ai 7.0.118 → 7.0.120

package/docs/03-ai-sdk-core/45-provider-management.mdx

@@ -79,7 +79,7 @@ import { customProvider, gateway } from 'ai';
 export const anthropic = customProvider({
   languageModels: {
     opus: gateway('anthropic/claude-opus-5.5'),
-    sonnet: gateway('anthropic/claude-sonnet-5'),
+    sonnet: gateway('anthropic/claude-sonnet-5.5'),
     haiku: gateway('anthropic/claude-haiku-4.5'),
   },
   fallbackProvider: gateway,
@@ -100,7 +100,7 @@ import {
 
 export const myProvider = customProvider({
   languageModels: {
-    'text-medium': gateway('anthropic/claude-sonnet-5'),
+    'text-medium': gateway('anthropic/claude-sonnet-5.5'),
     'text-small': gateway('openai/gpt-6-luna'),
     'reasoning-medium': wrapLanguageModel({
       model: gateway('openai/gpt-6-astra'),
@@ -154,7 +154,7 @@ const myOpenAI = customProvider({
 // custom provider with skills interface:
 const myAnthropic = customProvider({
   languageModels: {
-    sonnet: anthropic('claude-sonnet-5'),
+    sonnet: anthropic('claude-sonnet-5-5'),
   },
   skills: anthropic.skills(),
 });
@@ -330,7 +330,7 @@ import { createProviderRegistry, customProvider, uploadSkill } from 'ai';
 
 const registry = createProviderRegistry({
   anthropic: customProvider({
-    languageModels: { sonnet: anthropic('claude-sonnet-5') },
+    languageModels: { sonnet: anthropic('claude-sonnet-5-5') },
     skills: anthropic.skills(),
   }),
 });

package/docs/03-ai-sdk-harnesses/02-harness-agent.mdx

@@ -138,6 +138,39 @@ Harness runtimes execute their built-in tools internally. For those tools,
 lifecycle after the runtime reports the result. The callbacks are still
 delivered before the tool result is published to the result stream.
 
+## Runtime Context
+
+Set `runtimeContext` on `HarnessAgent` to make application context available to
+every turn. Use `prepareCall` to replace it for each new prompt turn. The
+prepared context is available through `generate()` and `stream()` and remains
+the same during in-memory continuations with `continueGenerate()` and
+`continueStream()`. Lifecycle callbacks receive the full context, and it is
+also available on each result's `finalStep.runtimeContext`.
+
+```ts
+const agent = new HarnessAgent({
+  harness: claudeCode,
+  runtimeContext: {
+    conversationId: 'conversation-123',
+    userId: 'user-456',
+  },
+  telemetry: {
+    includeRuntimeContext: {
+      conversationId: true,
+    },
+  },
+  onEnd: ({ runtimeContext }) => {
+    console.log(runtimeContext.userId);
+  },
+});
+```
+
+Runtime context is excluded from telemetry by default. Use
+`telemetry.includeRuntimeContext` to allowlist individual top-level properties;
+only properties set to `true` are sent to telemetry integrations. Lifecycle
+callbacks and results still receive the full context, so avoid storing secrets
+that those consumers should not access.
+
 ## Generate Structured Output
 
 Set `output` when constructing `HarnessAgent` to require the same typed output
@@ -227,7 +260,7 @@ HTTP routes that need multi-turn continuity.
 ## Change Settings Between Turns
 
