Hack to deduplicate reasoning

src/app/api/openrouter/[...path]/route.ts

@@ -5,7 +5,14 @@ import { generateProviderSpecificHash } from '@/lib/providerHash';
 import { extractPromptInfo, type MicrodollarUsageContext } from '@/lib/processUsage';
 import { validateFeatureHeader, FEATURE_HEADER } from '@/lib/feature-detection';
 import type { OpenRouterChatCompletionRequest } from '@/lib/providers/openrouter/types';
-import { applyProviderSpecificLogic, getProvider, openRouterRequest } from '@/lib/providers';
+import {
+  applyProviderSpecificLogic,
+  fixDuplicatedReasoning,
+  getProvider,
+  isLegacyRooBasedClient,
+  isModernOpenCodeBasedClient,
+  openRouterRequest,
+} from '@/lib/providers';
 import { debugSaveProxyRequest } from '@/lib/debugUtils';
 import { captureException, setTag, startInactiveSpan } from '@sentry/nextjs';
 import { getUserFromAuth } from '@/lib/user.server';
@@ -384,6 +391,10 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
     repairTools(requestBodyParsed);
   }
 
+  if (isModernOpenCodeBasedClient(fraudHeaders)) {
+    fixDuplicatedReasoning(requestBodyParsed);
+  }
+
   const toolsAvailable = getToolsAvailable(requestBodyParsed.tools);
   const toolsUsed = getToolsUsed(requestBodyParsed.messages);
 
@@ -397,11 +408,7 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
   );
 
   const response = customLlm
-    ? await customLlmRequest(
-        customLlm,
-        requestBodyParsed,
-        !!fraudHeaders.http_user_agent?.startsWith('Kilo-Code/')
-      )
+    ? await customLlmRequest(customLlm, requestBodyParsed, isLegacyRooBasedClient(fraudHeaders))
     : await openRouterRequest({
         path,
         search: url.search,

src/lib/providers/index.ts

@@ -3,6 +3,7 @@ import { debugSaveProxyResponseStream } from '../debugUtils';
 import { fetchWithBackoff } from '../fetchWithBackoff';
 import { captureException, captureMessage } from '@sentry/nextjs';
 import type {
+  MessageWithReasoning,
   OpenRouterChatCompletionRequest,
   OpenRouterGeneration,
 } from '@/lib/providers/openrouter/types';
@@ -40,6 +41,7 @@ import { isAnonymousContext } from '@/lib/anonymous';
 import { isOpenAiModel } from '@/lib/providers/openai';
 import { applyQwenModelSettings, isQwenModel } from '@/lib/providers/qwen';
 import type { ProviderId } from '@/lib/providers/provider-id';
+import { ReasoningDetailType } from '@/lib/custom-llm/reasoning-details';
 
 export type Provider = {
   id: ProviderId;
@@ -258,6 +260,46 @@ function applyPreferredProvider(
   }
 }
 
+export function isModernOpenCodeBasedClient(headers: { http_user_agent: string | null }) {
+  return headers.http_user_agent === 'opencode-kilo-provider';
+}
+
+export function isLegacyRooBasedClient(headers: { http_user_agent: string | null }) {
+  return headers.http_user_agent?.startsWith('Kilo-Code/') === true;
+}
+
+export function fixDuplicatedReasoning(requestToMutate: OpenRouterChatCompletionRequest) {
+  // workaround for several bugs in @openrouter/ai-sdk-provider v1 where it duplicates or doesn't merge reasoning correctly
+  for (const msg of requestToMutate.messages ?? []) {
+    if (msg.role !== 'assistant') {
+      continue;
+    }
+    const msgWithReasoning = msg as MessageWithReasoning;
+    if (!msgWithReasoning.reasoning_details || msgWithReasoning.reasoning_details.length <= 1) {
+      continue;
+    }
+    // it appears the last item is correct
+    const lastEncryptedItem = msgWithReasoning.reasoning_details.findLast(
+      rd => rd.type === ReasoningDetailType.Encrypted
+    );
+    const lastSignedItem = msgWithReasoning.reasoning_details.findLast(
+      rd => rd.type === ReasoningDetailType.Text && rd.signature
+    );
+    const lastTextItem = msgWithReasoning.reasoning_details.findLast(
+      rd => rd.type === ReasoningDetailType.Text
+    );
+    const prevailingItem = lastEncryptedItem ?? lastSignedItem ?? lastTextItem;
+    if (!prevailingItem) {
+      continue;
+    }
+    console.debug(
+      '[fixDuplicatedReasoning] removing all but one item',
+      msgWithReasoning.reasoning_details
+    );
+    msgWithReasoning.reasoning_details = [prevailingItem];
+  }
+}
+
 export function applyProviderSpecificLogic(
   provider: Provider,
   requestedModel: string,