Claude: "Auto" effort = no thinking

da7f97b66377a970f1accb890ffd865aeb269ffb

Cohee <18619528+Cohee1207@users.noreply.github.com>

src/endpoints/backends/chat-completions.js+4 -3 Viewed

             betaHeaders.push('prompt-caching-2024-07-31');
+        }
-        if (useThinking) {
+        const reasoningEffort = request.body.reasoning_effort;
+        const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
+        if (useThinking && Number.isInteger(budgetTokens)) {
             // No prefill when thinking
             voidPrefill = true;
-            const reasoningEffort = request.body.reasoning_effort;
-            const budgetTokens = calculateClaudeBudgetTokens(requestBody.max_tokens, reasoningEffort, requestBody.stream);
             const minThinkTokens = 1024;
             if (requestBody.max_tokens <= minThinkTokens) {
                 const newValue = requestBody.max_tokens + minThinkTokens;

src/prompt-converters.js+3 -2 Viewed

  * @param {number} maxTokens Maximum tokens
  * @param {string} reasoningEffort Reasoning effort
  * @param {boolean} stream If streaming is enabled
- * @returns {number} Budget tokens
+ * @returns {number?} Budget tokens
  */
 export function calculateClaudeBudgetTokens(maxTokens, reasoningEffort, stream) {
     let budgetTokens = 0;
     switch (reasoningEffort) {
+        case REASONING_EFFORT.auto:
+            return null;
         case REASONING_EFFORT.min:
             budgetTokens = 1024;
             break;
         case REASONING_EFFORT.low:
             budgetTokens = Math.floor(maxTokens * 0.1);
             break;
-        case REASONING_EFFORT.auto:
         case REASONING_EFFORT.medium:
             budgetTokens = Math.floor(maxTokens * 0.25);
             break;