 Use `callOptionsSchema` and `prepareCall` to derive `model`, `skills`,
-`instructions`, and `tools` for each new turn. This follows the same
+`instructions`, `tools`, and `runtimeContext` for each new turn. This follows the same
 call-options pattern as `ToolLoopAgent`:
 
 ```ts
@@ -253,6 +286,7 @@ const agent = new HarnessAgent({
     ...call,
     model: options.useCheaperModel ? 'claude-haiku-4-5' : undefined,
     instructions: `Work as the ${options.area} specialist.`,
+    runtimeContext: { area: options.area },
     skills: [options.area === 'frontend' ? frontendSkill : backendSkill],
     tools: options.enablePolicyTool ? { getPolicy } : undefined,
   }),
@@ -390,6 +424,7 @@ const session = await agent.createSession({
   sandboxSession,
   // Rebind this when the suspended turn used host-only tool context:
   toolsContext,
+  runtimeContext,
 });
 
 const result = await agent.continueStream({ session });
@@ -398,10 +433,12 @@ const result = await agent.continueStream({ session });
 Use `continueStream()` for incremental output, or `continueGenerate()` to drain
 the continued turn and return a `GenerateTextResult`.
 
-`toolsContext` is intentionally not stored in continuation state because it can
-contain credentials or non-serializable host objects. When a suspended turn
-used static or `prepareCall`-derived tool context, pass the same per-tool map to
-`createSession` to rebind it in the process that continues the turn.
+`toolsContext` and `runtimeContext` are intentionally not stored in
+continuation state because they can contain credentials or non-serializable
+host objects. When a suspended turn used context derived by `prepareCall`,
+pass the same values to `createSession` to rebind them in the process that
+continues the turn. After a process resume without rebinding, the continuation
+uses the agent's constructor-level `runtimeContext` and `toolsContext`.
 
 ## Stop After a Harness Step
 
@@ -481,9 +518,10 @@ const agent = new HarnessAgent({
 ```
 
 `workDir` is optional. When provided, it must be relative to the sandbox's
-default working directory and is used as the session working directory. When
-omitted, regular sessions use the default `<harnessId>-<sessionId>` directory,
-while `onBootstrap` receives the sandbox's default working directory.
+default working directory and is used as the session working directory. Set it
+to `'.'` to use the sandbox's default working directory itself. When omitted,
+regular sessions use the default `<harnessId>-<sessionId>` directory, while
+`onBootstrap` receives the sandbox's default working directory.
 
 ## Prepare Reusable Sandboxes
 
@@ -555,6 +593,8 @@ supplied, their values must match.
 - `stopWhen`: condition(s) for finishing a result slice after a completed
   harness tool step that can continue into another model step.
 - `tools`: AI SDK tools executed by the host when the harness calls them.
+- `runtimeContext`: application context forwarded to lifecycle callbacks and
+  results for every turn; `prepareCall` can replace it for a new prompt turn.
 - `activeTools`: allowlist of built-in and host-executed tools the harness can
   call.
 - `inactiveTools`: denylist of built-in and host-executed tools the harness
@@ -563,7 +603,9 @@ supplied, their values must match.
 - `permissionMode`: built-in tool permission mode.
 - `toolApproval`: approval status map for host-executed tools.
 - `sandboxConfig`: sandbox working-directory and lifecycle hook configuration.
-- `telemetry`, `debug`, and `onLog`: observability and diagnostics.
+- `telemetry`, `debug`, and `onLog`: observability and diagnostics. Use
+  `telemetry.includeRuntimeContext` to allowlist runtime context properties for
+  telemetry.
 
 Telemetry reports each turn's resolved model, instructions, and active
 host-defined tools. Skills are adapter context and do not have a corresponding

package/docs/04-ai-sdk-ui/02-chatbot.mdx

@@ -947,7 +947,7 @@ Check out the [stream protocol guide](/docs/ai-sdk-ui/stream-protocol) for more
 ## Reasoning
 
 Some models such as DeepSeek `deepseek-r1`
-and Anthropic `claude-sonnet-5` support reasoning tokens.
+and Anthropic `claude-sonnet-5-5` support reasoning tokens.
 These tokens are typically sent before the message content.
 You can forward them to the client with the `sendReasoning` option:
 

package/docs/07-reference/01-ai-sdk-core/61-wrap-image-model.mdx

@@ -61,4 +61,19 @@ const { image } = await generateImage({
 
 ### Returns
 
-A new `ImageModelV4` instance with middleware applied.
+A new `ImageModelV4` instance with middleware applied. The wrapped model
+preserves the original model's `supportsFileInputs` and `supportsMaskInputs`
+capabilities by default. Middleware can override them with
+`overrideSupportsFileInputs` and `overrideSupportsMaskInputs`. Returning
+`undefined` from an override intentionally marks that capability as unknown.
+
+```ts
+const model = wrapImageModel({
+  model: originalModel,
+  middleware: {
+    specificationVersion: 'v4',
+    overrideSupportsFileInputs: () => true,
+    overrideSupportsMaskInputs: () => false,
+  },
+});
+```

