Raise Anthropic default max tokens

server/src/llm/protocols/messages-api.ts

@@ -28,6 +28,45 @@ import type { ChatMessage } from "../types.js";
 const INTERNAL_CORRECTION =
   "Internal correction: the previous assistant message claimed it would run a tool, but no tool call was made. If the task needs an available tool, call it now. Otherwise provide the final answer directly without saying you will run a tool.";
 
+const DEFAULT_ANTHROPIC_MAX_TOKENS = 128_000;
+const MODEL_MAX_TOKENS_CACHE_MS = 24 * 60 * 60 * 1000;
+
+const modelMaxTokensCache = new Map<string, { maxTokens: number; expiresAt: number }>();
+
+function readMaxTokens(value: unknown) {
+  return Number.isSafeInteger(value) && (value as number) > 0 ? (value as number) : undefined;
+}
+
+function getModelInfoMaxTokens(modelInfo: any) {
+  return readMaxTokens(modelInfo?.max_tokens) ?? readMaxTokens(modelInfo?.maxTokens);
+}
+
+async function getMessagesMaxTokens(params: ToolAwareCompletionParams) {
+  if (params.maxTokens) return params.maxTokens;
+
+  const cached = modelMaxTokensCache.get(params.model);
+  if (cached && cached.expiresAt > Date.now()) return cached.maxTokens;
+
+  try {
+    const retrieve = params.client?.models?.retrieve;
+    if (typeof retrieve === "function") {
+      const modelInfo = await retrieve.call(params.client.models, params.model);
+      const maxTokens = getModelInfoMaxTokens(modelInfo);
+      if (maxTokens) {
+        modelMaxTokensCache.set(params.model, {
+          maxTokens,
+          expiresAt: Date.now() + MODEL_MAX_TOKENS_CACHE_MS,
+        });
+        return maxTokens;
+      }
+    }
+  } catch {
+    // Fall back to the documented max for Claude Opus 4.8 and related high-output models.
+  }
+
+  return DEFAULT_ANTHROPIC_MAX_TOKENS;
+}
+
 function toTools(tools: any[]) {
   return tools
     .map((tool) => {
@@ -160,11 +199,12 @@ function mergeUsage(acc: Required<ToolAwareUsage>, usage: any) {
 
 export async function completeWithMessagesApi(params: ToolAwareCompletionParams): Promise<ToolAwareCompletionResult> {
   const enabledTools = getEnabledChatTools(params);
+  const maxTokens = await getMessagesMaxTokens(params);
   if (!enabledTools.length) {
     const response = await params.client.messages.create({
       model: params.model,
       system: buildTopLevelSystemPrompt(params.messages, params.userLocation),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
       temperature: params.temperature,
       messages: buildBaseMessages(params),
     } as any);
@@ -192,7 +232,7 @@ export async function completeWithMessagesApi(params: ToolAwareCompletionParams)
     const response = await params.client.messages.create({
       model: params.model,
       system: buildTopLevelSystemPrompt(params.messages, params.userLocation, buildChatToolSystemPrompt(params)),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
       temperature: params.temperature,
       messages: conversation,
       tools: toTools(enabledTools),
@@ -248,6 +288,7 @@ export async function completeWithMessagesApi(params: ToolAwareCompletionParams)
 
 export async function* streamWithMessagesApi(params: ToolAwareCompletionParams): AsyncGenerator<ToolAwareStreamingEvent> {
   const enabledTools = getEnabledChatTools(params);
+  const maxTokens = await getMessagesMaxTokens(params);
   if (!enabledTools.length) {
     const rawResponses: unknown[] = [];
     const usageAcc: Required<ToolAwareUsage> = { inputTokens: 0, outputTokens: 0, totalTokens: 0 };
@@ -259,7 +300,7 @@ export async function* streamWithMessagesApi(params: ToolAwareCompletionParams):
     const stream = await params.client.messages.create({
       model: params.model,
       system: buildTopLevelSystemPrompt(params.messages, params.userLocation),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
       temperature: params.temperature,
       messages: buildBaseMessages(params),
       stream: true,
@@ -315,7 +356,7 @@ export async function* streamWithMessagesApi(params: ToolAwareCompletionParams):
     const stream = await params.client.messages.create({
       model: params.model,
       system: buildTopLevelSystemPrompt(params.messages, params.userLocation, buildChatToolSystemPrompt(params)),
-      max_tokens: params.maxTokens ?? 1024,
+      max_tokens: maxTokens,
       temperature: params.temperature,
       messages: conversation,
       tools: toTools(enabledTools),