v2.1.2

AI Code Conductor v2.1.2

Stable Release

A fix for Macs. When the app is opened from the Dock or Finder, it now finds Claude Code for the setup version check, Insights and Sentinel, which failed before.

Fixed

• On a Mac, opening the app from the Dock or Finder now finds Claude Code for the setup version check, Insights and Sentinel. They failed before when Claude Code was installed in a folder that only your login shell adds to PATH, such as the folders Homebrew and the Claude Code installer use, and setup could say Claude Code was installed and, on the same page, that its version could not be read. These runs now use the Claude Code that setup found. Windows is unchanged.

Full changelog: https://github.com/nubbymong/ai-code-conductor/blob/v2.1.2/CHANGELOG.md

---

Upgrading from 2.1.0-beta.5 or earlier? Download the installer from this page — that build's updater looked for the old file name.

Claude and Claude Code are Anthropic trademarks.

SHA-256 checksums are in CHECKSUMS.txt attached to this release.

v2.1.1

AI Code Conductor v2.1.1

Stable Release

Everything new since 2.1.0. Codex is now a full assistant beside Claude Code, with Codex accounts of its own that you add in the app, and its sessions, reviews, usage and tools work as Claude's do. GPT-6-Sol and GPT-6-Luna join the Codex model picker. After an update, What's New shows what is new since the version you came from. First-run setup, the Windows installer and SSH sessions get many fixes, and new sessions start on a computer without Node.js. On a Mac, 2.1.1 needs macOS 13 or later; on macOS 12, stay on 2.1.0. The project now lives at github.com/nubbymong/ai-code-conductor. The lines that start with Known issue or Notice say what to do; Known issues and workarounds in the Feature Guide has the full list, and the notes for 2.1.1-beta.1, beta.2 and beta.3 list each change in full.

Added

• Codex is a full assistant beside Claude Code, and no longer marked Beta. Settings, Accounts is one page for both: a Providers card turns Claude Code and Codex on or off, says whether each is installed and which version, and lists the install or update commands, each with Run it for me and Copy. The app can run with Codex alone. Codex sessions and reviews run on this computer only, not over SSH.
• Codex accounts of their own. Add more than one, with ChatGPT or an API key (the key goes to Codex, and this app never stores it). Each signs in inside its own folder in the app's resources folder, so sessions never mix identities, and a saved config picks the Codex account it runs under. A Codex account is always added in the app, as a Claude account is; if you used Codex in 2.1.0, see the known issue below.
• A Codex session works as a Claude session does. It reopens in the same conversation after the app restarts, can restart into an earlier conversation, moves to another Codex account with Switch Account and keeps the conversation, shows in Logs, and has the attention dot, the Session Watchdog, Plan mode and Compact.
• Code review now works both ways: a Codex session can ask for a Claude review, as a Claude session already could for a Codex review, each on your reviewer account. A Codex review receives the change itself and runs no commands, so it works on Windows too.
• Usage and spend for Codex. The Usage page, the usage strip at the foot of the window and Tokenomics cover both assistants, each account shown on its own and never merged. A Codex account signed in with ChatGPT shows its 5-hour and weekly allowance, or its credits; an API-key account is billed per token.
• The Agent Canvas, Vision and the in-app browser, Ask Conductor, Sentinel, Insights, Cloud Agents and the Memory page work for Codex too, and a Codex account can sign in to chatgpt.com inside the app, kept apart from every other account.
• Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 are in the model picker, and the opus, sonnet and haiku choices now mean them.
• Setup can install Claude Code and Codex for you. Every install or update command has Run it for me, which runs it in a visible terminal once you confirm it, and Copy. Setup checks again when the command ends and finds the new install without restarting the app.

Changed

• GPT-6-Sol and GPT-6-Luna are in the Codex model picker; Codex lists them from version 0.156.1, the newest Codex this release is tested with. The Codex models and efforts match the ones Codex offers in its own picker.
• After an update, What's New shows what is new since the version you came from, under a heading that names it, with a short summary and pages on the main changes; the full notes stay in the Feature Guide, What's New. A fresh install asks which assistants you use: Claude Code, Codex, or both.
• Windows: the installer and uninstaller look like the app, with a new Welcome page and the splash's dark theme and artwork.
• The app runs on Electron 44. On a Mac it needs macOS 13 (Ventura) or later; on macOS 12, stay on 2.1.0. Its Mac download now has a name ending in macos13.dmg, which the updater in 2.1.0 does not look for (see the known issue below), and from 2.1.1 on the updater never offers a Mac a version that needs a newer macOS than it runs.
• The project now lives at github.com/nubbymong/ai-code-conductor, and old links to it redirect there. The app finds updates there on its own; if Check for Updates says Up to date, see the known issue below.
• Known issue: if Check for Updates says Up to date when a newer release is out, restart the app once while online, then check again. The app looks for the project's new home when it starts, so after a start while offline, or in an app already running when the project moved on October 12, 2026, its checks read the old name until it restarts.
• Known issue: Check for Updates also says Up to date when it cannot reach GitHub (no network, a proxy or firewall in the way, or GitHub's hourly limit on checks reached). Restart the app and check again later, or see the newest release on the Releases page on GitHub.
• Known issue: a Mac on 2.1.0 is not offered 2.1.1 in the app, and Check for Updates there says Up to date. 2.1.1 needs macOS 13 (Ventura) or later. On macOS 12, stay on 2.1.0. On macOS 13 or later, download 2.1.1's .dmg from the Releases page on GitHub once by hand; from then on the updater offers new versions as before.
• Known issue: on an SSH host whose shell is zsh, such as a Mac (zsh is its default), and possibly on another host whose shell is zsh, at least when the app runs on Windows, the SSH session's partner terminal opens in your home folder instead of the config's remote folder. Type cd and the folder in the partner (see Known issues in the Feature Guide).
• Known issue: if you used Codex in 2.1.0, add your Codex account once. Nothing carries over from the earlier Codex setting: choose Yes, set up Codex on the one-time Do you use Codex? page and sign in, or later turn Codex on and choose Add Codex account in Settings, Accounts. The Codex CLI's own sign-in is left as it is, and earlier Codex conversations stay with the Codex CLI: codex resume in a terminal in the project folder picks them up.
• Known issue: GPT-5.5 retires from Codex with ChatGPT sign-in on October 14, 2026, OpenAI says. From then on, a Codex config set to GPT-5.5 on an account signed in with ChatGPT needs another model: edit the config and pick one your account offers, such as GPT-6-Sol where OpenAI makes it available, or Default: follows Codex.
• Known issue: with a Codex newer than 0.156.1, the newest this release is tested with, the Providers card in Settings, Accounts, the Set up Codex page and Sentinel (when it is on) show a notice. Codex still runs. A newer Codex can also offer a model the picker does not list yet, such as GPT-6.1-Sol, and Sentinel then says it is not in the model picker: choose Default: follows Codex in the config to use the model Codex picks. If anything misbehaves, install Codex 0.156.1 (with npm, npm install -g @openai/codex@0.156.1; in PowerShell on Windows, npm.cmd install -g @openai/codex@0.156.1) and restart the app.
• Known issue: after an in-app update on Windows (from 2.1.0, or from 2.1.1-beta.1 to beta.3), if Vision on the Conductor MCP page shows Stopped and Start browser says its debug port 9222 is in use, the browser the earlier version started for Vision is still running, and the app could not check which programs are running, so it did not end it. Sign out of Windows and back in, or restart the computer; Vision then starts as usual.
• Known issue: on Windows, the app does not find its data or resources folder when the full path of that folder holds a character other than a plain English letter, a digit or ordinary punctuation, such as an accented letter, and your saved configs, accounts and sessions can then look missing. Everything in the folder you chose is still there. The Feature Guide, Known issues has the workaround: move what is in the folder to a new folder on this computer's own disk (not a USB drive, a FAT or exFAT drive or a network share) whose full path holds only plain English letters, digits, spaces, dots, hyphens and underscores, then run the installer again and choose the folders in their new place.
• Notice: a Claude Code version that a saved config was set to use in an older release is no longer used. Its sessions and cloud agents start the Claude Code installed on this computer, and the setting stays in the config. To run a particular version, install it yourself, make it the only Claude Code on your PATH, and turn off Claude Code's automatic updates (see Known issues in the Feature Guide).

