Propagate abort signals to bindings/fetches

Fix request cancellation by passing abort signals through outbound calls. Update ai-gateway-provider to forward options.abortSignal to fetch and binding.run and widen the binding.run type to accept an options object. Update tanstack-ai create-fetcher utilities to forward init.signal to binding.run (gateway and workers binding paths) and to build/run options consistently. Update workers-ai-provider chat, embedding, and image models to include abortSignal when calling binding.run so canceled requests are properly aborted. Update tests to reflect the optional options object shape and ensure extraHeaders handling remains correct. Adds a changeset documenting the patch.

Author: threepointone

.changeset/fix-abort-signal-propagation.md

@@ -0,0 +1,13 @@
+---
+"ai-gateway-provider": patch
+"workers-ai-provider": patch
+"@cloudflare/tanstack-ai": patch
+---
+
+Fix request cancellation by propagating `abortSignal` to outbound network calls.
+
+**ai-gateway-provider**: Pass `abortSignal` to the `fetch` call (API path) and to `binding.run()` (binding path) so that cancelled requests are properly aborted.
+
+**workers-ai-provider**: Pass `abortSignal` to `binding.run()` for chat, embedding, and image models, matching the existing behavior in transcription, speech, and reranking models.
+
+**@cloudflare/tanstack-ai**: Pass `signal` through to `binding.run()` in both `createGatewayFetch` (AI Gateway binding path) and `createWorkersAiBindingFetch` (Workers AI binding path).

packages/ai-gateway-provider/src/index.ts

@@ -141,7 +141,9 @@ export class AiGatewayChatLanguageModel implements LanguageModelV3 {
 					...Object.fromEntries(headers.entries()),
 				},
 			}));
-			resp = await this.config.binding.run(updatedBody);
+			resp = await this.config.binding.run(updatedBody, {
+				signal: options.abortSignal,
+			});
 		} else {
 			headers.set("Content-Type", "application/json");
 			headers.set("cf-aig-authorization", `Bearer ${this.config.apiKey}`);
@@ -151,6 +153,7 @@ export class AiGatewayChatLanguageModel implements LanguageModelV3 {
 					body: JSON.stringify(body),
 					headers: headers,
 					method: "POST",
+					signal: options.abortSignal,
 				},
 			);
 		}
@@ -243,7 +246,7 @@ export type AiGatewayAPISettings = {
 };
 export type AiGatewayBindingSettings = {
 	binding: {
-		run(data: unknown): Promise<Response>;
+		run(data: unknown, options?: { signal?: AbortSignal }): Promise<Response>;
 	};
 	options?: AiGatewayOptions;
 };

packages/tanstack-ai/src/utils/create-fetcher.ts

@@ -3,7 +3,7 @@
 // ---------------------------------------------------------------------------
 
 export interface CloudflareAiGateway {
-	run(request: unknown): Promise<Response>;
+	run(request: unknown, options?: { signal?: AbortSignal }): Promise<Response>;
 }
 
 export interface AiGatewayBindingConfig {
@@ -245,7 +245,9 @@ export function createGatewayFetch(
 		}
 
 		if ("binding" in config) {
-			return config.binding.run(request);
+			return config.binding.run(request, {
+				signal: init?.signal ?? undefined,
+			});
 		}
 
 		return fetch(
@@ -335,10 +337,14 @@ export function createWorkersAiBindingFetch(
 		if (body.response_format) inputs.response_format = body.response_format;
 		if (stream) inputs.stream = true;
 
+		const runOptions: Record<string, unknown> = {};
+		if (options?.extraHeaders) runOptions.extraHeaders = options.extraHeaders;
+		if (init?.signal) runOptions.signal = init.signal;
+
 		const result = await binding.run(
 			model,
 			inputs,
-			options?.extraHeaders ? { extraHeaders: options.extraHeaders } : undefined,
+			Object.keys(runOptions).length > 0 ? runOptions : undefined,
 		);
 
 		if (stream && result instanceof ReadableStream) {

packages/tanstack-ai/test/binding-fetch.test.ts

@@ -571,8 +571,14 @@ describe("createWorkersAiBindingFetch", () => {
 		});
 
 		expect(binding.run).toHaveBeenCalledOnce();
-		const [, , options] = binding.run.mock.calls[0]!;
-		expect(options).toBeUndefined();
+		const [, , options] = binding.run.mock.calls[0]! as [
+			unknown,
+			unknown,
+			Record<string, unknown> | undefined,
+		];
+		if (options) {
+			expect(options).not.toHaveProperty("extraHeaders");
+		}
 	});
 
