Merge pull request #1519 from itzrnvr/fix/openai-compat-preserve-user-agent

fix(proxy): add Anthropic passthrough mode for coding-agent-only endpoints

src/proxy/profile-router.ts

@@ -12,6 +12,13 @@ export interface OpenAICompatProfileConfig {
   provider: DroidProvider;
   insecure?: boolean;
   forceOpenAIReasoningModel?: boolean;
+  /**
+   * When true, forward Anthropic-format requests to the upstream
+   * `/v1/messages` endpoint instead of translating to OpenAI chat
+   * completions. Required for providers that reject OpenAI-format
+   * requests (e.g. Kimi, Anthropic-API mirrors).
+   */
+  passthrough?: boolean;
   model?: string;
   opusModel?: string;
   sonnetModel?: string;
@@ -30,6 +37,13 @@ export interface OpenAICompatProfileEnv {
   CCS_DROID_PROVIDER?: string;
   CCS_OPENAI_PROXY_INSECURE?: string;
   CCS_OPENAI_REASONING_MODEL?: string;
+  /**
+   * Force Anthropic passthrough for this profile. When set, the
+   * proxy forwards incoming `/v1/messages` requests to the
+   * upstream `/v1/messages` endpoint without translating to
+   * OpenAI format.
+   */
+  CCS_OPENAI_PROXY_PASSTHROUGH?: string;
 }
 
 export function isOpenAICompatProvider(provider: DroidProvider | null): provider is DroidProvider {
@@ -72,6 +86,7 @@ export function resolveOpenAICompatProfileConfig(
     provider,
     insecure: isTruthyEnv(env.CCS_OPENAI_PROXY_INSECURE),
     forceOpenAIReasoningModel: isTruthyEnv(env.CCS_OPENAI_REASONING_MODEL),
+    passthrough: isTruthyEnv(env.CCS_OPENAI_PROXY_PASSTHROUGH),
     model: env.ANTHROPIC_MODEL?.trim() || undefined,
     opusModel: env.ANTHROPIC_DEFAULT_OPUS_MODEL?.trim() || undefined,
     sonnetModel: env.ANTHROPIC_DEFAULT_SONNET_MODEL?.trim() || undefined,

src/proxy/server/http-helpers.ts

@@ -58,6 +58,50 @@ export function readJsonBody(req: http.IncomingMessage): Promise<unknown> {
   });
 }
 
+/**
+ * Read the raw request body as a UTF-8 string, suitable for forwarding
+ * verbatim in passthrough mode. Rejects bodies larger than 10MB.
+ */
+export function readRawBody(req: http.IncomingMessage): Promise<string> {
+  return new Promise((resolve, reject) => {
+    const chunks: Buffer[] = [];
+    let total = 0;
+    let settled = false;
+
+    const resolveOnce = (payload: string) => {
+      if (!settled) {
+        settled = true;
+        resolve(payload);
+      }
+    };
+
+    const rejectOnce = (error: Error) => {
+      if (!settled) {
+        settled = true;
+        reject(error);
+      }
+    };
+
+    req.on('data', (chunk: Buffer) => {
+      total += chunk.length;
+      if (total > MAX_BODY_SIZE) {
+        req.pause();
+        rejectOnce(new Error('Request body too large (max 10MB)'));
+        return;
+      }
+      chunks.push(chunk);
+    });
+
+    req.on('end', () => {
+      resolveOnce(Buffer.concat(chunks).toString('utf8'));
+    });
+
+    req.on('error', (error) => {
+      rejectOnce(error instanceof Error ? error : new Error(String(error)));
+    });
+  });
+}
+
 export async function pipeWebResponseToNode(
   response: Response,
   res: http.ServerResponse

src/proxy/server/messages-route.ts

@@ -8,9 +8,9 @@ import {
   type ProxyOpenAIRequest,
 } from '../transformers/request-transformer';
 import { ProxySseStreamTransformer } from '../transformers/sse-stream-transformer';
-import { resolveOpenAIChatCompletionsUrl } from '../upstream-url';
+import { isAnthropicPassthroughProfile, resolveOpenAIChatCompletionsUrl } from '../upstream-url';
 import { createLogger } from '../../services/logging';
-import { pipeWebResponseToNode, readJsonBody, writeJson } from './http-helpers';
+import { pipeWebResponseToNode, readRawBody, writeJson } from './http-helpers';
 
 const REQUEST_TIMEOUT_MS = 600_000;
 const DIRECT_OPENAI_REASONING_CHAT_MODEL = /^(?:gpt-5|o[134])(?:[-.]|$)/;
@@ -23,12 +23,46 @@ class ProxyInputError extends Error {
   }
 }
 
