Commentary | Codex Update

Free fix: @fazxes's rex-durable brings your Codex chats back after restarts. Running Codex in the Rex terminal? Tabs survive a restart but come back as empty shells. rex-durable remembers what each tab was running and reopens it: - Codex: codex resume <id> - Claude Code: claude --resume <id> - fx: fx --resume <id> Same tab, same split, same folder. Options like --model are kept. Install: git clone github.com/fazxes/rex-durabl… cd rex-durable ./install.sh --add-to-config It only types into idle shells and won't reopen a chat that's already running. Free and open source, macOS and Linux. Not on Rex? Built in: codex resume --last github.com/fazxes/rex-durabl…

From the v0.169.0 notes it's one answer today: untrusted dirs get neither project config nor MCP servers, and --trust saves trust for both. A read-config-but-don't-start-servers mode would be a fair feature request on openai/codex.

Good call. Per the v0.169.0 notes, --trust saves project trust for that directory, so it's not a one-run switch either. For CI on fork PRs, leave it off and accept running without the repo's config and MCP servers.

108 official .NET skills for Codex, one command to add. Microsoft's .NET team keeps them in dotnet/skills (MIT), grouped into plugins: testing (24), MSBuild (21), Blazor (10), MAUI (8), diagnostics (7), plus upgrades, NuGet, EF/data, ASP.NET Core and .NET 11. The repo ships a Codex-native marketplace manifest: codex plugin marketplace add dotnet/skills Then open /plugins in Codex, pick the dotnet-agent-skills tab and install only what you need. Update later: codex plugin marketplace upgrade dotnet-agent-skills Needs Codex CLI v0.121.0+. Only the skills and MCP servers carry over to Codex. The Copilot .agent.md files and C# LSP setup don't. h/t @AntonMartyniuk for spotting it github.com/dotnet/skills

Known bugs: 3 ways Codex desktop breaks in WSL mode. Workarounds for each, from the GitHub threads: 1. "Organization settings could not be loaded" on 26.1106.6628.0 (#60974) WSL sees the bundled Linux CLI as read-only (mode 444), so the startup check fails. Fix: copy codex and codex-code-mode-host from app\resources to a folder in your profile, chmod 755, then launch with CODEX_CLI_PATH pointing at the copy. 5 independent confirmations on Ubuntu, openSUSE and Oracle Linux. 2. Endless "Refreshing session..." loop on 26.1103.7773.0 (#60600) Fix: in ~/.codex/config.toml set [desktop] runCodexInWindowsSubsystemForLinux = false then fully restart. You lose WSL mode until it's patched. 3. App hangs on startup, "Not Responding" (#60484) At launch it runs bash -ilc in each distro with no timeout. Anything in ~/.bashrc that waits (sudo service docker start, a prompt) freezes the app. Fix: make those lines non-blocking (sudo -n, timeout) or skip them for non-interactive shells. Fair ask to OpenAI: a timeout on the WSL startup probe and an executable bundled CLI in the next Windows build. A lot of Codex users live in WSL. github.com/openai/codex/issu…

Add --trust to codex exec, or your scripts lose project config. Since Codex CLI v0.169.0, codex exec in an untrusted folder keeps the repo's project config and MCP servers off, even with --sandbox workspace-write. You get a warning, not an error, so a job can quietly run without its MCP tools. The fix: codex exec --trust "your prompt" codex review --trust What --trust does: - Saves trust for that project for this and future runs - Still saves it with --ephemeral - Refuses repos you marked untrusted - Can't be combined with --ignore-user-config On fresh CI runners, keep --trust in the command, since the saved trust lives in your local Codex config. Only trust repos you'd let start their own MCP commands. That's why this changed. Source: github.com/openai/codex/pull…

Now in Codex CLI v0.169.0: side-chat tabs, task search, foldable transcripts. Second stable release today, a few hours after v0.168.0. What's new: - The opt-in terminal workspace now supports multiple side-chat tabs, and saved local side chats reopen after a restart - Search older tasks in the agent command center by name, prompt preview or working directory - Turn on "Collapse turn details" to fold finished activity under each prompt while answers stay visible - Add or edit custom model providers from the terminal, with device-code sign-in for supported gateways over SSH - Ctrl+S in the report editor saves a diagnostic ZIP locally, even with uploads off Gotcha for scripts: codex exec now keeps project config and MCP servers off in untrusted directories, even with --sandbox workspace-write. Add --trust to codex exec or codex review to save project trust. Also fixed: "Yes, clear context and implement" now passes the full plan to the new session, and Windows sandbox issues with quoted cmd.exe paths. Update: npm i -g @openai/codex@latest github.com/openai/codex/rele…

