Zed's Inline Assistant Did Nothing When the Model Emitted Reasoning Before Its Tool Call
GitHub·medium signal
Zed v1.18.0-pre (2026-08-26) fixes the inline assistant silently no-opping when a model produced reasoning tokens ahead of its tool call, a failure mode that gets more common as reasoning-by-default models spread. The same release adds GPT-5.6's 1M-token context window on Amazon Bedrock and Gemini 3.5 Flash-Lite to the Google AI models. The preceding v1.17.2 added low reasoning effort support for DeepSeek V4 Flash and V4 Pro.