-function buildUpstreamHeaders(profile: OpenAICompatProfileConfig): Record<string, string> {
-  return {
+/**
+ * Build the upstream request headers. Preserves the original User-Agent from
+ * the incoming request when available, since some providers (e.g. Kimi)
+ * reject requests from unknown User-Agents. Falls back to a stable
+ * `CCS-OpenAI-Compat-Proxy/<version>` User-Agent when the client did not
+ * provide one.
+ */
+function buildUpstreamHeaders(
+  profile: OpenAICompatProfileConfig,
+  incomingHeaders: http.IncomingHttpHeaders,
+  options: { preserveUserAgent?: boolean } = {}
+): Record<string, string> {
+  const headers: Record<string, string> = {
     'Content-Type': 'application/json',
     Authorization: `Bearer ${profile.apiKey}`,
     'User-Agent': 'CCS-OpenAI-Compat-Proxy/1.0',
   };
+
+  if (options.preserveUserAgent) {
+    // Preserve the original User-Agent (or x-stainless-user-agent fallback) so
+    // providers that require a recognized coding-agent identifier accept the
+    // passthrough request.
+    const rawUserAgent = pickHeaderValue(incomingHeaders, ['user-agent', 'x-stainless-user-agent']);
+    headers['User-Agent'] = rawUserAgent?.trim() || headers['User-Agent'];
+  }
+
+  return headers;
+}
+
+function pickHeaderValue(headers: http.IncomingHttpHeaders, names: string[]): string | undefined {
+  for (const name of names) {
+    const value = headers[name];
+    if (typeof value === 'string' && value.length > 0) {
+      return value;
+    }
+    if (Array.isArray(value) && value.length > 0 && typeof value[0] === 'string') {
+      return value[0];
+    }
+  }
+  return undefined;
 }
 
 function isKnownOpenAIReasoningChatModel(model: string | undefined): boolean {
@@ -152,8 +186,29 @@ function shapeUpstreamChatPayload(
 
 function buildUpstreamRequest(
   profile: OpenAICompatProfileConfig,
-  rawBody: unknown
+  rawBody: unknown,
+  options: {
+    passthrough?: boolean;
+    rawBodyText?: string;
+    route?: ReturnType<typeof resolveProxyRequestRoute>;
+  } = {}
 ): { body: string; route: ReturnType<typeof resolveProxyRequestRoute> } {
+  // In passthrough mode we forward the incoming Anthropic body verbatim to
+  // the upstream provider, skipping both the Anthropic→OpenAI translation
+  // and the per-provider payload shaping. The body is still serialized as
+  // a string so we can stream it on the wire.
+  if (options.passthrough) {
+    if (rawBody === undefined || rawBody === null) {
+      throw new ProxyInputError('Missing request body');
+    }
+    const body = options.rawBodyText ?? JSON.stringify(rawBody);
+    if (!body.trim()) {
+      throw new ProxyInputError('Missing request body');
+    }
+    const route = options.route ?? resolveProxyRequestRoute(profile, buildRoutingRequest(rawBody));
+    return { body, route };
+  }
+
   let transformed: ProxyOpenAIRequest;
   try {
     const transformer = new ProxyRequestTransformer();
@@ -174,6 +229,118 @@ function buildUpstreamRequest(
   return { body: JSON.stringify(body), route };
 }
 
