diff --git a/src/chat/inbound.ts b/src/chat/inbound.ts
index bdf7b30551..3916d9afa1 100644
--- a/src/chat/inbound.ts
+++ b/src/chat/inbound.ts
@@ -310,10 +310,16 @@ export function chatCompletionsToResponsesBody(raw: unknown): Rec {
         // Responses assistant item schema admits only output content blocks, so there
         // is no attachment point on the message itself, and the parser buffers a
         // reasoning item and prepends it to the NEXT assistant message. Emitting it
-        // here keeps that adjacency intact.
+        // here keeps that adjacency intact. OpenAI requires `summary` on replayed
+        // reasoning items; keep it empty because this is raw provider reasoning, not
+        // a caller-authorized user-visible summary.
         const reasoningText = assistantReasoningText(msg);
         if (reasoningText !== undefined) {
-          input.push({ type: "reasoning", content: [{ type: "reasoning_text", text: reasoningText }] });
+          input.push({
+            type: "reasoning",
+            summary: [],
+            content: [{ type: "reasoning_text", text: reasoningText }],
+          });
         }
         const blocks = assistantContentToBlocks(msg.content);
         if (blocks.length > 0) input.push({ type: "message", role: "assistant", content: blocks });
diff --git a/structure/data-planes/inbound-compat.md b/structure/data-planes/inbound-compat.md
index aa9aa15f52..e5eea065de 100644
--- a/structure/data-planes/inbound-compat.md
+++ b/structure/data-planes/inbound-compat.md
@@ -303,7 +303,9 @@ An assistant turn's `reasoning_content` or `reasoning_details` is carried into t
 projection as a `reasoning` input item emitted immediately before its assistant
 message, matching the parser's buffer-and-prepend adjacency. Only representable
 plaintext crosses: no signature, encrypted payload or provider item id is
-reconstructed, because those attest to content this proxy never received. Opaque
+reconstructed, because those attest to content this proxy never received. The item
+always carries `summary: []`, which satisfies the Responses replay shape without
+publishing raw provider reasoning as a user-visible summary. Opaque
 reasoning replay across a Chat boundary remains unimplemented by design.
 `presence_penalty` and `frequency_penalty` are carried too; per-model
 `noPenaltyModels` opt-outs still apply at the adapter.
diff --git a/tests/responses/chat-inbound-reasoning-replay.test.ts b/tests/responses/chat-inbound-reasoning-replay.test.ts
index de78b3eaaa..a5044fb600 100644
--- a/tests/responses/chat-inbound-reasoning-replay.test.ts
+++ b/tests/responses/chat-inbound-reasoning-replay.test.ts
@@ -35,6 +35,7 @@ describe("F6 assistant reasoning survives translation", () => {
     const idx = out.findIndex(i => i.type === "reasoning");
 
     expect(idx).toBeGreaterThanOrEqual(0);
+    expect(out[idx]!.summary).toEqual([]);
     expect(out[idx]!.content).toEqual([{ type: "reasoning_text", text: "prior analysis" }]);
     // Adjacency matters: the parser prepends a buffered reasoning item to the NEXT
     // assistant message, so it must sit immediately before it.
@@ -57,6 +58,9 @@ describe("F6 assistant reasoning survives translation", () => {
   test("no signature, encrypted payload or item id is forged", () => {
     const item = items(body([USER, { role: "assistant", content: "a", reasoning_content: "t" }])).find(i => i.type === "reasoning")!;
 
+    // Raw provider reasoning must not be reclassified as a visible summary merely to
+    // satisfy the Responses wire requirement.
+    expect(item.summary).toEqual([]);
     expect(item.signature).toBeUndefined();
     expect(item.encrypted_content).toBeUndefined();
     expect(item.id).toBeUndefined();