Fixed

• First-run setup is sturdier: a folder that was not saved keeps the folder step open and says which one, and closing the app during Claude Code setup stops the setup terminal. On Windows and Linux, the setup screens, the onboarding pages and the one-time pages after an upgrade have the window's own minimize, maximize and close buttons.
• SSH: a session's partner terminal now opens on the host, and its command buttons run there; remote SSH sessions raise the attention indicator when Claude waits on you (Linux and macOS hosts; an SSH Persistent session started before the update gets it once you End it and start it again); SSH sessions with tmux start on a host whose login shell is zsh; and the mouse wheel scrolls an SSH Persistent session's history.
• New sessions start on a computer without Node.js: Claude Code or Codex starts directly, without the resume list a new session otherwise opens. Install Node.js 22 or later to get the list back.
• Many smaller fixes across accounts, sessions and Windows. Among them: a Claude Code session now starts when Claude Code was installed with npm into a folder whose path has a space or brackets in it; failure text and the app's log hide credentials written inside URLs; and at each start the app makes the folder its session scripts run from private to your account.

Full changelog: https://github.com/nubbymong/ai-code-conductor/bl...

v2.1.1-beta.3

AI Code Conductor Beta v2.1.1-beta.3

Beta Release (pre-release)

Fixes from testing 2.1.1-beta.2. Codex accounts are now always added in the app, and anyone who chose the Codex sign-in already on this computer in 2.1.1-beta.2 gets one page, once, to replace it. What's New shows what is new since the release you came from, and every prerelease of 2.1.1 shows the whole 2.1.1 story. An SSH session's partner terminal now opens on the host, remote SSH sessions raise the attention indicator, new sessions start on a computer without Node.js, and the Windows installer looks like the app.

Added

• The update showcase has five pages for 2.1.1, for anyone updating from 2.1.0 or earlier: Codex is now a full assistant, Codex accounts of their own, a Codex session keeps its place, code review both ways, and usage and spend for Codex. The summary above them lists what 2.1.1 brought, and the Feature Guide, What's New still has the full notes.
• Remote SSH sessions now raise the attention indicator when Claude is waiting on you (a permission prompt, an idle prompt, or an agent asking for a decision) and clear it on the same events a local session does. The signal comes back over the session's own terminal, so no extra port forward is needed. It covers SSH, SSH Persistent and SSH container sessions on Linux and macOS hosts. It follows the HTTP Hooks Gateway switch in Settings, Hooks: turning the switch off stops remote sessions raising it at once (one already raised still clears as usual), and a session that connected while the switch was off gets it from its next connection. Windows remote hosts are not covered yet.

Changed

• If you chose Use this sign-in in 2.1.1-beta.2, that account keeps working as before, confirmed at each launch. Its row in Settings, Accounts and every launch confirmation on it now say it is being retired: add a Codex account, make it the default, then make This computer's Codex inactive and archive it. Open sessions can move with Switch account, which keeps the conversation. Archiving it is final (it has no Restore), and only forgets it in the app: the Codex CLI stays signed in.
• If you chose Use this sign-in for the Codex sign-in already on this computer in 2.1.1-beta.2, the app now shows one page, once: Replace this sign-in with a Codex account. Add a Codex account opens the usual add-account dialog; Not now leaves everything as it is. Nothing is signed out.
• When a session or saved config that used that sign-in starts again, its confirmation also lists your other Codex accounts. Pick one and choose Use this account: the session starts on it, and a saved config that used that sign-in moves to it too. A restored tab moved this way starts without its earlier conversation, and the confirmation says so: to keep the conversation, choose Launch, then use Switch account on the running session.
• Windows: the installer and uninstaller now look like the app. A new Welcome page opens the installer. The Welcome, Finish and folder pages use the splash's dark theme and artwork, and every other page gets a dark header band with the app's mark. The installer's Welcome and folder pages play a short ring-sweep animation, and the Data and Resources folder pages have proper titles. A fresh install opens with a brief fade-in of the splash; updates go straight to the wizard.
• While the HTTP Hooks Gateway switch is on, an SSH session to a Linux or macOS host now also gets the remote project's own Claude Code hooks added to its settings file, alongside your user hooks, as local sessions do. They come from the files Claude Code itself reads on the host: the .claude/settings.json and settings.local.json in the session's folder and, for a session in a subfolder of a git checkout or in a linked git worktree, the checkout's own .claude/settings.local.json (for a worktree, the main checkout's), which the app finds with git on the host. A file is added only when it and its .claude folder belong to you, are not links, and are not writable by other users; writable by your own per-user group (the one Ubuntu and the RHEL family give each account) is fine, but not by a group other accounts share, as on macOS. Claude Code still reads the project's own settings files itself.
• GPT-6-Sol and GPT-6-Luna are in the Codex model picker. Codex lists both from version 0.156.1, the newest Codex this release is tested with. OpenAI names them as the replacements for GPT-5.5, which retires from Codex with ChatGPT sign-in on October 14, 2026 (see Known issues in the Feature Guide).
• The project now lives at github.com/nubbymong/ai-code-conductor, and old links to it redirect there. The app finds updates there on its own; if no update shows, restart the app once.

