fix: bypass multimodal endpoint memory for gemini

src/server/proxy-core/surfaces/geminiSurface.ts

@@ -367,6 +367,11 @@ export async function geminiProxyRoute(app: FastifyInstance) {
     const isCountTokensAction = isGeminiCliDownstream
       ? options?.action === 'countTokens'
       : modelActionPath.endsWith(':countTokens');
+    const rawUrl = request.raw.url || request.url || '';
+    const wantsSseEnvelope = (
+      isStreamAction
+      && (isGeminiCliDownstream || /(?:^|[?&])alt=sse(?:&|$)/i.test(rawUrl))
+    );
     if (!requestedModel) {
       return reply.code(400).send({
         error: { message: 'Gemini model path is required', type: 'invalid_request_error' },
@@ -857,11 +862,18 @@ export async function geminiProxyRoute(app: FastifyInstance) {
           parsedUsage.completionTokens,
           parsedUsage.totalTokens,
         );
-        return reply.code(upstream.status).send(
+        const downstreamPayload = isGeminiCliDownstream
-          isGeminiCliDownstream
           ? { response: geminiResponse }
-            : geminiResponse,
+          : geminiResponse;
-        );
+        if (wantsSseEnvelope) {
+          // Some compatibility upstreams finish stream requests with a final JSON payload.
+          // Preserve Gemini streaming UX by wrapping that terminal payload back into one SSE event.
+          return reply
+            .code(upstream.status)
+            .type('text/event-stream; charset=utf-8')
+            .send(geminiGenerateContentTransformer.stream.serializeSsePayload(downstreamPayload));
+        }
+        return reply.code(upstream.status).send(downstreamPayload);
       } catch (error) {
         lastStatus = 502;
         lastContentType = 'application/json';

src/server/routes/proxy/gemini.test.ts

@@ -602,6 +602,81 @@ describe('gemini native proxy routes', () => {
     });
   });
 
+  it('serializes non-streaming generic upstream JSON into Gemini SSE when alt=sse is requested', async () => {
+    selectChannelMock.mockReturnValue({
+      channel: { id: 41, routeId: 22 },
+      site: { id: 77, name: 'openai-site', url: 'https://api.openai.com', platform: 'openai' },
+      account: { id: 37, username: 'openai-user@example.com' },
+      tokenName: 'default',
+      tokenValue: 'openai-access-token',
+      actualModel: 'gpt-4.1',
+    });
+    fetchMock.mockResolvedValue(new Response(JSON.stringify({
+      id: 'chatcmpl-openai-stream-1',
+      object: 'chat.completion',
+      created: 1_742_160_003,
+      model: 'gpt-4.1',
+      choices: [
+        {
+          index: 0,
+          message: {
+            role: 'assistant',
+            content: 'hello from openai stream fallback',
+          },
+          finish_reason: 'stop',
+        },
+      ],
+      usage: {
+        prompt_tokens: 5,
+        completion_tokens: 6,
+        total_tokens: 11,
+      },
+    }), {
+      status: 200,
+      headers: { 'content-type': 'application/json' },
+    }));
+
+    const response = await app.inject({
+      method: 'POST',
+      url: '/v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse',
+      headers: {
+        authorization: 'Bearer sk-managed-gemini',
+      },
+      payload: {
+        contents: [
+          {
+            role: 'user',
+            parts: [{ text: 'hello' }],
+          },
+        ],
+      },
+    });
+
+    expect(response.statusCode).toBe(200);
+    expect(response.headers['content-type']).toContain('text/event-stream');
+    expect(parseSsePayloads(response.body)).toEqual([
+      {
+        responseId: 'chatcmpl-openai-stream-1',
+        modelVersion: 'gpt-4.1',
+        candidates: [
+          {
+            index: 0,
+            content: {
+              role: 'model',
+              parts: [{ text: 'hello from openai stream fallback' }],
+            },
+            finishReason: 'STOP',
+          },
+        ],
+        usageMetadata: {
+          promptTokenCount: 5,
+          candidatesTokenCount: 6,
+          totalTokenCount: 11,
+        },
+      },
+    ]);
+  });
+
   it('exposes GeminiCLI downstream generateContent endpoint and wraps the downstream response payload', async () => {
     selectChannelMock.mockReturnValue({
       channel: { id: 42, routeId: 22 },

