OpenAI's Codex CLI gets an agents dashboard and per-session queuing
v0.149.0 lets you queue messages and juggle agent tasks from the terminal — plus redacted session links, an on-device piano model, and open diffusion speed.
Copy markdown
One dashboard for every Codex task
v0.149.0 adds an interactive agents dashboard to search, start, open, rename, and stop tasks with custom shortcuts, and you can now queue follow-up messages into a running local or remote session instead of waiting for it to finish.
Move Codex around your repo mid-session
New /cd, /pwd, and /cwd commands change the working directory without restarting, and Vim editing picks up cw, c$, and cc — handy for driving a multi-package monorepo from a single session.
codex doctor now debugs the broken install
The rebuilt codex doctor diagnoses endpoint-protection, proxy, network, and update-connectivity failures, alongside fixes to session resumption, sub-agent duplication, and WebRTC stability.
Share a Codex session as a redacted read-only link
The same August 20 Codex update adds shareable read-only chat snapshots with automatic secret redaction, so you can hand off a debugging session for review without leaking API keys.
Build this weekend: a 125M model that jams on piano
A new write-up trains a 125M on-device transformer to autocomplete piano from live MIDI at ~108 notes/sec on an iPhone 15 (INT8 Core ML), learned from ~300M public-domain note events; it ships free as the RollTab app, with the full tokenization and DPO recipe in the post.
Elsewhere: DiffusionGemma's report clocks ~1,500 tok/s
Resurfacing on Hacker News, Google DeepMind's DiffusionGemma report details how the open, Apache-2.0 model (26B MoE, 3.8B active) reaches roughly 1,500 tokens/sec on one H100 by denoising ~256 tokens in parallel — fine-tuned from Gemma 4 on under 10% of its budget. Weights are on Hugging Face.