fix(claude): reconcile max_tokens vs thinking budget and lift per-model ceiling (#2381)

On the translated OpenAI->Claude path, adjustMaxTokens capped max_tokens
before applyThinking set thinking.budget_tokens, so max-effort budget
(128000) could exceed a 64k-clamped max_tokens -> Anthropic 400.
prepareClaudeRequest now reconciles after the budget is known: prefer
raising max_tokens, only shrink budget when it meets/exceeds the ceiling.

Also lift the global 64000 cap: the ceiling is now the model's real
maxOutput, so high-output models (fable/mythos, opus-4.8/sonnet-4.6) get
their full budget. adjustMaxTokens gains an optional ceiling arg (default
unchanged, callers untouched); openai-to-claude passes the model maxOutput.

Native Claude Code passthrough is unaffected.

Co-Authored-By: Claude <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/translator/formats/claude.js

@@ -192,10 +192,27 @@ export function prepareClaudeRequest(body, provider = null, apiKey = null, conne
     delete body.output_config;
   }
 
-  // Clamp max_tokens to the model output ceiling (never above DEFAULT_MAX_TOKENS)
+  // Clamp max_tokens to the model's real output ceiling. Models whose caps
+  // declare a higher maxOutput (e.g. Opus 4.8 / Sonnet 4.6 = 128000) are allowed
+  // up to it, so max-effort thinking gets full budget; others fall back to the
+  // conservative 64000 default.
   if (body.max_tokens) {
-    const ceiling = Math.min(getCapabilitiesForModel(provider, body.model).maxOutput, DEFAULT_MAX_TOKENS);
+    const ceiling = getCapabilitiesForModel(provider, body.model).maxOutput || DEFAULT_MAX_TOKENS;
     if (body.max_tokens > ceiling) body.max_tokens = ceiling;
+
+    // Reconcile against thinking budget. applyThinking (thinkingUnified.js) runs
+    // AFTER adjustMaxTokens capped max_tokens, and the claude-budget format maps
+    // max effort → budget_tokens 128000 — larger than the clamped max_tokens.
+    // Anthropic requires max_tokens strictly greater than budget_tokens (else 400).
+    // Prefer raising max_tokens to preserve the requested thinking depth; if the
+    // budget alone meets/exceeds the ceiling, cap output and shrink the budget so
+    // some tokens remain for the answer.
+    if (body.thinking?.type === "enabled" && body.thinking.budget_tokens && body.thinking.budget_tokens >= body.max_tokens) {
+      body.max_tokens = Math.min(body.thinking.budget_tokens + 1024, ceiling);
+      if (body.thinking.budget_tokens >= body.max_tokens) {
+        body.thinking.budget_tokens = Math.max(1024, body.max_tokens - 1024);
+      }
+    }
   }
 
   // 1. System: remove all cache_control, add only to last block with ttl 1h

open-sse/translator/formats/maxTokens.js

@@ -3,9 +3,13 @@ import { DEFAULT_MAX_TOKENS, DEFAULT_MIN_TOKENS } from "../../config/runtimeConf
 /**
  * Adjust max_tokens based on request context
  * @param {object} body - Request body
+ * @param {number} [ceiling=DEFAULT_MAX_TOKENS] - Upper bound for max_tokens.
+ *   Callers with model context (e.g. openai-to-claude) pass the model's real
+ *   maxOutput so high-output models (Opus 4.8 = 128000) aren't pre-clamped to
+ *   the conservative 64000 default before the model-aware step sees them.
  * @returns {number} Adjusted max_tokens
  */
-export function adjustMaxTokens(body) {
+export function adjustMaxTokens(body, ceiling = DEFAULT_MAX_TOKENS) {
   let maxTokens = body.max_tokens || DEFAULT_MAX_TOKENS;
 
   // Auto-increase for tool calling to prevent truncated arguments (min never above max)
@@ -16,14 +20,14 @@ export function adjustMaxTokens(body) {
   }
 
   // Ensure max_tokens > thinking.budget_tokens (Claude API requirement)
-  // Claude API requires strictly greater, so add buffer instead of using DEFAULT_MAX_TOKENS
-  // which could equal budget_tokens when budget_tokens >= 64000
+  // Claude API requires strictly greater, so add buffer instead of using the
+  // ceiling which could equal budget_tokens when budget_tokens >= ceiling
   if (body.thinking?.budget_tokens && maxTokens <= body.thinking.budget_tokens) {
     maxTokens = body.thinking.budget_tokens + 1024;
   }
 
-  // Never exceed the global ceiling
-  if (maxTokens > DEFAULT_MAX_TOKENS) maxTokens = DEFAULT_MAX_TOKENS;
+  // Never exceed the ceiling
+  if (maxTokens > ceiling) maxTokens = ceiling;
 
   return maxTokens;
 }