src/server/routes/proxy/upstreamEndpoint.test.ts

@@ -196,6 +196,82 @@ describe('resolveUpstreamEndpointCandidates', () => {
     expect(order).toEqual(['responses', 'messages', 'chat']);
   });
 
+  it('does not apply runtime endpoint memory to image attachments', async () => {
+    recordUpstreamEndpointSuccess({
+      siteId: baseContext.site.id,
+      endpoint: 'responses',
+      downstreamFormat: 'openai',
+      modelName: 'gpt-5.3',
+      requestCapabilities: {
+        conversationFileSummary: {
+          hasImage: true,
+          hasAudio: false,
+          hasDocument: false,
+          hasRemoteDocumentUrl: false,
+        },
+      },
+    });
+
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'openai',
+      undefined,
+      {
+        conversationFileSummary: {
+          hasImage: true,
+          hasAudio: false,
+          hasDocument: false,
+          hasRemoteDocumentUrl: false,
+        },
+      },
+    );
+
+    expect(order).toEqual(['chat', 'messages', 'responses']);
+  });
+
+  it('does not apply runtime endpoint memory to document attachments', async () => {
+    recordUpstreamEndpointSuccess({
+      siteId: baseContext.site.id,
+      endpoint: 'messages',
+      downstreamFormat: 'openai',
+      modelName: 'gpt-5.3',
+      requestCapabilities: {
+        hasNonImageFileInput: true,
+        conversationFileSummary: {
+          hasImage: false,
+          hasAudio: false,
+          hasDocument: true,
+          hasRemoteDocumentUrl: false,
+        },
+      },
+    });
+
+    const order = await resolveUpstreamEndpointCandidates(
+      {
+        ...baseContext,
+        site: { ...baseContext.site, platform: 'new-api' },
+      },
+      'gpt-5.3',
+      'openai',
+      undefined,
+      {
+        hasNonImageFileInput: true,
+        conversationFileSummary: {
+          hasImage: false,
+          hasAudio: false,
+          hasDocument: true,
+          hasRemoteDocumentUrl: false,
+        },
+      },
+    );
+
+    expect(order).toEqual(['responses', 'messages', 'chat']);
+  });
+
   it('remembers the last successful endpoint per site capability profile', async () => {
     recordUpstreamEndpointSuccess({
       siteId: baseContext.site.id,

src/server/routes/proxy/upstreamEndpoint.ts

@@ -41,6 +41,8 @@ export type EndpointPreference = DownstreamFormat | 'responses';
 
 type EndpointCapabilityProfile = {
   preferMessagesForClaudeModel: boolean;
+  hasImageInput: boolean;
+  hasAudioInput: boolean;
   hasNonImageFileInput: boolean;
   hasRemoteDocumentUrl: boolean;
   wantsNativeResponsesReasoning: boolean;
@@ -711,22 +713,35 @@ function buildEndpointCapabilityProfile(input?: {
     wantsNativeResponsesReasoning?: boolean;
   };
 }): EndpointCapabilityProfile {
+  const conversationFileSummary = input?.requestCapabilities?.conversationFileSummary;
   return {
     preferMessagesForClaudeModel: (
       isClaudeFamilyModel(asTrimmedString(input?.modelName))
       || isClaudeFamilyModel(asTrimmedString(input?.requestedModelHint))
     ),
+    hasImageInput: conversationFileSummary?.hasImage === true,
+    hasAudioInput: conversationFileSummary?.hasAudio === true,
     hasNonImageFileInput: (
-      input?.requestCapabilities?.conversationFileSummary?.hasDocument === true
+      conversationFileSummary?.hasDocument === true
       || input?.requestCapabilities?.hasNonImageFileInput === true
     ),
     hasRemoteDocumentUrl: (
-      input?.requestCapabilities?.conversationFileSummary?.hasRemoteDocumentUrl === true
+      conversationFileSummary?.hasRemoteDocumentUrl === true
     ),
     wantsNativeResponsesReasoning: input?.requestCapabilities?.wantsNativeResponsesReasoning === true,
   };
 }
 
