Codex CLI gets a /app command, faster streams, and smarter image handling
The latest Codex CLI release adds Desktop handoff via /app, local image path exposure, Alt-key fallback shortcuts, and MCP/Ollama performance improvements.
The Codex CLI has shipped a release that meaningfully tightens the connection between your terminal and the Codex Desktop app, while fixing a handful of rough edges that have been quietly annoying power users for a while. Nothing here is a wholesale reinvention, but the cumulative effect is a noticeably smoother experience, particularly if you move between CLI and GUI, work with images, or run heavy MCP or Ollama sessions.
Here is what changed and why it matters.
Jump from the CLI to the Desktop app mid-session
The headline addition is a /app command that hands off your current CLI thread directly into Codex Desktop on macOS or native Windows. You no longer need to abandon your session and manually recreate context in the app. Type /app, and the thread moves across.
On macOS, Codex opens the workspace path automatically. On Windows, it prints the path you need to open, which removes the previous manual prompt step that had been blocking a smooth transition. There is also a new codex app terminal command that opens an installed Desktop app outright, or kicks off the installer if the app is not present yet.
For anyone who has been using the CLI for its speed but occasionally wanted the Desktop app’s parallel thread management, worktree support, or Git tooling, this is the missing bridge. You start a session in the terminal because it is fast, and you escalate to the app when the work grows complex enough to benefit from a visual interface. No thread context lost in the handoff.
Image workflows are more reliable
Two changes here, both addressing the same underlying problem. Local image attachments and standalone image generations now expose their saved file paths directly to the model. Previously, when you wanted to follow up on an attached image or edit a generated one, the model was effectively guessing the file path from conversation history. Unsurprisingly, that produced errors.
The fix means file references are passed accurately, so follow-up edits and image manipulations land on the right input. If you use Codex for anything involving image generation or image-based edits, this is a meaningful reliability improvement rather than a cosmetic one.
Reasoning effort selection gets more flexible
The terminal UI (TUI) now includes Alt-key fallback shortcuts for changing reasoning effort levels. Some terminals do not pass Alt-key bindings through correctly, which meant users in those environments could not adjust effort levels at all. The fallback shortcuts close that gap.
Separately, model-defined effort levels now flow through in the order the model itself advertises them, rather than being reordered by the CLI. These are small fixes, but reasoning effort control is central enough to the Codex workflow that getting it right across all terminal environments matters.
Faster processing for MCP, Ollama, and long histories
A single low-level optimisation to byte scanning has improved processing speed across three areas: large MCP server response streams, Ollama inference streams, and long local message histories. If you run Codex against a local model via Ollama or connect it to MCP servers that return substantial payloads, you should notice the difference in throughput. The optimisation applies uniformly, so longer conversation histories benefit too.
AGENTS.md loading is more accurate for edge-case workspaces
Codex discovers project configuration, including AGENTS.md custom instructions, by walking up from the working directory until it reaches a project root (typically the directory containing .git). For most setups this works cleanly, but remote and symlinked workspace layouts could cause the wrong AGENTS.md to be loaded, meaning your custom instructions were silently ignored or the wrong ones were applied.
This release fixes that resolution logic for those edge cases. Teams working with symlinked monorepos or remote workspace configurations should now see the correct instructions picked up consistently.
A handful of other fixes worth knowing about
A few smaller items round out the release:
Session resume performance. resume --last now queries the state database first to find the newest matching session, which speeds up restore when you have a large local session history.
Thread reset stability. Commands like /new, /clear, and /fork no longer drop cloud-managed requirements or feature flags when the TUI reloads configuration. This was a subtle bug that could change behaviour after a reset without any obvious signal.
MCP subagent warning routing. Startup warnings from MCP subagents now stay in the thread that owns them. Previously they could surface as duplicate alerts in the parent thread and leave spinners stuck in the TUI.
URL display in the TUI. Bare URLs containing a tilde (~) in the path were being truncated at that character instead of being fully linkified. Fixed.
CLI prompt argument handling. codex resume --last "..." and codex fork --last "..." now correctly treat the trailing argument as the initial prompt rather than misreading it as a session ID.
Who should care
If you are a Codex CLI user on macOS, the /app handoff is the most immediately useful change. On Windows, the removal of the manual prompt step makes the CLI-to-Desktop workflow viable in a way it was not before.
If you run local models through Ollama or connect to MCP servers with large response payloads, the stream optimisation will show up clearly in longer sessions.
If your team uses remote or symlinked repository layouts, check that AGENTS.md instructions are now loading as expected after this update. The previous behaviour was silent enough that some teams may not have realised the wrong instructions were being applied.
Codex is available to ChatGPT Plus, Pro, Business, Edu, and Enterprise subscribers. The CLI is open source on GitHub and built in Rust.