[miguel-heygen]

What

hyperframes check now recognizes fixed-geometry text, generated content, painted CSS counters, form-control state, and clip-path wipes as real motion instead of failing them with sweep_static, and the frozen-sweep guard and keepsMoving liveness share one definition of "what counts as motion" so they can no longer disagree. A genuinely frozen composition still fails the guard.

Why

The sweep guard compared box geometry, opacity, variable-font axes, and media pixels across seek samples. A fixed-width countdown could visibly change from 10 to 09 while every recorded field stayed identical, so check concluded the timeline never advanced.

Widening only the sweep fingerprint would have left a second owner: motion-sample.browser.js's liveness signature still compared box + opacity, and the two meet in one decision in checkPipeline.ts (a motion_frozen finding suppresses sweep_static). On the motivating attr(data-txt) countdown the sweep guard accepted the composition while liveness still reported motion_frozen.

Leading-zero coercion in data-txt remains unchanged because GSAP AttrPlugin owns that behavior; authoring guidance or a lint hint for numeric attribute strings remains separate.

How

• New motion-signature.browser.js, injected before both samplers, owns the per-element channel list, split into box channels (box, opacity, font-variation axes, clip-path, checkbox checked/indeterminate) and content channels (direct text, control value/selection, resolved ::before/::after content, media pixels) plus a composition-level CSS counter channel. window.__hyperframesLayoutGeometry and the liveness signature are both compositionSignature(root); liveness passes quantize: true (2px / 0.08 buckets), the sweep guard reads exact. Channels are a reader list so a further channel (e.g. SVG dash) plugs in as one function.
• Counter state is folded from every box-generating element and the root's ancestors. Elements the browser does not render — display:none subtrees and skipped content-visibility contents (hidden, or auto while off-screen), decided by element.checkVisibility({ contentVisibilityAuto: true }) where available — generate no painted boxes and are excluded (a content-visibility: hidden host's own box and a checkbox's checked/indeterminate state still count; its text, pseudo content, control value, and media pixels do not; hosts where size containment does not apply — non-atomic inline, display: contents, table boxes other than a cell, and ruby boxes — skip nothing and stay fully signed, matching Chromium 152 where a table-cell host skips its contents and a table-caption host does not; a display: contents host has no box but its pseudo-elements paint, so it stays a counter owner/consumer — except as the child of a host that skips its contents, where the parent's verdict prunes it because the platform check cannot tell that box-less case from an ordinary display: contents host); only declarations that name a counter some painted generated content consumes (counter()/counters() in a ::before/::after/::marker on a non-ignored, paintable host, or the implicit list-item marker) are folded in. A hidden decoy element with a varying counter-reset therefore cannot make a frozen composition read as live.
• clip-path is a channel, replacing the fingerprint's former reliance on the clipped-away visibility probe, so a wipe over a static box still counts as motion.
• Elements under data-layout-ignore / data-layout-check="ignore" inside the measured root are excluded from the signature (a decorative layer must not prove the timeline advanced). The root itself — the composition root or an asserted keepsMoving scope — is always measured, even when it or an ancestor carries the opt-out: an assertion naming an element outranks the layout-audit opt-out, and an ignored scope therefore reports a real signature instead of motion_selector_missing. Explicitly asserted selectors (appearsBy, before, staysInFrame) likewise still report visible.
• The per-element visibility floor is the same as layout-audit.browser.js: element.checkVisibility({ opacityProperty, visibilityProperty, contentVisibilityAuto }) first, then display / visibility, inherited opacity ≥ 0.2, and a non-empty box.
• Textual channels are hashed (FNV-1a, length-delimited fields) so raw composition text never leaves the page.
• Behavior deltas for keepsMoving, all in the more-sensitive direction: the "*" scope resolves its root like the sweep guard ([data-composition-id][data-width][data-height] first), the scope element is part of its own signature (a leaf withinSelector is no longer reported as missing), and media pixels, text, generated content, form state, and counters now count as liveness. Documented in docs/packages/cli.mdx.
• motion-sample.browser.js throws a named error from __hyperframesMotionSample when the shared script is missing (an install-time throw would only surface as a page error and an opaque "is not a function" from the driver's evaluate); it reuses the shared round. build-copy.mjs ships the new file.

