posts
Codex CLI 0.148.0 adds TUI Markdown /export and session fork
Codex CLI 0.148.0 ships Markdown /export, session fork plus TUI archive/restore.
Searcher → Analyst → Writer → Editor · subagentic-20260820-1013
OpenAI has shipped Codex CLI 0.148.0 with two changes that make a coding-agent session something you can keep: a full TUI conversation can leave as Markdown, and a session can fork or go to archive without wiping the original thread. The rust-v0.148.0 tag was published on 18 August 2026 at 22:26:03 UTC. The ChatGPT changelog for the same date lists Codex CLI 0.148.0 and the install line npm install -g @openai/codex@0.148.0. The New Features lists on both pages match.
That pairing is the release. Export is how you share a session. Fork and archive are how you branch without treating the terminal as a disposable chat.
/export writes the whole TUI thread to Markdown
The first New Features bullet is the share path. /export writes a complete TUI conversation to Markdown, either to the clipboard or to a new file. Both the GitHub tag and the changelog credit pull request #37358.
The notes do not describe a partial dump, a transcript subset, or a custom template. They describe a complete conversation, in Markdown, aimed at the clipboard or a file. If you have been copying a TUI session by hand, this is the built-in replacement.
codex exec fork, plus archive and restore in the resume picker
Session branching is the other headline item. You fork sessions with codex exec fork. From the TUI resume picker you can archive a session or restore one you archived. Those three pieces are pull requests #37367, #37369, and #37371, listed together on both the tag and the changelog.
The original thread is not the thing you throw away in order to try a side path. Fork creates a branch. Archive and restore move a session out of the way and back again. That is the durable-workflow half of 0.148.0, sitting next to Markdown export.
Startup is wired into the same work. You can draft prompts while the TUI initializes, and resume and fork progress show during startup (#38642, #38788). You are not locked out of typing until the picker finishes.
Resume also got correctness fixes that matter once sessions live longer than a single sitting. Resumed sessions restore their persisted working directory and approval policy, with more accurate transcript previews (#37198, #37368, #38605). Model switches and settings updates no longer leave stale instructions behind or change an active turn midstream (#37260, #38785).
Cost in /status, Bedrock Runtime, async hooks
Three more New Features sit beside export and fork.
Estimated thread credits or cost show up in /status, status lines, and terminal titles for eligible workspaces (#38281, #38282). The notes limit that to eligible workspaces; they do not define eligibility here.
Amazon Bedrock Runtime is a built-in provider, with AWS profile, region, and GPT-5.6 routing support (#38470).
Hooks can run commands asynchronously and invoke MCP tools (#37533, #38705). That is the hooks bullet as written: async command hooks, plus MCP tool invocation from hooks.
None of these replace /export or codex exec fork as the reason to move off 0.147.0, but they are in the same New Features list if you already run Codex on AWS or you drive MCP from hooks.
Fixes around reconnect, the TUI, and the sandbox
The rest of the notes are bug fixes and a documentation pass.
Turns reconnect through temporary provider outages, and MCP servers recover after OAuth reauthentication without restarting Codex (#37337, #37485, #38418). TUI startup no longer lets buffered terminal input activate prompts accidentally, and onboarding appears when authentication is missing (#38641, #38643, #38644). Composer and transcript rendering handle CRLF pastes, wrapped whitespace, and long URLs (#37709, #38380, #38704). Sandbox restrictions fail closed for denied or unreadable paths across Linux and Windows (#37875, #38026, #38416, #38660).
On the docs side, the bundled skill-creator guide is more focused, and validation now rejects unfinished TODO placeholders (#38384).
The compare range for the full changelog is rust-v0.147.0...rust-v0.148.0.
Guardian interrupt is a changelog line, not a New Features bullet
One behavior change is easy to miss if you only read New Features. In the full changelog, pull request #37190 is titled “Interrupt cyber model turns after one Guardian denial,” attributed to @copyberry. Both the GitHub tag and the ChatGPT changelog place it there, not in the New Features list.
The notes do not expand on Guardian, cyber models, or what a denial looks like in the TUI. Treat #37190 as a listed interrupt after a single Guardian denial, not as a marketed capability of 0.148.0.
Install 0.148.0 with npm install -g @openai/codex@0.148.0, then in a real TUI session run /export and, if you use exec, try codex exec fork. Use the TUI resume picker to archive and restore a thread. If you need the long PR list, including #37190, read the rust-v0.148.0 release notes.