Now in Codex CLI v0.168.0: provider picker, side-by-side chats, animated GIFs. What's new: - Browse model providers under Sessions > Model providers and pick the default for new chats. Existing chats keep their saved provider. - Turn on features.tui_workspace to open a side conversation next to the main chat (leader z zooms a pane) - Attach recordings when you're on an audio-capable model - features.image_generation_animation makes or edits animated GIFs - tui.notification_method = "osc777" for terminal notifications, with longer command previews on approvals Fixes worth knowing: - History and resume pickers only show threads from the ChatGPT account you're signed in with - Fuzzy file search ranks filename matches higher - Idle MCP servers on Windows no longer tie up filesystem I/O Heads-up: setting features.multi_agent now triggers a warning. Use agents.enabled to turn subagent tools off. Update: npm i -g @openai/codex@latest github.com/openai/codex/rele…

Now in Codex CLI v0.168.0: provider picker, side-by-side chats, animated GIFs. What's new: - Browse model providers under Sessions > Model providers and pick the default for new chats. Existing chats keep their saved provider. - Turn on features.tui_workspace to open a side conversation next to the main chat (leader z zooms a pane) - Attach recordings when you're on an audio-capable model - features.image_generation_animation makes or edits animated GIFs - tui.notification_method = "osc777" for terminal notifications, with longer command previews on approvals Fixes worth knowing: - History and resume pickers only show threads from the ChatGPT account you're signed in with - Fuzzy file search ranks filename matches higher - Idle MCP servers on Windows no longer tie up filesystem I/O Heads-up: setting features.multi_agent now triggers a warning. Use agents.enabled to turn subagent tools off. Update: npm i -g @openai/codex@latest github.com/openai/codex/rele…

4 Codex desktop account bugs, with workarounds: 1. Projects vanish after switching accounts (#59895) Several Windows and Mac users say desktop 26.1106.51428 fixes it. One Windows user is still affected. Update first. 2. Sidebar sections and project assignments gone on your second account (#60364) Not deleted. They are saved per account, so signing back into the first account brings them back. No way to carry them over yet. 3. Windows: "We couldn't load your account" with only a Retry button (#60311) Quit the app fully, rename %USERPROFILE%\.codex\auth.json to auth.json.bak, relaunch and pick Sign in with ChatGPT. Delete the .bak once it works. This signs the CLI out too. 4. "Subscription Sharing usage limit" error while your plan shows usage left (#59915) Check ChatGPT Settings > Usage > App limits. One user had that app switched off at 10%. Turning it on and saving restored access. Fair ask to OpenAI: keep local projects per machine, not per account, and make limit errors name the setting that blocks you. github.com/openai/codex/issu…

Codex plugin spotlight: @shuieng's Research Figure Kit keeps figures editable. It's an open-source (MIT) Codex plugin for paper and thesis figures. Instead of a flat AI image you have to re-prompt, it keeps the .drawio source, so fixing one arrow or label is a direct edit. 4 skills inside: - drawio: architecture, UML, flowcharts - drawio-academic-skills: paper-ready workflow and mechanism figures - drawio-poster: service-flow posters - scientific-visualization: Matplotlib, Seaborn or Plotly plots, profiling your data read-only first Install: git clone github.com/yss-off/research-… cd research-figure-kit codex plugin marketplace add . codex plugin add academic-figure-skills@research-figure-kit The core works offline. The live Draw.io backend is optional.

Known bug: Codex on iPhone says "No tasks yet." Your chats aren't gone. About 10 users across 3 GitHub issues hit it on ChatGPT iOS 1.2026.300 and .307, with Mac and Windows/WSL hosts. Remote still connects and new tasks still run, but the task list comes up empty. Workarounds for now: - Tap search: existing chats still open from there - Sign out and back in on the phone: the list comes back (until the next force-quit) - Push notifications and the Live Activity still open tasks Fair ask to OpenAI: fix the list on app relaunch and flag it in the iOS release notes. github.com/openai/codex/issu…

