diff --git a/apps/server/src/pullRequest/PullRequestService.test.ts b/apps/server/src/pullRequest/PullRequestService.test.ts
index 8e9262a081048c30e58d7fe380c26ae7bad06cb6..9d563d3965c33b05c18680291ea5f409aef6057f 100644
--- a/apps/server/src/pullRequest/PullRequestService.test.ts
+++ b/apps/server/src/pullRequest/PullRequestService.test.ts
@@ -5382,7 +5382,7 @@ it.effect("keeps the diff cached across a file being ticked off", () =>
     yield* service.diff(reference);
     yield* service.filesViewed(reference);
 
-    // The press forgets only the reader's own ticks: a diff of any size survives it.
+    // The press forgets only the reader's own ticks; cached diffs survive it.
     assert.strictEqual(diffReads, 1);
     assert.strictEqual(viewedReads, 2);
 
@@ -5397,6 +5397,72 @@ it.effect("keeps the diff cached across a file being ticked off", () =>
   }),
 );
 
+it.effect("returns large diff slices intact without retaining them in either cache", () =>
+  Effect.gen(function* () {
+    let reads = 0;
+    const patch = "\u{1f4bb}".repeat(140_000);
+    const service = yield* makeService({
+      projects: [project({ id: "p1", title: "web", workspaceRoot: "/a", repository: "acme/web" })],
+      providers: [
+        fakeProvider("github", {
+          getDiff: () =>
+            Effect.sync(() => {
+              reads += 1;
+              return { patch, truncated: false, nextCursor: "2" };
+            }),
+        }),
+      ],
+    });
+    const reference = { projectId: "p1" as ProjectId, repository: "acme/web", number: 1 };
+    for (const input of [
+      reference,
+      { ...reference, cursor: "2" },
+      { ...reference, commit: "a".repeat(40) },
+    ]) {
+      const before = reads;
+      assert.deepStrictEqual(yield* service.diff(input), {
+        patch,
+        truncated: false,
+        nextCursor: "2",
+      });
+      assert.deepStrictEqual(yield* service.diff(input), {
+        patch,
+        truncated: false,
+        nextCursor: "2",
+      });
+      assert.strictEqual(reads, before + 2);
+    }
+  }),
+);
+
+it.effect("caches a small replacement after releasing a large diff", () =>
+  Effect.gen(function* () {
+    let reads = 0;
+    const largePatch = "x".repeat(300_000);
+    const service = yield* makeService({
+      projects: [project({ id: "p1", title: "web", workspaceRoot: "/a", repository: "acme/web" })],
+      providers: [
+        fakeProvider("github", {
+          getDiff: () =>
+            Effect.sync(() => {
+              reads += 1;
+              return {
+                patch: reads === 1 ? largePatch : "@@ small replacement",
+                truncated: false,
+                nextCursor: null,
+              };
+            }),
+        }),
+      ],
+    });
+    const reference = { projectId: "p1" as ProjectId, repository: "acme/web", number: 1 };
+    assert.strictEqual((yield* service.diff(reference)).patch, largePatch);
+    assert.strictEqual((yield* service.diff(reference)).patch, "@@ small replacement");
+    assert.strictEqual((yield* service.diff(reference)).patch, "@@ small replacement");
+    assert.strictEqual(reads, 2);
+  }),
+);
+
 const environmentViewedProvider = (
   revisions: Map<string, string>,
   asked: Array<ReadonlyArray<string>>,
diff --git a/apps/server/src/pullRequest/PullRequestService.ts b/apps/server/src/pullRequest/PullRequestService.ts
index 2cc1180bb826633d2d097ea8b3a204bc33cfbf2d..a653ef94c713f5c76cc48543df19556911da79e5 100644
--- a/apps/server/src/pullRequest/PullRequestService.ts
+++ b/apps/server/src/pullRequest/PullRequestService.ts
@@ -151,6 +151,10 @@ const LIST_CACHE_CAPACITY = 64;
 const LIST_STATS_CACHE_CAPACITY = 32;
 const DETAIL_CACHE_CAPACITY = 128;
 const DIFF_CACHE_CAPACITY = 128;
+// Each diff cache can retain at most 64 MiB of patch text, counting UTF-16 storage.
+const MAX_CACHED_DIFF_PATCH_BYTES = 512 * 1024;
+const canCacheDiff = (value: PullRequestDiffResult) =>
+  value.patch.length * 2 <= MAX_CACHED_DIFF_PATCH_BYTES;
 const FILES_VIEWED_CACHE_CAPACITY = 128;
 const VIEWER_CACHE_CAPACITY = 32;
 
@@ -2507,6 +2511,7 @@ export const make = Effect.gen(function* () {
     const record = (key: string, value: PullRequestDiffResult) =>
       Effect.map(Clock.currentTimeMillis, (at) => {
         held.delete(key);
+        if (!canCacheDiff(value)) return;
         if (held.size >= DIFF_CACHE_CAPACITY) {
           const oldest = held.keys().next().value;
           if (oldest !== undefined) held.delete(oldest);
@@ -2977,7 +2982,21 @@ export const make = Effect.gen(function* () {
         ? (lastGoodSummary.peek(refCacheKey(input))?.updatedAt ?? null)
         : null,
     ]);
-    return staleDiff(key, Cache.get(diffCache, key));
+    const read = Cache.get(diffCache, key).pipe(
+      Effect.tap((value) =>
+        canCacheDiff(value)
+          ? Effect.void
+          : Cache.getSuccess(diffCache, key).pipe(
+              Effect.flatMap((current) =>
+                Option.isSome(current) && current.value === value
+                  ? Cache.invalidate(diffCache, key)
+                  : Effect.void,
+              ),
+              Effect.uninterruptible,
+            ),
+      ),
+    );
+    return staleDiff(key, read);
   };
 
   const filesViewedCache = yield* Cache.makeWith(