Fixed

• Codex accounts are now always added in the app. The Set up Codex page and Settings, Accounts no longer offer to use the Codex sign-in already on this computer (~/.codex): the app can never tell whose sign-in that is, so it had to be confirmed at every launch and could never run code reviews. Add a Codex account instead, as you would a Claude account.
• Archiving the Codex sign-in already on this computer now says plainly, before you confirm, that it is final: the account cannot be restored. The Codex CLI stays signed in.
• On the Codex sign-in already on this computer, Make inactive and Archive, which only take it out of the app, now come before Sign out, which is renamed Sign Codex out on this computer... because it signs the Codex CLI out everywhere on this computer. Sign in again is no longer offered there, so nothing on that row signs other apps out to sign in again.
• In Settings, Accounts on a narrow window, a row button such as This is still my account wraps inside its column instead of running past it, ChatGPT sign-in no longer breaks after the hyphen, and the reason behind Confirm each launch reaches screen reader users, not only a mouse pointer.
• What's New after an update now shows what is new since the version you came from, and its heading names that version (for example What's new since 2.1.0). An update within 2.1 no longer replays the whole 2.1 story and its nine pages. Every prerelease of a release shows that release's whole story, so 2.1.1-beta.1, beta.2, beta.3 and 2.1.1 itself all show what 2.1.1 brought, and an update with nothing new to tell shows one line about its fixes.
• The rename and roadmap page no longer opens every prerelease update for beta testers. It shows on a fresh install and to anyone updating from a build before the rename.
• The update pages no longer promise more than is true: the last one no longer says there is nothing to set up when the one-time Codex question follows it, the rename page and its summary line no longer say every account carries over or that nothing else changes (the earlier Codex setting does not carry over), and the Agent Canvas page no longer calls itself new.
• Hello Codex says where each Codex account's usage is: a ChatGPT-signed-in account shows its 5-hour and weekly allowance, or its credits, on the Usage page, and an API-key account is billed per token. Its launch page also names the one-time prompt to trust the app's hooks (Trust all and continue), which the attention dot needs.
• The Enable Multi Spawn page no longer comes back after every update. It appears once, on the update that brings Multi Spawn in, and if you already answered it on an earlier version you will not see it again. Its opening line also no longer calls Multi Spawn new in this update.
• On Windows, a data or resources folder whose path holds a percent sign (%), an exclamation mark (!) or a backtick is now saved exactly as you chose it. Before, the app left those characters out when it saved the folder, so after a restart it could look in a different folder.
• First-run setup no longer moves on as if a folder were set when it was not saved. If a folder is empty, is not a full path, or cannot be created or saved, the folder step stays open and says which folder, so you can choose another one.
• The pages that cover the whole window now have the window's own minimize, maximize and close buttons at the top right on Windows and Linux, and the bar along their top moves the window: the brief loading screen when the app starts, the first setup screens (choosing the app's folders, installing and setting up Claude Code), the onboarding pages, the release notes page, and the one-time pages shown after an upgrade (such as the Do you use Codex? question, the Multi Spawn page and the Codex introduction, including its replay). Before, these pages covered the window's own buttons, and the first setup screens had only Exit, which stays.
• Closing the app while Claude Code setup is running (Alt+F4, the taskbar, Cmd+Q, or the red button on a Mac) now stops the setup terminal. Before, only the setup screen's own buttons stopped it, and on a Mac it kept running after the window closed until you quit the app.
• First-run setup no longer stops a Claude Code install that is still finishing when Check again finds Claude Code. It says the install is finishing and carries on by itself when the install ends. On a Mac or Linux, Stop it and continue is there for a terminal left at a prompt.
• On Windows, the install screen no longer tells you to press Back at a prompt, which stopped the install. It says setup checks again by itself when the install ends, and when to press Check again.
• When the setup screen shows a PATH hint, Check again now says it checked again instead of changing nothing on screen.
• On a Mac or Linux, when Claude Code or Codex is in ~/.local/bin but your login shell's PATH does not include it, Copy now gives a command to run once that adds the folder to the shell file your login shell reads. Before, it gave a line that only changed the terminal you pasted it into.
• On Windows, when your PATH names a folder through a variable set after the app started (such as nvm-windows' NVM_SYMLINK), the app now notes, beside its not-found ...

v2.1.1-beta

AI Code Conductor Beta v2.1.1

Beta Release (pre-release)

Everything new since 2.1.0. Codex is now a full assistant beside Claude Code, with Codex accounts of its own that you add in the app, and its sessions, reviews, usage and tools work as Claude's do. GPT-6-Sol and GPT-6-Luna join the Codex model picker. After an update, What's New shows what is new since the version you came from. First-run setup, the Windows installer and SSH sessions get many fixes, and new sessions start on a computer without Node.js. On a Mac, 2.1.1 needs macOS 13 or later; on macOS 12, stay on 2.1.0. The project now lives at github.com/nubbymong/ai-code-conductor. The lines that start with Known issue or Notice say what to do; Known issues and workarounds in the Feature Guide has the full list, and the notes for 2.1.1-beta.1, beta.2 and beta.3 list each change in full.

Added

• Codex is a full assistant beside Claude Code, and no longer marked Beta. Settings, Accounts is one page for both: a Providers card turns Claude Code and Codex on or off, says whether each is installed and which version, and lists the install or update commands, each with Run it for me and Copy. The app can run with Codex alone. Codex sessions and reviews run on this computer only, not over SSH.
• Codex accounts of their own. Add more than one, with ChatGPT or an API key (the key goes to Codex, and this app never stores it). Each signs in inside its own folder in the app's resources folder, so sessions never mix identities, and a saved config picks the Codex account it runs under. A Codex account is always added in the app, as a Claude account is; if you used Codex in 2.1.0, see the known issue below.
• A Codex session works as a Claude session does. It reopens in the same conversation after the app restarts, can restart into an earlier conversation, moves to another Codex account with Switch Account and keeps the conversation, shows in Logs, and has the attention dot, the Session Watchdog, Plan mode and Compact.
• Code review now works both ways: a Codex session can ask for a Claude review, as a Claude session already could for a Codex review, each on your reviewer account. A Codex review receives the change itself and runs no commands, so it works on Windows too.
• Usage and spend for Codex. The Usage page, the usage strip at the foot of the window and Tokenomics cover both assistants, each account shown on its own and never merged. A Codex account signed in with ChatGPT shows its 5-hour and weekly allowance, or its credits; an API-key account is billed per token.
• The Agent Canvas, Vision and the in-app browser, Ask Conductor, Sentinel, Insights, Cloud Agents and the Memory page work for Codex too, and a Codex account can sign in to chatgpt.com inside the app, kept apart from every other account.
• Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 are in the model picker, and the opus, sonnet and haiku choices now mean them.
• Setup can install Claude Code and Codex for you. Every install or update command has Run it for me, which runs it in a visible terminal once you confirm it, and Copy. Setup checks again when the command ends and finds the new install without restarting the app.