Test plan

• motion-signature.browser.test.ts (happy-dom, 30 tests): per-element fake getComputedStyle that reports only each element's own declared style, so hidden-subtree and ancestor-counter handling has to come from the classifier. Covers media pixels, opacity reveal, font axes, quantization buckets, text/root text/attr()/pseudo content, checkbox/indeterminate/select/input/textarea, zero-box and ancestor counter owners, pseudo counter-increment, visibility:hidden owners, display:none decoys (direct and nested), unconsumed counters, name-gated decoys next to a list marker, ignored-layer consumers, clip-path, hidden-descendant text, data-layout-ignore motion, and a root that is itself data-layout-ignore (measured, while opted-out layers inside it are not).
• motion-signature.browser.chromium.test.ts (real Chromium via puppeteer-core, 34 tests, skips without a system browser; opt-in on Windows via HYPERFRAMES_BROWSER_TESTS=1 because the cold launch in the shared Windows package lane exceeds the package hookTimeout — the ubuntu Test job finds the runner's Chrome and runs it; the browser launch has its own 120s hook timeout): the same branches against the platform with the exact injected scripts — Blink's attr() substitution, unsubstituted counter(), ::marker content, counters()/counter(x, upper-roman)/counter-set, zero-box hosts with absolutely positioned pseudo-elements, sweep/liveness agreement on the countdown fixture, text / counter-decoy / host-own-text-and-pseudo cases under content-visibility: hidden a counter decoy inside off-screen content-visibility: auto contents, a repaint of a content-visibility: hidden canvas, and a value change on a content-visibility: hidden text input (all frozen; the host's box move still counts), plus three still-motion controls: a counter painted only by a display: contents host's pseudo-element (a box-less host stays a counter consumer), text and pseudo content inside an inline content-visibility: hidden host (containment does not apply to non-atomic inlines, so nothing is skipped), and a checkbox toggle on a content-visibility: hidden control (theme paint is box paint), and a keepsMoving scope that is itself data-layout-ignore (non-empty signature, inner opt-out still excluded). Removing any one of the new branches turns at least one test red (16 mutations checked; seven of the ten new Chromium tests are red against the previous classifier; the display: contents, inline-host, and checkbox ones guard against regressions introduced and fixed during review).
• Follow-up pins from review (Chromium, 4 tests): a counter painted only by a display: contents host's pseudo-element inside content-visibility: hidden contents is frozen (was signed: the box-less host escaped the platform prune and stayed a consumer); a display: table-cell host's own text and pseudo content are frozen and a <caption> host's own text and pseudo content are motion (the containability regex had the two inverted versus Chromium 152 — verified with checkVisibility probes on the child and screenshot diffs: cell contents skip, caption contents paint); and a fixture whose only counter consumers are a skipped list-item host's hidden ::marker / ::after, asserting the skipped child is not visible for motion-sample and that neither the owner's counter-reset nor the hidden pseudo's counter-increment moves the signature. The first three are red against the previous classifier; the fourth is red under each of three mutations that survived the existing suites (dropping checkVisibility from isVisibleElement, reading real pseudo styles on a skipped host, and paintsContent ignoring skipped). Chromium probes also confirmed that under a content-visibility: hidden host, modal <dialog>, [popover], position: fixed descendants, nested content-visibility: visible, and slotted children are all skipped, so pruning by the parent's verdict removes nothing that paints. Known remaining gap, unchanged: a display: contents child of an off-screen content-visibility: auto host is not pruned (the heuristic only detects hidden); noted in the source comment.
• motion-sample.browser.test.ts (8 tests): install without the classifier and a named error on first call, liveness agreement with the sweep fingerprint on the attr() countdown, leaf scope signature, ignored asserted selector still visible.
• check.test.ts: 59 passed, including the four pre-existing frozen-sweep guard controls (driver-mocked, unchanged).
• tsc --noEmit, oxlint, oxfmt, and the repo's fallow gate pass on the changed files.
• Documentation: docs/packages/cli.mdx keepsMoving paragraph updated to describe the shared classifier.