package/docs/07-reference/04-ai-sdk-workflow/01-workflow-agent.mdx

@@ -17,7 +17,7 @@ import { tool } from 'ai';
 import { z } from 'zod';
 
 const agent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
   instructions: 'You are a helpful assistant.',
   tools: {
     weather: tool({
@@ -71,7 +71,7 @@ To see `WorkflowAgent` in action, check out [these examples](#examples).
       type: 'LanguageModel',
       isRequired: true,
       description:
-        "The language model to use. A string compatible with the Vercel AI Gateway (e.g., 'anthropic/claude-sonnet-5') or a provider instance (e.g., `openai('gpt-6-astra')`).",
+        "The language model to use. A string compatible with the Vercel AI Gateway (e.g., 'anthropic/claude-sonnet-5.5') or a provider instance (e.g., `openai('gpt-6-astra')`).",
     },
     {
       name: 'instructions',
@@ -808,7 +808,7 @@ Infers the UI message type for a `WorkflowAgent` instance. Optionally accepts a
 import { WorkflowAgent, InferWorkflowAgentUIMessage } from '@ai-sdk/workflow';
 
 const agent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
   tools: { weather: weatherTool },
 });
 
@@ -835,7 +835,7 @@ import { tool } from 'ai';
 import { z } from 'zod';
 
 const agent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
   instructions: 'You are a helpful assistant.',
   tools: {
     weather: tool({
@@ -890,7 +890,7 @@ export async function chat(messages: UIMessage[]) {
   const modelMessages = await convertToModelMessages(messages);
 
   const agent = new WorkflowAgent({
-    model: 'anthropic/claude-sonnet-5',
+    model: 'anthropic/claude-sonnet-5.5',
     instructions: 'You are a flight booking assistant.',
     tools: {
       searchFlights: tool({
@@ -937,7 +937,7 @@ import { WorkflowAgent, Output } from '@ai-sdk/workflow';
 import { z } from 'zod';
 
 const analysisAgent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
 });
 
 const result = await analysisAgent.stream({
@@ -977,7 +977,7 @@ import { tool } from 'ai';
 import { z } from 'zod';
 
 const agent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
   experimental_toolApprovalSecret: {
     environmentVariable: 'TOOL_APPROVAL_SECRET',
   },
@@ -1009,7 +1009,7 @@ serialize it. A stream-level reference overrides the constructor value.
 import { WorkflowAgent } from '@ai-sdk/workflow';
 
 const agent = new WorkflowAgent({
-  model: 'anthropic/claude-sonnet-5',
+  model: 'anthropic/claude-sonnet-5.5',
   tools: { weather: weatherTool },
 
   // Agent-wide callbacks

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai",
-  "version": "7.0.118",
+  "version": "7.0.120",
   "type": "module",
   "description": "AI SDK by Vercel - build apps like ChatGPT, Claude, Gemini, and more with a single interface for any model using the Vercel AI Gateway or go direct to OpenAI, Anthropic, Google, or any other model provider.",
   "license": "Apache-2.0",
@@ -42,20 +42,20 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/gateway": "4.0.96",
-    "@ai-sdk/provider": "4.0.18",
-    "@ai-sdk/provider-utils": "5.0.49"
+    "@ai-sdk/gateway": "4.0.98",
+    "@ai-sdk/provider": "4.0.19",
+    "@ai-sdk/provider-utils": "5.0.50"
   },
   "devDependencies": {
-    "@ai-sdk/amazon-bedrock": "5.0.97",
-    "@ai-sdk/deepseek": "3.0.54",
-    "@ai-sdk/google": "4.0.82",
-    "@ai-sdk/groq": "4.0.50",
-    "@ai-sdk/huggingface": "2.0.57",
-    "@ai-sdk/moonshotai": "3.0.58",
-    "@ai-sdk/openai": "4.0.78",
+    "@ai-sdk/amazon-bedrock": "5.0.99",
+    "@ai-sdk/deepseek": "3.0.55",
+    "@ai-sdk/google": "4.0.84",
+    "@ai-sdk/groq": "4.0.51",
+    "@ai-sdk/huggingface": "2.0.58",
+    "@ai-sdk/moonshotai": "3.0.59",
+    "@ai-sdk/openai": "4.0.79",
     "@ai-sdk/test-server": "2.0.2",
-    "@ai-sdk/xai": "5.0.10",
+    "@ai-sdk/xai": "5.0.11",
     "@edge-runtime/vm": "^5.0.0",
     "@smithy/eventstream-codec": "^4.3.3",
     "@smithy/util-utf8": "^4.3.3",

