feat(ai): add OpenAI-compatible endpoint

[jgilman1337]

Summary

Adds an OpenAI-compatible AI provider so JobSync can use a custom base URL and optional API key (LM Studio, llama.cpp, vLLM, and similar local servers).

This continues #79. The original contributor went MIA, so this rebases the work onto dev and includes the review requests from that PR.

Changes

• Register an OpenAI-compatible provider with a configurable base URL and optional API key
• List models from {baseURL}/v1/models and run inference through the OpenAI SDK client
• Verify dual credentials ({ baseURL, apiKey } vs a plain string) without widening other providers to any
• Require an explicit model selection instead of calling the API with an empty model name
• Cover credential resolution, verification, and model listing in unit tests

Related Issues

Follows up #79

Testing

• npx eslint on the changed files — pass
• npx tsc --noEmit — pass
• Added tests for dual-credential verification, openai-compatible-key resolution, getModel, and the models route
• Configure the provider in Settings with a local server URL, optional key, then select a model under AI Provider

Made with Cursor

[jgilman1337]

An additional fix is available in this PR. Calling gemma4-e4b-it using this new provider was failing. OpenAI-compatible chat used the Responses API. Local servers only implement Chat Completions, so Gemma’s stream skipped text-start and the UI aborted.

The OpenAI-compatible provider was calling the Responses API (/v1/responses). Local servers like llama-swap speak Chat Completions (/v1/chat/completions).

createOpenAI()(model) in the AI SDK defaults to Responses. Gemma’s stream then came back as text-delta chunks with msg_… ids and no text-start. The UI stream requires text-start first, so chat failed with:

Received text-delta for missing text part…

1cdbb02 switches that provider to .chat(model) so it uses Chat Completions, which those local servers actually implement.