 	it("should pass response_format to binding for structured output", async () => {

packages/tanstack-ai/test/workers-ai-adapter.test.ts

@@ -998,7 +998,7 @@ describe("WorkersAiTextAdapter config modes", () => {
 
 		expect(binding.run).toHaveBeenCalledOnce();
 		const [, , options] = binding.run.mock.calls[0]!;
-		expect(options).toEqual({
+		expect(options).toMatchObject({
 			extraHeaders: { "x-session-affinity": "my-session-id" },
 		});
 	});
@@ -1015,7 +1015,13 @@ describe("WorkersAiTextAdapter config modes", () => {
 		);
 
 		expect(binding.run).toHaveBeenCalledOnce();
-		const [, , options] = binding.run.mock.calls[0]!;
-		expect(options).toBeUndefined();
+		const [, , options] = binding.run.mock.calls[0]! as [
+			unknown,
+			unknown,
+			Record<string, unknown> | undefined,
+		];
+		if (options) {
+			expect(options).not.toHaveProperty("extraHeaders");
+		}
 	});
 });

packages/workers-ai-provider/src/workersai-chat-language-model.ts

@@ -178,10 +178,11 @@ export class WorkersAIChatLanguageModel implements LanguageModelV3 {
 
 		const output = await this.config.binding.run(
 			args.model as keyof AiModels,
-			// Content arrays for vision are valid at runtime but not in the
-			// binding's strict TypeScript definitions (which expect string content).
 			inputs as AiModels[keyof AiModels]["inputs"],
-			runOptions,
+			{
+				...runOptions,
+				signal: options.abortSignal,
+			} as AiOptions,
 		);
 
 		if (output instanceof ReadableStream) {
@@ -228,7 +229,10 @@ export class WorkersAIChatLanguageModel implements LanguageModelV3 {
 		const response = await this.config.binding.run(
 			args.model as keyof AiModels,
 			inputs as AiModels[keyof AiModels]["inputs"],
-			runOptions,
+			{
+				...runOptions,
+				signal: options.abortSignal,
+			} as AiOptions,
 		);
 
 		// If the binding returned a stream, pipe it through the SSE mapper

packages/workers-ai-provider/src/workersai-embedding-model.ts

@@ -52,7 +52,10 @@ export class WorkersAIEmbeddingModel implements EmbeddingModelV3 {
 		this.config = config;
 	}
 
-	async doEmbed({ values }: EmbeddingModelV3CallOptions): Promise<EmbeddingModelV3Result> {
+	async doEmbed({
+		values,
+		abortSignal,
+	}: EmbeddingModelV3CallOptions): Promise<EmbeddingModelV3Result> {
 		if (values.length > this.maxEmbeddingsPerCall) {
 			throw new TooManyEmbeddingValuesForCallError({
 				maxEmbeddingsPerCall: this.maxEmbeddingsPerCall,
@@ -76,8 +79,9 @@ export class WorkersAIEmbeddingModel implements EmbeddingModelV3 {
 			},
 			{
 				gateway: this.config.gateway ?? gateway,
+				signal: abortSignal,
 				...passthroughOptions,
-			},
+			} as AiOptions,
 		);
 
 		return {

packages/workers-ai-provider/src/workersai-image-model.ts

@@ -31,6 +31,7 @@ export class WorkersAIImageModel implements ImageModelV3 {
 		size,
 		aspectRatio,
 		seed,
+		abortSignal,
 	}: Parameters<ImageModelV3["doGenerate"]>[0]): Promise<
 		Awaited<ReturnType<ImageModelV3["doGenerate"]>>
 	> {
@@ -47,12 +48,19 @@ export class WorkersAIImageModel implements ImageModelV3 {
 		}
 
 		const generateImage = async () => {
-			const output = (await this.config.binding.run(this.modelId as keyof AiModels, {
-				height,
-				prompt: prompt ?? "",
-				seed,
-				width,
-			})) as unknown;
+			const output = (await this.config.binding.run(
+				this.modelId as keyof AiModels,
+				{
+					height,
+					prompt: prompt ?? "",
+					seed,
+					width,
+				},
+				{
+					gateway: this.config.gateway,
+					signal: abortSignal,
+				} as AiOptions,
+			)) as unknown;
 
 			return toUint8Array(output);
 		};