Changed

• GPT-6-Sol and GPT-6-Luna are in the Codex model picker; Codex lists them from version 0.156.1, the newest Codex this release is tested with. The Codex models and efforts match the ones Codex offers in its own picker.
• After an update, What's New shows what is new since the version you came from, under a heading that names it, with a short summary and pages on the main changes; the full notes stay in the Feature Guide, What's New. A fresh install asks which assistants you use: Claude Code, Codex, or both.
• Windows: the installer and uninstaller look like the app, with a new Welcome page and the splash's dark theme and artwork.
• The app runs on Electron 44. On a Mac it needs macOS 13 (Ventura) or later; on macOS 12, stay on 2.1.0. Its Mac download now has a name ending in macos13.dmg, which the updater in 2.1.0 does not look for (see the known issue below), and from 2.1.1 on the updater never offers a Mac a version that needs a newer macOS than it runs.
• The project now lives at github.com/nubbymong/ai-code-conductor, and old links to it redirect there. The app finds updates there on its own; if Check for Updates says Up to date, see the known issue below.
• Known issue: if Check for Updates says Up to date when a newer release is out, restart the app once while online, then check again. The app looks for the project's new home when it starts, so after a start while offline, or in an app already running when the project moved on October 12, 2026, its checks read the old name until it restarts.
• Known issue: Check for Updates also says Up to date when it cannot reach GitHub (no network, a proxy or firewall in the way, or GitHub's hourly limit on checks reached). Restart the app and check again later, or see the newest release on the Releases page on GitHub.
• Known issue: a Mac on 2.1.0 is not offered 2.1.1 in the app, and Check for Updates there says Up to date. 2.1.1 needs macOS 13 (Ventura) or later. On macOS 12, stay on 2.1.0. On macOS 13 or later, download 2.1.1's .dmg from the Releases page on GitHub once by hand; from then on the updater offers new versions as before.
• Known issue: on an SSH host whose shell is zsh, such as a Mac (zsh is its default), and possibly on another host whose shell is zsh, at least when the app runs on Windows, the SSH session's partner terminal opens in your home folder instead of the config's remote folder. Type cd and the folder in the partner (see Known issues in the Feature Guide).
• Known issue: if you used Codex in 2.1.0, add your Codex account once. Nothing carries over from the earlier Codex setting: choose Yes, set up Codex on the one-time Do you use Codex? page and sign in, or later turn Codex on and choose Add Codex account in Settings, Accounts. The Codex CLI's own sign-in is left as it is, and earlier Codex conversations stay with the Codex CLI: codex resume in a terminal in the project folder picks them up.
• Known issue: GPT-5.5 retires from Codex with ChatGPT sign-in on October 14, 2026, OpenAI says. From then on, a Codex config set to GPT-5.5 on an account signed in with ChatGPT needs another model: edit the config and pick one your account offers, such as GPT-6-Sol where OpenAI makes it available, or Default: follows Codex.
• Known issue: with a Codex newer than 0.156.1, the newest this release is tested with, the Providers card in Settings, Accounts, the Set up Codex page and Sentinel (when it is on) show a notice. Codex still runs. A newer Codex can also offer a model the picker does not list yet, such as GPT-6.1-Sol, and Sentinel then says it is not in the model picker: choose Default: follows Codex in the config to use the model Codex picks. If anything misbehaves, install Codex 0.156.1 (with npm, npm install -g @openai/codex@0.156.1; in PowerShell on Windows, npm.cmd install -g @openai/codex@0.156.1) and restart the app.
• Known issue: after an in-app update on Windows (from 2.1.0, or from 2.1.1-beta.1 to beta.3), if Vision on the Conductor MCP page shows Stopped and Start browser says its debug port 9222 is in use, the browser the earlier version started for Vision is still running, and the app could not check which programs are running, so it did not end it. Sign out of Windows and back in, or restart the computer; Vision then starts as usual.
• Known issue: on Windows, the app does not find its data or resources folder when the full path of that folder holds a character other than a plain English letter, a digit or ordinary punctuation, such as an accented letter, and your saved configs, accounts and sessions can then look missing. Everything in the folder you chose is still there. The Feature Guide, Known issues has the workaround: move what is in the folder to a new folder on this computer's own disk (not a USB drive, a FAT or exFAT drive or a network share) whose full path holds only plain English letters, digits, spaces, dots, hyphens and underscores, then run the installer again and choose the folders in their new place.
• Notice: a Claude Code version that a saved config was set to use in an older release is no longer used. Its sessions and cloud agents start the Claude Code installed on this computer, and the setting stays in the config. To run a particular version, install it yourself, make it the only Claude Code on your PATH, and turn off Claude Code's automatic updates (see Known issues in the Feature Guide).

Fixed

• First-run setup is sturdier: a folder that was not saved keeps the folder step open and says which one, and closing the app during Claude Code setup stops the setup terminal. On Windows and Linux, the setup screens, the onboarding pages and the one-time pages after an upgrade have the window's own minimize, maximize and close buttons.
• SSH: a session's partner terminal now opens on the host, and its command buttons run there; remote SSH sessions raise the attention indicator when Claude waits on you (Linux and macOS hosts; an SSH Persistent session started before the update gets it once you End it and start it again); SSH sessions with tmux start on a host whose login shell is zsh; and the mouse wheel scrolls an SSH Persistent session's history.
• New sessions start on a computer without Node.js: Claude Code or Codex starts directly, without the resume list a new session otherwise opens. Install Node.js 22 or later to get the list back.
• Many smaller fixes across accounts, sessions and Windows. Among them: a Claude Code session now starts when Claude Code was installed with npm into a folder whose path has a space or brackets in it; failure text and the app's log hide credentials written inside URLs; and at each start the app makes the folder its session scripts run from private to your account.

Full changelog: https://github.com/nubbymong/ai-...

v2.1.1-beta.2

AI Code Conductor Beta v2.1.1-beta.2

Beta Release (pre-release)

Codex becomes a full second assistant beside Claude Code. One Accounts page covers both: turn each on or off, see whether it is installed and up to date, and keep more than one Codex account, each with its own sign-in. A saved config picks the Codex account it runs under, a Codex session can restart into an earlier conversation, and each assistant can ask the other to review its work. A new install asks which assistants you use, an update asks once whether you use Codex, and the app can run with Codex alone. The Usage page, the usage strip and Tokenomics cover both assistants, with each account's allowance and spend shown apart. A Codex session now works as a Claude session does: it keeps its conversation through a restart, moves to another Codex account with Switch Account, shows in Logs, and has the attention dot, the Session Watchdog, Plan mode and Compact. Sentinel watches Codex too. The Agent Canvas, Vision and the in-app browser, Ask Conductor, Cloud Agents and the Memory page work for Codex too, and Codex is no longer marked Beta.

