Codex CLI 0.149.0 Adds Agent Dashboard
This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.
@openai’s Codex changelog for Codex CLI 0.149.0 is an official release note for OpenAI Codex tooling, published August 20, 2026. Managing long-running local and remote Codex sessions without losing context, approvals, or working-directory state. That 0.149.0 moves the Codex CLI toward a session control surface, not just a prompt box. For Codex CLI github readers, the release also points back to the openai/codex repo for the implementation trail and issue discussion.
Open the agents dashboard before starting another thread
The Codex CLI is a terminal interface for running OpenAI Codex against a repository or a remote coding session. In 0.149.0, @openai added an interactive Codex agents dashboard for searching, starting, opening, renaming, and stopping tasks, with configurable shortcuts.
That matters because the hard part of agent work is often not asking the first question. It is remembering which task is still alive, which one has pending output, and which one should be stopped before it edits the same files as another session.
Picture a repo like billing-api with three threads: one investigating flaky Jest tests, one updating invoice copy, and one trying a database migration. A dashboard gives those sessions a visible place to live. Don't treat visibility as cleanup. If task names are vague, a dashboard only makes the mess easier to see.
Send follow-ups with codex queue
The release adds codex queue, a command for sending messages to existing local or remote sessions. This is the small feature that feels boring until it saves you from opening a duplicate terminal and making the same agent rediscover the same repo state.
The changelog also says queued messages now wake idle sessions more reliably. Duplicate session names resolve more usefully, and pasted or deferred command semantics are preserved. In plain English: queueing is meant to behave like a continuation, not like a half-new chat wearing an old name.
Use it when the session already owns the task. For example, if a Codex agent is halfway through replacing a deprecated feature flag, queue the follow-up there: ask it to include the migration test, not to start a fresh investigation. Avoid queueing stateful instructions when you are unsure of the working directory or permission profile.
Check location and permissions before trusting a resume
Codex CLI 0.149.0 adds /cd, /pwd, and /cwd commands for managing the working directory in TUI sessions. That is a nice quality-of-life change, but it is also a safety feature. A coding agent that is in the wrong directory can produce very confident nonsense.
The release also fixes resumed and forked threads so they restore their active permission profile instead of silently falling back to current defaults. That is a sharp bug fix. If you paused a session under a restrictive profile, you do not want a later resume to inherit a looser default without making noise.
There are adjacent fixes for duplicate sub-agent activity, TUI routing for sub-agent notifications and approvals, and bounded replay buffers for inactive TUI threads. None of this is glamorous. It is the plumbing that makes a codex agent feel less spooky during real work.
Run codex doctor when the environment feels haunted
codex doctor now diagnoses endpoint protection, network and proxy failures, desktop app state, and update connectivity. That list reads like a week of actual support tickets compressed into one command.
This helps when Codex behaves inconsistently across laptops. One developer might be blocked by endpoint protection, another by a proxy, and another by update connectivity. A doctor command gives the first pass a shared shape before everyone starts guessing.
It does not replace debugging. It tells you where to look first. If a WebRTC sideband connection drops, 0.149.0 also includes a reconnect fix for unexpected transport loss without dropping pending output, but network policies and desktop state can still be outside Codex’s control.
Try 0.149.0 in one existing repo
Start with one repo where a failed experiment will be annoying, not catastrophic. If you are mapping this into your own Codex CLI workflows, keep the first pass narrow: one branch, one task, one verification loop.
Install the exact version from the changelog:
npm install -g @openai/codex@0.149.0
codex doctor
codex
Inside the TUI, try the working-directory commands before asking Codex to edit anything:
/pwd
/cwd
/cd packages/api
Then add a tiny repo-local instruction so the agent has a boundary. A practical AGENTS.md for a Node service might look like this:
# AGENTS.md
- Before editing, confirm the working directory with /pwd in the Codex TUI.
- Reuse the existing Codex session for follow-up requests when it owns the task.
- After changes, run: npm test -- --runInBand
- If MCP access is needed, use read-only GitHub and Jira tools unless the task explicitly asks for writes.
That MCP note is intentionally plain. Codex MCP access is powerful because it can connect work to GitHub, Jira, docs, or private systems. The boundary should say what kind of access is expected before the agent discovers tools mid-task.
If you came from Codex CLI github, treat GitHub as the discussion trail and the official changelog as the behavior list. The two answer different questions: one shows how the project is maintained, the other says what shipped.
A related pattern shows up in cliclaw Runs Coding CLIs From Telegram: once coding sessions become controllable from somewhere else, naming, wakeups, and handoffs matter more than the chat transcript.
Checklist: what to try first in one repo
- Install the exact release with
npm install -g @openai/codex@0.149.0. - Run
codex doctorbefore starting a real task. - Open one known repo on a throwaway branch.
- Confirm the TUI working directory with
/pwdor/cwd. - Start one task, give it a boring name, and look for it in the agents dashboard.
- Use
codex queueonly for a follow-up that belongs to that same task. - Resume or fork a session, then confirm the active permission profile before edits.
- Add or update
AGENTS.mdwith the repo’s test command and MCP access boundary. - Stop the task from the dashboard when you are done, so stale sessions do not keep pretending to be active work.