fix google vertex unsigned tool call replay closes #4032

packages/ai/CHANGELOG.md

@@ -12,6 +12,7 @@
 
 ### Fixed
 
+- Fixed Google Vertex Gemini 3 tool call replay by no longer sending the non-Vertex `skip_thought_signature_validator` sentinel for unsigned tool calls ([#4032](https://github.com/badlogic/pi-mono/issues/4032)).
 - Updated `@anthropic-ai/sdk` to `^0.91.1` to clear GHSA-p7fg-763f-g4gf audit findings ([#3992](https://github.com/badlogic/pi-mono/issues/3992)).
 - Fixed DeepSeek V4 Flash `xhigh` thinking support so requests preserve `xhigh` and map it to DeepSeek's `max` reasoning effort ([#3944](https://github.com/badlogic/pi-mono/issues/3944)).
 - Fixed Anthropic streams that end before `message_stop` to be treated as errors instead of successful partial responses ([#3936](https://github.com/badlogic/pi-mono/issues/3936)).

packages/ai/src/providers/google-shared.ts

@@ -51,11 +51,6 @@ export function retainThoughtSignature(existing: string | undefined, incoming: s
 // Thought signatures must be base64 for Google APIs (TYPE_BYTES).
 const base64SignaturePattern = /^[A-Za-z0-9+/]+={0,2}$/;
 
-// Sentinel value that tells the Gemini API to skip thought signature validation.
-// Used for unsigned function call parts (e.g. replayed from providers without thought signatures).
-// See: https://ai.google.dev/gemini-api/docs/thought-signatures
-const SKIP_THOUGHT_SIGNATURE = "skip_thought_signature_validator";
-
 function isValidThoughtSignature(signature: string | undefined): boolean {
 	if (!signature) return false;
 	if (signature.length % 4 !== 0) return false;
@@ -161,18 +156,13 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 					}
 				} else if (block.type === "toolCall") {
 					const thoughtSignature = resolveThoughtSignature(isSameProviderAndModel, block.thoughtSignature);
-					// Gemini 3 requires thoughtSignature on all function calls when thinking mode is enabled.
-					// Use the skip_thought_signature_validator sentinel for unsigned function calls
-					// when replayed from providers without thought signatures.
-					const isGemini3 = model.id.toLowerCase().includes("gemini-3");
-					const effectiveSignature = thoughtSignature || (isGemini3 ? SKIP_THOUGHT_SIGNATURE : undefined);
 					const part: Part = {
 						functionCall: {
 							name: block.name,
 							args: block.arguments ?? {},
 							...(requiresToolCallId(model.id) ? { id: block.id } : {}),
 						},
-						...(effectiveSignature && { thoughtSignature: effectiveSignature }),
+						...(thoughtSignature && { thoughtSignature }),
 					};
 					parts.push(part);
 				}

packages/ai/test/google-shared-gemini3-unsigned-tool-call.test.ts

@@ -2,15 +2,17 @@ import { describe, expect, it } from "vitest";
 import { convertMessages } from "../src/providers/google-shared.js";
 import type { Context, Model } from "../src/types.js";
 
-const SKIP_THOUGHT_SIGNATURE = "skip_thought_signature_validator";
-
-function makeGemini3Model(id = "gemini-3-pro-preview"): Model<"google-generative-ai"> {
+function makeGemini3Model<TApi extends "google-generative-ai" | "google-vertex">(
+	api: TApi,
+	provider: Model<TApi>["provider"],
+	id = "gemini-3-pro-preview",
+): Model<TApi> {
 	return {
 		id,
 		name: "Gemini 3 Pro Preview",
-		api: "google-generative-ai",
-		provider: "google",
-		baseUrl: "https://generativelanguage.googleapis.com",
+		api,
+		provider,
+		baseUrl: "https://example.com",
 		reasoning: true,
 		input: ["text"],
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
@@ -19,154 +21,96 @@ function makeGemini3Model(id = "gemini-3-pro-preview"): Model<"google-generative
 	};
 }
 
