Add Gemini and OpenAI chat providers

[xPoleStarx]

Summary

This PR adds first-class OpenAI and Gemini chat support alongside the existing Claude integration.

The main goal was to keep the macOS app's voice flow unchanged while making the chat layer provider-agnostic. The app can now route multimodal screenshot + transcript requests through the Cloudflare Worker to Claude, OpenAI, or Gemini using a shared request/response shape.

What Changed

• Added a provider-agnostic chat client on the macOS side
• Expanded the model picker to support:
  • claude-sonnet-4-6
  • claude-opus-4-6
  • gpt-5.4
  • gemini-2.5-flash
• Updated CompanionManager to use a model-agnostic chat pipeline instead of a Claude-specific one
• Reworked the Cloudflare Worker /chat route to support:
  • Anthropic Messages API
  • OpenAI Responses API
  • Gemini generateContent
• Normalized all provider responses into the SSE shape already expected by the macOS client
• Updated the standalone OpenAI helper to use the current Responses API
• Updated docs for the new worker secrets and multi-provider chat architecture

Why

The existing app flow was tightly coupled to Claude. This change keeps the user-facing interaction model the same while allowing Clicky to switch between major multimodal providers with minimal client-side branching.

That makes it easier to compare model behavior, keep provider flexibility, and continue evolving the app without hard-coding one vendor into the primary chat path.

Notes

New worker secrets required:

• OPENAI_API_KEY
• GEMINI_API_KEY

Existing secrets still required:

• ANTHROPIC_API_KEY
• ASSEMBLYAI_API_KEY
• ELEVENLABS_API_KEY

Testing

I did not run xcodebuild because the repo instructions explicitly say not to run it from the terminal due to TCC permission resets.

Static validation performed:

• reviewed integration points in the macOS client
• updated the worker routing logic end-to-end
• checked staged changes and commit integrity
• ran git diff --check

Follow-ups

Potential follow-up work, if desired:

• add provider-specific fallback/error messaging in the UI
• move more legacy direct-provider helpers behind the worker
• add a lightweight smoke-test path for worker chat providers

[qodo-ai-reviewer]

Hi, The Worker buffers the full provider response and emits a single SSE content_block_delta event, so the macOS streaming parser only receives one chunk at the end. This removes progressive delivery and can increase perceived latency (and defeats SSE streaming semantics) compared to a real token/chunk stream.

Severity: remediation recommended | Category: reliability

How to fix: Stream provider output as SSE

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

The /chat route emits a single SSE chunk containing the entire model response, which eliminates true streaming behavior.

Issue Context

CompanionChatAPI.analyzeImageStreaming() is implemented as an SSE line parser that can handle multiple incremental content_block_delta events.

Fix Focus Areas

• worker/src/index.ts[74-90]
• worker/src/index.ts[380-406]

What to change

• Implement per-provider streaming adapters (Anthropic/OpenAI/Gemini) that read each provider’s streaming response and re-emit incremental normalized SSE content_block_delta events.
• Alternatively, if intentional, document that streaming is “single-chunk SSE” and consider renaming client-side APIs or adjusting UX/timeouts accordingly.

---

Found by Qodo code review. FYI, Qodo is free for open-source.