Added

• Settings, Accounts is now one page for both assistants. A Providers card at the top turns Claude Code and Codex on or off, says whether each one is installed and which version, and, when Codex is missing or too old, lists the install or update commands from OpenAI's own instructions, each with Run it for me and Copy, with Check again. When the app can tell how the Codex it found was installed, it lists the update for that way (one that OpenAI's own installer put there is updated by running that installer again); when it cannot, it lists every update. Your Claude accounts and your Codex accounts are listed below it. At least one provider always stays on, and a provider cannot be switched off while anything of it is still running.
• Claude Code and Codex can both be installed from setup. Every install or update command has Run it for me, which runs it in a visible terminal once you confirm it, and Copy. The tool's own installer comes first (Anthropic's from claude.ai, OpenAI's from chatgpt.com), and its confirmation says it downloads a script from there and runs it; npm comes second, and says so when Node.js is not found. Setup checks again when the command ends, also when it fails or you press Ctrl+C on Windows, and finds the new install without restarting the app. Anthropic's installer does not put its folder on your PATH: setup then says Claude Code is installed there, and on Windows Add it to PATH for me adds that one folder to your PATH; on a Mac or Linux it shows the line to add to your shell file. Settings, Accounts and the CLI help at the foot of the window offer the same. The setup screens where you choose the app's folders and install and set up Claude Code have Exit; the pages after them do not.
• More than one Codex account. Add one with ChatGPT or an API key; the key goes to Codex, and this app never stores it. Each Codex account signs in inside its own folder, which the app creates in its resources folder, so sessions never mix identities, and the app never reads the sign-in Codex keeps there. Each account row has a menu: Make default, Make reviewer, Sign in again, Check sign-in, Sign out, Make inactive and Archive.
• The Codex sign-in already on your computer can be used too, but only when you choose it: Use this sign-in on the Set up Codex page, or Use this computer's Codex sign-in in Settings, Accounts. The app then asks Codex whether it is signed in, and lists it only if it is. Because the app did not create it, you confirm it at each launch, and it never runs code reviews; if it is signed out, run codex login in a terminal, then Check sign-in.
• New saved config picks the Codex account a config runs under, with the default listed first; an account signed in a different way than before cannot be picked until you confirm it. A Codex session header has a Restart menu: Restart carries on with the same conversation (a new one if the session had none yet), and Restart and pick a conversation lists recent ones in the terminal to pick from, the conversations of the project's other git worktrees included, each tagged with its worktree and started in it, and led by the name you gave a session. A Codex session that was open when the app closed reopens in the same conversation, under the same account.
• Code review works both ways. A Claude session can ask for a Codex review, and a Codex session can ask for a Claude review. Each review is a separate, one-off, read-only reviewer in the project, on your reviewer account: the reviewer default, or the default account when none is set. Settings, General, Built-in Tools has a switch for each direction that names the account reviews will use and, when a review cannot run, says why. A review tool is now offered only while a review could actually run. On macOS, Claude reviews use your normal Claude sign-in.
• A fresh install asks which assistants you will use: Claude Code, Codex, or both. A computer without Claude Code can choose Use Codex only on the setup screen instead of stopping. Choosing Codex leads to a Set up Codex page that checks for the Codex CLI, lists OpenAI's installer first and then npm (and Homebrew on a Mac), each with Run it for me, which runs it in a visible terminal tab once you confirm it, and Copy, and signs you in.
• After updating, the app asks once whether you use Codex: Yes, set up Codex, or No, I don't use Codex. Codex now has accounts of its own in this app, so nothing carries over from your earlier Codex setting, and until you answer, Codex is not set up and nothing of it starts. Yes opens the Set up Codex page, where you add a Codex account or choose Use this sign-in for the Codex sign-in already on this computer; that sign-in is used only after you choose it. No keeps Codex off, and you can set it up later in Settings, Accounts.
• The Usage page covers both assistants. With Codex on, it has a Claude Code section and a Codex section. Each Codex account shows its 5-hour and weekly allowance and its plan, marked Updated when the figure comes from an open session or a fresh check, or As of, with its age, when it comes from that account's latest session. An account with no session yet says its allowance shows after the first one, and an API-key account says it is billed per token, with no plan allowance. The rail shows the page once you have two or more accounts across both assistants, and its foot links to Tokenomics for token use and cost.
• A closed Codex account signed in with ChatGPT is checked afresh when you open the Usage page, press Refresh or use a card's Retry: the app asks Codex to run its own usage check once, in that account's folder, which takes about a second. It never checks an API-key account, the sign-in Codex keeps for this computer, or an account a session or review is using, and it checks one account at a time. Closing the page stops it, and if the check fails the card shows the last reading from that account's latest session, with its time.
• The usage strip at the foot of the window shows one pill per person, grouped by provider: someone with a Claude Code account and a Codex account sees both, and percentages are never merged across providers. Settings, Status Line lists the bars to show for each provider. A limit window that has reset since the last reading says no reading since instead of showing an old figure, on the Usage page, in the strip and in a session's status line, and a Codex session's cost keeps its API eq label, with a tooltip that says API-equivalent estimate, or Estimate at API list prices for an API-key account.
• Tokenomics splits by provider and account. A Provider filter appears once you have both Claude Code and Codex usage, and an Account filter lists each account under its provider, with This computer's sign-in for the Codex folder on this computer and Not recorded for usage with no account. Each KPI shows Claude Code and Codex apart when both have usage and no filter narrows the view, and the cost chart draws a line for each when both have spend in the range; the sessions table has an Account column and each provider's mark. Codex usage is sorted by the folder it was written in; Claude usage by the account profile a local session runs under, from this version on, while SSH sessions and sessions with no account profile read Not recorded, except that a local session later resumed in the app under an account profile moves its earlier usage to that profile's account. The first start after updating sorts your earlier Codex history once, with a notice while it runs; the totals are complete throughout.
• A Codex session's card and tab now behave like a Claude session's. The attention dot lights up when Codex asks you to approve something, and a minute after a turn ends if you have not replied; the busy sweep shows while Codex works, the working pill names Codex, and the sleeping moon shows as it does for Claude. The Session Watchdog now covers Codex sessions too, under the same one switch in Settings, which stays off by default: it reads Codex's own usage-limit and server-error messages, waits out the reset, and types the retry only into Codex's empty input box. The safeguard check is Claude's only, so a Codex session's menu shows it as not available. For the attention dot, Codex asks once per Codex account to review the hooks this app gives it (Hooks need review): choose Trust all and continue. Declining leaves the session working as normal, without the dot on that account.
• The Logs page covers Codex. A local Codex session's conversation is indexed as a Claude session's is, from that session's own Codex account folder: it shows in the Logs page with its history and search, and the Logs button on a Codex tab opens it, live, with Codex's turns marked as Codex's. A Codex config has the same Index conversation logs switch. A name you give a Codex session is kept next to its co...

v2.1.1-beta.1

AI Code Conductor Beta v2.1.1-beta.1

Beta Release (pre-release)

The mouse wheel scrolls a remote session the way it scrolls a local one, instead of typing arrow keys at Claude. Otherwise a maintenance release: the runtime and its dependencies refreshed, and the largest source files split into focused modules with no change in behaviour.

Changed

• Electron moves to 43.7.1, with the terminal backend (node-pty), zod, marked and the transcript database typings refreshed. Two development-only dependencies with published advisories (vitest, js-yaml) and the MCP server library (hono) are on their patched versions.
• Internal: the four largest source files were split into focused modules as pure moves, each verified against its origin and pinned by new regression tests. No change in behaviour.

