it is stored write-only under apiKeyEnv and never enters cordis.patch.yml. An exported INFRON_API_KEY works too.

What it does

• Streams text and reasoning deltas, and validates streamed tool_calls before completion.
• Sends images to vision models; text-only models get the harness placeholder text.
• Reports prompt, completion, and cached-prompt tokens per step.
• Aborts with the turn and bounds a stalled stream with an idle watchdog.
• Fetches GET /v1/models at mount, on a configurable interval, and on demand from the card. Only streamable LLM rows with an openai endpoint are advertised.
• Surfaces effort as the gateway hint reasoning: { effort }.

Limits

• OpenAI Chat Completions only; routes that publish just an Anthropic or Responses endpoint are not served.
• No provider replay state: assistant content is sent provider-neutrally.
• Tool-result images become a following user message, because a tool message only carries text.
• The Models card keys on the loader entry id llm-infron; renaming that entry hides the card.

Develop

Tests import the harness packages this plugin extends, so point the checkout at a DSH installation first:

mkdir -p node_modules
ln -sfn /path/to/dsh/node_modules/@deepseek-ai node_modules/@deepseek-ai
node --test test/

License

MIT. See LICENSE.