OpenAI Codex CLI Adds an Interactive Agents Dashboard and GitLab Support
OpenAI shipped a run of Codex updates across three days in mid-August. Codex CLI 0.148.0 (August 18) added conversation export to Markdown, session forking, Amazon Bedrock support, and asynchronous hook execution — letting teams run Codex against Bedrock-hosted models and fork an in-progress session to explore an alternate approach without losing the original conversation. The same day, ChatGPT for iOS 1.2026.223 added MCP form support and improved voice reliability.
On August 19, GitLab support reached beta across all ChatGPT plans, letting users connect a GitLab project to Codex cloud, create an environment for it, and start tasks directly from GitLab issues — extending a workflow that was previously GitHub-only. Codex CLI 0.149.0 followed on August 20 with the headline feature of this batch: an interactive agents dashboard for searching, starting, opening, renaming, and stopping tasks, addressing a real pain point for anyone running several Codex sessions in parallel. The same release added /cd, /pwd, and /cwd commands for directory management, a codex queue command for messaging existing sessions without interrupting them, enhanced Vim-mode editing, expanded codex doctor diagnostics covering endpoint protection and network issues, and SDK support for exact CLI config overrides and reasoning-effort selection.
Together these updates push Codex further toward being a session-management tool as much as a coding agent — the dashboard and queue command in particular are aimed at developers juggling multiple concurrent agent tasks rather than a single linear conversation.
Read more — OpenAI
Docker Sandboxes Bring Isolated Agent Execution to GitHub Actions
Docker detailed how Docker Sandboxes — isolated microVM environments originally built for running coding agents safely on a developer's laptop — now integrate with GitHub Agentic Workflows (gh-aw) as of version 0.82.9, letting teams run AI agents directly inside GitHub Actions with the same isolation guarantees. Each sandbox gets a dedicated microVM with its own kernel, filesystem, and network stack, plus a private Docker daemon so the agent has full root privileges inside the sandbox without ever touching the host runner's Docker daemon.
The integration supports Testcontainers by passing the sandbox's Docker socket through to nested containers, so an agent running inside the sandbox can spin up a database container and run real integration tests rather than being limited to unit tests. Docker's worked example shows an agent reading requirements for a Java 21 registration service, running baseline tests, finding a case-sensitivity bug in email handling, fixing it with a one-line change, and opening a draft pull request restricted to src/** — all inside a hosted ubuntu-24.04 runner in roughly 11 minutes.
Configuration is declarative: workflows specify a network allowlist, sandbox runtime, and safe-outputs policy (such as create-pull-request with protected-files: blocked and an allowed-files scope) directly in the workflow file, and gh aw compile generates the corresponding lock file. For teams already experimenting with autonomous coding agents, this closes a real gap — running agents in CI with meaningfully weaker isolation than a local sandbox has been a common shortcut, and this gives that setup a supported, auditable path via Docker AI Governance.
Read more — Docker Blog
Liquid AI's DSpark Draft Models Deliver Up to 3.2x Faster Local Inference
Liquid AI released DSpark draft-model checkpoints for three models in its LFM2.5 family — LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B — adding a speculative decoding path that trades a small memory increase for a substantial decoding speedup without changing output quality. Speculative decoding works by having a lightweight draft model propose candidate tokens, which the full target model then verifies in a single forward pass, amortizing the cost of loading weights across all verified tokens at once instead of one token at a time.
The gains varied meaningfully by hardware and architecture: up to 3.18x throughput improvement on GPU (H100) and up to 2.87x on-device. Liquid AI noted that LFM2.5-8B-A1B, a mixture-of-experts model, showed the strongest GPU speedup but the weakest on-device gain — only 1.18x mean on an M4 Max (90 to 106 tok/s) — because of how MoE models currently execute on Metal in llama.cpp. The DSpark checkpoints ship with day-one support for llama.cpp and SGLang, and run on Apple silicon through the Metal backend.
For teams deploying local or on-device LLM inference, this is a practical example of the draft-model pattern maturing beyond research demos into shipped, quantized (GGUF) checkpoints that plug into existing inference stacks with no architecture changes on the serving side — only the choice of which checkpoint to load.
Read more — Hugging Face / Liquid AI
