refactor: extract google ai logic to windmill-common and use native gemini api in chat proxy

[centdix]

Summary

Extracts Google AI (Gemini) types and conversion logic from the worker into windmill-common, mirroring the existing Bedrock pattern. The API chat proxy now uses the native Gemini API directly instead of routing through Google's OpenAI-compatibility shim (/openai suffix).

Changes

• windmill-common/src/ai_google.rs (new): shared Gemini types, request/response structs, SSE event types, and conversion functions (openai_messages_to_gemini, openai_tools_to_gemini, parse_gemini_sse_event, parse_data_url, find_gemini_function_name)
• windmill-api/src/google.rs (new): native Gemini handler for the API chat proxy — converts OpenAI-format requests to Gemini format, calls streamGenerateContent?alt=sse or generateContent, and converts responses back to OpenAI format for the frontend
• windmill-api/src/ai.rs: route GoogleAI + chat/completions to the new native handler; remove the /openai suffix hack from prepare_request; expose HTTP_CLIENT, KEEPALIVE_INTERVAL_SECS, inject_keepalives as pub(crate)
• windmill-worker/src/ai/providers/google_ai.rs: import Gemini types from windmill-common instead of defining them locally; worker-specific S3 image handling stays in place
• windmill-worker/src/ai/sse.rs: remove local Gemini SSE type definitions; GeminiSSEParser.parse_event_data now delegates to parse_gemini_sse_event from common
• windmill-api/Cargo.toml: add eventsource-stream dependency for HTTP SSE parsing in the new handler

Test plan

• Configure a Google AI resource in workspace settings and verify chat completions work (streaming and non-streaming)
• Verify the AI agent (worker path) still works correctly with Google AI models
• Verify other providers (OpenAI, Anthropic, Bedrock) are unaffected

---

Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying windmill with    Cloudflare Pages

Latest commit:	50191c4
Status:	✅  Deploy successful!
Preview URL:	https://863606f0.windmill.pages.dev
Branch Preview URL:	https://ai-logic-share.windmill.pages.dev

View logs

[claude]

Claude finished @centdix's task in 3m 37s —— View job

---

PR Review: Refactor Google AI Logic

• Read all changed files and understand the refactoring
• Review windmill-common/src/ai_google.rs (shared types & conversions)
• Review windmill-api/src/google.rs (native Gemini handler)
• Review windmill-api/src/ai.rs (routing changes)
• Review worker-side changes
• Check for bugs, security, performance issues
• Post detailed review with inline comments

---

Overall Assessment

Well-structured refactoring that cleanly extracts shared Google AI types and conversion logic into windmill-common, following the same pattern used for Bedrock. The move from Google's OpenAI-compatibility shim to the native Gemini API is a good architectural decision — it gives full control over the request/response format and avoids the limitations of the compatibility layer.

What looks good

• Clean separation of concerns: Shared types, SSE parsing, and format conversion live in windmill-common. The API proxy handler and worker each have their own domain-specific logic.
• Consistent error handling: Both streaming and non-streaming paths handle API errors consistently with status code checking and body extraction.
• SSE keepalive injection: Reusing the existing inject_keepalives mechanism for the Gemini streaming proxy prevents connection drops.
• Well-documented code: Module-level docs, doc comments on public functions, and inline comments explain design decisions clearly.
• Thought signature support: Proper handling of Gemini 3+ thoughtSignature for multi-turn agent conversations is preserved through the refactoring.

Issues found (see inline comments)

1. Potential bug — double models/ prefix (google.rs:139): The model listing returns IDs like models/gemini-2.5-flash, but the chat/streaming endpoints prepend models/ again when building the URL. This would produce …/models/models/gemini-2.5-flash:streamGenerateContent.

2. Empty API key fallback (ai.rs:815): unwrap_or("") silently proceeds with an empty key instead of returning a clear error.

3. Incomplete finish_reason mapping (ai_google.rs:563): Gemini's MAX_TOKENS becomes "max_tokens" instead of OpenAI's expected "length".

4. Missing annotations in non-streaming response (google.rs:297): Grounding/web search citations are parsed but not included in the OpenAI-format non-streaming response.

5. Code duplication (ai_google.rs:363-448): openai_messages_to_gemini and the worker's convert_messages_to_gemini share ~80% of logic. Not blocking but worth noting for future maintenance.

Testing instructions

To test the changes: go to a workspace's Settings > AI page. Configure a Google AI provider by adding a resource with a valid Gemini API key. Then navigate to any Script editor (or the AI Chat in the workspace) and start a conversation using a Google AI / Gemini model. Verify that both streaming responses (real-time token output) and non-streaming responses work correctly. Also test with tool calling if possible (e.g., a script with AI agent capabilities). Finally, verify that other providers (OpenAI, Anthropic, Bedrock) still work normally — their code paths should be unaffected by this change.