codec/geminiapi targets generateContent and streamGenerateContent. The request body is identical in both modes; the route selects streaming.

Content

System text becomes systemInstruction. User and model turns use contents with roles user and model; tool results are a user turn carrying a functionResponse. Images become inline base64 or fileData; tool calls use functionCall with raw object arguments. Thinking is a part with thought: true when the model advertises thinking and the request asks for it.

body, err := geminiapi.EncodeRequest(req)
if err != nil {
	return err
}
fmt.Println(string(body))

Route and stream

route.GeminiGenerateContent builds POST {base}/models/{name}:generateContent for invoke and POST {base}/models/{name}:streamGenerateContent?alt=sse for stream. The codec does not add a stream JSON flag. DecodeResponse reads candidates[0]. Input is promptTokenCount less cachedContentTokenCount, plus toolUsePromptTokenCount, which Gemini reports separately; output is candidates plus thoughts. totalTokenCount is validated as a well-formed count but is not reconciled against those components. The codec does not emit a request cache marker; the only cache evidence it consumes is the response usage field.

Source and proof

Run go test ./codec/geminiapi ./route.