What shipped
OpenAI released Codex CLI 0.149.0 on 20 August 2026, per the release notes on the openai/codex GitHub repository. It's a session and task-management release rather than a model or skills update: the headline feature is an interactive dashboard for juggling multiple running Codex tasks, alongside working-directory commands, a session-messaging command, expanded Vim editing, and diagnostic improvements to codex doctor.
The codex agents dashboard
The release notes describe it plainly: an interactive codex agents dashboard "for searching, starting, opening, renaming, and stopping tasks, with configurable shortcuts." If you've run more than one Codex task at a time before, the previous experience was tracking sessions yourself, remembering which terminal or window had which task, with no single view of what's actually running. The dashboard replaces that with one screen where you can search across tasks, start a new one, jump into an existing one, rename it for your own bookkeeping, or stop it outright, all through keyboard shortcuts you can configure rather than a fixed set.
This matters most for anyone running Codex against several repositories or several independent pieces of work simultaneously, the same use case that makes a dashboard useful in any tool managing concurrent background jobs. A single task running in the foreground doesn't benefit much from a dashboard built for managing many.
A worked scenario makes the difference concrete. Say you've kicked off three Codex tasks over the course of a morning: a long-running migration in one repository, a smaller bug fix in another, and a documentation pass you started and then set aside. Before 0.149.0, keeping track of all three meant remembering which terminal tab or window each one lived in, and there was no single place to check what state any of them were in without switching to each in turn. With the dashboard, codex agents opens one screen listing all three by name, lets you search if you've lost track of which is which, jump straight into the one that needs attention, rename the documentation task something more specific once you remember what it's actually for, and stop the migration task if it turns out to be stuck, all without hunting across windows first.
Working directory commands: /cd, /pwd, /cwd
0.149.0 adds three new TUI commands specifically for managing where a session operates from, without leaving it:
/pwd
/cwd
/cd path/to/directory
/pwd and /cwd report the session's current working directory. /cd changes it. Before this, redirecting a running session at a different part of a monorepo, or a sibling project entirely, generally meant starting a new session pointed at the right path from the outset. These commands make that a mid-session action instead.
codex queue: messaging a session from outside it
The release notes list codex queue as new in this release: "Added codex queue for sending messages to existing local or remote sessions." This is distinct from starting a new task through the dashboard. It targets a session that's already running and queues a message into it from outside, useful for handing off a follow-up instruction to a long-running task without switching into its own window or tab to type it directly.
Expanded Vim editing
For anyone who edits Codex prompts with Vim keybindings, 0.149.0 "expanded Vim editing with character replacement and more change motions such as cw, c$, and cc." These are standard Vim change operators: cw changes a word, c$ changes to the end of the line, cc changes the whole line. Their addition closes a gap for anyone whose muscle memory already assumes they work, rather than adding a new capability outside standard Vim conventions.
codex doctor gets more diagnostic coverage
codex doctor now checks a wider range of failure points. Per the release notes, it "diagnoses endpoint protection, network/proxy failures, desktop app state, and update connectivity." If Codex CLI is behaving oddly and you suspect the environment rather than your own configuration, codex doctor is the first thing worth running, and 0.149.0 makes it cover more of the actual causes of that kind of failure than before, network and proxy issues in particular, which are a common source of confusing, hard-to-diagnose behaviour in any CLI tool that talks to a remote API.
SDK changes: exact config overrides and reasoning effort
For anyone building on top of Codex rather than using the CLI directly, the release notes state that "SDK users can now pass exact CLI config overrides and select max or ultra reasoning effort." That gives SDK-based integrations the same level of control over configuration and reasoning effort that a person typing commands into the CLI directly already had, closing a gap where the SDK path was previously less flexible than the interactive one.
Bug fixes worth knowing about
A handful of the fixes in 0.149.0 address specific, previously confusing behaviour rather than general polish:
- Queued messages now wake idle sessions reliably, resolve duplicate session names more usefully, and preserve pasted or deferred command semantics, according to the release notes. If a queued message previously seemed to vanish against an idle session, this is the fix for that.
- Resumed and forked threads now restore their active permission profile instead of silently falling back to current defaults, so a thread you resume keeps behaving the way it did when you left it rather than reverting to whatever the defaults happen to be today.
- Duplicate sub-agent activity was fixed, alongside tightened TUI routing for sub-agent notifications and approvals, relevant if you've seen the same sub-agent event surfaced more than once.
- Realtime WebRTC sideband connections now reconnect after unexpected transport loss without dropping pending output, and inactive TUI thread replay buffers are now bounded, preventing excessive retained output from piling up on a thread you're not actively using.
