Field lesson from translating a real Motion timeline: the two returned
encodings window durations differently, and keyframes at times ~0.9999
are loop-wrap resets, not authored motion. Hand-normalizing across
encodings and inventing visible returns produced a render that diverged
from Figma. The skill now mandates verbatim single-encoding translation,
wrap-via-repeat, and a frame-grid comparison against export_video ground
truth before completion.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
slugify("3D Object - Headphones") produced id="3d-object-headphones" —
valid HTML, but querySelector("#3d-…") throws (CSS idents cannot start
with a digit), which kills GSAP targeting and figma-motion translation
against imported components. uniqueSlug now prefixes digit-leading slugs
("n3d-object-headphones"). Found translating a real Figma Motion timeline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both found running the brand-loop guide end-to-end against the Simple
Design System:
- nodeToHtml subtracted the ROOT origin from every node's absolute bounds,
but CSS absolute positioning resolves against the nearest positioned
ancestor — every nesting level re-added its ancestors' offsets, drifting
nested content down-right and pushing deep children off-frame (hero
buttons invisible, pricing grid collapsed to one card). Children now
subtract their PARENT's box; regression test with a two-level tree.
- trackCommandFailures asserted unknown flags against the command group's
own (flagless) arg table even when the group was delegating to a
subcommand, so `figma component <ref> --name x` imported and THEN threw
"Unknown flag: --name". The assertion is now skipped when the first
positional names a subcommand; leaf and non-delegating behavior is
unchanged and covered by tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
applyPositionEdits(doc) guarded each element with `instanceof HTMLElement` —
but `doc` is frequently an iframe's document (the SDK's edit preview, any host
embedding a composition), whose elements are HTMLElement instances of THAT
frame's realm, never this module's. The check silently no-ops on every single
element cross-realm, so bulk position edits never apply inside an iframe.
Use the document's own realm's HTMLElement constructor (doc.defaultView);
duck-type on `.style` when defaultView is unavailable (a detached/synthetic
document). The single-element applyPositionEditToElement was already
realm-safe — only the bulk wrapper had the bug.
Added a regression test using a real jsdom iframe, confirmed it fails on the
old `instanceof HTMLElement` check and passes with the fix.
Address unanimous review feedback on PR #2095 (Miga, Magi REQUEST_CHANGES,
Rames D Jusso): the router's process.env.HF_DE_PARALLEL_STREAM mutation
was only cleared on the DE self-verify-retry branch — every other exit
(happy path, any non-DE-verify error, abort) left it set, leaking the
parallel-streaming opt-in into the next render sharing the same process
(a regression/benchmark harness, or any batch host).
Capture the prior value before mutating and restore it in the outer
`finally` (executeRenderJob's own top-level try/finally, which runs on
every exit path by construction), not just the narrow retry branch. Note
`deParallelRouter` and the new `deParallelStreamEnvBefore` had to move
above the outer `try` — a `let` declared inside `try {}` is not visible
in the sibling `finally {}` block in JS, so the original placement
alongside the other DE state would not have compiled once referenced
from the finally.
Also: renamed the shared `preInversionWorkerCount` local to
`preRoutingWorkerCount` (Miga + Rames nit — it now serves both the
inversion and the router), and pinned worker count 3 explicitly overrides
calibration by design (documented per Miga/Rames's question, not a bug).
Verified end-to-end (not just unit tests): ran two executeRenderJob calls
back-to-back in one process, router-eligible then not — env is restored
to undefined after render 1 and stays clean through render 2, the exact
leak scenario the reviews described. 117 orchestrator tests pass (3 new,
covering the restoreEnv primitive directly).
Blocker (flagged by all three reviewers, still open after the CI fix):
- Composition.serialize() interface in types.ts never got the { stripRuntime? }
param the implementation already accepts, so a consumer holding a
Composition-typed ref (exactly pacific's case) got a strict-TS arity error
calling comp.serialize({ stripRuntime: true }). Widened the interface.
Also addresses:
- getRootElements() now cached like elementsCache (same 3 invalidation sites) —
cheap insurance if a layer panel calls it every render tick.
- getElementTimings' resolver now uses the already-parsed expr.value for the
absolute-number case instead of silently re-parsing via parseFloat, via a
small resolveReferenceStart helper split out to keep resolveStart's own
branching low.
