v0.1.33

🚀 Release Highlights — v0.1.33 (2026-01-02)

🔄 Reactive Context Compression

• Automatic Recovery: Conversation automatically compressed when context length errors occur
• Seamless Continuation: Agents resume work after compression without losing progress
• Configurable Ratio: Set compression_target_ratio (0-1) to control how much context to preserve

📦 Streaming Buffer System

• Response Tracking: Tracks partial agent responses during streaming for compression recovery
• Backend Integration: Works across all supported backends (OpenAI, Claude, Gemini, Grok)

🛡️ MCP Tool Protections

• File Overwrite Protection: write_file tool refuses to overwrite existing files, preventing accidental data loss
• Task Plan Duplicate Prevention: create_task_plan blocks duplicate plan creation after compression recovery

🐛 Bug Fixes

• Grok MCP Tools: Fixed MCP tool visibility by adjusting tool handling in chat completions
• Gemini Vote-Only Mode: Fixed vote_only parameter handling in Gemini backend streaming
• GPT-5 Model Behavior: System prompt adjustments and default reasoning set for newer models
• Circuit Breaker: Improved debugging output with shorter ultimate timeout for faster failure detection

---

📖 Getting Started

• Quick Start Guide: Try the new features today
• Try These Examples:
  • test_reactive_compression.yaml - Test reactive compression with filesystem operations

What's Changed

• feat: Reactive Context Compression, Tool Result Eviction, Coordination Enhancements, Bug Fixes by @ncrispino in #697
• docs: docs for v0.1.33 by @Henry-811 in #729
• feat: v0.1.33 by @Henry-811 in #724

Full Changelog: v0.1.32...v0.1.33