diff --git a/src/prune.ts b/src/prune.ts
index 7f23bef..1331afb 100644
--- a/src/prune.ts
+++ b/src/prune.ts
@@ -97,6 +97,31 @@ const ANSWER_PARTS = new Set(['tool-result', 'tool-error']);
 const anyParts = (message: ModelMessage): Part[] =>
   Array.isArray(message.content) ? (message.content as Part[]) : [];
 
+/**
+ * Every part again with its provider `itemId` gone, so the history goes out inline
+ * rather than as `item_reference` entries pointing at provider-side storage.
+ *
+ * A reference only resolves while the provider still holds that item; once it does
+ * not, the request is rejected with 404 "Item with id '...' not found" and no retry
+ * of the same history can succeed. The content is already in the local history, so
+ * inlining loses nothing.
+ */
+export function detachProviderItems(messages: ModelMessage[]): ModelMessage[] {
+  return messages.map((message) => {
+    const parts = anyParts(message);
+    if (parts.length === 0) return message;
+
+    let changed = false;
+    const next = parts.map((part) => {
+      if (itemId(part) === undefined) return part;
+      changed = true;
+      return withoutItemId(part);
+    });
+
+    return changed ? ({ ...message, content: next } as ModelMessage) : message;
+  });
+}
+
 /**
  * Drops tool results whose tool call is gone.
  *
@@ -178,17 +203,21 @@ export type FitOptions = {
  * request that will be rejected for size.
  */
 export function pruneToFit({ messages, threshold, estimate }: FitOptions): ModelMessage[] {
-  const withoutReasoning = prunePreservingItems({ messages, reasoning: 'all', emptyMessages: 'remove' });
+  const withoutReasoning = detachProviderItems(
+    prunePreservingItems({ messages, reasoning: 'all', emptyMessages: 'remove' }),
+  );
   if (estimate(withoutReasoning) <= threshold) return withoutReasoning;
 
   let narrowest = withoutReasoning;
   for (const keep of KEEP_LADDER) {
-    narrowest = prunePreservingItems({
-      messages,
-      reasoning: 'all',
-      toolCalls: `before-last-${keep}-messages`,
-      emptyMessages: 'remove',
-    });
+    narrowest = detachProviderItems(
+      prunePreservingItems({
+        messages,
+        reasoning: 'all',
+        toolCalls: `before-last-${keep}-messages`,
+        emptyMessages: 'remove',
+      }),
+    );
     if (estimate(narrowest) <= threshold) return narrowest;
   }
   return narrowest;
diff --git a/test/prune.test.ts b/test/prune.test.ts
index fce2ef2..f43e1d4 100644
--- a/test/prune.test.ts
+++ b/test/prune.test.ts
@@ -423,6 +423,21 @@ test('reasoning is always dropped, whatever the threshold', () => {
   expect(JSON.stringify(fitted)).not.toContain('deciding');
 });
 
+test('compaction removes a plain assistant item reference without inline reasoning', () => {
+  const messages: ModelMessage[] = [
+    { role: 'user', content: `question ${'x'.repeat(2000)}` },
+    {
+      role: 'assistant',
+      content: [{ type: 'text', text: 'answer', providerOptions: { openai: { itemId: 'msg_plain' } } }],
+    },
+  ];
+
+  const fitted = pruneToFit({ messages, threshold: 1, estimate });
+
+  expect(itemIds(fitted)).toEqual([]);
+  expect(JSON.stringify(fitted)).toContain('answer');
+});
+
 test('the user prompt survives even the narrowest rung', () => {
   const messages = transcript(200, 4000);
   const fitted = pruneToFit({ messages, threshold: 100, estimate });