+function parseJsonBodyText(rawBodyText: string): unknown {
+  const raw = rawBodyText.trim();
+  if (!raw) {
+    return {};
+  }
+
+  try {
+    return JSON.parse(raw);
+  } catch {
+    throw new ProxyInputError('Invalid JSON in request body');
+  }
+}
+
+function asRecord(value: unknown): Record<string, unknown> | null {
+  return typeof value === 'object' && value !== null && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : null;
+}
+
+function routingTextContent(content: unknown): string {
+  if (typeof content === 'string') {
+    return content;
+  }
+  if (!Array.isArray(content)) {
+    return '';
+  }
+
+  return content
+    .map((part) => {
+      const parsed = asRecord(part);
+      if (!parsed) {
+        return '';
+      }
+      if (parsed.type === 'text' && typeof parsed.text === 'string') {
+        return parsed.text;
+      }
+      if (parsed.type === 'image') {
+        return '[image]';
+      }
+      return '';
+    })
+    .filter(Boolean)
+    .join('\n');
+}
+
+function routingMessages(messages: unknown): ProxyOpenAIRequest['messages'] {
+  if (!Array.isArray(messages)) {
+    return [];
+  }
+
+  return messages.flatMap((message): ProxyOpenAIRequest['messages'] => {
+    const parsed = asRecord(message);
+    const role = parsed?.role;
+    if (role !== 'system' && role !== 'user' && role !== 'assistant' && role !== 'tool') {
+      return [];
+    }
+
+    return [
+      {
+        role,
+        content: routingTextContent(parsed?.content),
+      },
+    ];
+  });
+}
+
+function routingTools(tools: unknown): ProxyOpenAIRequest['tools'] | undefined {
+  if (!Array.isArray(tools)) {
+    return undefined;
+  }
+
+  const normalized = tools.flatMap((tool): NonNullable<ProxyOpenAIRequest['tools']> => {
+    const parsed = asRecord(tool);
+    const fn = asRecord(parsed?.function);
+    if (parsed?.type !== 'function' || typeof fn?.name !== 'string') {
+      return [];
+    }
+    return [
+      {
+        type: 'function',
+        function: {
+          name: fn.name,
+          description: typeof fn.description === 'string' ? fn.description : undefined,
+          parameters: asRecord(fn.parameters) ?? {},
+        },
+      },
+    ];
+  });
+
+  return normalized.length > 0 ? normalized : undefined;
+}
+
+function buildRoutingRequest(rawBody: unknown): ProxyOpenAIRequest {
+  const parsed = asRecord(rawBody);
+  const model = typeof parsed?.model === 'string' ? parsed.model : undefined;
+  const thinking = asRecord(parsed?.thinking);
+  const thinkingEnabled = thinking?.type === 'enabled' || thinking?.type === 'adaptive';
+
+  return {
+    model,
+    stream: parsed?.stream === true,
+    messages: routingMessages(parsed?.messages),
+    tools: routingTools(parsed?.tools),
+    reasoning: thinkingEnabled
+      ? {
+          enabled: true,
+          effort: 'high',
+        }
+      : undefined,
+  };
+}
+
 export function extractIncomingProxyToken(headers: http.IncomingHttpHeaders): string | null {
   const xApiKey = headers['x-api-key'];
   if (typeof xApiKey === 'string' && xApiKey.trim().length > 0) {
@@ -206,11 +373,13 @@ function buildFetchInit(
   profile: OpenAICompatProfileConfig,
   body: string,
   signal: AbortSignal,
+  incomingHeaders: http.IncomingHttpHeaders,
+  preserveUserAgent: boolean,
   insecureDispatcher?: Dispatcher
 ): RequestInit {
   const init: RequestInit = {
     method: 'POST',
-    headers: buildUpstreamHeaders(profile),
+    headers: buildUpstreamHeaders(profile, incomingHeaders, { preserveUserAgent }),
     body,
     signal,
   };
@@ -320,9 +489,20 @@ export async function handleProxyMessagesRequest(
 
