A demo of Codex CLI, Qwen 3.6, stream_interval=23 (simulate large multi-token like from speculative decoding, MTP, or future diffusion) before the new qwen3 parser in PR https://github.com/vllm-project/vllm/pull/45413
The main thing of note is how it struggles to keep even the most basic agentic loop going as tool call parsing failures, dropped tokens, and similar happen that continue to break the loop.