Add LiteLLM proxy support for ChatGPT subscription AI access

- Add ai_provider setting: 'openai' (API key) or 'litellm' (ChatGPT subscription proxy)
- Auto-strip max_tokens/max_completion_tokens for chatgpt/ prefix models
  (ChatGPT subscription backend rejects token limit fields)
- LiteLLM mode: dummy API key when none configured, base URL required
- isOpenAIConfigured() checks base URL instead of API key for LiteLLM
- listAvailableModels() returns manualEntry flag for LiteLLM (no models.list)
- Settings UI: conditional fields, info banner, manual model input with
  chatgpt/ prefix examples when LiteLLM selected
- All 7 AI services work transparently via buildCompletionParams()

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/components/settings/ai-settings-form.tsx

@@ -4,7 +4,7 @@ import { useForm } from 'react-hook-form'
 import { zodResolver } from '@hookform/resolvers/zod'
 import { z } from 'zod'
 import { toast } from 'sonner'
-import { Cog, Loader2, Zap, AlertCircle, RefreshCw, SlidersHorizontal } from 'lucide-react'
+import { Cog, Loader2, Zap, AlertCircle, RefreshCw, SlidersHorizontal, Info } from 'lucide-react'
 import { trpc } from '@/lib/trpc/client'
 import { Button } from '@/components/ui/button'
 import { Input } from '@/components/ui/input'
@@ -67,7 +67,10 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
     },
   })
 
-  // Fetch available models from OpenAI API
+  const watchProvider = form.watch('ai_provider')
+  const isLiteLLM = watchProvider === 'litellm'
+
+  // Fetch available models from OpenAI API (skip for LiteLLM — no models.list support)
   const {
     data: modelsData,
     isLoading: modelsLoading,
@@ -76,6 +79,7 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
   } = trpc.settings.listAIModels.useQuery(undefined, {
     staleTime: 5 * 60 * 1000, // Cache for 5 minutes
     retry: false,
+    enabled: !isLiteLLM,
   })
 
   const updateSettings = trpc.settings.updateMultiple.useMutation({
@@ -182,32 +186,50 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
                   </SelectTrigger>
                 </FormControl>
                 <SelectContent>
-                  <SelectItem value="openai">OpenAI</SelectItem>
+                  <SelectItem value="openai">OpenAI (API Key)</SelectItem>
+                  <SelectItem value="litellm">LiteLLM Proxy (ChatGPT Subscription)</SelectItem>
                 </SelectContent>
               </Select>
               <FormDescription>
-                AI provider for smart assignment suggestions
+                {field.value === 'litellm'
+                  ? 'Route AI calls through a LiteLLM proxy connected to your ChatGPT Plus/Pro subscription'
+                  : 'Direct OpenAI API access using your API key'}
               </FormDescription>
               <FormMessage />
             </FormItem>
           )}
         />
 