-describe("google-shared convertMessages — Gemini 3 unsigned tool calls", () => {
-	it("uses skip_thought_signature_validator for unsigned tool calls on Gemini 3", () => {
-		const model = makeGemini3Model();
-		const now = Date.now();
-		const context: Context = {
-			messages: [
-				{ role: "user", content: "Hi", timestamp: now },
-				{
-					role: "assistant",
-					content: [
-						{
-							type: "toolCall",
-							id: "call_1",
-							name: "bash",
-							arguments: { command: "ls -la" },
-							// No thoughtSignature: simulates replay from a provider without thought signatures.
-						},
-					],
-					api: "google-generative-ai",
-					provider: "google",
-					model: "claude-sonnet-4-6",
-					usage: {
-						input: 0,
-						output: 0,
-						cacheRead: 0,
-						cacheWrite: 0,
-						totalTokens: 0,
-						cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+function makeContext(model: { api: string; provider: string; id: string }, thoughtSignature?: string): Context {
+	const now = Date.now();
+	return {
+		messages: [
+			{ role: "user", content: "Hi", timestamp: now },
+			{
+				role: "assistant",
+				content: [
+					{
+						type: "toolCall",
+						id: "call_1",
+						name: "bash",
+						arguments: { command: "echo hi" },
+						...(thoughtSignature && { thoughtSignature }),
 					},
-					stopReason: "stop",
-					timestamp: now,
+					{
+						type: "toolCall",
+						id: "call_2",
+						name: "bash",
+						arguments: { command: "ls -la" },
+					},
+				],
+				api: model.api,
+				provider: model.provider,
+				model: model.id,
+				usage: {
+					input: 0,
+					output: 0,
+					cacheRead: 0,
+					cacheWrite: 0,
+					totalTokens: 0,
+					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
 				},
-			],
-		};
+				stopReason: "toolUse",
+				timestamp: now,
+			},
+		],
+	};
+}
 
-		const contents = convertMessages(model, context);
+describe("google-shared convertMessages — Gemini 3 unsigned tool calls", () => {
+	it("does not add skip_thought_signature_validator for unsigned Google Gen AI tool calls", () => {
+		const model = makeGemini3Model("google-generative-ai", "google");
+		const contents = convertMessages(model, makeContext({ ...model, id: "other-model" }));
 
 		const modelTurn = contents.find((c) => c.role === "model");
 		expect(modelTurn).toBeTruthy();
 
-		// Should be a structured functionCall, NOT text fallback
-		const fcPart = modelTurn?.parts?.find((p) => p.functionCall !== undefined);
-		expect(fcPart).toBeTruthy();
-		expect(fcPart?.functionCall?.name).toBe("bash");
-		expect(fcPart?.functionCall?.args).toEqual({ command: "ls -la" });
-		expect(fcPart?.thoughtSignature).toBe(SKIP_THOUGHT_SIGNATURE);
+		const functionCallParts = modelTurn?.parts?.filter((p) => p.functionCall !== undefined) ?? [];
+		expect(functionCallParts).toHaveLength(2);
+		expect(functionCallParts[0]?.thoughtSignature).toBeUndefined();
+		expect(functionCallParts[1]?.thoughtSignature).toBeUndefined();
+		expect(JSON.stringify(modelTurn)).not.toContain("skip_thought_signature_validator");
 
-		// No text fallback should exist
 		const textParts = modelTurn?.parts?.filter((p) => p.text !== undefined) ?? [];
 		const historicalText = textParts.filter((p) => p.text?.includes("Historical context"));
 		expect(historicalText).toHaveLength(0);
 	});
 