- bareId's `?? scopedId` fallback gets a comment: it's unreachable at runtime
(split() always returns >=1 element) but required by noUncheckedIndexedAccess.
- serialize({ stripRuntime }) docblock generalized past "the editing iframe" —
it's for any host driving its own clock.
- Documented (and pinned with a test) the bare-id reference resolution's
cross-scope behavior: a sub-composition element referencing a colliding bare
id resolves to the canonical top-level match, same as the runtime's own
resolver — consistent, but a real authoring footgun worth calling out.
- New tests: chained (A->B->C) references, a direct self-reference cycle, a
mutual A<->B cycle, the cross-scope bare-id collision above, and an import
assertion that RUNTIME_BOOTSTRAP_ATTR is actually reachable from
@hyperframes/core and matches the marker generators stamp.
422/422 sdk tests passing (417 + 5 new). Full workspace build (incl. studio)
verified clean.
* fix(cli): map zh to espeak-ng's cmn for Kokoro TTS synthesis
espeak-ng 1.52.0 recognizes Mandarin Chinese as the ISO 639-3 code "cmn",
not Kokoro's own voice-prefix convention "zh". `hyperframes tts --lang zh`
was forwarding "zh" straight through to kokoro_onnx.Kokoro.create(), which
failed with "language zh is not supported by espeak backend". Translate
only at the Python/espeak boundary; the public --lang value stays "zh"
since that matches Kokoro's own docs and voice-ID prefixes.
* test(cli): cover Kokoro zh language override
Promotes the opt-in HF_DE_PARALLEL_STREAM mechanism (#2056) into the
auto-routing decision, gated behind its own default-off flag
(HF_DE_PARALLEL_ROUTER). This is the next step from the 2026-07-08
parallel-DE benchmark verdict: par3/single 1.16-1.36x on real-work
comps >=2,000 frames, no comp anywhere losing to single-worker.
shouldPreferParallelDrawElement mirrors shouldPreferSingleWorkerDrawElement
(#2026) but takes priority over it when both are eligible — its higher
default threshold (HF_DE_PARALLEL_MIN_FRAMES=2000 vs the inversion's 900)
means it only ever picks up the long tail the inversion's own benchmark
didn't cover. Fixed at 3 workers (benchmark-validated; not calibration-
derived), same shape as the inversion pinning to a fixed 1.
resolveParallelRouterRetryPlan mirrors resolveInversionRetryPlan for the
self-verify-failure rollback path: falls back to the ordinary (non-DE)
parallel-disk path at the pre-router worker count. The caller must clear
HF_DE_PARALLEL_STREAM before recomputing useStreamingEncode or the retry
would keep resolving to the parallel-streaming shape.
New telemetry (de_parallel_router, de_pre_router_workers) tags which
render used the router, separate from de_worker_inversion — needed
before the planned telemetry soak can segment revert-rate and
de_verify_min_db to the parallel cohort specifically; today there's no
way to tell those apart from ordinary single-worker DE renders.
Verified end-to-end: HF_DE_PARALLEL_ROUTER=true on a 2,381-frame comp
resolves to 3 workers with 3 separate drawElement sessions and renders
successfully; without the flag, behavior is unchanged (falls through to
the existing single-worker inversion, workerCount=1) — no regression to
current production routing. 114 orchestrator tests pass (15 new).
Closes gaps surfaced by pacific#30298 (hyperframes layer panel), where consumer
code had to hand-roll fixes for things the SDK/core already solve or nearly solve:
- getRootElements(): getElements() flattens the tree, so every descendant also
appears as its own top-level entry. buildRoots() already computes true roots
internally; this exposes it directly instead of making consumers re-derive
roots by filtering out descendant ids.
- Export isNewHostBoundary + bareId from @hyperframes/sdk: both already existed
internally (engine/model.ts) but weren't exported, so consumers were
duplicating sub-composition-boundary detection and scoped-id-to-DOM-leaf
conversion by hand.