package/src/middleware/wrap-image-model.ts

@@ -48,6 +48,8 @@ const doWrap = ({
     overrideProvider,
     overrideModelId,
     overrideMaxImagesPerCall,
+    overrideSupportsFileInputs,
+    overrideSupportsMaskInputs,
   },
   modelId,
   providerId,
@@ -71,11 +73,22 @@ const doWrap = ({
       ? maxImagesPerCallRaw.bind(model)
       : maxImagesPerCallRaw;
 
+  const supportsFileInputs =
+    overrideSupportsFileInputs !== undefined
+      ? overrideSupportsFileInputs({ model })
+      : model.supportsFileInputs;
+  const supportsMaskInputs =
+    overrideSupportsMaskInputs !== undefined
+      ? overrideSupportsMaskInputs({ model })
+      : model.supportsMaskInputs;
+
   return {
     specificationVersion: 'v4',
     provider: providerId ?? overrideProvider?.({ model }) ?? model.provider,
     modelId: modelId ?? overrideModelId?.({ model }) ?? model.modelId,
     maxImagesPerCall,
+    supportsFileInputs,
+    supportsMaskInputs,
     async doGenerate(
       params: ImageModelV4CallOptions,
     ): Promise<ImageModelV4Result> {

package/src/telemetry/tracing-channel-publisher.ts

@@ -136,6 +136,7 @@ export function openTelemetryChannelSpanContext({
   completion: PromiseLike<unknown>;
 }): TracingChannelContext | undefined {
   if (!isNodeRuntime()) {
+    Promise.resolve(completion).catch(() => {});
     return undefined;
   }
 

package/src/test/mock-image-model-v4.ts

@@ -6,6 +6,8 @@ export class MockImageModelV4 implements ImageModelV4 {
   readonly provider: ImageModelV4['provider'];
   readonly modelId: ImageModelV4['modelId'];
   readonly maxImagesPerCall: ImageModelV4['maxImagesPerCall'];
+  readonly supportsFileInputs: ImageModelV4['supportsFileInputs'];
+  readonly supportsMaskInputs: ImageModelV4['supportsMaskInputs'];
 
   doGenerate: ImageModelV4['doGenerate'];
 
@@ -13,16 +15,22 @@ export class MockImageModelV4 implements ImageModelV4 {
     provider = 'mock-provider',
     modelId = 'mock-model-id',
     maxImagesPerCall = 1,
+    supportsFileInputs,
+    supportsMaskInputs,
     doGenerate = notImplemented,
   }: {
     provider?: ImageModelV4['provider'];
     modelId?: ImageModelV4['modelId'];
     maxImagesPerCall?: ImageModelV4['maxImagesPerCall'];
+    supportsFileInputs?: ImageModelV4['supportsFileInputs'];
+    supportsMaskInputs?: ImageModelV4['supportsMaskInputs'];
     doGenerate?: ImageModelV4['doGenerate'];
   } = {}) {
     this.provider = provider;
     this.modelId = modelId;
     this.maxImagesPerCall = maxImagesPerCall;
+    this.supportsFileInputs = supportsFileInputs;
+    this.supportsMaskInputs = supportsMaskInputs;
     this.doGenerate = doGenerate;
   }
 }

package/src/ui/chat.ts

@@ -263,6 +263,9 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
   private pendingApprovalMessageId: string | undefined;
   private activeResponse: ActiveResponse<UI_MESSAGE> | undefined = undefined;
   private activeResumeRequest: ActiveResumeRequest | undefined = undefined;
+  private resumableStreamState:
+    | StreamingUIMessageState<UI_MESSAGE>
+    | undefined = undefined;
   private jobExecutor = new SerialJobExecutor();
 