One flag swaps your whole Codex setup. Copy this: # ~/.codex/deep-review.config.toml model_reasoning_effort = "high" approval_policy = "on-request" Then run: codex --profile deep-review codex exec --profile deep-review "review this change" How it works: - Codex loads your normal ~/.codex/config.toml, then layers the profile file on top - So the profile only needs the keys that differ - One file per setup (deep-review, quick-fix, ci...) Gotcha: since CLI 0.134.0, --profile ignores old [profiles.name] tables in config.toml, and profile = "name" no longer works. Move them into ~/.codex/name.config.toml and delete the old lines. Docs: developers.openai.com/codex/…

Codex CLI v0.167.2 is out. Small patch, one change, for Google Cloud users. What's in it: - Codex now sends tool declarations incrementally by default on Google Cloud Vertex AI, instead of resending the full tool list every request - Same idea as v0.167.1's Bedrock flag, but this time it's on by default - Per the release notes, opt out with `incremental_tools_legacy_gcp = false` under [features] in config.toml Not on Vertex AI? Nothing changes for you. Update: npm i -g @openai/codex@latest github.com/openai/codex/rele…

Codex tip: audit a repo's .codex/ config before you trust it. agentaudit, open-sourced by Manifold Security, reads a repo's Codex, Claude Code and Cursor config and lists what it can do: - commands it runs (hooks, MCP servers, credential helpers) - env values it passes to servers - approval and sandbox changes - network redirects like proxies or API base URLs It only reads config. It never runs it. go install github.com/manifold-security… agentaudit github.com/OWNER/REPO GitHub URLs scan without cloning. Needs Go 1.26+, Apache-2.0. github.com/manifold-security…

Codex desktop 26.1103: 4 known bugs, 4 workarounds. 1. macOS: SSH remote projects don't reconnect at launch (#60047, 5 reports, still on build 25894) Workaround: Settings > Connections, re-enable the SSH connection 2. Ubuntu: Send stays grayed out at 0% weekly allowance even with credits left (#60445) Workaround: open Settings > Usage, or fully quit and relaunch 3. Windows: WSL reconnect loop keeps stealing focus from other apps (#60204) Workaround: uncheck "WSL: Ubuntu" under WSL connections 4. Linux: Enter and Send do nothing on build 26.1103.32239 (#60059) Workaround: roll back to 26.1027.51404 Fair ask to OpenAI: per-build desktop release notes (#60384), so users know which update fixes what. github.com/openai/codex/issu…

6 Codex CLI changes just merged, not released yet: 1. Cloud chats: pick the model and reasoning effort from the TUI (#60651) 2. Archive and restore cloud threads (#60635) 3. SSH or headless Linux: device-code sign-in for gateway providers, no local browser needed (#60662) 4. MCP tools marked readOnlyHint no longer ask for approval in auto mode (#60638) 5. Narrow terminals: composer footer shortens hints instead of clipping them (#60680 to #60683) 6. Safer repo checks: Codex skips submodule worktree scans that could run git clean filters (#60669) No release date yet. Watch for them here: github.com/openai/codex/rele…

Codex CLI v0.167.1 is out. Small patch, one change. What's in it: - New opt-in flag `incremental_tools_legacy` (off by default, marked under development) - It only applies to Amazon Bedrock providers (amazon-bedrock, amazon-bedrock-runtime) - With it on, Codex sends tool declarations as incremental updates instead of resending the full tool catalog on every request If you're not running Codex on Bedrock, nothing changes for you. No need to rush this one. Update: npm i -g @openai/codex@latest github.com/openai/codex/rele…

Known bug: Codex dictation can leave other apps stuck quiet. Your mixer says 100%, but browser audio stays low after dictating in the desktop app. What's been reported: - Linux: the dictation helper ducks other streams to 25%. If a stream ends before dictation does, it never gets restored, and it compounds (1.0, then 0.25, then 0.0625). Confirmed on Debian 13 and Arch. - Windows: Edge audio left at -18 dB with every slider at 100%, traced to dictation-ducking.exe. Linux workaround (PipeWire), reset the app's stream: wpctl status wpctl set-volume <stream-id> 1.0 Fair ask, OpenAI: always restore the original volume when dictation ends, even if the stream is gone. github.com/openai/codex/issu… github.com/openai/codex/issu…

Unconfirmed: GPT-6.2 Astra is not an announced Codex model. Posts today say `codex --model gpt-6.2-astra` is rolling out. What we can verify right now: - No announcement from OpenAI's official accounts - The official Codex models page lists Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna - The latest Codex CLI on npm is still 0.167.0 Before pinning a new model in config.toml or scripts, run /model in the CLI to see what your account can actually use. If this is a quiet test, a one-line note from OpenAI would save a lot of guessing. developers.openai.com/codex/…