feat(gateway): add Anthropic Messages API support to openrouter proxy

package.json

@@ -41,6 +41,7 @@
     "@ai-sdk/mistral": "^3.0.21",
     "@ai-sdk/openai": "^3.0.37",
     "@ai-sdk/openai-compatible": "^2.0.31",
+    "@anthropic-ai/sdk": "^0.78.0",
     "@aws-sdk/client-s3": "^3.980.0",
     "@aws-sdk/s3-request-presigner": "^3.980.0",
     "@chat-adapter/slack": "^4.15.0",

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

@@ -7,6 +7,7 @@ import { validateFeatureHeader, FEATURE_HEADER } from '@/lib/feature-detection';
 import type {
   OpenRouterChatCompletionRequest,
   GatewayResponsesRequest,
+  GatewayMessagesRequest,
   GatewayRequest,
 } from '@/lib/providers/openrouter/types';
 import { applyProviderSpecificLogic, getProvider, openRouterRequest } from '@/lib/providers';
@@ -69,6 +70,7 @@ import { applyResolvedAutoModel, isKiloAutoModel } from '@/lib/kilo-auto-model';
 import { fixOpenCodeDuplicateReasoning } from '@/lib/providers/fixOpenCodeDuplicateReasoning';
 import type { MicrodollarUsageContext, PromptInfo } from '@/lib/processUsage.types';
 import { extractResponsesPromptInfo } from '@/lib/processUsage.responses';
+import { extractMessagesPromptInfo } from '@/lib/processUsage.messages';
 import { getMaxTokens, hasMiddleOutTransform } from '@/lib/providers/openrouter/request-helpers';
 import { isKiloAffiliatedUser } from '@/lib/isKiloAffiliatedUser';
 
@@ -82,13 +84,17 @@ const PROMOTION_MODEL_LIMIT_REACHED = 'PROMOTION_MODEL_LIMIT_REACHED';
 function validatePath(
   url: URL
 ):
-  | { path: '/chat/completions' | '/responses' }
+  | { path: '/chat/completions' | '/responses' | '/messages' }
   | { errorResponse: ReturnType<typeof invalidPathResponse> } {
   const pathSuffix =
     stripRequiredPrefix(url.pathname, '/api/gateway') ??
     stripRequiredPrefix(url.pathname, '/api/openrouter');
 
-  if (pathSuffix === '/chat/completions' || pathSuffix === '/responses') {
+  if (
+    pathSuffix === '/chat/completions' ||
+    pathSuffix === '/responses' ||
+    pathSuffix === '/messages'
+  ) {
     return { path: pathSuffix };
   }
   return { errorResponse: invalidPathResponse() };
@@ -113,6 +119,9 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
       // Inject or merge stream_options.include_usage = true
       body.stream_options = { ...(body.stream_options || {}), include_usage: true };
       requestBodyParsed = { kind: 'chat_completions', body };
+    } else if (path === '/messages') {
+      const body: GatewayMessagesRequest = JSON.parse(requestBodyText);
+      requestBodyParsed = { kind: 'messages', body };
     } else {
       const body: GatewayResponsesRequest = JSON.parse(requestBodyText);
       body.store = false;
@@ -236,13 +245,13 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
   }
 
   if (
-    requestBodyParsed.kind === 'responses' &&
+    ['messages', 'responses'].includes(requestBodyParsed.kind) &&
     !isKiloAffiliatedUser(maybeUser, organizationId ?? null)
   ) {
     return NextResponse.json(
       {
         error: {
-          message: 'The Responses API is experimental and not yet available to all users.',
+          message: `The ${requestBodyParsed.kind} API is experimental and not yet available to all users.`,
         },
       },
       { status: 403 }
@@ -309,7 +318,9 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
   const promptInfo: PromptInfo =
     requestBodyParsed.kind === 'chat_completions'
       ? extractPromptInfo(requestBodyParsed.body)
-      : extractResponsesPromptInfo(requestBodyParsed.body);
+      : requestBodyParsed.kind === 'messages'
+        ? extractMessagesPromptInfo(requestBodyParsed.body)
+        : extractResponsesPromptInfo(requestBodyParsed.body);
 
   const usageContext: MicrodollarUsageContext = {
     api_kind: requestBodyParsed.kind,
@@ -387,14 +398,23 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
     return dataCollectionRequiredResponse();
   }
 
-  if (taskId) {
-    requestBodyParsed.body.prompt_cache_key = generateProviderSpecificHash(
-      user.id + taskId,
-      provider
-    );
+  const userId = generateProviderSpecificHash(user.id, provider);
+  if (requestBodyParsed.kind === 'messages') {
+    requestBodyParsed.body.metadata = { user_id: userId };
+    requestBodyParsed.body.user = userId;
+    if (taskId) {
+      requestBodyParsed.body.session_id = generateProviderSpecificHash(user.id + taskId, provider);
+    }
+  } else {
+    if (taskId) {
+      requestBodyParsed.body.prompt_cache_key = generateProviderSpecificHash(
+        user.id + taskId,
+        provider
+      );
+    }
+    requestBodyParsed.body.safety_identifier = userId;
+    requestBodyParsed.body.user = userId; // deprecated, but this is what OpenRouter uses
   }
-  requestBodyParsed.body.safety_identifier = generateProviderSpecificHash(user.id, provider);
-  requestBodyParsed.body.user = requestBodyParsed.body.safety_identifier; // deprecated, but this is what OpenRouter uses
 
   if (requestBodyParsed.kind === 'chat_completions') {
     if (ENABLE_TOOL_REPAIR) {
@@ -422,9 +442,11 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
 
