pycodebr/heditor 2.20.0 went up on a 2026-11-04 commit (8b8add8). heditor is a package that provides a vertical-video editing skill to AI agents such as Claude Code, Codex, OpenCode, Cursor, and Hermes. This version has two main axes. The rendering engine HyperFrames was upgraded from 0.8.147 to 0.9.13, and `heditor study`, which measures a reference video locally and turns it into a style, was newly added. The commit touched 18 files, adding 584 lines and deleting 20 lines.
What disappeared in HyperFrames 0.9.0 is the Plan v1 delivery method for distributed rendering on Lambda or Cloud Run, and heditor does not use that path. The render flags heditor passes are unchanged as well. The CHANGELOG records that, before bumping the version, two pilot pieces were rendered with each of the two versions and compared with `hyperframes diff`. A 240-frame physics example using three.js with bloom and depth of field had an average SSIM of 0.9957 (p05 0.9954, worst-section p05 0.991), and a 1,228-frame example using GSAP SplitText, DrawSVG, MorphSVG, MotionPath, subtitles, and a mixed soundtrack had an average SSIM of 0.9946 (p05 0.9937, worst-section p05 0.9876). The lengths were the same and the sound offset was 0 ms. Existing installations must re-run `heditor setup` to receive the new pinned version.
`heditor study` is a command that measures a chosen reference video only on this computer, without login or credits. `measure probe`, `measure cuts` (which catches the jump cuts and punch-ins that scene scores miss, through frame-to-frame differences), `measure events --map` (appearance, disappearance, change, stillness, strong sound onset), `measure look`, `measure face`, `audio --quality local` (sound types, loudness, silence, start points, tempo, sections), and a summary sheet divide up the commands. The results are saved as raw JSON, and as study.json and study.md containing shot lengths, cuts per 10 seconds (broken into thirds), the first cut, on-screen events per 10 seconds, the share of still sections, sound starts per 10 seconds, loudness, the share of silence, look, and face size, and `heditor study compare <ref> <yours>` sets the pilot and the reference side by side. `audio` never runs together with `--quality best`, which uploads to HeyGen and can incur charges, and tests block calls to paid commands. CI also runs test_study.py.
The limitation is that the SSIM figures are the repository owner's own records. The two pieces are pilots, and the comparison reference, `hyperframes diff`, is also a standard defined by that same repository. What `heditor study` produces are numbers like cuts, events, sound, and look, but how to carry those numbers on into a style depends on the procedure written in `references/reference-study.md`. The paid command `reverse-any-video` runs only when the user has approved the plan and the quote, is used only on video the user has the rights to, and is explicitly stated not to include other people's video, faces, logos, or music in the result.