   constructor({
@@ -723,6 +726,10 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
     trigger: 'submit-message' | 'resume-stream' | 'regenerate-message';
     messageId?: string;
   } & ChatRequestOptions) {
+    if (trigger !== 'resume-stream') {
+      this.resumableStreamState = undefined;
+    }
+
     const abortController = new AbortController();
     const activeResumeRequest =
       trigger === 'resume-stream' ? { abortController } : undefined;
@@ -760,6 +767,7 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
           await reconnect?.cancel().catch(() => {});
           if (isCurrentRequest()) {
             this.setStatus({ status: 'ready' });
+            this.resumableStreamState = undefined;
           }
           clearActiveResumeRequest();
           return;
@@ -767,6 +775,7 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
 
         if (reconnect == null) {
           this.setStatus({ status: 'ready' });
+          this.resumableStreamState = undefined;
           clearActiveResumeRequest();
           return; // no active stream found, so we do not resume
         }
@@ -779,6 +788,7 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
         ) {
           if (isCurrentRequest()) {
             this.setStatus({ status: 'ready' });
+            this.resumableStreamState = undefined;
           }
           clearActiveResumeRequest();
           return;
@@ -820,13 +830,17 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
 
     try {
       const response = {
-        state: createStreamingUIMessageState({
-          lastMessage:
-            trigger === 'resume-stream' || trigger === 'regenerate-message'
-              ? undefined
-              : this.state.snapshot(responseMessage),
-          messageId: this.generateId(),
-        }),
+        state:
+          trigger === 'resume-stream' && this.resumableStreamState != null
+            ? this.resumableStreamState
+            : createStreamingUIMessageState({
+                lastMessage:
+                  trigger === 'resume-stream' ||
+                  trigger === 'regenerate-message'
+                    ? undefined
+                    : this.state.snapshot(responseMessage),
+                messageId: this.generateId(),
+              }),
         abortController,
       } as ActiveResponse<UI_MESSAGE>;
 
@@ -916,12 +930,18 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
       if (isAbort) {
         if (isCurrentRequest()) {
           this.setStatus({ status: 'ready' });
+          if (this.resumableStreamState === response.state) {
+            this.resumableStreamState = undefined;
+          }
         }
         return null;
       }
 
       if (isCurrentRequest()) {
         this.setStatus({ status: 'ready' });
+        if (this.resumableStreamState === response.state) {
+          this.resumableStreamState = undefined;
+        }
       }
     } catch (err) {
       // Ignore abort errors as they are expected.
@@ -929,6 +949,9 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
         isAbort = true;
         if (isCurrentRequest()) {
           this.setStatus({ status: 'ready' });
+          if (this.resumableStreamState === activeResponse?.state) {
+            this.resumableStreamState = undefined;
+          }
         }
         return null;
       }
@@ -948,6 +971,12 @@ export abstract class AbstractChat<UI_MESSAGE extends UIMessage> {
         isDisconnect = true;
       }
 
+      if (isDisconnect) {
+        this.resumableStreamState = activeResponse?.state;
+      } else if (this.resumableStreamState === activeResponse?.state) {
+        this.resumableStreamState = undefined;
+      }
+
       if (this.onError && err instanceof Error) {
         this.onError(err);
       }

package/src/ui/process-ui-message-stream.ts

@@ -31,8 +31,6 @@ import {
   type UIMessage,
   type UIMessagePart,
 } from './ui-messages';
-import { warnIfUIMessageHasDeprecatedRawInput } from './warn-if-ui-message-has-deprecated-raw-input';
-
 export type StreamingUIMessageState<UI_MESSAGE extends UIMessage> = {
   message: UI_MESSAGE;
   activeTextParts: Record<string, TextUIPart>;
@@ -732,15 +730,12 @@ export function processUIMessageStream<UI_MESSAGE extends UIMessage>({
                   toolCallId: chunk.toolCallId,
                   toolName: chunk.toolName,
                   state: 'output-error',
-                  input: undefined,
-                  rawInput: chunk.input,
+                  input: chunk.input,
                   errorText: chunk.errorText,
                   providerExecuted: chunk.providerExecuted,
                   providerMetadata: chunk.providerMetadata,
                   toolMetadata: chunk.toolMetadata,
                 });
-
-                warnIfUIMessageHasDeprecatedRawInput([state.message]);
               }
 
