Summary

Completes the Gemini (Palm2 replacement) API support for the JS SDK as requested in #279.

/claim #279

What’s done

1. Enhanced GeminiAI class (src/ai/src/lib/gemini/gemini.ts)

• Multi-turn conversation support via messages parameter (array of {role, content})
• Proper generationConfig with temperature, maxOutputTokens, topP, topK, responseMimeType
• Safety settings support (safetySettings array)
• New chatText() convenience method that extracts text content directly
• Full TypeScript types for all request/response shapes

2. Exported TypeScript types (src/ai/src/types/index.ts)

• All Gemini types re-exported: GeminiModel, GeminiRole, GeminiResponse, GeminiChatOptions, etc.

3. Unit tests in testcases/palm2 folder (src/ai/src/tests/palm2/geminiEndpoints.test.ts)

• Constructor tests (API key from param, env var, missing key warning)
• Chat tests: simple prompt, custom model, generation config params, multi-turn messages, safety settings, retry on failure
• ChatText tests: text extraction, empty candidates, multi-part concatenation
• Uses mocked axios (same pattern as OpenAI tests)

4. Jsonnet example (examples/chat-with-gemini/)

• Prompts defined in jsonnet/main.jsonnet (not hardcoded)
• API key in jsonnet/secrets.jsonnet
• Hono server with /chat POST endpoint
• Follows same structure as chat-with-llm example

Acceptance criteria from issue

• Classes and TypeScript types for Gemini API
• Unit testcases in testcases folder named “palm2”
• Example with jsonnet prompts (not hardcoded)

Implementation notes

• Uses native HTTP calls via axios (not Google’s @google/generative-ai SDK), matching the existing OpenAI implementation pattern
• Gemini API replaces deprecated Palm2 as confirmed by maintainer @sandys
• Backward compatible: existing GeminiAI.chat() API preserved, new features are additive