[miguel-heygen]

What

Release v0.9.32: a version bump of all packages and plugin manifests, the changelog entry and releases/v0.9.32.md.

It carries the seventeen changes since v0.9.31 and nothing else:

• fix(producer): size the first-frame stall window from the probe's browser start-up #7517: a long film that is slow to open no longer fails before its first frame. The wait for the first frame follows how long the browser took to start (at least 5 minutes, twice the measured start-up, at most 15 minutes).
• fix(lint): flag a script that calls gsap before GSAP loads, and say so when a render fails on it #7514: lint flags a script that calls GSAP before GSAP loads, and a render that fails on it says so.
• feat(producer): the render stream's complete event names the producer version that rendered it #7515: the render stream's complete event names the producer version that rendered it.
• fix(producer,engine): harden render paths, render slots and local listeners #7508: producer and engine render paths, render slots and local listeners are hardened. A render request with 0, fractional or negative workers now gets a 400 instead of a 500, and a rejected request frees its render slot.
• fix(studio-server): harden preview paths, cache folders and local servers #7510: Studio server preview paths, cache folders and local servers are hardened. Nothing is written through a link in .hyperframes or .thumbnails; if .hyperframes itself is a link, project history stays off for that project, with a warning.
• fix(cli): keep project folders from choosing the programs and config the CLI runs #7511: a project folder can no longer choose the programs, imports and configuration the CLI and skill scripts run. Helper programs come from PATH, not the working directory; the documented .hyperframes/bin and @ffmpeg-installer FFmpeg lookups stay.
• fix(cli): harden lambda destroy, snapshot media, upgrade and project file serving #7509: lambda destroy, snapshot media, upgrade and project file serving are hardened. lambda commands refuse a state file recorded for a different stack name.
• fix: harden docs snippets, SVG sanitizer, lint regexes, GSAP writers and registry pins #7512: docs snippets, the SVG sanitizer, lint regexes, the GSAP writers and registry pins are hardened.
• perf(docs): load showcase film tiles only when they are near the screen #7507: the docs showcase loads film tiles only when they are near the screen.
• fix(share): publish fallback locks with their owner #7516: on file systems that refuse hard links, a share or export killed while taking its lock no longer leaves a lock that later attempts cannot clear. An older client that meets an active fallback lock may report a file system error. The share package's size budget grows to fit.
• docs(readme): show the GitHub rank and skills downloads badges #7519: the README shows the GitHub rank and skills downloads badges.
• fix(studio): respect composition and cancellation in text inputs #6011: in Studio's text inputs, Enter or Escape while an input method is still composing text no longer commits it, and Escape then leaving the field no longer saves the cancelled edit.
• test(inspector): decode file URLs in HTTP contract fixtures #6010: the inspector's HTTP contract tests pass from a checkout whose path has spaces or non-Latin letters (tests only).
• feat: opt-in 4:4:4 chroma for SDR MP4 masters #5820: SDR MP4 renders can keep full colour detail with --chroma-subsampling 444, for masters that will be re-encoded; the default is unchanged.
• fix(studio): run the composition hover preview in an opaque origin #7522: Studio's composition hover preview runs in an opaque origin (sandbox="allow-scripts"), so a composition's scripts cannot reach into Studio.
• docs: list the hub's /blog, /vs and /tools sitemaps in robots.txt #7527: the hub's robots.txt lists its /blog, /vs and /tools sitemaps (docs).
• fix(share): guide recovery when an export replay is refused #7531: when a share export replay is refused, the message guides recovery (share package only).

Why

Ships the first-frame wait fix for long, mostly still films, the share lock fix, opt-in 4:4:4 chroma, and the hardening changes merged since v0.9.31.

Related work

Follows chore: release v0.9.31 (#7506).

How

bun run changelog:draft 0.9.32 --write --from v0.9.31, then the generated TODO summary was written and the style-bar line removed in both notes files, and the #7517, #7516, #6011, #5820 and #7522 lines were put in plain words. Then bun run release:prepare 0.9.32 --from v0.9.31 --no-tag set 15 packages and 5 plugin manifests to 0.9.32. The changed file list matches the v0.9.31 release commit (22 files). No tag is pushed; the release workflow tags the merge SHA.

Release gate

The full release gate (Desktop's full walk with these packed packages) runs on the previous cut, 9fc0a130, which holds every change here except #7531. #7531 changes only packages/share (reviewExport.mjs, reviewVersionSentences.mjs, their test and README), so on this head the checks that cover it are: this pull request's Build, the share package's own tests, and a render with the packed CLI. Links below as they finish.

• Full gate on 9fc0a130: pending
• Build on this head: pending
• Share package tests on this head: pending
• Packed CLI render on this head: pending

No visible change

Version bump and changelog only. Each change was captured in its own PR.

Test plan

• Unit tests added/updated (release commit only, no code)
• Manual testing performed: diff checked to be only version strings plus the two notes files
• Documentation updated: docs/changelog.mdx and releases/v0.9.32.md, summary written by hand
• Comments follow CONTRIBUTING.md "Comments"

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
hyperframes	🟢 Ready	View Preview	Nov 11, 2026, 1:41 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[heygengenesis]

Requesting changes at a40893ba. The release would publish code that its notes leave out and its gate never ran.