   let timeoutMs = REQUEST_TIMEOUT_MS;
   try {
-    const rawBody = await readJsonBody(req);
-    logger.stage('transform', 'request.transform.start', 'Transforming inbound proxy body');
-    const upstream = buildUpstreamRequest(profile, rawBody);
+    const rawBodyText = await readRawBody(req);
+    const rawBody = parseJsonBodyText(rawBodyText);
+    const initialRoute = resolveProxyRequestRoute(profile, buildRoutingRequest(rawBody));
+    const passthrough = isAnthropicPassthroughProfile(initialRoute.profile.baseUrl, {
+      forcePassthrough: initialRoute.profile.passthrough === true,
+    });
+    logger.stage('transform', 'request.transform.start', 'Transforming inbound proxy body', {
+      passthrough,
+    });
+    const upstream = buildUpstreamRequest(profile, rawBody, {
+      passthrough,
+      rawBodyText,
+      route: passthrough ? initialRoute : undefined,
+    });
     logger.stage('route', 'request.routed', 'Resolved proxy upstream route', {
       profileName: upstream.route.profile.profileName,
       provider: upstream.route.profile.provider,
@@ -371,17 +551,34 @@ export async function handleProxyMessagesRequest(
         profileName: profile.profileName,
         routedProfileName: upstream.route.profile.profileName,
         insecureTls: dispatcher !== undefined,
+        passthrough,
+      });
+      const upstreamUrl = resolveOpenAIChatCompletionsUrl(upstream.route.profile.baseUrl, {
+        passthrough,
       });
       const upstreamResponse = await fetch(
-        resolveOpenAIChatCompletionsUrl(upstream.route.profile.baseUrl),
-        buildFetchInit(upstream.route.profile, upstream.body, controller.signal, dispatcher)
+        upstreamUrl,
+        buildFetchInit(
+          upstream.route.profile,
+          upstream.body,
+          controller.signal,
+          req.headers,
+          passthrough,
+          dispatcher
+        )
       );
       logger.stage('upstream', 'upstream.response', 'Received upstream response', {
         profileName: profile.profileName,
         routedProfileName: upstream.route.profile.profileName,
         status: upstreamResponse.status,
       });
-      const response = await transformer.transform(upstreamResponse);
+      // In passthrough mode the upstream already returns Anthropic-format
+      // bytes, so we just pipe the response through unchanged. The SSE
+      // transformer is only used for OpenAI-mode responses that need to
+      // be re-encoded as Anthropic SSE.
+      const response = passthrough
+        ? upstreamResponse
+        : await transformer.transform(upstreamResponse);
       await pipeWebResponseToNode(response, res);
       logger.stage('respond', 'request.respond', 'Proxy response written', undefined, {
         latencyMs: Date.now() - startedAt,

src/proxy/upstream-url.ts

@@ -1,3 +1,23 @@
+/**
+ * Resolve upstream URLs for OpenAI-compat proxy requests.
+ *
+ * Two modes are supported:
+ *
+ * 1. Default (OpenAI mode) - appends `/chat/completions` or `/models` to the
+ *    base URL so requests are routed to the OpenAI-compatible endpoint of the
+ *    upstream provider.
+ *
+ * 2. Anthropic passthrough - some providers (e.g. Kimi coding endpoints)
+ *    expose an Anthropic-compatible `/v1/messages` endpoint and reject
+ *    OpenAI-format requests. For these profiles we preserve the incoming
+ *    Anthropic body and forward it directly to the provider's
+ *    `/v1/messages` endpoint.
+ *
+ * Passthrough is enabled when:
+ *   - `CCS_OPENAI_PROXY_PASSTHROUGH=1` is set in the profile env, OR
+ *   - The base URL is a known Anthropic-style provider host.
+ */
+
 function normalizePathname(pathname: string): string {
   const trimmed = pathname.replace(/\/+$/, '');
   return trimmed || '';
@@ -9,16 +29,71 @@ function ensureSupportedProtocol(parsed: URL): void {
   }
 }
 
+function isAnthropicPassthroughHost(hostname: string): boolean {
+  const normalized = hostname.toLowerCase();
+  return (
+    normalized === 'api.anthropic.com' ||
+    normalized.endsWith('.anthropic.com') ||
+    normalized === 'api.kimi.com' ||
+    normalized.endsWith('.kimi.com')
+  );
+}
+
 function isOpenRouterHost(hostname: string): boolean {
   const normalized = hostname.toLowerCase();
   return normalized === 'openrouter.ai' || normalized.endsWith('.openrouter.ai');
 }
 
-function buildResolvedUrl(baseUrl: string, suffix: string): string {
+/**
+ * Returns true if the profile should pass Anthropic-format requests through
+ * directly to the upstream provider, skipping the OpenAI translation.
+ */
+export function isAnthropicPassthroughProfile(
+  baseUrl: string,
+  options: { forcePassthrough?: boolean } = {}
+): boolean {
+  if (options.forcePassthrough) {
+    return true;
+  }
+  let parsed: URL;
+  try {
+    parsed = new URL(baseUrl);
+  } catch {
+    return false;
+  }
+  ensureSupportedProtocol(parsed);
+  return isAnthropicPassthroughHost(parsed.hostname);
+}
+
+/**
+ * Build the upstream URL for either the messages path or the models path.
+ * When passthrough is enabled, the messages path is appended verbatim
+ * (i.e. `/v1/messages`); otherwise the OpenAI suffix is appended.
+ */
+function buildResolvedUrl(
+  baseUrl: string,
+  suffix: string,
+  options: { passthrough?: boolean } = {}
+): string {
   const parsed = new URL(baseUrl);
   ensureSupportedProtocol(parsed);
 
