fix: sanitize Gemini 3.1 Google reasoning payloads

This commit is contained in:
Peter Steinberger
2026-02-25 01:36:36 +00:00
parent 039713c3e7
commit b35d00aaf8
4 changed files with 360 additions and 0 deletions

View File

@@ -6,9 +6,13 @@ import { isTruthyEnvValue } from "../infra/env.js";
import { applyExtraParamsToAgent } from "./pi-embedded-runner.js";
const OPENAI_KEY = process.env.OPENAI_API_KEY ?? "";
const GEMINI_KEY = process.env.GEMINI_API_KEY ?? "";
const LIVE = isTruthyEnvValue(process.env.OPENAI_LIVE_TEST) || isTruthyEnvValue(process.env.LIVE);
const GEMINI_LIVE =
isTruthyEnvValue(process.env.GEMINI_LIVE_TEST) || isTruthyEnvValue(process.env.LIVE);
const describeLive = LIVE && OPENAI_KEY ? describe : describe.skip;
const describeGeminiLive = GEMINI_LIVE && GEMINI_KEY ? describe : describe.skip;
describeLive("pi embedded extra params (live)", () => {
it("applies config maxTokens to openai streamFn", async () => {
@@ -62,3 +66,170 @@ describeLive("pi embedded extra params (live)", () => {
expect(outputTokens ?? 0).toBeLessThanOrEqual(20);
}, 30_000);
});
describeGeminiLive("pi embedded extra params (gemini live)", () => {
function isGoogleModelUnavailableError(raw: string | undefined): boolean {
const msg = (raw ?? "").toLowerCase();
if (!msg) {
return false;
}
return (
msg.includes("not found") ||
msg.includes("404") ||
msg.includes("not_available") ||
msg.includes("permission denied") ||
msg.includes("unsupported model")
);
}
function isGoogleImageProcessingError(raw: string | undefined): boolean {
const msg = (raw ?? "").toLowerCase();
if (!msg) {
return false;
}
return (
msg.includes("unable to process input image") ||
msg.includes("invalid_argument") ||
msg.includes("bad request")
);
}
async function runGeminiProbe(params: {
agentStreamFn: typeof streamSimple;
model: Model<"google-generative-ai">;
apiKey: string;
oneByOneRedPngBase64: string;
includeImage?: boolean;
prompt: string;
onPayload?: (payload: Record<string, unknown>) => void;
}): Promise<{ sawDone: boolean; stopReason?: string; errorMessage?: string }> {
const userContent: Array<
{ type: "text"; text: string } | { type: "image"; mimeType: string; data: string }
> = [{ type: "text", text: params.prompt }];
if (params.includeImage ?? true) {
userContent.push({
type: "image",
mimeType: "image/png",
data: params.oneByOneRedPngBase64,
});
}
const stream = params.agentStreamFn(
params.model,
{
messages: [
{
role: "user",
content: userContent,
timestamp: Date.now(),
},
],
},
{
apiKey: params.apiKey,
reasoning: "high",
maxTokens: 64,
onPayload: (payload) => {
params.onPayload?.(payload as Record<string, unknown>);
},
},
);
let sawDone = false;
let stopReason: string | undefined;
let errorMessage: string | undefined;
for await (const event of stream) {
if (event.type === "done") {
sawDone = true;
stopReason = event.reason;
} else if (event.type === "error") {
stopReason = event.reason;
errorMessage = event.error?.errorMessage;
}
}
return { sawDone, stopReason, errorMessage };
}
it("sanitizes Gemini 3.1 thinking payload and keeps image parts with reasoning enabled", async () => {
const model = getModel(
"google",
"gemini-3.1-pro-preview",
) as unknown as Model<"google-generative-ai">;
const agent = { streamFn: streamSimple };
applyExtraParamsToAgent(agent, undefined, "google", model.id, undefined, "high");
const oneByOneRedPngBase64 =
"iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR4nGP4zwAAAgIBAJBzWgkAAAAASUVORK5CYII=";
let capturedPayload: Record<string, unknown> | undefined;
const imageResult = await runGeminiProbe({
agentStreamFn: agent.streamFn,
model,
apiKey: GEMINI_KEY,
oneByOneRedPngBase64,
includeImage: true,
prompt: "What color is this image? Reply with one word.",
onPayload: (payload) => {
capturedPayload = payload;
},
});
expect(capturedPayload).toBeDefined();
const thinkingConfig = (
capturedPayload?.config as { thinkingConfig?: Record<string, unknown> } | undefined
)?.thinkingConfig;
expect(thinkingConfig?.thinkingBudget).toBeUndefined();
expect(thinkingConfig?.thinkingLevel).toBe("HIGH");
const imagePart = (
capturedPayload?.contents as
| Array<{ parts?: Array<{ inlineData?: { mimeType?: string; data?: string } }> }>
| undefined
)?.[0]?.parts?.find((part) => part.inlineData !== undefined)?.inlineData;
expect(imagePart).toEqual({
mimeType: "image/png",
data: oneByOneRedPngBase64,
});
if (!imageResult.sawDone && !isGoogleModelUnavailableError(imageResult.errorMessage)) {
expect(isGoogleImageProcessingError(imageResult.errorMessage)).toBe(true);
}
const textResult = await runGeminiProbe({
agentStreamFn: agent.streamFn,
model,
apiKey: GEMINI_KEY,
oneByOneRedPngBase64,
includeImage: false,
prompt: "Reply with exactly OK.",
});
if (!textResult.sawDone && isGoogleModelUnavailableError(textResult.errorMessage)) {
// Some keys/regions do not expose Gemini 3.1 preview. Fall back to a
// stable model to keep live reasoning verification active.
const fallbackModel = getModel(
"google",
"gemini-2.5-pro",
) as unknown as Model<"google-generative-ai">;
const fallback = await runGeminiProbe({
agentStreamFn: agent.streamFn,
model: fallbackModel,
apiKey: GEMINI_KEY,
oneByOneRedPngBase64,
includeImage: false,
prompt: "Reply with exactly OK.",
});
expect(fallback.sawDone).toBe(true);
expect(fallback.stopReason).toBeDefined();
expect(fallback.stopReason).not.toBe("error");
return;
}
expect(textResult.sawDone).toBe(true);
expect(textResult.stopReason).toBeDefined();
expect(textResult.stopReason).not.toBe("error");
}, 45_000);
});