fix(sse): require Responses-shaped body before native OpenAI-compatible passthrough (#12129)

The internal context-handoff/universal-handoff summary request is built in Chat
Completions shape and dispatched through the same handleSingleModel closure
that still carries the original client request's /responses endpoint. This
made resolveChatCoreRequestFormat resolve sourceFormat to "openai-responses"
purely from the inherited endpoint path, and
shouldUseNativeOpenAICompatibleResponsesPassthrough then skipped chat->responses
translation entirely for any openai-compatible-* connection with
apiType: "responses", regardless of the actual body shape. The upstream ended
up receiving the raw Chat Completions body (messages) on /v1/responses with no
input, so the request was rejected outright.

shouldUseNativeOpenAICompatibleResponsesPassthrough now takes the request body
and requires it to actually look Responses-shaped (input present, messages
absent) before allowing the native-passthrough fast path. A genuine client
Responses-API request still gets the zero-translation shortcut; an
internally-synthesized chat-shaped body is now routed through the normal
chat->responses translation layer instead.

changelog.d/fixes/12129-context-handoff-native-passthrough-shape.md

@@ -0,0 +1 @@
+- fix(sse): require Responses-shaped body before native OpenAI-compatible passthrough (#12129)

open-sse/handlers/chatCore.ts

@@ -731,6 +731,7 @@ export async function handleChatCore({
       sourceFormat,
       endpointPath,
       providerSpecificData: credentials?.providerSpecificData,
+      body,
     });
   const responsesInputItems = Array.isArray(body?.input) ? body.input : [];
   const customToolNames = collectCustomToolNamesForSourceFormat(

open-sse/handlers/chatCore/passthroughHelpers.ts

@@ -53,19 +53,35 @@ export function stampNativeResponsesPassthroughBody(
   return { ...body, _nativeOpenAICompatibleResponsesPassthrough: true };
 }
 
+// A body only qualifies for the native-Responses passthrough fast path when it is
+// actually shaped like a Responses API request (`input`, no `messages`). Endpoint
+// path alone is not sufficient: an internally-synthesized Chat Completions-shaped
+// body (e.g. the context-handoff summary request) can be dispatched through a
+// closure that still carries the original client request's `/responses` endpoint,
+// which otherwise makes `sourceFormat` resolve to "openai-responses" even though
+// the body itself was never translated. See issue #12129.
+function isResponsesShapedBody(body: unknown): boolean {
+  if (!body || typeof body !== "object") return false;
+  const candidate = body as Record<string, unknown>;
+  return candidate.input !== undefined && candidate.messages === undefined;
+}
+
 export function shouldUseNativeOpenAICompatibleResponsesPassthrough({
   provider,
   sourceFormat,
   endpointPath,
   providerSpecificData,
+  body,
 }: {
   provider?: string | null;
   sourceFormat?: string | null;
   endpointPath?: string | null;
   providerSpecificData?: unknown;
+  body?: unknown;
 }): boolean {
   if (!provider?.startsWith("openai-compatible-")) return false;
   if (sourceFormat !== FORMATS.OPENAI_RESPONSES) return false;
+  if (body !== undefined && !isResponsesShapedBody(body)) return false;
   if (providerSpecificData && typeof providerSpecificData === "object") {
     const psd = providerSpecificData as Record<string, unknown>;
     if (psd.apiType === "responses" || psd._omnirouteForceResponsesUpstream === true) {

tests/unit/context-handoff-native-passthrough-bug.test.ts

@@ -0,0 +1,86 @@
+// Regression test for issue #12129: an internal context-handoff summary request (built in
+// Chat Completions shape -- `messages`, no `input`) is dispatched through the SAME
+// handleSingleModel closure that carries the ORIGINAL client request's endpoint.
+// When that original endpoint matched `/responses` and the resolved handoff-model
+// provider is an openai-compatible-* connection configured with apiType "responses",
+// the pipeline used to decide the body was already native-Responses-shaped and skip
+// chat->responses translation entirely (`_nativeOpenAICompatibleResponsesPassthrough`),
+// so the upstream received `messages` on `/v1/responses` and rejected it with zero input.
+//
+// Fix: `shouldUseNativeOpenAICompatibleResponsesPassthrough` now requires the body to
+// actually look Responses-shaped (`input` present, `messages` absent) before allowing
+// the passthrough fast path, so an internally-synthesized chat-shaped body is routed
+// through the normal chat->responses translation layer instead.
+import assert from "node:assert/strict";
+import { test } from "node:test";
+
+import { resolveChatCoreRequestFormat } from "../../open-sse/handlers/chatCore/requestFormat.ts";
+import { shouldUseNativeOpenAICompatibleResponsesPassthrough } from "../../open-sse/handlers/chatCore/passthroughHelpers.ts";
+
+test("internal chat-shaped handoff body is no longer treated as native Responses passthrough", () => {
+  const clientRawRequest = {
+    endpoint: "/v1/responses",
+    headers: new Headers(),
+  };
+
+  const summaryBody = {
+    model: "some-handoff-model",
+    messages: [{ role: "user", content: "Summarize this conversation." }],
+    stream: false,
+    max_tokens: 800,
+    temperature: 0.1,
+    _omnirouteSkipContextRelay: true,
+    _omnirouteInternalRequest: "context-handoff",
+  };
+
+  const { sourceFormat, endpointPath } = resolveChatCoreRequestFormat({
+    clientRawRequest,
+    body: summaryBody,
+    provider: "openai-compatible-responses-cliproxy",
+    userAgent: null,
+  });
+
+  assert.equal(sourceFormat, "openai-responses");
+  assert.equal(endpointPath, "/v1/responses");
+
+  const providerSpecificData = { apiType: "responses" };
+
+  const nativePassthrough = shouldUseNativeOpenAICompatibleResponsesPassthrough({
+    provider: "openai-compatible-responses-cliproxy",
+    sourceFormat,
+    endpointPath,
+    providerSpecificData,
+    body: summaryBody,
+  });
+
+  assert.equal(
+    nativePassthrough,
+    false,
+    "fixed: chat-shaped internal body must not take the native-Responses passthrough shortcut"
+  );
+
+  assert.equal((summaryBody as Record<string, unknown>).input, undefined);
+  assert.ok(Array.isArray(summaryBody.messages) && summaryBody.messages.length > 0);
+});
+
+test("genuine Responses-shaped body still takes the native passthrough fast path", () => {
+  const genuineResponsesBody = {
+    model: "gpt-5.6-sol",
+    input: [{ role: "user", content: [{ type: "input_text", text: "Hello" }] }],
+    stream: false,
+  };
+
+  const nativePassthrough = shouldUseNativeOpenAICompatibleResponsesPassthrough({
+    provider: "openai-compatible-responses-cliproxy",
+    sourceFormat: "openai-responses",
+    endpointPath: "/v1/responses",
+    providerSpecificData: { apiType: "responses" },
+    body: genuineResponsesBody,
+  });
+
+  assert.equal(
+    nativePassthrough,
+    true,
+    "a genuine Responses-shaped client body must keep the zero-translation fast path"
+  );
+});