Codex CLI ships a fullscreen TUI and turns voice on by default
A coordinated night for terminal agents: Copilot CLI and Claude Code adopt Opus 5.5, Ollama adds offline model transfers, llama.cpp speeds up Intel GPUs.

Copy markdown
Voice, no longer behind a flag
Codex CLI 0.156.0 promotes voice conversations — experimental in 0.155 — to on by default, with an F8 toggle and bundled audio for Linux and Windows. You can talk through a change hands-free while the diff scrolls, no setup required.
A terminal that scrolls, searches, and draws
The same release adds an optional fullscreen TUI (/tui) with transcript search, mouse selection, right-click copy, six themes, and native Mermaid and equation rendering. Worktree sessions now default on — fan out parallel agents and filter tasks by status — and /usage surfaces token, plugin, and skill totals.
Opus 5.5 lands in two more terminals
GitHub Copilot CLI v1.0.89 adds claude-opus-5.5 to its model list, and Claude Code 2.1.280 makes Opus 5.5 the default Opus (1M context, $4/$20 per Mtok, $0.20 cache reads) — and flips Pro and Team Standard plans from a Sonnet default to Opus. Check /model before your next run so you aren't paying Opus rates by accident.
Ollama learns to move models offline
New ollama export and ollama import commands ship content-addressed, offline model transfers: copy a pulled model to an air-gapped or bandwidth-capped box without re-downloading from a registry. Built for homelabs and locked-down networks.
llama.cpp speeds up Intel GPUs
This week's llama.cpp builds add Vulkan flash-attention for Intel Xe and OpenCL binary kernels for quantized models, plus a fix for a server router eviction race. Local inference on Intel Arc and Xe gets faster and steadier under concurrent load.