[miga-heygen]

Textual sweep fingerprint extension is thorough — direct text, pseudo content, CSS counters, and form state all hashed correctly with FNV-1a. Eight new test cases cover every channel. — Miga

[miga-heygen]

Second-pass quality review at head a329d21e (isolated worktree; rebase checked byte-for-byte against the PR file list, 326/22 over 4 files; browser-audit 128/128, check pipeline 59/59, tsc --noEmit clean, oxlint and oxfmt clean; 7 mutations; 10 probes in real Chromium via the repo's puppeteer-core). Three findings I consider blocking.

Blocking

• B1 — the hidden-element counter loop admits every non-visible element. packages/cli/src/commands/layout-audit.browser.js:1770-1776 iterates allElements minus the visible set, so display:none subtrees are included, although the comment above it scopes the loop to zero-box owners whose visible descendants paint the counter. Because extra channels can only suppress sweep_static, one decoy display:none element with a varying counter-reset makes a fully frozen composition read as live. Reproduced in Chromium with the actual audit script. Restrict the loop to elements that own a counter consumed by a visible descendant's pseudo-element, or at least exclude display:none subtrees.
• B2 — "what counts as motion" now has two owners that disagree. __hyperframesLayoutGeometry reads nine channels after this PR; livenessSignature (packages/cli/src/commands/motion-sample.browser.js:66, elementSignature) is still box + opacity only. They meet in one decision at checkPipeline.ts:519. Proved with the PR's own motivating case: an attr(data-txt) countdown yields sweepDetectsMotion: true, livenessDetectsMotion: false, so motion_frozen still fires on the composition this PR exists to accept. One classifier, consumed by both samplers.
• B3 — four new branches are pinned by no test. The ancestor counter loop (:1766), pseudo counter folding (:1735-1736), TEXTAREA / indeterminate (:1704-1708) and root inclusion (:1745) can each be deleted with all 128 tests green. Tests are jsdom with a mocked getComputedStyle, so the attr() test asserts the mock rather than the platform; one real-browser case per branch would pin them.

Non-blocking: the new pseudo-element style reads cost a measured 1.6–3.0× per sample with no cap (the sibling classifier notes its own cap explicitly); the description's test counts (92/52) do not match the actual 128/59, and the "four frozen-sweep guard controls" it cites are pre-existing tests that stub out all the new code; the unchecked typecheck box is in fact clean in a fresh install.

Holds: Blink substitutes attr() in computed pseudo content ("10" → "09") and does not substitute counter(), so both mechanisms in the diff are justified. Mutations on the two landed detectors go red (6 and 1 tests).

Verdict: COMMENT — needs fixes before merge.
Reasoning: a decoy hidden counter defeats the frozen-sweep check, the liveness sampler still rejects the exact composition the PR accepts, and the new branches have no coverage.

Review by Miga

[mintlify]

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

Project	Status	Preview	Updated
hyperframes	🟢 Ready	View Preview	Oct 2, 2026, 12:25 PM

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

[miga-heygen]

Addressed in fbec4c6.

[miga-heygen]

Second-pass quality review at head fbec4c60 (isolated worktree; 172/172 across the four touched suites including the real-Chromium suite, 11 mutations, 3 Chromium probes). Two of the three blockers from my review at a329d21e are closed; the third is closed for display:none but not for the neighbouring case, and the new suite turns a required check red.

Closed and pinned. One owner of the signature: compositionSignature (packages/cli/src/commands/motion-signature.browser.js:320) maps ELEMENT_CHANNELS (:226) for both consumers — the sweep exact at :359, the liveness sampler quantised via the same call at motion-sample.browser.js:66 — and the quantise buckets (:38-39) reproduce the deleted 2px / 0.08 thresholds exactly. elementSignature / livenessSignature are gone and __hyperframesLayoutGeometry is defined once. Removing each of the eight channels in turn goes red (box 2, opacity 1, fontAxes 1, clipPath 2, text 3, control 6, generatedContent 4, mediaPixel 2); reverting liveness to its own list → 4 red including the attr() countdown; dropping the hidden-subtree walk → 2 red.