- Export stripEmbeddedRuntimeScripts + RUNTIME_BOOTSTRAP_ATTR from
@hyperframes/core, and wire serialize({ stripRuntime: true }) on the SDK
session: a proper tokenizing implementation already existed in
compiler/htmlDocument.ts (handles more runtime-script marker variants than a
naive regex), just never exported. The SDK itself imports these via narrow
subpaths (./runtime/start-expression, ./compiler/html-document) rather than
the wide ./compiler barrel, matching the SDK's existing import convention and
avoiding pulling Node-only compiler code (fs/path) into browser bundles.
- Fix getElementTimings(): data-start can be a relative-reference expression
("intro", "intro + 2" — see parseStartExpression's grammar), not just an
absolute number. The old code did a raw parseFloat() on it, which silently
resolved any reference expression to 0. Now resolves references recursively
against the target element's own resolved start + duration, Node-safe (no
live GSAP timeline needed for this case).
14 new tests (session.timings.test.ts, session.subcomp.test.ts). Full sdk
suite: 417/417 passing. Full workspace build (incl. studio) verified clean.
Address PR #2093 review feedback (Miga, Rames D Jusso):
- The walker treated a repeating nested timeline (total > single) as an
opaque interval and never descended into it, so a tl.call() living
inside one would slip past hasTimelineCall detection entirely — the
"any tl.call() disqualifies" claim wasn't quite literal. Now recurses
for detection purposes even when the span is already opaque; the
parent-level interval still dominates for frame-animated-marking, so
this only widens what counts as "has a call()," never narrows the
existing interval coverage.
- Restored the totalDuration() vs duration() rationale comment that got
dropped when the tl.call() detection comment was added above it.
A drag that both moved a clip in time and restacked it fired two writers on
the same file via separate queues (a targeted z-index patch and a full-file
timing overwrite), so the overwrite could clobber the just-persisted z-index
and the restack silently vanished after reload. The move now awaits the
z-index commit before persisting timing, giving the file one ordered writer
per gesture. Adds a regression test that gates the commit and asserts the
timing write waits.
Real bug report: a mono count span driven by a GSAP tl.call() (a counter
going "0 sur 0" -> "1 sur 1" at a later beat) rendered the LATER value
baked in from frame 0 of an EARLIER, unrelated static-hold span, despite
the dedup log reporting "verified".
Root cause: computeStaticFrameSet's tween walker only tracks property
tweens, so a call()-driven textContent mutation carries no tracked
interval and the span around it looks fully static. verifyStaticFramesSafe
does catch genuine drift WITHIN a run it's checking, but a call() is a
one-shot side effect wired as both onComplete and onReverseComplete (GSAP
has no separate "undo" — crossing it in either direction fires the SAME
forward mutation). Verifying a LATER run forward-seeks past the call(),
permanently mutating the live page; an EARLIER run already passed its own
check before that happened, so nothing re-verifies it afterward. Real
capture then starts on the same corrupted page and bakes the wrong value
into the earlier span's reused buffer.
No reliable way to tell a DOM-mutating call() from a harmless one
(analytics ping, class toggle) without executing it, so this disqualifies
the whole comp on ANY call() — conservative, costs some dedup perf on
comps that use call() harmlessly, but correctness over speed.
The ruler now generates ticks across the visible width instead of stopping
at the composition end, so a zoomed-out ruler stays labelled to the right
edge. The tick interval stays pps-driven, so spacing is unchanged. Logic
lives in a new generateVisibleTicks helper to keep Timeline.tsx under the
600-line cap.
When the composition is narrower than the panel (zoomed out) the track
lanes, ruler bar, and context header now fill the whole timeline instead
of stopping at the content width and leaving a flat dark void. Content
stays in composition coordinates, so ticks, clips, and the playhead are
unaffected.
The structural refresh now force-syncs data-active on freshly queried clip
nodes instead of diffing. A clip that changes lanes on a reorder remounts as
a new DOM node; if it stayed under the playhead the diff skipped it and it
rendered as inactive.
Clips share a lane only when they carry the same z-index and don't overlap
in time. Previously a non-overlapping clip packed into the first
time-compatible row regardless of z, so restacking changed the z-index but
the clip's row never followed.
Lane/overlap resolution now uses the clip's authored time span instead
of the live dragged start. A diagonal drag that drifts the clip out of
overlap no longer flips the placement from restack to lane-join, so the
two axes never fight.