Fixed

• The mouse wheel scrolls the history of an SSH Persistent session, the same as a local one. It used to send arrow-key presses to Claude instead -- so a remote session could not be scrolled back at all, and every notch typed an arrow key at whatever was running, which in a full-screen program could move the cursor or change what was selected. Claude runs inside tmux on an SSH Persistent session and the terminal has no scrollback of its own there, so the wheel now drives tmux’s own scrollback instead. Selecting text with the mouse is unchanged. Scrolling up opens the scrollback view and the first key you type closes it and returns you to the live bottom, so you can scroll back to read something and just start replying.
• When no login shell is set in the environment, the check for the Claude CLI and the session launch now agree on the shell instead of assuming zsh for one and bash for the other: zsh on macOS, the platform default, and elsewhere the first of bash, zsh and /bin/sh that is installed. A Linux machine without zsh is no longer told the CLI is missing when it is installed, and one without bash can still launch a session.

Full changelog: https://github.com/nubbymong/claude-command-center/blob/v2.1.1-beta.1/CHANGELOG.md

---

Upgrading from 2.1.0-beta.5 or earlier? Download the installer from this page — that build's updater looked for the old file name.

Claude and Claude Code are Anthropic trademarks.

SHA-256 checksums are in CHECKSUMS.txt attached to this release.

v2.1.0

AI Code Conductor v2.1.0

Stable Release

The first stable release under the app’s new name, AI Code Conductor -- and if you are coming from 2.0.0, everything below arrives at once. Remote work stops being second-class: an SSH session survives a dropped connection, reports its account and full usage exactly like a local one, and can run inside a Docker or Podman container. The Agent Canvas arrives as a full visual review surface, with plans, mockups and a Testing mode that saves evidence. claude.ai runs inside the app, signed in per account. The terminal’s glyph corruption is root-caused and fixed, with GPU rendering on by default. A Session Watchdog can wait out a rate limit and type the retry itself. And there is a Linux build.

Added

• The app is now AI Code Conductor, with a new name, icon, start-up animation and session setup dialog. Your existing installation upgrades in place and keeps every config, account and saved session -- nothing needs re-creating. The first launch after upgrading explains what changed and previews what 2.2 is aiming at: support for more coding agents beside Claude Code and Codex.
• A remote session is now a full session. SSH sessions run inside tmux on the remote, so closing the lid, losing wifi or switching networks no longer kills the work -- reconnecting picks it back up where it was. Local and remote sessions report their status over the same channel, so an SSH session shows its account and its full usage, per-model bars included, exactly like one on your own machine. That includes Windows remotes, which never reported at all before. A remote you left running waits in a Remote Resumable section at the foot of the Running tab, one click from the same Claude and the same conversation.
• The session config dialog answers "where does this run?" properly. Connection is a choice of three -- Local, SSH, SSH Persistent -- and a new Runtime section runs the session inside a Docker or Podman container without you maintaining a shell one-liner. Entering a container is confirmed from inside the container before anything is launched there, and checked again before Claude itself starts, so a container that stopped in between gets nothing typed onto the host.
• The Agent Canvas is a place your agent shows you work and you mark it up. It handles plans, mockups and the built site, anchored notes go on the thing you are pointing at, and a review is a decision that settles: approving a version owes the agent nothing, and an approved artefact signs itself off. A Testing mode lets you click through the build your agent is serving and save evidence notes that lock the screen, the page state and how you got there together into a pass-or-fail pack. Unfinished canvas work outlives the session that made it -- any session on the project can pick it up -- and a Canvas Explained page draws the whole model.
• claude.ai comes into the app. Each account holds its own web session in a dedicated per-account browser, kept fully separate, and the whole sign-in can be routed into the in-app pane instead of a separate window -- which is also what gets past Cloudflare’s "verify you are human" loop. The browser pane has a pinned "claude.ai -- your artifacts" entry that opens signed in as the session’s own account.
• A Session Watchdog waits out a rate limit and types the retry itself. Its three auto-retry checks -- rate-limit resume, API overload and safeguard -- are separate switches in Settings, and each session can turn them on or off for itself from its right-click menu, taking effect immediately. A Watchdog pill in the session header says what is live for that session. It is careful never to type over you, and the sleeping-session indicator only ever reports and never types.
• Session cards show life at a glance. A session actively producing output gets a green "working" pill and a sweep across its context bar; one that has gone quiet gets a moon with how long it has been asleep. Clicking a sleeping session does not wake its moon -- only real output does.
• The left panel has two modes, Saved configs and Running sessions, with a Quick Start. The command bar under each session is one row: the fixed tools, then your Global buttons, then this config’s Session buttons, and the bar knows what kind of session it is in -- a Codex session says Codex, an SSH session says which computer each button runs on. A config can be marked Allow Multi Spawn, so the ones you genuinely want several copies of get a copy-count control and a Select mode that launches a batch, and everything else is protected from a stray second launch.
• Every page -- Tokenomics, Usage, Settings and the new Feature Guide -- opens as a tab beside your sessions instead of taking over the window. Ask Conductor is a slim help session that answers questions about the app itself, and it can install a helper skill so every Claude session on this machine can answer them too, without opening the Ask tab.
• Experimental Linux support: AI Code Conductor runs on Linux as an AppImage, alongside Windows and macOS. It needs glibc 2.39 or newer (Ubuntu 24.04+, Rocky 10+), and the vision browser needs a deb or rpm Chrome rather than a snap.

Changed

• Resuming your work is more reliable. A conversation that was running in a git worktree comes back correctly even after that worktree is cleaned up, instead of opening a fresh empty session, and the name you gave a session travels with the conversation so the resume list uses your own name for it. The Usage page loads account by account in place, so a slow account no longer holds the whole page blank, and Codex spend is counted correctly.
• The runtime moved to Electron 43, with the terminal backend and the local database updated to match. Ctrl+V pastes into terminals, which also makes voice dictation and text expanders work, and Check for Updates can install the update it finds. Your own Claude hooks run in the app’s sessions, and each config can set its own permission mode and CLI arguments.

Fixed

• The terminal corruption is understood and gone. GPU rendering shared one cache of character images across every open session, so one session rebuilding that cache wiped the text out of all the others -- which is why it looked random, why it got worse the more sessions you had, and why resizing the window brought it back. That cache is a limit of the drawing library and is still shared, so the repair is elsewhere: only the terminal you are looking at draws on the GPU at all, and a terminal about to lose its characters now drops its own drawing model first and then repaints -- which is exactly what resizing the window did by hand. GPU rendering is on by default as a result, and stale glyphs left after scrolling are cleared.
• A security line through the whole release: several local-attacker vulnerabilities were found and fixed across the local tools server, the session-launch path, the updater and the bundled dependencies, each with an advisory published alongside the release that carried it, and every open dependency alert on the project was cleared. Security-sensitive changes now go through an independent adversarial review before they can merge.
• A long tail of paper cuts from the release-candidate pass: startup pages take strict turns so a tour and a restore prompt no longer wait for each other and show neither; a right-click menu near the bottom of the window opens upward instead of running off the edge; on a Mac, Close sessions closes the sessions and quitting asks first; hovering a link no longer flickers the terminal; and a setup page taller than your screen scrolls instead of hiding its buttons.