   const pathname = normalizePathname(parsed.pathname);
+
+  if (options.passthrough) {
+    // For Anthropic passthrough, the suffix is always `/v1/messages` or
+    // `/v1/models`. If the base URL already ends in `/v1` we drop the
+    // duplicated prefix; otherwise we append the suffix verbatim.
+    if (pathname.endsWith('/v1') && suffix.startsWith('/v1/')) {
+      parsed.pathname = `${pathname}${suffix.slice(3)}`;
+      parsed.search = '';
+      return parsed.toString();
+    }
+    parsed.pathname = pathname ? `${pathname}${suffix}` : suffix;
+    parsed.search = '';
+    return parsed.toString();
+  }
+
   if (pathname.endsWith(suffix)) {
     return parsed.toString();
   }
@@ -37,10 +112,32 @@ function buildResolvedUrl(baseUrl: string, suffix: string): string {
   return parsed.toString();
 }
 
-export function resolveOpenAIChatCompletionsUrl(baseUrl: string): string {
+/**
+ * Resolve the upstream URL for a chat completions request.
+ *
+ * For OpenAI-compatible providers this resolves to
+ * `<base>/v1/chat/completions`. For Anthropic passthrough providers this
+ * resolves to `<base>/v1/messages`.
+ */
+export function resolveOpenAIChatCompletionsUrl(
+  baseUrl: string,
+  options: { passthrough?: boolean } = {}
+): string {
+  if (options.passthrough) {
+    return buildResolvedUrl(baseUrl, '/v1/messages', { passthrough: true });
+  }
   return buildResolvedUrl(baseUrl, '/chat/completions');
 }
 
-export function resolveOpenAIModelsUrl(baseUrl: string): string {
+/**
+ * Resolve the upstream URL for the models endpoint.
+ */
+export function resolveOpenAIModelsUrl(
+  baseUrl: string,
+  options: { passthrough?: boolean } = {}
+): string {
+  if (options.passthrough) {
+    return buildResolvedUrl(baseUrl, '/v1/models', { passthrough: true });
+  }
   return buildResolvedUrl(baseUrl, '/models');
 }

tests/unit/proxy/messages-route.test.ts

@@ -103,6 +103,12 @@ beforeEach(() => {
       ANTHROPIC_MODEL: 'MiniMax-M2.7',
       CCS_DROID_PROVIDER: 'openai',
     }),
+    kimic: writeSettings('kimic', {
+      ANTHROPIC_BASE_URL: 'https://api.kimi.com/coding/',
+      ANTHROPIC_AUTH_TOKEN: 'kimi_token',
+      ANTHROPIC_MODEL: 'kimi-k2p7-coding',
+      CCS_DROID_PROVIDER: 'generic-chat-completion-api',
+    }),
   };
 
   fs.writeFileSync(
@@ -166,6 +172,152 @@ describe('attachDisconnectAbortHandlers', () => {
 });
 