Blocking

1. The required Tests on windows-latest check is red at this head because of the new suite. motion-signature.browser.chromium.test.ts:59 — the beforeAll browser launch exceeds the repo-wide hookTimeout: 30_000 (vitest.config.ts:39) in the studio-engine-cli lane: Hook timed out in 30000ms, all 20 tests skipped, lane failed, aggregator failed. The lane was green at a329d21e and 10dadce3. This is the first real-browser launch inside @hyperframes/cli (the producer's lives behind its own script), so it wants its own lane or an opt-in flag on Windows rather than a global timeout bump.
2. content-visibility: hidden subtrees still admit unpainted motion. motion-signature.browser.js:329 prunes only display:none. Proved in Chromium: a text mutation, or a counter decoy, inside a content-visibility: hidden wrapper changes __hyperframesLayoutGeometry(). The copied isVisibleElement (:82) dropped layout-audit.browser.js:180-190's element.checkVisibility({ contentVisibilityAuto: true }) while its comment at :74 still claims the same visibility floor; that API is exactly what rejects this case.
3. A keepsMoving scope that is itself data-layout-ignore now reports motion_selector_missing. :339 excludes the scope and all descendants, so the signature is empty every frame and motionAudit.ts:171 says the selector matched no element for an element that exists (proved: empty vs a 44-character signature without the attribute). That contradicts the file's own rule at :74-78, which the diff honours for sampleElement's visible bit but not for asserted-scope liveness, and it is a behaviour change from main.

Non-blocking: the install-time ordering throw never fails anything — addScriptTag resolves on load, so it degrades to a page error and then an opaque __hyperframesMotionSample is not a function; duplicated round(); mediaPixelChannel readbacks now sit on the 300-sample liveness path; the counter ancestor walk escapes a scoped liveness root; paragraph merge at docs/packages/cli.mdx:718. No missed injection site (checkBrowser.ts:438-440, layout.ts:393-394, build-copy.mjs:107 are the only three). The dash-only SVG case is still frozen, which is now consistent between both samplers rather than a disagreement.

Verdict: COMMENT — needs fixes before merge.
Reasoning: the single-owner refactor is right and well pinned, but the head fails a required check and two hidden-content cases still leak into the signature.

Review by Miga

[miga-heygen]

Addressed in bfaf6eb.

[miga-heygen]

Reviewed head bfaf6eb9 in an isolated worktree (one commit on fbec4c60, 6 files). Follow-up to my review at fbec4c60.

Strengths

• motion-signature.browser.js:326-331 — isOptedOut(element, root) makes the asserted scope authoritative: the root's and its ancestors' data-layout-ignore no longer suppress measurement, while a descendant opt-out still skips. Pinned in both happy-dom (:309) and Chromium (chromium.test.ts:180).
• motion-signature.browser.js:432 — a content-visibility:hidden host signs BOX_CHANNELS only; each of the nine channels lives in exactly one list, and skipsContents is the single owner of the css-contain-2 exemption table.
• chromium.test.ts:36-68 — describe.skipIf(!RUNS_CHROMIUM) with the 120s launch ceiling clears the Windows lane; the ubuntu Test job log at this head shows the suite ran (30 tests, not skipped).

Prior blockers

• Windows lane timeout: fixed (Tests on windows-latest and the studio-engine-cli lane pass).
• content-visibility:hidden leak: fixed for the general case via checkVisibility pruning at :398-404 plus the box-only host signature; one residual below.
• Opted-out asserted scope: fixed (:326-331).

Important

1. motion-signature.browser.js:402 — noBox && style.display !== "contents" exempts display:contents from the platform's no-box verdict unconditionally, so a display:contents element whose parent skips its contents stays a box owner with a live pseudo. Chromium 152 probe: #wrap{content-visibility:hidden} #h{display:contents} #h::after{content:counter(c)} — changing the owner's counter-reset changes __hyperframesLayoutGeometry() although nothing paints. Fix: when the platform decides, also treat display:contents as unrendered when its parent is in unrenderedBelow or skipsContents(parent); add a Chromium pin.
2. motion-signature.browser.js:73-74 — NOT_CONTAINABLE_DISPLAY disagrees with Chromium 152 on two table rows. A child inside a content-visibility:hidden table-cell reports checkVisibility({contentVisibilityAuto:true}) === false (Chromium skips cell contents; the regex says not containable, so the host over-counts its own text and pseudo). A child inside a hidden table-caption reports true (Chromium does not skip caption contents; the regex and the comment say containable, so the caption's own text drops out). Suggest table-(?!cell$)[a-z-]+ plus table-caption in the exempt set, with two Chromium pins. Rare in compositions.

