Gemma 4 model integration, language TTS, auto config wrangler URL

[MathisZerbib]

📝 Summary

• Gemma 4 & Local Model Integration: Integrated support for local models through LM Studio via a repurposed OpenAIAPI.swift client pointing to 127.0.0.1:1234. If a selected model name doesn't contain "claude", the app routes Vision and Chat requests locally entirely free of charge. Fetches available local models dynamically.
• Localization & Language TTS: Added a language selection dropdown. The selected language code is injected directly into system prompts across all pipelines (CRITICAL: Respond EXCLUSIVELY in the following language code: ...), ensuring the TTS and generated text match the user's localized preference.
• Auto-Config Wrangler URL: Introduced WorkerEnvironment, which performs a fast 0.5s pre-flight check to localhost:8787 on startup. If a local Cloudflare worker is running, it auto-connects to it; otherwise, it seamlessly falls back to the production worker URL.
• Settings UI Overhaul: All of these configurations (Language, LM Studio API Key, Model Selection) are now tucked neatly behind a new inline Settings view gear icon on the Companion Panel. All preferences persist via UserDefaults.

---

🗂️ Files Changed

• CompanionManager.swift: Added @Published settings for selectedLanguageCode, lmStudioAPIKey, and availableModels. Added fetchAvailableModels() to pull models from LM Studio on init. Updated analyzeImageStreaming() pipelines to route appropriately to either Claude or the OpenAI SDK depending on the model, injecting the target language code into the prompt.
• CompanionPanelView.swift: Introduced an animated gear icon resolving to a new settingsView. Features a secure API key field with a visibility toggle, the new language picker, and the model picker rows smoothly tucked away.
• WorkerEnvironment.swift: New actor dedicated to sniffing out the http://localhost:8787 worker connection and failing over to the production URL.
• ElevenLabsTTSClient.swift & OpenAIAPI.swift: Refactored to accept mutable base URLs. Implemented analyzeImageStreaming within the OpenAI client using SSE to conform to the same progressive data streaming expectations as the Claude API.

---

✅ Test Plan

• UI Verification: Verify the new gear icon appears in the menu bar panel after onboarding and accurately expands/collapses the new settings rows.
• Language Test: Select a language other than English (e.g., French or Spanish), run a push-to-talk query, and ensure the model replies and reads back the response in the correct language.
• Local Model Test: Start LM Studio with Gemma 4 running locally on port 1234. Enter a local API key (if applicable), pick the Gemma model from the dropdown, and fire a vision query. Ensure responses stream back properly using the local backend.
• Wrangler Fall-over: Run npx wrangler dev in the worker folder, launch the app, and verify it prints 🌐 Auto-detected local worker. Kill the worker, relaunch the app, and verify it falls back to parsing the prod route.
• Persistence Test: Restart the app to ensure selectedLanguageCode, lmStudioAPIKey, and selectedGemmaModel state values survive termination gracefully via UserDefaults.

[MathisZerbib]

Api keys revoqued

[MathisZerbib]