Qwen Code 0.24.0 lands a Linux sandbox and a breaking hooks change
A bwrap sandbox and web-search caps top a busy day for coding CLIs; Codex draws Mermaid in the terminal, and DeepSeek V4.1-Flash still owns trending.
Copy markdown
A kernel-isolated sandbox for Linux tool calls
Qwen Code 0.24.0 adds a bwrap (bubblewrap) sandbox backend that runs tool calls inside kernel-level isolation on Linux, so an agent's shell commands can't wander outside the box. If you run Qwen Code on a Linux dev machine, this is the upgrade to take.
The hooks change that can break your config
The release flags a breaking change: bash now expands project-directory variables inside command hooks (#11864), and web_search is capped per session (#11852) to stop runaway search loops. Re-check any custom hooks before upgrading, since variable references may resolve differently.
Codex draws Mermaid diagrams in the terminal
OpenAI merged PR #46054, rendering Mermaid code blocks as actual diagrams inside the Codex TUI, and Kimi Code shipped the same feature the same week. Ask your agent for an architecture diagram and see it drawn in the terminal instead of raw markup.
DeepSeek V4.1-Flash still owns Hugging Face trending
A week after launch, the MIT-licensed MoE is still the #1 trending model: it activates just 8B params on prefill and 16B on decode, and carries a 1M-token context at roughly 890 bytes of KV cache per token. It's the cheap long-context workhorse you can self-host today.
Heads-up: capacity errors and runaway agent loops
Builders are reportedly hitting a rough patch across coding CLIs this week, with recurring 'selected model is at capacity' errors on Codex and infinite agent loops or request storms on several tools. If your agent hangs or spams requests, it's not just you.