+function shouldUseEndpointRuntimeMemory(capabilityProfile: EndpointCapabilityProfile): boolean {
+  // Attachment-capable requests are not protocol-equivalent across chat/messages/responses.
+  // A transient 200 on one endpoint should not bias later multimodal requests onto a lossy path.
+  return (
+    !capabilityProfile.hasImageInput
+    && !capabilityProfile.hasAudioInput
+    && !capabilityProfile.hasNonImageFileInput
+  );
+}
+
 function buildEndpointRuntimeStateKey(input: {
   siteId: number;
   downstreamFormat: EndpointPreference;
@@ -866,17 +881,19 @@ export function recordUpstreamEndpointSuccess(input: {
     wantsNativeResponsesReasoning?: boolean;
   };
 }): void {
+  const capabilityProfile = buildEndpointCapabilityProfile({
+    modelName: input.modelName,
+    requestedModelHint: input.requestedModelHint,
+    requestCapabilities: input.requestCapabilities,
+  });
+  if (!shouldUseEndpointRuntimeMemory(capabilityProfile)) return;
   if (!shouldRememberSuccessfulEndpoint(input)) return;
 
   const nowMs = Date.now();
   const key = buildEndpointRuntimeStateKey({
     siteId: input.siteId,
     downstreamFormat: input.downstreamFormat,
-    capabilityProfile: buildEndpointCapabilityProfile({
+    capabilityProfile,
-      modelName: input.modelName,
-      requestedModelHint: input.requestedModelHint,
-      requestCapabilities: input.requestCapabilities,
-    }),
   });
   const state = getOrCreateEndpointRuntimeState(key, nowMs);
   state.preferredEndpoint = input.endpoint;
@@ -898,17 +915,19 @@ export function recordUpstreamEndpointFailure(input: {
     wantsNativeResponsesReasoning?: boolean;
   };
 }): void {
+  const capabilityProfile = buildEndpointCapabilityProfile({
+    modelName: input.modelName,
+    requestedModelHint: input.requestedModelHint,
+    requestCapabilities: input.requestCapabilities,
+  });
+  if (!shouldUseEndpointRuntimeMemory(capabilityProfile)) return;
   if (!shouldBlockEndpointByError(input.status, input.errorText)) return;
 
   const nowMs = Date.now();
   const key = buildEndpointRuntimeStateKey({
     siteId: input.siteId,
     downstreamFormat: input.downstreamFormat,
-    capabilityProfile: buildEndpointCapabilityProfile({
+    capabilityProfile,
-      modelName: input.modelName,
-      requestedModelHint: input.requestedModelHint,
-      requestCapabilities: input.requestCapabilities,
-    }),
   });
   const state = getOrCreateEndpointRuntimeState(key, nowMs);
   state.blockedUntilMsByEndpoint[input.endpoint] = nowMs + ENDPOINT_RUNTIME_BLOCK_TTL_MS;
@@ -1009,7 +1028,9 @@ export async function resolveUpstreamEndpointCandidates(
     capabilityProfile,
   });
   const applyRuntimePreference = (candidates: UpstreamEndpoint[]) => (
-    applyEndpointRuntimePreference(candidates, runtimeStateKey)
+    shouldUseEndpointRuntimeMemory(capabilityProfile)
+      ? applyEndpointRuntimePreference(candidates, runtimeStateKey)
+      : candidates
   );
   const conversationFileSummary = requestCapabilities?.conversationFileSummary ?? {
     hasImage: false,