Add proactive context window management to prevent AI API errors

apps/web/src/app/api/ai/chat/route.ts

@@ -71,6 +71,13 @@ import {
 } from '@/lib/ai/core/stream-abort-registry';
 import { validateUserMessageFileParts, hasFileParts } from '@/lib/ai/core/validate-image-parts';
 import { hasVisionCapability } from '@/lib/ai/core/model-capabilities';
+import {
+  determineMessagesToInclude,
+  getContextWindowSize,
+  estimateSystemPromptTokens,
+  estimateToolDefinitionTokens,
+} from '@pagespace/lib/ai-context-calculator';
+import { isContextLengthError } from '@/lib/ai/shared/error-messages';
 
 
 // Allow streaming responses up to 5 minutes for complex AI agent interactions
@@ -741,13 +748,10 @@ export async function POST(request: Request) {
       });
     }
 
-    // Convert UIMessages to ModelMessages for the AI model
-    // First sanitize messages to remove tool parts without results (prevents "input-available" state errors)
+    // Sanitize messages to remove tool parts without results (prevents "input-available" state errors)
     // NOTE: We use database-loaded messages, NOT messages from client
+    // modelMessages is computed after system prompt is built so we can apply context truncation
     const sanitizedMessages = sanitizeMessagesForModel(conversationHistory);
-    const modelMessages = convertToModelMessages(sanitizedMessages, {
-      tools: filteredTools  // Use original tools - no wrapping needed
-    });
 
     // Fetch user personalization for AI system prompt injection
     const personalization = await getUserPersonalization(userId);
@@ -818,8 +822,43 @@ export async function POST(request: Request) {
     }
 
     loggers.ai.debug('AI Chat API: Tools configured for Page AI', { toolCount: Object.keys(filteredTools).length });
+
+    // Context-length guard: proactively truncate oldest messages to fit within the model's context window.
+    // This prevents AI_APICallError from providers when a conversation grows too long.
+    // We build modelMessages here (after system prompt) so we have accurate token budgeting.
+    const fullSystemPrompt = systemPrompt + timestampSystemPrompt + pageTreePrompt;
+    const contextWindow = getContextWindowSize(currentModel, currentProvider);
+    const systemPromptTokens = estimateSystemPromptTokens(fullSystemPrompt);
+    // Cast needed because filteredTools is a ToolSet (Vercel AI SDK type) but calculator expects plain object
+    const toolTokens = estimateToolDefinitionTokens(filteredTools as Record<string, unknown>);
+    // Reserve 25% headroom for output tokens and tokenizer inaccuracies
+    const inputBudget = Math.floor(contextWindow * 0.75);
+    const { includedMessages, wasTruncated } = determineMessagesToInclude(
+      sanitizedMessages,
+      inputBudget,
+      systemPromptTokens,
+      toolTokens
+    );
+
+    if (wasTruncated) {
+      loggers.ai.warn('AI Chat API: Conversation truncated to fit context window', {
+        originalMessageCount: sanitizedMessages.length,
+        includedMessageCount: includedMessages.length,
+        model: currentModel,
+        provider: currentProvider,
+        contextWindow,
+        inputBudget,
+        systemPromptTokens,
+        toolTokens,
+      });
+    }
+
+    const modelMessages = convertToModelMessages(includedMessages, {
+      tools: filteredTools  // Use original tools - no wrapping needed
+    });
+
     loggers.ai.info('AI Chat API: Starting streamText for Page AI', { model: currentModel, pageName: page.title });
-    
+
     // Create UI message stream with visual content injection support
     // This handles the case where tools return visual content that needs to be injected into the stream
     let result;
@@ -1199,8 +1238,15 @@ export async function POST(request: Request) {
     });
 
     // Return a proper error response
-    return NextResponse.json({ 
-      error: 'Failed to process chat request. Please try again.' 
+    const errorMsg = error instanceof Error ? error.message : '';
+    if (isContextLengthError(errorMsg)) {
+      return NextResponse.json(
+        { error: 'context_length_exceeded', details: errorMsg },
+        { status: 413 }
+      );
+    }
+    return NextResponse.json({
+      error: 'Failed to process chat request. Please try again.'
     }, { status: 500 });
   }
 }

