Don't send the developer role from the Ollama provider slot

[warmwind]

Fixes #36.

A model added through the Ollama provider slot against a self-hosted OpenAI-compatible endpoint fails on the very first message:

400 {"message":"developer is not one of ['system', 'assistant', 'user', 'tool', 'function'] - 'messages.[0].role'","type":"invalid_request_error"}

pi chooses the system-prompt role with

const useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;

and detectCompat() infers supportsDeveloperRole from the provider id and base URL. ollama is not in its isNonStandard list, and a self-hosted base URL matches none of its hostname checks, so the endpoint is assumed to accept role: "developer". The branch also hardcodes reasoning: true, which supplies the other half of the condition — so the developer role went out on every request.

This slot is the only one in AddModelModal that pairs a user-supplied base URL with the chat-completions API (openai uses openai-responses, google uses google-generative-ai, cloudflare is account-scoped), so it is what vLLM, LM Studio, llama.cpp and self-hosted gateways get added through, not just Ollama itself. Any of them that validates message roles strictly was unusable.

workersAiCompat() in the same file already pins this flag; this brings the ollama branch in line.

Verification

Against an internal OpenAI-compatible gateway serving a DeepSeek model, before the change every chat failed with the 400 above. After it:

• plain chat ✅
• streaming, including reasoning_content deltas rendered as thinking ✅
• executeCode tool calls ✅
• multi-turn with tool-result replay ✅ (asked it to compute the sum of squares 1..100 in code; it ran executeCode and returned 338350)

pnpm --filter @gadgets/workshop-backend types:check passes.

I also checked with curl that the endpoint itself accepts everything else pi sends — non-streaming, streaming, assistant-message replay with and without reasoning_content, and function calling all return 200. The developer role was the only rejected part.

🤖 Generated with Claude Code

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[warmwind]

I have read the CLA Document and I hereby sign the CLA

[kentonv]

Hi @warmwind,

Apologies but, per our contributing guide, we prefer not to receive external pull requests. We'd rather receive issue reports, which we can then prompt our own agents to solve. Could you please file an issue instead?

I realize this one may seem trivial, but I think it's not really -- presumably Pi is using the developer role for a reason here, and by accepting this change we might lose somtehing. The ollama provider is not intended to be a general "OpenAI compatible" provider; it is specifically intended for ollama, and in my testing it does in fact work correctly against ollama.