AI CLI Version

/ai-cli-version — CLI version and update check

Run the platform launcher from this skill's directory: ./run.sh on macOS or Linux (sh run.sh if it is not executable); py -3 run.py on Windows, falling back to python run.py if py is unavailable. Reformat its output into the report below, then follow the optional update flow.

The launcher is check-only and must never install anything. During the check phase, do not run claude update, codex update, or grok update without --check.

The first output line is ### ai-cli-version <version>. Leave it out of the report; quote it only when the user asks which version they are on or reports a bug. Adding --version to the platform command prints the version alone.

What is checked

Only the latest version and the currently installed version — no history.

• claude update and codex update install immediately and have no check-only flag, so their latest version comes from the npm registry (queried over HTTP; npm is not needed). Grok Build has --check.
• Installed at is the folder's creation time (Windows, macOS) or modified time (Linux). If that path is missing — an npm or Homebrew install, or the macOS/Linux Grok installer, which stores builds without a version in the name — the real executable behind the command is used instead, then its folder.
• Tested on Windows and macOS. On Linux, if the script fails or a value looks wrong, report it as-is and mention that this platform has not been tested yet.
• A check failed (... CERTIFICATE_VERIFY_FAILED ...) on macOS means a python.org Python whose Install Certificates.command was never run. Say so; do not work around it. Timestamps before 2015 are discarded (npm extracts files with a fixed 1985 date).
• Grok release time is approximate. The official changelog (x.ai/build/changelog) blocks scripted requests, so the time comes from the GitHub release of the third-party NixOS package timoteuszelle/x.ai-grok, which auto-detects new versions. The official release is no later than that time.

Each tool line is tab-separated: name installed=… installed_at=… latest=… released=… released_note=… channel=… status=… update_cmd=…, or name<TAB>not installed, or name<TAB>check failed (<reason>). Times are local MM/DD HH:MM; ? means unknown.

Report

CLI VERSIONS — 09-17

  Tool          Installed               Latest                             Status
  Claude Code   2.1.274 (09/17 11:20)   2.1.274 (released 09/17 06:36)     ✓ up to date      channel: latest
  Codex         0.153.4 (09/07 11:03)   0.154.0 (released 09/10 06:40)     ↑ update → codex update
  Grok Build    1.0.34  (09/17 09:28)   1.0.34  (released ≤ 09/17 03:32*)  ✓ up to date      channel: stable

  * Grok release time is approximate (third-party mirror); official: x.ai/build/changelog
  → 1 update available: Codex.

  Update the 1 outdated CLI now?
  [y] Update all    [Esc] Skip

Rules:

• Show every tool. not installed and check failed (...) get their own row with the reason; never drop a row, and never fill in a version or time from memory.
• Show ? for an unknown time; do not guess.
• Grok's release time always gets ≤ and the footnote.
• ↑ update rows end with the update command.
• ahead of channel means the installed build is newer than the channel tag — show it as-is, not as an error.
• Last report line: how many updates are available and for which tools, or All up to date.
• Match the user's language for labels.

Optional update

Only when one or more rows have status=update available, append a two-choice prompt in the user's language and stop. Keep the keys exactly y and Esc:

Update the N outdated CLIs now?
[y] Update all    [Esc] Skip

• /ai-cli-version authorizes the check, not an update. Do not run any updater in the same turn as the initial report.
• A subsequent exact y or Y authorizes every update_cmd from the immediately preceding report whose status was update available. Run those commands separately, so one failure does not prevent the remaining updates.
• Esc or Escape skips all updates. Confirm that no updates were run.
• Any other reply is not authorization. Repeat the two valid choices without updating.
• After y, show success or failure for every selected tool, then run the check-only launcher again and show the refreshed report. Do not show another prompt for a tool whose updater failed during that same update attempt.
• If all tools are current, or the user explicitly asked for check-only behavior, do not show the prompt.
• No preamble, recap, or closing offer outside the report, prompt, and update results.