Full changelog: https://github.com/nubbymong/claude-command-center/blob/v2.1.0/CHANGELOG.md

---

Upgrading from 2.1.0-beta.5 or earlier? Download the installer from this page — that build's updater looked for the old file name.

Claude and Claude Code are Anthropic trademarks.

SHA-256 checksums are in CHECKSUMS.txt attached to this release.

v2.1.0-rc.17

AI Code Conductor Beta v2.1.0-rc.17

Beta Release (pre-release)

The Session Watchdog is no longer all-or-nothing. Its three auto-retry checks -- rate-limit resume, API overload and safeguard -- are now separate switches in Settings, and each session can turn them on or off for itself from its right-click menu, taking effect immediately. A new Watchdog pill in the session header says what is live for that session.

Added

• Each session can switch the watchdog’s auto-retry off for itself, without touching any other session. Right-click the session and use the Watchdog auto-retry block to turn the rate-limit resume, API overload and safeguard checks on or off. It applies straight away to the running session -- a check you switch off stops watching and never types, and if it was already waiting to retry, that retry is dropped. This is for the current run only: relaunching the session brings back whatever Settings says. Turning a check off never affects the sleeping-session indicator, which only ever reports and never types.
• The session header shows a Watchdog pill, immediately right of the account, so you can see at a glance what the watchdog will do in that session: green when all three checks are on, "partial" when some are off, and "off" when none will type. There is no pill when no watchdog is running for that session.
• Settings now lists the watchdog’s three checks separately instead of one switch. Rate-limit resume waits out a usage-limit reset and then continues; API overload backs off and retries on server errors; Safeguard retries once a flagged message clears. Two of these already existed but could not be reached from the app. These decide what a newly launched session starts with.

Fixed

• A startup page is no longer interrupted by the sessions it is describing. Restoring your sessions starts them straight away, and each one that asks which account to use put its own prompt on top of whatever was still on screen -- so the Enable Multi Spawn page could be covered before you had finished with it. Those prompts now wait until the last startup page is dismissed, and none of them is answered for you in the meantime.
• A first launch that has both saved sessions and a guided tour to show no longer gets stuck showing neither. The tour and the restore prompt each waited for the other, so nothing appeared and the saved sessions could not be restored at all. Startup pages now take strict turns, and the one being shown is always the one that was chosen.
• A session right-click menu opened near the bottom of the window stays on screen. It used to open at the pointer and run off the edge, which hid the lower entries with no way to scroll to them -- most visibly the account list under Switch Account, where the last few accounts simply could not be reached. The menu now opens upward when there is more room above, and scrolls when it is taller than the window either way.

Full changelog: https://github.com/nubbymong/claude-command-center/blob/v2.1.0-rc.17/CHANGELOG.md

---

Upgrading from 2.1.0-beta.5 or earlier? Download the installer from this page — that build's updater looked for the old file name.

Claude and Claude Code are Anthropic trademarks.

SHA-256 checksums are in CHECKSUMS.txt attached to this release.

v2.1.0-rc.16

AI Code Conductor Beta v2.1.0-rc.16

Beta Release (pre-release)

Fixes from an independent stability review of rc.15. Entering a container is confirmed from inside it before anything is launched there, and checked again before Claude itself starts; two more account-refresh races are closed and capturing a second account can no longer put a spent token back on the first; on a Mac, Close sessions closes the sessions; a remote left running is never pruned because a tmux upgrade broke the client; and a rejected plan no longer hides from the canvas resume list behind an approved design.

Fixed

• A container entry is confirmed from inside the container before anything is launched there. The app enters a container with a command that prints a one-time confirmation from inside it before the shell starts, and only that confirmation makes the session count as inside: a prompt that merely looks different, a quiet pane, a cancelled sudo prompt or a host prompt the app cannot recognise (a zsh prompt ending in a percent sign) never does. Launching re-checks the attached shell first, and again before Claude itself is started, so a container you detached from or that stopped in between gets nothing typed onto the host. A container configured to Start, or an after-connecting command that is not a plain docker or podman exec, cannot be confirmed this way: the app says so and asks before launching. Configs written as a free-text exec line get the same confirmation and keep the shell they name, and the Runtime section has a "Shell inside the container" choice for images without bash.
• Two sessions on one account can no longer capture a half-written sign-in. The backup that follows a token rotation was checked once per session instead of once per account, so with two sessions open a sign-in could be backed up between its two writes and restored later as a mixed account.
• A session started while its account's token is being refreshed now waits for the refresh to finish before it starts, as background runs already did, so it never reads a token that is about to be replaced. Closing the tab during the wait cancels the start; anything typed during the wait is dropped rather than queued into the shell.
• Capturing a second account that was signed into mid-session no longer reinstalls a token the first account may already have spent. The original account gets its identity back and shows Sign in; signing in once brings it back, instead of it failing silently later on a dead token.
• Cancelling an agent before it has started really cancels it. An agent cancelled while it was still waiting for an account refresh or a Claude install used to start anyway, labelled cancelled, with no control left to stop it.
• On a Mac, Close sessions closes the sessions. Choosing Close sessions on the close dialog discarded the cards but left the sessions running invisibly behind the Dock icon; they are ended first now, and anything left over is swept when the last window closes.
• Closing the window before answering the resume prompt keeps everything. With both saved sessions and left-running remotes, closing before choosing Resume or Don't open cleared the saved file, remotes included; the file is left as it is now, and the next launch asks again.
• A remote left running is never pruned because its tmux cannot answer. A tmux client that exists but cannot talk to its server (after a tmux upgrade, or a permission problem on the socket) used to read as "no sessions" and the entry was removed as dead; the check now tells that apart from a genuine empty answer and keeps the entry, marked as not verified.
• Output from a restarted session's old process no longer reaches the new one. Late output from the replaced process was still shown in the new terminal and fed to its watchdog; it is dropped now, as its exit already was.
• A rejected plan no longer hides from the canvas resume list behind an approved design. Resume judged a canvas by its newest artefact only; it now looks at every artefact and at every note still owed, so a canvas with a rejected plan that was never reworked, an unsubmitted note, or an answered note you have not looked at stays resumable, and Mark complete says which artefact was rejected and not yet reworked.

Full changelog: https://github.com/nubbymong/claude-command-center/blob/v2.1.0-rc.16/CHANGELOG.md

---

Upgrading from 2.1.0-beta.5 or earlier? Download the installer from this page — that build's updater looked for the old file name.

Claude and Claude Code are Anthropic trademarks.

SHA-256 checksums are in CHECKSUMS.txt attached to this release.

v2.1.0-rc.15

AI Code Conductor Beta v2.1.0-rc.15

Beta Release (pre-release)

Fixes from the rc.14 feedback and from an independent stability review of rc.14. The in-app browser pane no longer sits in front of a page tab; the SSO Sign-in browser choice is back on managed machines with a per-user Chrome; on a Mac, quitting asks before it stops anything and reopening from the Dock works; a restarted session is no longer shown as exited; and the remote sessions you leave running are remembered even when they were the only thing open. The Usage page loads account by account and reuses the figures your open sessions already report (the primary account aside), a background run can no longer be signed out by a token refresh, and editing a config that has a session running warns you first.