+        {isLiteLLM && (
+          <Alert>
+            <Info className="h-4 w-4" />
+            <AlertDescription>
+              <strong>LiteLLM Proxy Mode</strong> — AI calls will be routed through your LiteLLM proxy
+              using your ChatGPT subscription. Token limits are automatically stripped (not supported by ChatGPT backend).
+              Make sure your LiteLLM proxy is running and accessible.
+            </AlertDescription>
+          </Alert>
+        )}
+
         <FormField
           control={form.control}
           name="openai_api_key"
           render={({ field }) => (
             <FormItem>
-              <FormLabel>API Key</FormLabel>
+              <FormLabel>{isLiteLLM ? 'API Key (Optional)' : 'API Key'}</FormLabel>
               <FormControl>
                 <Input
                   type="password"
-                  placeholder={settings.openai_api_key ? '••••••••' : 'Enter API key'}
+                  placeholder={isLiteLLM
+                    ? 'Optional — leave blank for default'
+                    : (settings.openai_api_key ? '••••••••' : 'Enter API key')}
                   {...field}
                 />
               </FormControl>
               <FormDescription>
-                Your OpenAI API key. Leave blank to keep the existing key.
+                {isLiteLLM
+                  ? 'LiteLLM proxy usually does not require an API key. Leave blank to use default.'
+                  : 'Your OpenAI API key. Leave blank to keep the existing key.'}
               </FormDescription>
               <FormMessage />
             </FormItem>
@@ -219,16 +241,26 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
           name="openai_base_url"
           render={({ field }) => (
             <FormItem>
-              <FormLabel>API Base URL (Optional)</FormLabel>
+              <FormLabel>{isLiteLLM ? 'LiteLLM Proxy URL' : 'API Base URL (Optional)'}</FormLabel>
               <FormControl>
                 <Input
-                  placeholder="https://api.openai.com/v1"
+                  placeholder={isLiteLLM ? 'http://localhost:4000' : 'https://api.openai.com/v1'}
                   {...field}
                 />
               </FormControl>
               <FormDescription>
-                Custom base URL for OpenAI-compatible providers. Leave blank for OpenAI.
+                {isLiteLLM ? (
-                Use <code className="text-xs bg-muted px-1 rounded">https://openrouter.ai/api/v1</code> for OpenRouter (access Claude, Gemini, Llama, etc.)
+                  <>
+                    URL of your LiteLLM proxy. Typically{' '}
+                    <code className="text-xs bg-muted px-1 rounded">http://localhost:4000</code>{' '}
+                    or your server address.
+                  </>
+                ) : (
+                  <>
+                    Custom base URL for OpenAI-compatible providers. Leave blank for OpenAI.
+                    Use <code className="text-xs bg-muted px-1 rounded">https://openrouter.ai/api/v1</code> for OpenRouter.
+                  </>
+                )}
               </FormDescription>
               <FormMessage />
             </FormItem>
@@ -242,7 +274,7 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
             <FormItem>
               <div className="flex items-center justify-between">
                 <FormLabel>Model</FormLabel>
-                {modelsData?.success && (
+                {!isLiteLLM && modelsData?.success && !modelsData?.manualEntry && (
                   <Button
                     type="button"
                     variant="ghost"
@@ -256,7 +288,13 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
                 )}
               </div>
 
-              {modelsLoading ? (
+              {isLiteLLM || modelsData?.manualEntry ? (
+                <Input
+                  value={field.value}
+                  onChange={(e) => field.onChange(e.target.value)}
+                  placeholder="chatgpt/gpt-5.2"
+                />
+              ) : modelsLoading ? (
                 <Skeleton className="h-10 w-full" />
               ) : modelsError || !modelsData?.success ? (
                 <div className="space-y-2">
@@ -303,7 +341,15 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
                 </Select>
               )}
               <FormDescription>
-                {form.watch('ai_model')?.startsWith('o') ? (
+                {isLiteLLM ? (
+                  <>
+                    Enter the model ID with the{' '}
+                    <code className="text-xs bg-muted px-1 rounded">chatgpt/</code> prefix.
+                    Examples:{' '}
+                    <code className="text-xs bg-muted px-1 rounded">chatgpt/gpt-5.2</code>,{' '}
+                    <code className="text-xs bg-muted px-1 rounded">chatgpt/gpt-5.2-codex</code>
+                  </>
+                ) : form.watch('ai_model')?.startsWith('o') ? (
                   <span className="flex items-center gap-1 text-purple-600">
                     <SlidersHorizontal className="h-3 w-3" />
                     Reasoning model - optimized for complex analysis tasks

src/lib/openai.ts

@@ -8,6 +8,33 @@ const globalForOpenAI = globalThis as unknown as {
   openaiInitialized: boolean
 }
 
+// ─── Provider Detection ─────────────────────────────────────────────────────
+
+/**
+ * Get the configured AI provider from SystemSettings.
+ * Returns 'openai' (default) or 'litellm' (ChatGPT subscription proxy).
+ */
+export async function getConfiguredProvider(): Promise<'openai' | 'litellm'> {
+  try {
+    const setting = await prisma.systemSettings.findUnique({
+      where: { key: 'ai_provider' },
+    })
+    const value = setting?.value || 'openai'
+    return value === 'litellm' ? 'litellm' : 'openai'
+  } catch {
+    return 'openai'
+  }
+}
+
+/**
+ * Check if a model ID indicates LiteLLM ChatGPT subscription routing.
+ * Models like 'chatgpt/gpt-5.2' use the chatgpt/ prefix.
+ * Used by buildCompletionParams (sync) to strip unsupported token limit fields.
+ */
+export function isLiteLLMChatGPTModel(model: string): boolean {
+  return model.toLowerCase().startsWith('chatgpt/')
+}
+
 // ─── Model Type Detection ────────────────────────────────────────────────────
 