Blocker: main has moved past the release branch's base. This branch sits on 7ff8ece46 (#7517). Since then, two more commits have landed on main:

• b37e4b026 fix(share): publish fallback locks with their owner (#7516). This is a code change: packages/share/src/projectLock.mjs +151/−41, new tests, and a scripts/package-size-budgets.json change.
• 717d3d02c docs(readme): GitHub rank and skills downloads badges (#7519).

publish.yml checks out and tags github.event.pull_request.merge_commit_sha, and that commit is built on current main. So if this merges as is, v0.9.32 publishes @hyperframes/share with #7516 in it, while:

• the PR body, releases/v0.9.32.md and docs/changelog.mdx say the release "carries the nine changes since v0.9.31 and nothing else", and #7516 is not in any of them;
• the release gate and the render off the packed packages ran on a tree without #7516, so the published share package and the size budgets were never checked in this release;
• the notes' v0.9.31...v0.9.32 compare link will list 11 commits against 9 notes.

To clear it: rebase onto current main, then re-run changelog:draft 0.9.32 --from v0.9.31 and release:prepare. Add #7516 under Fixes (Share:) and decide where #7519 goes (Docs, or leave it out as the README-only change it is). Update the PR body's count, then re-run the release gate and the packed-package render on the new head.

Otherwise checked, and fine:

• The diff outside the notes is exactly 20 "version": "0.9.31" → "0.9.32" lines across 15 packages and 5 plugin manifests.
• The nine notes entries match v0.9.31..7ff8ece46 one to one, with correct commit SHAs and PR links.
• The #7517 line matches what that PR does: at least 5 min, twice the start-up, at most 15 min.

[heygengenesis]

Approving at 9fc0a130. The blocker from my last review is cleared.

What I checked

• Up to date with main. The release commit's parent is 499a17e67 (#7527), which is current origin/main. v0.9.31..main has 16 commits, and the release adds exactly one commit on top.
• The notes cover every change. All 16 PRs appear with correct commit SHAs and links in both releases/v0.9.32.md and docs/changelog.mdx: #7527, #5820, #7522, #6010, #6011, #7519, #7516, #7517, #7514, #7515, #7507, #7508, #7510, #7512, #7511 and #7509. The PR body lists the same 16.
• The version diff is clean. Outside the notes, the diff is exactly 20 "version": "0.9.31" → "0.9.32" lines, 22 files in all, the same shape as the v0.9.31 release.
• Spot checks. The --chroma-subsampling 444 flag in the #5820 line exists in packages/cli/src/commands/render.ts ("420" | "444"). #6010 is filed under Internal, and #7519 and #7527 under Docs & Examples.

Conditions

• CI, CodeQL, Docs and the Lambda arm64 image were still running on this head when I reviewed, and the merge state is BLOCKED until they finish. This approval assumes they and the release gate go green.
• publish.yml tags the merge commit. If anything else lands on main before this merges, the merge commit will contain changes the notes leave out, so cut the release again rather than merging on this approval.

Nit, not blocking: the summary paragraph at the top of the notes mentions only the first-frame fix and the hardening work. The share lock fix and opt-in 4:4:4 are in the PR's "Why" but not in that paragraph.

[heygengenesis]

Approving at e4167d1c, reviewed as the delta from 9fc0a130, which I approved last time. No blockers.

The delta is what the PR says it is

• Up to date with main. The release commit's parent is 941049a84 (#7531), which is current origin/main. The only commit on main since the previous cut's base (499a17e67) is #7531.
• #7531 is share-only: packages/share/src/reviewExport.mjs, reviewVersionSentences.mjs, README.md and tests/reviewPersistence.test.mjs. It touches no engine, producer, CLI or package manifest. It merged through the merge queue with CI [merge_group], regression and Windows render verification green.
• Release commit versus the last cut: range-diff shows the only changes are the #7531 bullet under Fixes in both releases/v0.9.32.md and docs/changelog.mdx, and the changelog tags (adds Share, drops README). The rest of the diff is still exactly 20 "version": "0.9.31" → "0.9.32" lines across 22 files.
• Share tests on this head: node --test tests/*.test.mjs in packages/share passes 57/57 locally.

Gate state, as Actions shows it

• On 9fc0a130, CI was cancelled once this head was pushed, after 10 jobs passed, including Release checks. Two Catalog jobs were cancelled. regression, Docs, Lambda arm64 image, preview-regression and Windows render verification all passed there. I can't see the full release gate you mention from Actions, so this approval relies on your report that it passes on that cut.
• On e4167d1c, regression, PR captures, Windows render verification, preview-regression and both player perf runs are green. CI, Docs, Lambda arm64 image and CodeQL are still running. This approval assumes they finish green.

Nit, not blocking: the #7531 bullet is the raw commit title ("Guide recovery when an export replay is refused"). Your other share lines are in plain words, saying what the user sees.

Same standing condition as before: if anything else lands on main before this merges, the merge commit publish.yml tags will contain a change the notes leave out, so cut it again.