fix(ai): normalize tool-result image data URLs

packages/ai/src/api/transform-messages.ts

@@ -62,11 +62,25 @@ function replaceImagesWithPlaceholder(content: (TextContent | ImageContent)[], p
 	return result;
 }
 
+function normalizeToolResultMediaData(content: (TextContent | ImageContent)[]): (TextContent | ImageContent)[] {
+	let changed = false;
+	const normalized = content.map((block) => {
+		if (block.type !== "image") return block;
+		const prefix = `data:${block.mimeType};base64,`;
+		if (!block.data.startsWith(prefix)) return block;
+		changed = true;
+		return { ...block, data: block.data.slice(prefix.length) };
+	});
+	return changed ? normalized : content;
+}
+
 function downgradeUnsupportedImages<TApi extends Api>(messages: Message[], model: Model<TApi>): Message[] {
 	const supportsImages = model.input.includes("image");
 	const supportsVideo = model.input.includes("video");
 	if (supportsImages && supportsVideo) {
-		return messages;
+		return messages.map((msg) =>
+			msg.role === "toolResult" ? { ...msg, content: normalizeToolResultMediaData(msg.content) } : msg,
+		);
 	}
 
 	return messages.map((msg) => {
@@ -86,7 +100,7 @@ function downgradeUnsupportedImages<TApi extends Api>(messages: Message[], model
 		}
 
 		if (msg.role === "toolResult") {
-			let content = msg.content;
+			let content = normalizeToolResultMediaData(msg.content);
 			if (!supportsVideo) {
 				content = replaceMediaWithPlaceholder(content, NO_VIDEO_TOOL_PLACEHOLDER, (b) =>
 					isVideoMimeType(b.mimeType),

packages/ai/src/changes.md

@@ -1,5 +1,24 @@
 ## PR #1304 review fixes: shared auth-miss prefix, login merge, sentinel repair (2026-09-10)
 
+## 2026-09-11 - Normalize canonical tool-result image data URLs before provider serialization
+
+### What changed
+
+- `api/transform-messages.ts`: canonical `data:${mimeType};base64,` prefixes on tool-result image blocks are stripped back to raw base64 at the shared message-normalization boundary. Raw base64 and mismatched/arbitrary data URLs remain unchanged.
+- `test/issue-1260-tool-result-image-data-url.test.ts`: covers raw base64 and already-prefixed canonical image data through the Responses serializer and pins exactly one wire prefix.
+
+### Why
+
+- Issue #1260: `ImageContent.data` is an internal raw-base64 contract, but tool results can arrive already prefixed. Passing that value through made Responses adapters prepend a second data-URL prefix and emit an invalid `image_url`.
+
+### Why an extension could not handle it
+
+- This is the shared provider-normalization boundary used before protocol serializers. Fixing one provider or extension would leave the malformed internal representation visible to other adapters.
+
+### Expected merge conflict zones
+
+- LOW: `api/transform-messages.ts` around tool-result media normalization and `downgradeUnsupportedImages()`.
+
 ### What changed
 
 - `packages/ai/src/auth/resolve.ts`: `PROVIDER_NOT_CONFIGURED_PREFIX` / `providerNotConfiguredMessage()` export the exact auth-miss wording every resolution site throws; `packages/ai/src/models.ts` re-exports both and throws through the helper. Consumers keying recovery decisions off that message (the coding-agent session layer and the credential-pool classifier) can never drift from the throw sites.

packages/ai/test/issue-1260-tool-result-image-data-url.test.ts

@@ -0,0 +1,87 @@
+import { describe, expect, it } from "vitest";
+import { convertResponsesMessages } from "../src/providers/openai-responses-shared.ts";
+import type { Context, Model } from "../src/types.ts";
+
+const model = {
+	id: "gpt-5.5",
+	name: "GPT-5.5",
+	provider: "openai-codex",
+	api: "openai-codex-responses",
+	baseUrl: "https://chatgpt.com/backend-api/codex",
+	reasoning: true,
+	input: ["text", "image"],
+	cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+	contextWindow: 1_000_000,
+	maxTokens: 128_000,
+} as Model<"openai-codex-responses">;
+
+function zeroUsage() {
+	return {
+		input: 0,
+		output: 0,
+		cacheRead: 0,
+		cacheWrite: 0,
+		totalTokens: 0,
+		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+	};
+}
+
+function convertToolImage(data: string) {
+	const context: Context = {
+		messages: [
+			{
+				role: "assistant",
+				content: [
+					{
+						type: "toolCall",
+						id: "call_image|fc_image",
+						name: "image_tool",
+						arguments: {},
+					},
+				],
+				api: "openai-codex-responses",
+				provider: "openai-codex",
+				model: "gpt-5.5",
+				usage: zeroUsage(),
+				stopReason: "toolUse",
+				timestamp: 1,
+			},
+			{
+				role: "toolResult",
+				toolCallId: "call_image|fc_image",
+				toolName: "image_tool",
+				content: [{ type: "image", data, mimeType: "image/png" }],
+				isError: false,
+				timestamp: 2,
+			},
+		],
+		tools: [],
+	};
+
+	return convertResponsesMessages(model, context, new Set(["openai-codex"]));
+}
+
+describe("issue #1260 Responses tool-result image data URLs", () => {
+	it("adds exactly one data URL prefix to raw base64", () => {
+		expect(convertToolImage("QUJD")).toMatchObject([
+			{ type: "function_call", call_id: "call_image", name: "image_tool" },
+			{
+				type: "function_call_output",
+				call_id: "call_image",
+				output: [{ type: "input_image", detail: "auto", image_url: "data:image/png;base64,QUJD" }],
+			},
+		]);
+	});
+
+	it("preserves an already-prefixed canonical image data URL", () => {
+		const dataUrl = "data:image/png;base64,QUJD";
+		expect(convertToolImage(dataUrl)).toMatchObject([
+			{ type: "function_call", call_id: "call_image", name: "image_tool" },
+			{
+				type: "function_call_output",
+				call_id: "call_image",
+				output: [{ type: "input_image", detail: "auto", image_url: dataUrl }],
+			},
+		]);
+	});
+});