diff --git a/packages/llm/src/model.middleware.test.ts b/packages/llm/src/model.middleware.test.ts
index 7be7c8e..0578e2d 100644
--- a/packages/llm/src/model.middleware.test.ts
+++ b/packages/llm/src/model.middleware.test.ts
@@ -1,10 +1,17 @@
+import { dir } from "@synstack/fs";
 import { fsCache } from "@synstack/fs-cache";
 import { MockLanguageModelV1, simulateReadableStream } from "ai/test";
 import assert from "node:assert/strict";
-import { describe, it } from "node:test";
+import { afterEach, describe, it } from "node:test";
 import type { Llm } from "./llm.types.ts";
 import { cacheCalls, includeAssistantMessage } from "./model.middleware.ts";
 
+const TMP_DIR = dir(import.meta.dirname).to("tmp");
+
+afterEach(async () => {
+  await TMP_DIR.rm();
+});
+
 describe("includeAssistantMessage", () => {
   it("adds the last assistant message to a generated output", async () => {
     const model = new MockLanguageModelV1({
@@ -155,4 +162,61 @@ describe("cache", () => {
 
     assert.equal(output, " world!");
   });
+
+  it("caches a streamed response on miss", async () => {
+    const cache = fsCache(TMP_DIR.path).key(["stream-miss"]);
+
+    let callCount = 0;
+    const model = new MockLanguageModelV1({
+      doStream: () => {
+        callCount++;
+        return Promise.resolve({
+          rawCall: { rawPrompt: null, rawSettings: {} },
+          stream: simulateReadableStream({
+            chunks: [
+              { type: "text-delta", textDelta: "Hello" },
+              { type: "text-delta", textDelta: " " },
+              { type: "text-delta", textDelta: "world" },
+              {
+                type: "finish",
+                finishReason: "stop",
+                usage: { promptTokens: 10, completionTokens: 20 },
+              },
+            ] satisfies Array<Llm.Model.Stream.Part>,
+            chunkDelayInMs: 0,
+          }),
+        });
+      },
+    });
+
+    const wrappedModel = cacheCalls(cache)(model);
+
+    const params: Llm.Model.Stream.Options = {
+      inputFormat: "messages",
+      mode: { type: "regular" },
+      prompt: [
+        {
+          role: "user",
+          content: [
+            {
+              type: "text",
+              text: "Hello",
+            },
+          ],
+        },
+      ],
+    };
+
+    const res1 = await wrappedModel.doStream(params);
+
+    let output1 = "";
+    for await (const chunk of res1.stream) {
+      if (chunk.type !== "text-delta") continue;
+      output1 += chunk.textDelta;
+    }
+
+    assert.equal(output1, "Hello world");
+    // Ensure the stream can be consumed without throwing and collects all parts
+    assert.equal(callCount, 1);
+  });
 });
diff --git a/packages/llm/src/model.middleware.ts b/packages/llm/src/model.middleware.ts
index a71a795..91e8b3f 100644
--- a/packages/llm/src/model.middleware.ts
+++ b/packages/llm/src/model.middleware.ts
@@ -138,7 +138,7 @@ export const cacheCalls =
           // We need to collect the parts and then set the cache
           const res = await doStream();
           const collector: Array<Llm.Model.Stream.Part> = [];
-          res.stream.pipeThrough(
+          const piped = res.stream.pipeThrough(
             new TransformStream<Llm.Model.Stream.Part, Llm.Model.Stream.Part>({
               transform(chunk, controller) {
                 collector.push(chunk);
@@ -150,7 +150,10 @@ export const cacheCalls =
             ...res,
             stream: collector,
           });
-          return res;
+          return {
+            ...res,
+            stream: piped,
+          };
         },
       },
     });