src/provider/gemini.rs

This file is your Google Gemini client, GeminiClient, and it runs about 3,720 lines with tests. You reach it under the gemini feature.

Key items

You build a client with from_env(), which reads GEMINI_API_KEY or GOOGLE_API_KEY, or with the builder, where with_include_thoughts starts false.

Gemini puts the model in the path, so you call :streamGenerateContent?alt=sse or :generateContent. Your API key rides the x-goog-api-key header and never one of those URLs.

The client folds your system prompt into systemInstruction, because the API rejects an inline system role and omits the field entirely when it is empty.

generationConfig carries your choices. You treat thinkingConfig.includeThoughts as an opt-in, and/or you set responseMimeType plus responseJsonSchema for formats.

The default for thoughts is off, which keeps your replies lean until you ask for reasoning.

The emitter synthesizes MessageStart, since Gemini chunks carry no ids. It latches usage from the latest non-empty usageMetadata and counts thoughts into your output.

A chunk with finishReason becomes your terminal event, and it fires once, so a proxy that re-emits cannot fake a second end.

Behavior notes

include_thoughts(true) on a non-reasoning model earns you a 400, so match the flag to the model you pick. Parsing still routes thought: true parts to your thinking lane even when you leave the flag at its default of false.

Gemini delivers tool calls complete rather than fragmented: open, one JSON delta, close. Your downstream sees the standard shape.

Errors map on status. Codes 429, 503 and 529, or the statuses RESOURCE_EXHAUSTED and UNAVAILABLE, become your rate limit.

Everything else becomes an API error. MAX_TOKENS becomes MaxTokens, and every other finish reason, SAFETY included, becomes EndTurn for you.

Deep dive: Gemini.