CI file-size check and fallow audit caught two issues the --no-verify
commits skipped: Timeline.tsx sat at 602 (max 600) and
postRootDurationToPreview was an unused export (only called in-file).
Trim a comment to hit 600; drop the export.
Extending a clip past the video end used to force the server-fallback
path that fully remounts the preview iframe (the SDK fast path can't
express the root composition's data-duration, and the runtime bakes+drops
data-duration at load so it can't be patched live). On a large comp that
remount is a visible hitch.
Add a runtime control-bridge action set-root-duration -> clock.setDuration,
so the studio can grow the transport length in place. On an extend the
studio now posts it (and patches the clip's own timing live) instead of
reloading; it only reloads when a GSAP source rewrite actually happened
(the gsap-mutation endpoints now report a mutated flag). Non-animated
extends — the common case — commit as fast as a normal edit.
Verified: bridge dispatch + studio no-reload/post-message paths unit-
tested; core/studio/studio-server typecheck + suites green; the built
runtime artifact carries the handler; E2E confirms the extend no longer
remounts the preview and still persists.
Dragging a clip past the video end fed back on itself: the growing preview
length shrank the fit-to-width zoom, which remapped the pointer, moved the
clip, and grew the length again (visible jumping). Split the committed
basis duration (drives zoom) from the displayed effective duration (adds
the live preview for ruler + width). Zoom holds fixed through the gesture
and the extra length scrolls; it re-fits once on drop. Duration math
extracted to pure tested helpers in timelineLayout.
Timeline moves and resizes previously snapped only to the beat grid; they
now also snap the dragged edge to other clips' start/end, the playhead,
and the composition bounds, whichever is nearest within the existing
threshold. Beats fold in as one target kind, so beat snapping is
unchanged. Snap logic lives in a new pure timelineSnapTargets module with
unit tests; a snapped edge shows a guide line (beat highlight, else a thin
accent line) for both drag and resize.
A vertical drag onto a lane whose clips overlap the dragged clip in time
used to resolve to the "nearest valid lane", which included the clip's own
lane — so a small drag snapped straight back and felt like the editor
refused the move. Overlapping clips can't share a row (a row is 1-D in
time), but the drag should still restack, never reject.
Now a conflicting onto-drop converts to an edge insertion adjacent to the
target by drag direction (up -> above, down -> below): the clip lands on
its own lane just in front of / behind the target, with its time
unchanged. Non-overlapping drops still join the lane. Deletes the
nearest-valid-placement search (~90 lines) this replaces.
Dragging a timeline clip (or its right resize edge) past the end of the
video now extends the composition duration on drop, instead of clamping
the clip at the current end. Extend-only and undoable.
- Relax the move/resize horizontal clamps that pinned a clip's end at the
current duration; effectiveDuration now folds in the active drag/resize
preview so the ruler and track width grow live as you drag past the end.
- On drop, extend the root composition data-duration (and the store) when
the clip's new end exceeds it, via a shared extendRootDurationInSource
helper extracted from the block installer (now the single owner of that
logic). An extending edit routes through the server persist path since
the SDK setTiming op can't express the root composition's own duration.
Rework the timeline row model from one-row-per-stacking-layer to NLE
"lanes": a row holds a time-sequence of non-overlapping clips, ordered
top = front. Stacking only matters between clips that overlap in time.
- Lane packing: per stacking context, sort clips by effective z-index
desc (DOM order tiebreak) and greedily pack each onto the first lane
whose members don't overlap it in time; else open a new lane. Lane's
representative z is the max member z. Audio stays one clip per lane.
- Lane-aware drag: dropping onto a lane joins it (clip takes the lane's
z) when there's no time-conflict; a conflicting drop is rejected and
the preview snaps to the nearest valid lane or a new-lane insertion.
Between/above/below still create a new lane at that stacking level.
- Overlap checks use the drag-preview start/duration, so conflict is
judged by where the clip lands, not where it started.
Only explicit drags write z-index; authored z is preserved and
data-track-index is never rewritten.
L4 polish for the stacking-layer timeline:
- Sub-composition layers group under a slim "Inside: <comp>" header with a green
accent, so it reads that restacking there is scoped to that context.
