Google Gemini SDK → OrcaRouter Lite

Lite serves the Gemini API natively (AI Studio style, /v1beta), so the official google-genai SDK connects without changes.

google-genai (Python)

from google import genai
from google.genai.types import HttpOptions

client = genai.Client(
    api_key="sk-orca-...",                                  # your Lite key
    http_options=HttpOptions(base_url="http://localhost:8000"),
)

resp = client.models.generate_content(
    model="auto",                       # or any catalog model
    contents="Hello!",
)
print(resp.text)

# streaming
for chunk in client.models.generate_content_stream(
    model="gemini-1.5-flash", contents="Tell me a story",
):
    print(chunk.text, end="")

curl

# blocking
curl "http://localhost:8000/v1beta/models/auto:generateContent" \
  -H "x-goog-api-key: sk-orca-..." \
  -H "content-type: application/json" \
  -d '{"contents":[{"role":"user","parts":[{"text":"Hello!"}]}]}'

# streaming (SSE)
curl -N "http://localhost:8000/v1beta/models/auto:streamGenerateContent?alt=sse" \
  -H "x-goog-api-key: sk-orca-..." \
  -H "content-type: application/json" \
  -d '{"contents":[{"role":"user","parts":[{"text":"Hello!"}]}]}'

Auth is accepted as x-goog-api-key (preferred) or ?key= — note the query-param form puts the key in URLs, which typically end up in access logs; the header form doesn't.

Known limitations (v1)

• candidateCount > 1, cachedContent, Files API (fileData), and server-side tools (googleSearch, codeExecution, urlContext) return 400. Function declarations are fully supported (uppercase proto type enums are normalized automatically).
• safetySettings, thinkingConfig, and topK are dropped; thought / thoughtSignature parts in history are dropped the same way.
• stopSequences beyond the OpenAI wire cap of 4 are truncated to the first 4 (a structlog warning is emitted).
• Streaming keeps text and functionCall parts in the model's order across chunks, but within ONE upstream chunk that carries both, the text part is always emitted first: the internal OpenAI delta has no ordering between content and tool_calls, so a call that preceded text in the same chunk cannot be told apart from one that followed it.
• temperature (0–2), topP (0–1) and maxOutputTokens (>= 1) are range-checked and rejected with a 400 INVALID_ARGUMENT rather than forwarded — an upstream rejection would reach you as a retryable 500.
• functionCallingConfig.mode: ANY with several allowedFunctionNames restricts the tools sent upstream to that subset, so the model cannot call a function you excluded.
• responseMimeType: "application/json" maps to JSON mode; responseSchema itself is not yet enforced.
• streamGenerateContent without ?alt=sse returns the full JSON array in one response (the SDKs all use alt=sse).