Documentation changes
Two smaller documentation updates shipped alongside the code changes: a clarification that external contributions should go through issues and design discussion rather than pull requests directly, and new documentation of DNS exfiltration risks and trust limitations for secure devcontainers, worth reading if you run Codex inside a devcontainer and rely on its sandboxing assumptions.
Why a dashboard matters more as agentic coding scales
The dashboard is worth reading as a signal of how Codex CLI is being used, not just a UI addition. Running one coding agent against one repository doesn't need a dashboard; a single terminal window covers that case fine. A dashboard for searching, starting, opening, renaming, and stopping tasks only earns its place once people are routinely running several Codex tasks concurrently, whether that's parallel work across independent repositories, several long-running background tasks against the same monorepo, or a mix of interactive and unattended sessions started at different points in a day. Configurable shortcuts on top of that suggest the same direction: a tool built for someone switching between many running tasks quickly, rather than babysitting one.
This mirrors a pattern showing up across agentic coding tools more broadly this year. Claude Code shipped its own Monitor tool for streaming background script output without polling, and separately agent teams for coordinating multiple subagents at once. The common thread across both ecosystems is the same: as a single person runs more concurrent agent work, the tooling gap shifts from "can the agent do the task" to "can I actually keep track of everything it's doing," and 0.149.0's dashboard is Codex CLI's answer to that second question specifically.
How this compares to Claude Code's session management
Claude Code doesn't currently ship an equivalent single-screen dashboard for managing multiple concurrent tasks the way codex agents does; its comparable mechanisms are spread across /resume, the agents view, and cross-session messaging via SendMessage and ListAgents, which let you address and message other running sessions individually rather than see and act on all of them from one interactive view. codex queue, similarly, gives Codex a way to message an existing session from outside it, which is conceptually close to what Claude Code's SendMessage already does for cross-session communication, just under different naming and mechanics. Neither tool's approach is strictly more capable than the other's here; they reflect two different design choices for the same underlying problem of juggling more than one running agent session at a time, and which one fits better depends on whether you'd rather have one consolidated dashboard view or address specific sessions directly by name.
What this release doesn't touch
It's worth being explicit about scope, since this site's usual focus is Agent Skills: OpenAI's own release notes for 0.149.0 don't mention Agent Skills, SKILL.md, or the skills directories at all. Codex CLI's existing skills support, reading from ~/.codex/skills/ and .codex/skills/, loading at session start, and supporting explicit $skill-name invocation, is unaffected by this release. See Claude Code vs Codex CLI for Agent Skills for how that side of Codex CLI works, and Codex CLI can now import your Cursor skills for the most recent change that did touch skills specifically, in the prior 0.147.0 release.
Checking your version
codex --version
All of the features above require 0.149.0 or later. If you're on an earlier version, none of the dashboard, working-directory commands, codex queue, the expanded Vim motions, or the codex doctor improvements will be present, and there's no partial rollout to troubleshoot around, it's a version boundary, not a flag you need to enable.
Troubleshooting
The codex agents dashboard doesn't open. Confirm your version with codex --version first; it's new in 0.149.0 and doesn't exist on earlier releases.
/cd doesn't seem to change anything for the session's actual behaviour. It changes the working directory a session operates from, not its loaded context or history. If you need the session to also pick up different project-level configuration (a different AGENTS.md or skills directory, for instance), that's a separate consideration from the working directory alone.
codex queue says the target session can't be found. It targets a specific existing local or remote session, not a new one. Confirm the session you're trying to message is actually still running and that you have its correct identifier from the dashboard.
Vim motions like cw or cc don't behave the way they do in a real Vim install. These are Codex CLI's own implementation of the standard Vim change motions inside its prompt editor, not a full Vim emulation layer. Confirm you're on 0.149.0 or later, since the release notes describe this as newly expanded functionality rather than a pre-existing feature.
codex doctor reports a network or proxy failure you don't think is real. Since 0.149.0 specifically expanded diagnosis of network and proxy failures, a report here is more likely to reflect an actual environment issue than in earlier versions, worth checking your proxy configuration and endpoint protection settings before assuming it's a false positive.
Where to go next
For the release notes in full, including every bug fix not covered above, see OpenAI's own rust-v0.149.0 release page on GitHub. For how Codex CLI compares to Claude Code specifically for running Agent Skills, see Claude Code vs Codex CLI for Agent Skills, and for the most recent skills-specific change in Codex CLI, Codex CLI can now import your Cursor skills. Browse the current Codex CLI catalogue at getclaudeskills.com/platforms/codex-cli.
Verified 23 August 2026 directly against Codex CLI's own 0.149.0 release notes on the openai/codex GitHub repository.