   let response: Response;
   if (customLlm) {
-    if (requestBodyParsed.kind === 'responses') {
+    if (requestBodyParsed.kind === 'responses' || requestBodyParsed.kind === 'messages') {
       return NextResponse.json(
-        { error: 'This model is not yet available on the Responses API' },
+        {
+          error: `This model is not available on the ${requestBodyParsed.kind} API`,
+        },
         { status: 404 }
       );
     }
@@ -548,9 +570,13 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
     (isKiloFreeModel(originalModelIdLowerCased) ||
       isActiveReviewPromo(botId, originalModelIdLowerCased))
   ) {
-    return requestBodyParsed.kind === 'chat_completions'
-      ? rewriteFreeModelResponse_ChatCompletions(response, originalModelIdLowerCased)
-      : rewriteFreeModelResponse_Responses(response, originalModelIdLowerCased);
+    if (requestBodyParsed.kind === 'chat_completions') {
+      return rewriteFreeModelResponse_ChatCompletions(response, originalModelIdLowerCased);
+    }
+    if (requestBodyParsed.kind === 'responses') {
+      return rewriteFreeModelResponse_Responses(response, originalModelIdLowerCased);
+    }
+    // messages kind: pass through as-is (free models don't currently use the Messages API)
   }
 
   return wrapInSafeNextResponse(response);

src/lib/abuse-service.ts

@@ -21,7 +21,7 @@ import 'server-only';
 import { getMaxTokens, hasMiddleOutTransform } from '@/lib/providers/openrouter/request-helpers';
 
 /**
- * Extract full prompts from a GatewayRequest (chat completions or responses API).
+ * Extract full prompts from a GatewayRequest (chat completions, responses, or messages API).
  * Unlike extractPromptInfo (which truncates to 100 chars), this returns full content for abuse analysis.
  */
 function extractFullPrompts(request: GatewayRequest): {
@@ -31,6 +31,30 @@ function extractFullPrompts(request: GatewayRequest): {
   if (request.kind === 'responses') {
     return extractFullPromptsFromResponses(request.body);
   }
+  if (request.kind === 'messages') {
+    const systemContent = request.body.system;
+    const systemPrompt =
+      typeof systemContent === 'string'
+        ? systemContent
+        : Array.isArray(systemContent)
+          ? systemContent.map(b => b.text).join('\n')
+          : null;
+    const lastUserMessage = request.body.messages.filter(m => m.role === 'user').at(-1);
+    let userPrompt: string | null = null;
+    if (lastUserMessage) {
+      const content = lastUserMessage.content;
+      if (typeof content === 'string') {
+        userPrompt = content;
+      } else if (Array.isArray(content)) {
+        userPrompt =
+          content
+            .filter(c => c.type === 'text')
+            .map(c => ('text' in c ? c.text : ''))
+            .join('\n') || null;
+      }
+    }
+    return { systemPrompt: systemPrompt || null, userPrompt };
+  }
   return extractFullPromptsFromChatCompletions(request.body);
 }
 

src/lib/kilo-auto-model.ts

@@ -214,15 +214,26 @@ export function applyResolvedAutoModel(
 ) {
   const resolved = resolveAutoModel(model, modeHeader);
   request.body.model = resolved.model;
-  if (resolved.reasoning) request.body.reasoning = resolved.reasoning;
-  if (resolved.verbosity) {
-    if (request.kind === 'chat_completions') {
-      request.body.verbosity = resolved.verbosity as OpenRouterChatCompletionRequest['verbosity'];
+  if (resolved.reasoning) {
+    if (request.kind === 'messages') {
+      request.body.thinking = { type: resolved.reasoning.enabled ? 'adaptive' : 'disabled' };
     } else {
+      request.body.reasoning = resolved.reasoning;
+    }
+  }
+  if (resolved.verbosity) {
+    if (request.kind === 'messages') {
+      request.body.output_config = {
+        ...request.body.output_config,
+        effort: resolved.verbosity,
+      };
+    } else if (request.kind === 'responses') {
       request.body.text = {
         ...request.body.text,
         verbosity: resolved.verbosity as OpenAI.Responses.ResponseTextConfig['verbosity'],
       };
+    } else {
+      request.body.verbosity = resolved.verbosity as OpenRouterChatCompletionRequest['verbosity'];
     }
   }
 }

src/lib/o11y/api-metrics.server.ts

@@ -81,6 +81,13 @@ export function getToolsAvailable(request: GatewayRequest): string[] {
     });
   }
 
+  if (request.kind === 'messages') {
+    return request.body.tools.map((tool): string => {
+      const name = typeof tool.name === 'string' ? tool.name.trim() : '';
+      return name ? `function:${name}` : 'function:unknown';
+    });
+  }
+
   return request.body.tools.map((tool): string => {
     if (tool.type === 'function') {
       const toolName = typeof tool.function?.name === 'string' ? tool.function.name.trim() : '';
@@ -116,6 +123,21 @@ export function getToolsUsed(request: GatewayRequest): string[] {
     return used;
   }
 
+  if (request.kind === 'messages') {
+    const used = new Array<string>();
+    for (const message of request.body.messages) {
+      if (message.role !== 'assistant') continue;
+      const content = Array.isArray(message.content) ? message.content : [];
+      for (const block of content) {
+        if (block.type === 'tool_use') {
+          const name = typeof block.name === 'string' ? block.name.trim() : '';
+          used.push(name ? `function:${name}` : 'function:unknown');
+        }
+      }
+    }
+    return used;
+  }
+
   if (!Array.isArray(request.body.messages)) return [];
 
   const used = new Array<string>();