[pcbeingused333]

When an assistant message makes parallel tool calls, OpenAI clients answer with one tool message per call. OpenAIReqToGemini turned each of them into its own user content holding one functionResponse. Gemini expects the responses to a parallel function call turn together in a single content, with as many functionResponse parts as there were functionCall parts, and rejects the request when they are split.

A tool message that directly follows another tool message now adds its functionResponse part to the previous content, in call order. A single tool result, and tool results of different turns (separated by other messages), still produce one content each.

Tests: openai_gemini_tool_results_test.go covers two parallel results in one content, and a single result plus a later user turn and second tool round staying apart; the first fails on master. go vet and go test ./internal/... pass. Not run against the live Gemini API.

This change was prepared with AI assistance (Claude Code).