               write();
@@ -821,7 +816,8 @@ export function processUIMessageStream<UI_MESSAGE extends UIMessage>({
                     providerExecuted: chunk.providerExecuted,
                     providerMetadata: chunk.providerMetadata,
                     title: toolInvocation.title,
-                    toolMetadata: toolInvocation.toolMetadata,
+                    toolMetadata:
+                      chunk.toolMetadata ?? toolInvocation.toolMetadata,
                   },
                   toolInvocation,
                 );
@@ -837,7 +833,8 @@ export function processUIMessageStream<UI_MESSAGE extends UIMessage>({
                     preliminary: chunk.preliminary,
                     providerMetadata: chunk.providerMetadata,
                     title: toolInvocation.title,
-                    toolMetadata: toolInvocation.toolMetadata,
+                    toolMetadata:
+                      chunk.toolMetadata ?? toolInvocation.toolMetadata,
                   },
                   toolInvocation as ToolUIPart<InferUIMessageTools<UI_MESSAGE>>,
                 );
@@ -861,7 +858,8 @@ export function processUIMessageStream<UI_MESSAGE extends UIMessage>({
                     providerExecuted: chunk.providerExecuted,
                     providerMetadata: chunk.providerMetadata,
                     title: toolInvocation.title,
-                    toolMetadata: toolInvocation.toolMetadata,
+                    toolMetadata:
+                      chunk.toolMetadata ?? toolInvocation.toolMetadata,
                   },
                   toolInvocation,
                 );
@@ -877,7 +875,8 @@ export function processUIMessageStream<UI_MESSAGE extends UIMessage>({
                     providerExecuted: chunk.providerExecuted,
                     providerMetadata: chunk.providerMetadata,
                     title: toolInvocation.title,
-                    toolMetadata: toolInvocation.toolMetadata,
+                    toolMetadata:
+                      chunk.toolMetadata ?? toolInvocation.toolMetadata,
                   },
                   toolInvocation as ToolUIPart<InferUIMessageTools<UI_MESSAGE>>,
                 );

package/src/util/write-to-server-response.ts

@@ -6,6 +6,10 @@ type FlushableServerResponse = ServerResponse & {
 
 /**
  * Writes the content of a stream to a server response.
+ *
+ * When the client disconnects before the stream has been fully written
+ * (premature `close`), the stream is cancelled so that upstream resources
+ * can be released, and no further chunks are written.
  */
 export function writeToServerResponse({
   response,
@@ -32,11 +36,36 @@ export function writeToServerResponse({
   }
 
   const reader = stream.getReader();
+
+  // Detect client disconnects. `close` also fires after a regular `end()`;
+  // `writableFinished` distinguishes the two cases.
+  let clientDisconnected = false;
+  let onDisconnect: (() => void) | undefined;
+  const handleClose = () => {
+    if (response.writableFinished) {
+      return;
+    }
+
+    clientDisconnected = true;
+    onDisconnect?.();
+
+    // cancelling the reader resolves a pending read with `done: true`:
+    reader.cancel(new Error('Client disconnected.')).catch(() => {});
+  };
+  response.once('close', handleClose);
+
+  // The response may have been destroyed before the close listener was
+  // registered. Cancel immediately instead of waiting indefinitely for the
+  // next stream chunk.
+  if (response.destroyed) {
+    handleClose();
+  }
+
   const read = async () => {
     try {
       while (true) {
         const { done, value } = await reader.read();
-        if (done) break;
+        if (done || clientDisconnected || response.destroyed) break;
 
         // Respect backpressure: if write() returns false, wait for 'drain' event
         const canContinue = response.write(value);
@@ -47,12 +76,22 @@ export function writeToServerResponse({
 
         if (!canContinue) {
           await new Promise<void>(resolve => {
+            // don't wait for `drain` on a disconnected response:
+            onDisconnect = resolve;
             response.once('drain', resolve);
           });
+          onDisconnect = undefined;
         }
       }
     } finally {
-      response.end();
+      response.off('close', handleClose);
+
+      if (clientDisconnected || response.destroyed) {
+        // release the stream; ending a destroyed response is not possible:
+        reader.cancel().catch(() => {});
+      } else {
+        response.end();
+      }
     }
   };