atomr-infer-runtime-gemini

Google Gemini runtime — both AI Studio (API key in query string) and Vertex AI (OAuth2 access token via the operator's preferred credential source).

Two variants, one runner

use inference_runtime_gemini::{GeminiConfig, GeminiRunner, GeminiVariant};
use inference_runtime_gemini::config::SecretRef;

// AI Studio
let cfg = GeminiConfig {
    variant: GeminiVariant::AiStudio { /* default endpoint */ ..Default::default() },
    credential: SecretRef::Env { name: "GOOGLE_API_KEY".into() },
    safety: vec![],
    ..GeminiConfig::default_for_aistudio()
};

// Vertex
let cfg = GeminiConfig {
    variant: GeminiVariant::Vertex {
        project: "my-gcp-project".into(),
        region: "us-central1".into(),
    },
    credential: SecretRef::Adc,                // Application Default Credentials
    safety: vec![],
    ..Default::default()
};

OAuth2 — pluggable, not bundled

We don't pull a full oauth2 stack into the workspace root. Instead, inference_remote_core::session::CredentialProvider is a trait with one async method:

async fn token(&self) -> InferenceResult<SecretString>;

For Vertex, supply a provider that calls gcloud auth print-access-token (or your preferred token source) and refreshes on a timer. For AI Studio, StaticApiKey is enough — the runner appends it as a ?key=... query param.

SSE format

Gemini's streamGenerateContent?alt=sse emits the same JSON schema as the unary generateContent, one chunk per SSE data: line. The runner reuses inference_remote_core::sse::decode_sse_stream and deserializes each chunk via serde_json::from_str::<GenerateContentResponse>.

Safety settings

GeminiConfig::safety: Vec<SafetySetting> is forwarded verbatim. The error classifier upgrades a FAILED_PRECONDITION whose message mentions "safety" or "blocked" to InferenceError::ContentFiltered, so the RequestActor knows not to retry.

Pricing