Added

• The Usage page loads one account at a time, in place. Opening it shows a row per account straight away and fills each one in as its figure arrives, so a slow account no longer holds the whole page blank, and an account whose fetch fails says so on its own row instead of stalling the rest.

Changed

• An account with an open session shows the usage its session already reported, instead of asking Anthropic again; the primary account is the exception and is always fetched. Accounts with no session open are fetched one at a time, so opening the page no longer fires a request per account (which was drawing rate-limit errors), and it never refreshes a signed-in account's token behind a running session. The one exception: an account on extra usage (paid credits) still makes a single request for the credits figure the status line cannot carry, with the token it already has.
• Editing a config that has a session running warns you first. Opening the editor for an SSH config with a live or left-running session shows what is affected and reminds you that changes apply on the next launch, not to the running session, and that changing where it connects can break resuming the one you left running. Cancel, or Edit anyway; it never blocks.

Fixed

• A signed-off canvas is done, not resumable. Once you approve (or dismiss) a canvas, it drops out of the resume list and the purple "resumable" dot on the Canvas button, instead of lingering there until you formally mark it complete. A canvas you sent back (rejected) stays resumable, because a rejection asks for another round, and one still awaiting your decision stays too.
• The terminal, browser, Agent Canvas and logs pane are now one at a time. Opening the canvas while the browser was open did nothing — the browser held the foreground and the canvas never appeared. Opening any of the browser, the canvas or the logs pane now switches to it and closes the others; clicking the pane you are already in returns you to the terminal.
• A token refresh can no longer sign out an account that a background run was using. The Usage page refreshes an account's token only while nothing is using it, but headless runs (Insights, cross-account synthesis, Sentinel), cloud agents and plain shells under an account were invisible to that check, so a refresh could rotate the token under them and the next start of that run redeemed an already-spent token and logged the account out. Every such run now counts as using the account for as long as it runs, and one that starts while a refresh is in flight waits for the new token before it reads anything.
• Editing a saved SSH config can no longer point a left-running session at the wrong host. A session you leave running is remembered with the host, user, port, path and container it was left on; if you then change where that config connects, the app no longer probes or ends a session of that name on the new host. The old session shows as retargeted in Remote Resumable, with Remove as its only action, and tells you how to end it on the host it is really on.
• A saved sudo password is never typed into a container. The password kept for a container entry that needs sudo answers the host's own sudo prompt, before the container shell appears; a sudo-shaped prompt printed inside the container afterwards, by a start-up script or by something you ran, gets nothing.
• Quitting no longer hangs when the app's window has crashed. If the window's renderer died while the save-or-cancel question was up, the quit waited forever for an answer that could not come; it proceeds now. Pressing Quit a second time while the question is showing also goes through, as a second close already did.
• A left-running remote cannot be revived by text a host prints on connect. The check that asks a host which sessions still exist reads only what the host printed between the check's own markers, so a login banner or a message of the day can no longer make a finished session look alive in Remote Resumable.
• Quitting on a Mac asks before it stops anything. Cmd+Q or Quit used to shut down every session and background service first and show the save-or-cancel dialog afterwards, so Cancel handed you back an app that had already stopped your work. The question now comes first: Cancel keeps everything running, and Save shuts down only after your sessions are written.
• Reopening from the Dock works on a Mac. Closing the last window and then clicking the app's Dock icon could bring back a blank, hidden window, or take the app down, because the window's controls were being registered a second time. They are registered once now.
• Restarting a session no longer shows it as exited. The old process's exit could arrive after the new one had started and mark the healthy replacement as ended, so Ask Conductor treated the session as dead and started yet another one. An exit from a process that has already been replaced is ignored.
• Remote sessions you left running are remembered even when they were the only thing open. Leaving the last SSH session running and then closing the app forgot its Remote Resumable entry, so the next launch of that config started new work instead of reattaching. The list of left-running remotes is now kept on its own, including when you decline the resume prompt or close without saving.
• Left-running remotes on a Mac with a Homebrew tmux are no longer pruned as dead. The liveness check could not see a tmux installed outside the login PATH, took its empty answer as proof, and removed live entries; it now looks in the same places the End command does.
• Signing in again for an expired account waits for the actual sign-in. The prompt used to report success after four seconds whether or not you had signed in, because the account's email was still on disk from before. It completes only once the credentials have really changed.
• Capturing a second account signed into mid-session no longer risks the first one. The app restores the original account from a backup that was refreshed only when the session ended, so a token that had rotated during the session could be replaced by a stale one that no longer worked. The backup now follows each rotation that happens while a session on that account is open.
• A session's transcript-indexing opt-out survives save and restore. A restored session dropped the setting and was indexed regardless.
• A container that cannot be entered no longer puts Claude on the host. When the container was stopped or missing, or the engine was down, the failed entry dropped back to the host shell and the app took that for the inside of the container, so Launch Claude set up and started Claude on the host instead. The failure is recognised now, from the engine's or sudo's message or from the host's own prompt coming back: the session says so, Launch Claude refuses the host, Run again re-enters the container once it is running, and Skip remains the explicit way onto the host shell.
• The saved SSH password is no longer typed into a sudo prompt. After a key-based login the saved password was still waiting for a prompt, and a sudo prompt shaped as a bare "Password:" (the macOS form) matched it first, so the wrong secret went to sudo and the first attempt failed even with a sudo password saved. Once the login has completed, such a prompt is treated as sudo's, whether or not the session runs a post-connect command.
• Closing a container session tab ends the Claude inside the container. Container sessions never showed the End-or-Leave dialog, and the plain close only killed the local connection, which does not reliably stop the process inside the container; abandoned Claudes accumulated there. Close now ends that session's process in the container first, without touching another session sharing it.
• The browser pane stays behind whatever you open on top of it. The pane, and the claude.ai artifacts view that can live in it, are drawn by the app itself rather than by the page, so nothing on the page can sit over them: they have to be told to step aside. They stepped aside when you switched sessions, but not when you opened a page tab, so opening Settings (or Tokenomics, Logs, the Feature Guide) with artifacts open left the artifacts view in front of the page you asked for. Every page tab now hides the pane while it is in front, and so does every window-covering dialog, the guided tour and the welcome walkthrough; the pane comes straight back when you return to the session, and a page that arrives while something is on top waits until the pane may show rather than flashing over it.
• The SSO Sign-in browser picker is back on machines with a per-user Chrome. On a managed workstation where you cannot install software for everyone, Chrome (and sometimes Edge) lives in your own AppData folder rather than under Program Files, and the app only ever looked under Program Files. A machine with both browsers therefore looked like it had one, and the Sign-in browser picker on an SSO account, which only appears when there is a genuine choice to make, stayed hidden: Edge could not be chosen. Both per-user locations are now checked as well, after the system-wide ones. The same lookup drives the v...