 /**
@@ -168,6 +195,12 @@ export function buildCompletionParams(
     params.response_format = { type: 'json_object' }
   }
 
+  // LiteLLM ChatGPT subscription models reject token limit fields
+  if (isLiteLLMChatGPTModel(model)) {
+    delete params.max_tokens
+    delete params.max_completion_tokens
+  }
+
   return params
 }
 
@@ -209,8 +242,12 @@ async function getBaseURL(): Promise<string | undefined> {
  */
 async function createOpenAIClient(): Promise<OpenAI | null> {
   const apiKey = await getOpenAIApiKey()
+  const provider = await getConfiguredProvider()
 
-  if (!apiKey) {
+  // LiteLLM proxy may not require a real API key
+  const effectiveApiKey = apiKey || (provider === 'litellm' ? 'sk-litellm' : null)
+
+  if (!effectiveApiKey) {
     console.warn('OpenAI API key not configured')
     return null
   }
@@ -218,11 +255,11 @@ async function createOpenAIClient(): Promise<OpenAI | null> {
   const baseURL = await getBaseURL()
 
   if (baseURL) {
-    console.log(`[OpenAI] Using custom base URL: ${baseURL}`)
+    console.log(`[OpenAI] Using custom base URL: ${baseURL} (provider: ${provider})`)
   }
 
   return new OpenAI({
-    apiKey,
+    apiKey: effectiveApiKey,
     ...(baseURL ? { baseURL } : {}),
   })
 }
@@ -259,6 +296,12 @@ export function resetOpenAIClient(): void {
  * Check if OpenAI is configured and available
  */
 export async function isOpenAIConfigured(): Promise<boolean> {
+  const provider = await getConfiguredProvider()
+  if (provider === 'litellm') {
+    // LiteLLM just needs a base URL configured
+    const baseURL = await getBaseURL()
+    return !!baseURL
+  }
   const apiKey = await getOpenAIApiKey()
   return !!apiKey
 }
@@ -270,8 +313,20 @@ export async function listAvailableModels(): Promise<{
   success: boolean
   models?: string[]
   error?: string
+  manualEntry?: boolean
 }> {
   try {
+    const provider = await getConfiguredProvider()
+
+    // LiteLLM proxy for ChatGPT subscription doesn't support models.list()
+    if (provider === 'litellm') {
+      return {
+        success: true,
+        models: [],
+        manualEntry: true,
+      }
+    }
+
     const client = await getOpenAI()
 
     if (!client) {

src/server/routers/settings.ts

@@ -201,8 +201,8 @@ export const settingsRouter = router({
         clearStorageProviderCache()
       }
 
-      // Reset OpenAI client if API key or base URL changed
+      // Reset OpenAI client if API key, base URL, model, or provider changed
-      if (input.settings.some((s) => s.key === 'openai_api_key' || s.key === 'openai_base_url' || s.key === 'ai_model')) {
+      if (input.settings.some((s) => s.key === 'openai_api_key' || s.key === 'openai_base_url' || s.key === 'ai_model' || s.key === 'ai_provider')) {
         const { resetOpenAIClient } = await import('@/lib/openai')
         resetOpenAIClient()
       }
@@ -247,6 +247,15 @@ export const settingsRouter = router({
   listAIModels: superAdminProcedure.query(async () => {
     const result = await listAvailableModels()
 
+    // LiteLLM mode: manual model entry, no listing available
+    if (result.manualEntry) {
+      return {
+        success: true,
+        models: [],
+        manualEntry: true,
+      }
+    }
+
     if (!result.success || !result.models) {
       return {
         success: false,