fix: add provider-specific max_tokens cap (#711)

open-sse/config/constants.ts

@@ -66,6 +66,15 @@ export const DEFAULT_MAX_TOKENS = 64000;
 // Minimum max tokens for tool calling (to prevent truncated arguments)
 export const DEFAULT_MIN_TOKENS = 32000;
 
+export const PROVIDER_MAX_TOKENS: Record<string, number> = {
+  groq: 16384,            // Groq strict per-model enforcement
+  openai: 16384,          // GPT-4/4o standard
+  anthropic: 65536,       // Claude models
+  gemini: 65536,          // Gemini Studio
+};
+
+export const DEFAULT_PROVIDER_MAX_TOKENS = 32000;
+
 // HTTP status codes
 export const HTTP_STATUS = {
   BAD_REQUEST: 400,

open-sse/handlers/chatCore.ts

@@ -15,7 +15,7 @@ import { getModelTargetFormat, PROVIDER_ID_TO_ALIAS } from "../config/providerMo
 import { resolveModelAlias } from "../services/modelDeprecation.ts";
 import { getUnsupportedParams } from "../config/providerRegistry.ts";
 import { createErrorResult, parseUpstreamError, formatProviderError } from "../utils/error.ts";
-import { HTTP_STATUS } from "../config/constants.ts";
+import { HTTP_STATUS, PROVIDER_MAX_TOKENS } from "../config/constants.ts";
 import { classifyProviderError, PROVIDER_ERROR_TYPES } from "../services/errorClassifier.ts";
 import { updateProviderConnection } from "@/lib/db/providers";
 import { logAuditEvent } from "@/lib/compliance";
@@ -797,11 +797,7 @@ export async function handleChatCore({
   // Provider-specific max_tokens caps (#711)
   // Some providers reject requests when max_tokens exceeds their API limit.
   // Cap before sending to avoid upstream HTTP 400 errors.
-  const PROVIDER_MAX_OUTPUT_TOKENS: Record<string, number> = {
-    groq: 16384,
-    cerebras: 8192,
-  };
-  const providerCap = PROVIDER_MAX_OUTPUT_TOKENS[provider];
+  const providerCap = PROVIDER_MAX_TOKENS[provider];
   if (providerCap) {
     for (const field of ["max_tokens", "max_completion_tokens"] as const) {
       if (typeof translatedBody[field] === "number" && translatedBody[field] > providerCap) {