apps/web/src/components/layout/right-sidebar/ai-assistant/SidebarChatTab.tsx

@@ -1,5 +1,6 @@
 import React, { useEffect, useState, useRef, useMemo, useCallback } from 'react';
 import { UIMessage } from 'ai';
+import { getAIErrorMessage } from '@/lib/ai/shared/error-messages';
 import { usePathname } from 'next/navigation';
 import { Button } from '@/components/ui/button';
 import { ChatInput, type ChatInputRef } from '@/components/ai/chat/input';
@@ -787,16 +788,7 @@ const SidebarChatTab: React.FC = () => {
         {error && showError && (
           <div className="p-2 bg-red-50 dark:bg-red-900/20 border border-red-200 dark:border-red-800 rounded text-xs flex items-center justify-between">
             <p className="text-red-700 dark:text-red-300">
-              {error.message?.includes('Unauthorized') || error.message?.includes('401')
-                ? 'Authentication failed. Please refresh the page and try again.'
-                : (error.message?.toLowerCase().includes('rate') ||
-                   error.message?.toLowerCase().includes('limit') ||
-                   error.message?.includes('429') ||
-                   error.message?.includes('402') ||
-                   error.message?.includes('Failed after') ||
-                   error.message?.includes('Provider returned error'))
-                ? 'Free tier rate limit hit. Please try again in a few seconds or subscribe for premium models and access.'
-                : 'Something went wrong. Please try again.'}
+              {getAIErrorMessage(error.message)}
             </p>
             <button
               onClick={() => setShowError(false)}

apps/web/src/lib/ai/shared/error-messages.ts

@@ -13,6 +13,11 @@ export function getAIErrorMessage(errorMessage: string | undefined): string {
     return 'Authentication failed. Please refresh the page and try again.';
   }
 
+  // Context length errors
+  if (isContextLengthError(errorMessage)) {
+    return 'The conversation is too long for this model\'s context window. Older messages have been trimmed to fit — try sending your message again.';
+  }
+
   // Rate limit errors
   if (
     errorMessage.toLowerCase().includes('rate') ||
@@ -36,6 +41,26 @@ export function isAuthenticationError(errorMessage: string | undefined): boolean
   return errorMessage.includes('Unauthorized') || errorMessage.includes('401');
 }
 
+/**
+ * Check if error is a context length / token limit error
+ */
+export function isContextLengthError(errorMessage: string | undefined): boolean {
+  if (!errorMessage) return false;
+  const msg = errorMessage.toLowerCase();
+  return (
+    msg.includes('context_length') ||     // API error key: context_length_exceeded
+    msg.includes('context length') ||     // Human-readable variant
+    msg.includes('context window') ||
+    msg.includes('maximum context') ||
+    msg.includes('token limit') ||
+    msg.includes('tokens exceeds') ||
+    msg.includes('too many tokens') ||
+    errorMessage.includes('413') ||
+    // OpenRouter / provider-specific phrasing
+    (msg.includes('maximum') && msg.includes('tokens'))
+  );
+}
+
 /**
  * Check if error is a rate limit error
  */

packages/lib/src/monitoring/ai-context-calculator.ts

@@ -217,8 +217,33 @@ export function getContextWindowSize(model: string, provider?: string): number {
     return 128_000; // Default for older MiniMax models
   }
 
-  // OpenRouter or unknown
-  return 200_000; // Conservative default
+  // OpenRouter - use model-specific limits where known, else 200k conservative default
+  if (providerLower === 'openrouter') {
+    // Claude models via OpenRouter
+    if (modelLower.includes('claude')) return 200_000;
+    // Gemini models via OpenRouter
+    if (modelLower.includes('gemini-2.5')) return 1_000_000;
+    if (modelLower.includes('gemini-2.0') || modelLower.includes('gemini-1.5')) return 1_000_000;
+    // GPT models via OpenRouter
+    if (modelLower.includes('gpt-4o') || modelLower.includes('gpt-4-turbo')) return 128_000;
+    // DeepSeek models - commonly 64k or 128k
+    if (modelLower.includes('deepseek-r1') || modelLower.includes('deepseek-v3')) return 128_000;
+    if (modelLower.includes('deepseek')) return 64_000;
+    // Qwen models
+    if (modelLower.includes('qwen-2.5') || modelLower.includes('qwq')) return 128_000;
+    if (modelLower.includes('qwen')) return 32_000;
+    // Llama models
+    if (modelLower.includes('llama-3') || modelLower.includes('llama3')) return 128_000;
+    if (modelLower.includes('llama')) return 32_000;
+    // Mistral models
+    if (modelLower.includes('mistral-large') || modelLower.includes('mistral-nemo')) return 128_000;
+    if (modelLower.includes('mistral')) return 32_000;
+    // OpenRouter platform hard cap is 400k for many endpoints - use 200k as safe default
+    return 200_000;
+  }
+
+  // Unknown provider/model - conservative default
+  return 200_000;
 }
 
 /**