heyhuynhgiabuu/proxypal v0.1.9

v0.1.9 - Extended Thinking Fix for Claude via Vertex AI

on GitHub

What's New

🧠 Extended Thinking/Reasoning Mode Fix

Fixed extended thinking mode for Claude models accessed through Vertex AI/Antigravity backend (e.g., gemini-claude-opus-4-5-thinking).

Root causes fixed:

• CLIProxyAPI: geminiToAntigravity() was incorrectly deleting maxOutputTokens and resetting thinkingBudget=-1 for all non-Gemini-3 models, including Claude
• CLIProxyAPI: ensureMaxTokensForThinking() only checked snake_case budget_tokens, missing camelCase budgetTokens
• ProxyPal: OpenCode config used budget_tokens but AI SDK expects budgetTokens (camelCase)
• ProxyPal: Output limit was too low (16384) for thinking models with budgetTokens=16000

Changes:

• Added isClaude check to preserve thinking config for Claude models via Antigravity
• Support both budget_tokens and budgetTokens formats
• Increased output limit to 32000 for -thinking suffix models
• Rebuilt CLIProxyAPI binary with fixes

This enables OpenCode and other AI tools to properly display Claude's thinking blocks when using ProxyPal with Vertex AI backend.