antigravity-claude-proxy 2.9.0 → 2.9.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "antigravity-claude-proxy",
-  "version": "2.9.0",
+  "version": "2.9.1",
   "description": "Proxy server to use Antigravity's Claude models with Claude Code CLI",
   "main": "src/index.js",
   "type": "module",

package/src/constants.js

@@ -202,8 +202,8 @@ export const STRATEGY_LABELS = {
     'hybrid': 'Hybrid (Smart Distribution)'
 };
 
-// Gemini-specific limits
-export const GEMINI_MAX_OUTPUT_TOKENS = 16384;
+// Gemini-specific limits (Gemini 3.8 / Gemini 4 Argon support up to 65536 output tokens)
+export const GEMINI_MAX_OUTPUT_TOKENS = 65536;
 
 // Gemini signature handling
 // Sentinel value to skip thought signature validation when Claude Code strips the field

package/src/format/request-converter.js

@@ -202,13 +202,24 @@ export function convertAnthropicToGoogle(anthropicRequest) {
             googleRequest.generationConfig.thinkingConfig = thinkingConfig;
         } else if (isGeminiModel) {
             // Gemini thinking config (uses camelCase)
-            // Clamp budget to model-specific max (e.g., Gemini 2.5 Flash max is 24,576)
+            // For callers that did not explicitly request thinking and asked for small max_tokens,
+            // use minimal thinking budget (32) and reserve output headroom so thinking doesn't
+            // consume the entire maxOutputTokens budget and return empty text.
+            const requestedBudget = thinking?.budget_tokens
+                ? clampGeminiThinkingBudget(modelName, thinking.budget_tokens)
+                : (max_tokens && max_tokens <= 1024 ? 32 : clampGeminiThinkingBudget(modelName, undefined));
             const thinkingConfig = {
-                includeThoughts: true,
-                thinkingBudget: clampGeminiThinkingBudget(modelName, thinking?.budget_tokens)
+                includeThoughts: Boolean(thinking?.type === 'enabled' || thinking?.budget_tokens || !max_tokens || max_tokens > 256),
+                thinkingBudget: requestedBudget
             };
             logger.debug(`[RequestConverter] Gemini thinking enabled with budget: ${thinkingConfig.thinkingBudget}`);
 
+            const currentMaxTokens = googleRequest.generationConfig.maxOutputTokens;
+            if (currentMaxTokens && currentMaxTokens <= thinkingConfig.thinkingBudget) {
+                const adjustedMaxTokens = Math.min(thinkingConfig.thinkingBudget + Math.max(currentMaxTokens, 8192), GEMINI_MAX_OUTPUT_TOKENS);
+                logger.debug(`[RequestConverter] Adjusting Gemini maxOutputTokens from ${currentMaxTokens} to ${adjustedMaxTokens} to exceed thinkingBudget (${thinkingConfig.thinkingBudget})`);
+                googleRequest.generationConfig.maxOutputTokens = adjustedMaxTokens;
+            }
 
             googleRequest.generationConfig.thinkingConfig = thinkingConfig;
         }

package/src/format/thinking-utils.js

@@ -4,7 +4,7 @@
  */
 
 import { MIN_SIGNATURE_LENGTH } from '../constants.js';
-import { getCachedSignatureFamily } from './signature-cache.js';
+import { getCachedSignature, getCachedSignatureFamily } from './signature-cache.js';
 import { logger } from '../utils/logger.js';
 
 // ============================================================================
@@ -447,6 +447,11 @@ function messageHasValidThinking(message) {
     const content = message.content || message.parts || [];
     if (!Array.isArray(content)) return false;
     return content.some(block => {
+        if (block && (block.type === 'tool_use' || block.functionCall)) {
+            if (block.thoughtSignature && block.thoughtSignature.length >= MIN_SIGNATURE_LENGTH) return true;
+            if (block.id && getCachedSignature(block.id)) return true;
+            return false;
+        }
         if (!isThinkingPart(block)) return false;
         // Check for valid signature (Anthropic style)
         if (block.signature && block.signature.length >= MIN_SIGNATURE_LENGTH) return true;