 describe('handleProxyMessagesRequest', () => {
+  it('auto-passes through Kimi requests and preserves the coding-agent User-Agent', async () => {
+    const activeProfile = buildProfile('kimic');
+    let capturedInput: RequestInfo | URL | undefined;
+    let capturedInit: RequestInit | undefined;
+    const rawBody = JSON.stringify({
+      model: 'kimi-k2p7-coding',
+      max_tokens: 50,
+      messages: [{ role: 'user', content: 'Say hi in 2 words' }],
+    });
+
+    globalThis.fetch = (async (input: RequestInfo | URL, init?: RequestInit) => {
+      capturedInput = input;
+      capturedInit = init;
+      return new Response(
+        JSON.stringify({
+          id: 'msg_1',
+          type: 'message',
+          role: 'assistant',
+          content: [{ type: 'text', text: 'Hello there!' }],
+          model: 'kimi-k2p7-coding',
+          stop_reason: 'end_turn',
+        }),
+        { status: 200, headers: { 'content-type': 'application/json' } }
+      );
+    }) as typeof globalThis.fetch;
+
+    const req = new FakeRequest({
+      'x-api-key': 'local-token',
+      'user-agent': 'claude-cli/2.1.170',
+    });
+    const res = new FakeResponse();
+    const pending = handleProxyMessagesRequest(
+      req as never,
+      res as never,
+      activeProfile,
+      'local-token'
+    );
+    req.end(rawBody);
+    await pending;
+
+    expect(String(capturedInput)).toBe('https://api.kimi.com/coding/v1/messages');
+    expect(capturedInit?.body).toBe(rawBody);
+    expect((capturedInit?.headers as Record<string, string>)['User-Agent']).toBe(
+      'claude-cli/2.1.170'
+    );
+    expect(res.statusCode).toBe(200);
+  });
+
+  it('uses routed-profile passthrough for explicit profile:model selectors', async () => {
+    const activeProfile = buildProfile('hf');
+    let capturedInput: RequestInfo | URL | undefined;
+    let capturedInit: RequestInit | undefined;
+    const rawBody = JSON.stringify({
+      model: 'kimic:kimi-k2p7-coding',
+      max_tokens: 50,
+      messages: [{ role: 'user', content: 'Say hi' }],
+    });
+
+    globalThis.fetch = (async (input: RequestInfo | URL, init?: RequestInit) => {
+      capturedInput = input;
+      capturedInit = init;
+      return new Response(
+        JSON.stringify({
+          id: 'msg_2',
+          type: 'message',
+          role: 'assistant',
+          content: [{ type: 'text', text: 'Hi' }],
+          model: 'kimi-k2p7-coding',
+          stop_reason: 'end_turn',
+        }),
+        { status: 200, headers: { 'content-type': 'application/json' } }
+      );
+    }) as typeof globalThis.fetch;
+
+    const req = new FakeRequest({
+      'x-api-key': 'local-token',
+      'x-stainless-user-agent': 'claude-cli/2.1.170',
+    });
+    const res = new FakeResponse();
+    const pending = handleProxyMessagesRequest(
+      req as never,
+      res as never,
+      activeProfile,
+      'local-token'
+    );
+    req.end(rawBody);
+    await pending;
+
+    expect(String(capturedInput)).toBe('https://api.kimi.com/coding/v1/messages');
+    expect(capturedInit?.body).toBe(rawBody);
+    expect((capturedInit?.headers as Record<string, string>)['User-Agent']).toBe(
+      'claude-cli/2.1.170'
+    );
+    expect(res.statusCode).toBe(200);
+  });
+
+  it('keeps default OpenAI-compatible requests on chat completions with CCS User-Agent', async () => {
+    const activeProfile = buildProfile('hf');
+    let capturedInput: RequestInfo | URL | undefined;
+    let capturedInit: RequestInit | undefined;
+
+    globalThis.fetch = (async (input: RequestInfo | URL, init?: RequestInit) => {
+      capturedInput = input;
+      capturedInit = init;
+      return new Response(
+        JSON.stringify({
+          id: 'chatcmpl_1',
+          object: 'chat.completion',
+          created: 1,
+          model: 'hf-default',
+          choices: [{ index: 0, message: { role: 'assistant', content: 'ok' } }],
+        }),
+        { status: 200, headers: { 'content-type': 'application/json' } }
+      );
+    }) as typeof globalThis.fetch;
+
+    const req = new FakeRequest({
+      'x-api-key': 'local-token',
+      'user-agent': 'claude-cli/2.1.170',
+    });
+    const res = new FakeResponse();
+    const pending = handleProxyMessagesRequest(
+      req as never,
+      res as never,
+      activeProfile,
+      'local-token'
+    );
+    req.end(
+      JSON.stringify({
+        model: 'hf-default',
+        messages: [{ role: 'user', content: 'stay translated' }],
+      })
+    );
+    await pending;
+
+    expect(String(capturedInput)).toBe('https://router.huggingface.co/v1/chat/completions');
+    expect(JSON.parse(String(capturedInit?.body))).toMatchObject({
+      model: 'hf-default',
+      messages: [{ role: 'user', content: 'stay translated' }],
+    });
+    expect((capturedInit?.headers as Record<string, string>)['User-Agent']).toBe(
+      'CCS-OpenAI-Compat-Proxy/1.0'
+    );
+    expect(res.statusCode).toBe(200);
+  });
+
   it('uses a per-request insecure dispatcher for routed profiles and closes it on failure', async () => {
     const activeProfile = buildProfile('hf');
     const sharedDispatcher = { name: 'shared-insecure-dispatcher' } as never;

