Gemini generateContent / streamGenerateContent (v1beta).
Messages become contents, where an assistant turn is role: "model". System
instructions are a top-level systemInstruction, sampling and output controls
live in generationConfig, and reasoning maps to thinkingConfig.
Gemini identifies a tool result by function name, not by a call id, so
LlmCall.ToolResult.call_id is not represented on this protocol. A request
without a function name on the result is rejected.
Streaming changes the method in the path (:streamGenerateContent) and adds
alt=sse to the query; the transport still owns SSE framing.
provider_options accepts :gemini and its :google alias. The alias is
collapsed per source before merging, so a request value wins over a config
default even when the two use different names, and nested native objects are
deep-merged (a request can extend a defaulted generationConfig).
finishReason maps into the canonical union: CONTINUATION is a resumable
token-limit stop, so it reads as :max_tokens; the content-policy reasons (the
safety, recitation, blocklist, language, image-safety, and escalation family)
read as :content_filter; and the reasons that mean the generation failed
(MALFORMED_FUNCTION_CALL, UNEXPECTED_TOOL_CALL, TOO_MANY_TOOL_CALLS,
MISSING_THOUGHT_SIGNATURE, MALFORMED_RESPONSE, NO_IMAGE, IMAGE_OTHER,
PUP_LIMITED_DISABLED) read as :error. The candidate's continuationToken
has no canonical field and stays in Response.raw.