-	it("preserves valid thoughtSignature when present (same provider/model)", () => {
-		const model = makeGemini3Model();
-		const now = Date.now();
-		// Valid base64 signature (16 bytes = 24 chars base64)
-		const validSig = "AAAAAAAAAAAAAAAAAAAAAA==";
-		const context: Context = {
-			messages: [
-				{ role: "user", content: "Hi", timestamp: now },
-				{
-					role: "assistant",
-					content: [
-						{
-							type: "toolCall",
-							id: "call_1",
-							name: "bash",
-							arguments: { command: "echo hi" },
-							thoughtSignature: validSig,
-						},
-					],
-					api: "google-generative-ai",
-					provider: "google",
-					model: "gemini-3-pro-preview",
-					usage: {
-						input: 0,
-						output: 0,
-						cacheRead: 0,
-						cacheWrite: 0,
-						totalTokens: 0,
-						cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-					},
-					stopReason: "stop",
-					timestamp: now,
-				},
-			],
-		};
-
-		const contents = convertMessages(model, context);
+	it("does not add skip_thought_signature_validator for unsigned Vertex tool calls", () => {
+		const model = makeGemini3Model("google-vertex", "google-vertex");
+		const contents = convertMessages(model, makeContext(model));
 		const modelTurn = contents.find((c) => c.role === "model");
-		const fcPart = modelTurn?.parts?.find((p) => p.functionCall !== undefined);
+		const functionCallParts = modelTurn?.parts?.filter((p) => p.functionCall !== undefined) ?? [];
 
-		expect(fcPart).toBeTruthy();
-		expect(fcPart?.thoughtSignature).toBe(validSig);
+		expect(functionCallParts).toHaveLength(2);
+		expect(functionCallParts[0]?.thoughtSignature).toBeUndefined();
+		expect(functionCallParts[1]?.thoughtSignature).toBeUndefined();
+		expect(JSON.stringify(modelTurn)).not.toContain("skip_thought_signature_validator");
 	});
 
-	it("does not add sentinel for non-Gemini-3 models", () => {
-		const model: Model<"google-generative-ai"> = {
-			id: "gemini-2.5-flash",
-			name: "Gemini 2.5 Flash",
-			api: "google-generative-ai",
-			provider: "google",
-			baseUrl: "https://generativelanguage.googleapis.com",
-			reasoning: true,
-			input: ["text"],
-			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-			contextWindow: 128000,
-			maxTokens: 8192,
-		};
-		const now = Date.now();
-		const context: Context = {
-			messages: [
-				{ role: "user", content: "Hi", timestamp: now },
-				{
-					role: "assistant",
-					content: [
-						{
-							type: "toolCall",
-							id: "call_1",
-							name: "bash",
-							arguments: { command: "ls" },
-							// No thoughtSignature
-						},
-					],
-					api: "google-generative-ai",
-					provider: "google",
-					model: "claude-sonnet-4-6",
-					usage: {
-						input: 0,
-						output: 0,
-						cacheRead: 0,
-						cacheWrite: 0,
-						totalTokens: 0,
-						cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-					},
-					stopReason: "stop",
-					timestamp: now,
-				},
-			],
-		};
+	it("preserves valid thoughtSignature when present for the same provider and model", () => {
+		const model = makeGemini3Model("google-generative-ai", "google");
+		const validSig = "AAAAAAAAAAAAAAAAAAAAAA==";
+		const contents = convertMessages(model, makeContext(model, validSig));
+		const modelTurn = contents.find((c) => c.role === "model");
+		const functionCallParts = modelTurn?.parts?.filter((p) => p.functionCall !== undefined) ?? [];
 
-		const contents = convertMessages(model, context);
+		expect(functionCallParts).toHaveLength(2);
+		expect(functionCallParts[0]?.thoughtSignature).toBe(validSig);
+		expect(functionCallParts[1]?.thoughtSignature).toBeUndefined();
+	});
+
+	it("does not add a thoughtSignature for non-Gemini-3 models", () => {
+		const model = makeGemini3Model("google-generative-ai", "google", "gemini-2.5-flash");
+		const contents = convertMessages(model, makeContext({ ...model, id: "other-model" }));
 		const modelTurn = contents.find((c) => c.role === "model");
 		const fcPart = modelTurn?.parts?.find((p) => p.functionCall !== undefined);
 
 		expect(fcPart).toBeTruthy();
-		// No sentinel, no thoughtSignature at all
 		expect(fcPart?.thoughtSignature).toBeUndefined();
 	});
 });