tests/unit/proxy/profile-router.test.ts

@@ -93,4 +93,28 @@ describe('resolveOpenAICompatProfileConfig', () => {
     expect(result).not.toBeNull();
     expect(result?.provider).toBe('generic-chat-completion-api');
   });
+
+  it('honors CCS_OPENAI_PROXY_PASSTHROUGH=1 for Anthropic-style endpoints', () => {
+    const result = resolveOpenAICompatProfileConfig('kimic', '/tmp/kimic.settings.json', {
+      ANTHROPIC_BASE_URL: 'https://api.kimi.com/coding/',
+      ANTHROPIC_AUTH_TOKEN: 'sk-kimi',
+      ANTHROPIC_MODEL: 'kimi-k2p7-coding',
+      CCS_OPENAI_PROXY_PASSTHROUGH: 'true',
+    });
+
+    expect(result).not.toBeNull();
+    expect(result?.passthrough).toBe(true);
+  });
+
+  it('defaults passthrough to false when CCS_OPENAI_PROXY_PASSTHROUGH is unset', () => {
+    const result = resolveOpenAICompatProfileConfig('gateway', '/tmp/gateway.settings.json', {
+      ANTHROPIC_BASE_URL: 'https://gateway.example.com/v1',
+      ANTHROPIC_AUTH_TOKEN: 'gateway-token',
+      ANTHROPIC_MODEL: 'gpt-4.1',
+      CCS_DROID_PROVIDER: 'generic-chat-completion-api',
+    });
+
+    expect(result).not.toBeNull();
+    expect(result?.passthrough).toBeFalsy();
+  });
 });

tests/unit/proxy/upstream-url.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from 'bun:test';
 import {
+  isAnthropicPassthroughProfile,
   resolveOpenAIChatCompletionsUrl,
   resolveOpenAIModelsUrl,
 } from '../../../src/proxy/upstream-url';
@@ -29,3 +30,56 @@ describe('OpenAI-compatible upstream URL resolution', () => {
     );
   });
 });
+
+describe('Anthropic passthrough URL resolution', () => {
+  it('resolves to /v1/messages when forcePassthrough is set', () => {
+    expect(
+      resolveOpenAIChatCompletionsUrl('https://api.kimi.com/coding/', {
+        passthrough: true,
+      })
+    ).toBe('https://api.kimi.com/coding/v1/messages');
+  });
+
+  it('drops a duplicated /v1 prefix when the base URL already ends in /v1', () => {
+    expect(
+      resolveOpenAIChatCompletionsUrl('https://api.kimi.com/coding/v1', {
+        passthrough: true,
+      })
+    ).toBe('https://api.kimi.com/coding/v1/messages');
+  });
+
+  it('routes /v1/models in passthrough mode', () => {
+    expect(
+      resolveOpenAIModelsUrl('https://api.kimi.com/coding/v1', {
+        passthrough: true,
+      })
+    ).toBe('https://api.kimi.com/coding/v1/models');
+  });
+
+  it('auto-detects known Anthropic-style hosts', () => {
+    expect(isAnthropicPassthroughProfile('https://api.kimi.com/coding/')).toBe(true);
+    expect(isAnthropicPassthroughProfile('https://api.kimi.com/coding/v1')).toBe(true);
+    expect(isAnthropicPassthroughProfile('https://api.anthropic.com')).toBe(true);
+  });
+
+  it('does not treat generic /v1 roots as Anthropic-style passthrough', () => {
+    expect(isAnthropicPassthroughProfile('https://example.test/v1')).toBe(false);
+    expect(isAnthropicPassthroughProfile('https://example.test/v1/')).toBe(false);
+    expect(isAnthropicPassthroughProfile('https://example.test/api/v1')).toBe(false);
+    expect(isAnthropicPassthroughProfile('https://api.openai.com/v1')).toBe(false);
+    expect(isAnthropicPassthroughProfile('https://api.minimax.io/v1')).toBe(false);
+  });
+
+  it('does not auto-detect OpenAI-style base URLs as Anthropic-style', () => {
+    expect(isAnthropicPassthroughProfile('https://api.fireworks.ai/inference')).toBe(false);
+    expect(isAnthropicPassthroughProfile('https://api.openai.com/v1')).toBe(false);
+  });
+
+  it('honors explicit force passthrough for unknown mirrors', () => {
+    expect(
+      isAnthropicPassthroughProfile('https://anthropic-mirror.example.test/v1', {
+        forcePassthrough: true,
+      })
+    ).toBe(true);
+  });
+});