Skip to content

takeone v0.2.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 Oct 20:42
v0.2.0
f17ed99
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The planner is now AI-first, can use an AI account you are already signed in to, and a plan edit follows its beat when the take is cut again.

Install

Needs Node.js 22.19 or newer and ffmpeg on PATH.

npm install -g --allow-remote=all https://github.com/umeranjum17/takeone/releases/latest/download/takeone.tgz
takeone doctor

New

  • AI plans the film. The planner decides at each beat which camera rule applies, in batches, and picks a theme, pace, opening, ending and camera style for the whole take. A --set key, an explicit --theme or a saved theme still wins over its choice. It also writes an AI title and a caption for each beat.
  • A long take now gets a film. It is planned in ordered chunks instead of falling back to the base look.
  • Plan with an AI account you already have. With no saved planner choice, takeone looks for an installed, signed-in Claude or Codex CLI, then a signed-in ChatGPT or OpenCode Go plan, then a stored key. You can also plan with an installed pi CLI. takeone stores no key of its own for these routes; see discovery order.
  • A signed-in CLI that refuses stops takeone make with exit 3, names the sign-in and the check to run, and renders nothing, instead of quietly falling back.
  • Edit the plan per beat. analysis/plan.json lists each beat's shot, idle and follow choice; a value you set is kept and applied on render. Each entry has a stable key, so an edit follows its beat when the take is cut again, an edit whose beat is gone is reported as dropped, and an entry carrying another beat's key is refused.
  • A plan edit the render cannot honour now stops with one error naming the beat, its kind and the value, instead of being dropped silently.
  • takeone render <take> --preview renders the same film at about 480p and 30 fps with no motion blur, for a quick look at a theme, aspect or --set change. The 1080p proxy it needs is made once beside the take.
  • Screenshot films can be planned locally while AI image planning is off.
  • TAKEONE_RENDER_THREADS=<n> holds a render or export to a fixed ffmpeg thread count.
  • macOS desktop recording preview: video only, through ScreenCaptureKit with a one-time Screen Recording approval, and takeone doctor reports the grant. It was built on Linux and has not been run on a Mac.

Improved

  • The default frame has a blue-to-violet backdrop and a softer shadow, and the drawn cursor is larger and smoother.
  • The title band is more compact in landscape and keeps the text-safe inset.
  • The camera holds a crop that cuts no text when it comes to rest, and eases the drag and the follow release more calmly.
  • Blurred wide camera renders are faster.
  • Android recordings: the immersive-mode confirmation is documented and automated.

Known issues

  • Signed-in Claude paths that skip a login made with an API key, Bedrock or Vertex, and the stop on a lapsed subscription, are tested against fakes only. They have not run against a real Bedrock or Vertex login or a real lapsed subscription.
  • The camera can still cut text while it moves slowly, for example during the follow and drag eases.
  • Desktop recording on Linux works on Wayland with Hyprland only. macOS desktop recording is the preview above.
  • iOS recording works with the iOS Simulator only, not a real iPhone.
  • Videos are silent: there is no audio track.

Guide: https://github.com/umeranjum17/takeone/blob/main/docs/guide.md

Assets 3
Loading