v0.5.2
CLI reliability fix (#102): --json now works with reasoning models (Qwen3.5, Kimi-K2.6, DeepSeek-style) over the openai-compatible provider. Sends response_format json_object on the OpenAI-style path (verified safe on Nvidia, Fireworks, Groq, Cerebras), hardens JSON extraction (strips think/reasoning, fence + balanced-brace recovery, reasoning_content fallback, key normalization, truncation hint), and fixes a Windows node crash on the error path. See CHANGELOG.md.