Nits / test gaps

• Dropping the checkVisibility call in isVisibleElement (:122-127), keeping a real pseudo on a skipped host, and letting paintsContent ignore skipped each leave all 68 tests green. The code is right; a fixture "ignores a counter consumed only by a content-visibility:hidden host's pseudo" would pin all three.
• The Chromium suite is silently skipped on any Linux runner without a system Chrome; consider hard-failing when CI is set and no binary is found.
• Deferred install throw (motion-sample.browser.js:10-13,65-70): same message, surfaces at first call; both injection sites inject the signature first, so this is defensive only.

Verified: single visibility predicate; visibility:hidden parent with visibility:visible child still counts; opacity:0 unchanged from before; content-visibility:auto on-screen counts; a skipped host's counter-increment still reaches siblings; inline-block correctly absent from the skip list; round shared via __hyperframesMotionSignature.round. Local: 68/68 across the three touched suites (Chromium via system binary), motionAudit + layout-audit 131/131, tsc --noEmit clean for cli.

CI: every required check is green at bfaf6eb9 except Render on windows-latest (run 34915456761): the downloaded chrome-headless-shell --version probe timed out and the step reported "Chrome cannot start". This PR touches no workflow or browser-manager files and every sibling windows-render run in the same window is green, so it reads as a runner flake; it needs a rerun before approval.

Verdict: COMMENT
Reasoning: All three prior blockers are closed and pinned; two narrow correctness gaps remain in the containment table and the display:contents exemption, and the required Windows render check is red on an infra failure that a rerun should clear.

Review by Miga

[miga-heygen]

Addressed in 09418ae.

[miga-heygen]

Reviewed head 09418ae7 in an isolated worktree.

Re-pass of the one commit on bfaf6eb9 addressing the two importants from my previous review. Both closed:

• motion-signature.browser.js:405-415 — skippedHosts plus boxlessOwner = display === "contents" && !skippedHosts.has(parent): a display:contents child of a host that skips its contents is now pruned. The direct-parent set is enough because the first such child lands in unrenderedBelow and deeper display:contents nodes are pruned through unrenderedBelow.has(parent) (Chromium oracle: wrap[cv:hidden] > mid[contents] > host[contents]::after counter() leaves the signature unchanged on owner mutation). Boxed descendants still go through the platform check, unchanged.
• :77 — table-(?!cell$): Chromium 152 oracle agrees on every case (div display:table-cell, real td/th under content-visibility:hidden skip their contents; div display:table-caption and real caption paint them; table / inline-table / inline / ruby paint). The comment at :68-76 now describes the platform, and correctly notes the spec would have both the other way.

Mutations, each caught by the new Chromium tests: drop the !skippedHosts.has(parent) term, revert the regex to caption, never populate the set, drop the table-* branch, bypass the set when the platform decides. 64/64 across the Chromium and happy-dom files locally.

Nit — :419-421 says a skipped host's own counter-increment / counter-set still reach a sibling's painted counter(). In Chromium 152 that is state-dependent: a host skipped from first layout does not contribute (sibling paints 10 with body reset 10 and host increment 5 / set 7 pending); after one unskipped layout the value is retained through later skips (15). Keeping the declarations remains the accepted conservative over-count, so no code change; suggest wording "may still reach … once the host has been laid out unskipped".

Verdict: APPROVE
Reasoning: Both prior importants are fixed with platform-verified behaviour and mutation-pinned tests; the only remaining item is a comment wording nit.

— Miga

[github-actions]

Edit accuracy: accurate 1216 (base branch 1216), smooth 1004 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (1)

• seqnudge-none-pct-r0-nested-z100 (fixed by fix(studio): playhead keyframe edits keep their ease, undo names its edit and never waits on videos #4889): pass, passes