- Audio clips render in a distinct bottom lane: separator border, muted row bg,
music glyph in the gutter, beat strip — clearly not part of the z-order stack.
- During a vertical drag, a drop affordance shows join-vs-insert: an "onto"
target highlights the row (join that layer / same z); "between"/"above"/
"below" shows an insertion line (new layer here).
New pure helper timelineDropIndicator.ts (+ test) maps placement -> indicator;
row-group rendering extracted into TimelineLayerGroupHeader / TimelineLayerGutter
/ TimelineDropInsertionLine / TimelineDragGhost to keep files under the 600 cap.
applyTimelineStackingReorder resolved each z-index change by looking the clip up
in the top-level timelineElements list, but sub-composition children live only
in the expanded list, so the lookup missed them and the reorder silently bailed.
Carry the element's locator (domId/selector/sourceFile) on each z-index change
so the commit resolves the live element directly from the preview DOM. Verified
E2E: dragging a clip inside a sub-composition restacks it and patches the
sub-comp source, while the parent composition is untouched.
Adds timelineEditingHelpers.test.ts (locator-based commit + audio no-op).
applyTimelineStackingReorder resolves the live clip from the preview IFRAME,
then gated it with `element instanceof HTMLElement` against the MAIN window's
constructor. Cross-realm instanceof is always false, so every timeline z-index
commit silently bailed ("element not live in iframe") — the drag resolved the
right z but never wrote it. Use the element's own-realm HTMLElement constructor
(matching timelineDOM.ts). Verified end-to-end: dragging a card down now lowers
its z-index and reorders the row, leaving sibling z-indexes untouched.
Unit tests missed this because the happy-dom test iframe shares a realm; caught
via a real-browser Puppeteer E2E drag.
Re-architect the timeline row model from data-track-index rows to stacking
layers. Rows represent stacking layers per context: explicit-z clips merge onto
one track when they share a z and don't overlap in time; auto-z clips stay one
row each (DOM order); audio is pulled into its own bottom lanes. Rows keyed by a
stable layer id, not data-track-index.
Vertical drag always writes z-index, never track: drop onto a layer joins it
(same z), between layers interpolates a new z, past the ends creates a new
front/back layer. data-track-index is never rewritten; #958 holds. Adds
hasExplicitZIndex capture (computed z != auto).
L1: hasExplicitZIndex on the element model
L2: buildStackingTimelineLayers (layer-based rows)
L3: layer-aware vertical drag (join / interpolate / new-extreme)
createTimelineElementFromManifestClip used `clip.zIndex ?? computed`, but the
runtime reports inline-only z-index (0 for CSS-rule authored z-index), and
`0 ?? x` keeps the 0 — so every CSS-styled clip collapsed to a z=0 tie. The
timeline then ordered rows by DOM position instead of true stacking, and the
first vertical drag renumbered the whole tie group, clobbering the author's
z-index. Prefer the effective computed read from the live element (the same
read the reorder commit uses); fall back to the runtime value only when the
element isn't live.
The element stacking key (element.key ?? id) was recomputed in four places
(reorder-intent generation, row ordering, the commit-time sibling lookup via a
threaded keyOf param, and resolveTimelineMove). Any drift would silently break
the sibling lookup and no-op the reorder. Route all of them through the existing
getTimelineElementIdentity owner, share one toStackingOrderItem mapper between
row ordering and reorder intent, and drop the keyOf parameter.
Also enforce the audio side-effect invariant in the single mutation owner
(applyTimelineStackingReorder): dragging an audio clip has no visual layer to
restack, so it never writes z-index. Covered by a new hook test.
Timeline rows now order by scoped stacking (z-index per stacking context)
instead of data-track-index, and dragging a clip up/down commits a targeted
z-index change through the same shared path the layers panel uses. Both panels
stay consistent and moving a clip actually changes front/back. data-track-index
is demoted to time-overlap layout only; no bulk z-index injection (#958 intact).
Also restores beat-snapping on keyframe retiming (re-wires snapKeyframePctToBeat,
orphaned when keyframe dragging was removed) which surfaced while unifying the
model. Extracts pure track-ordering logic to timelineTrackOrder.ts, the stacking
reorder commit + deleteSelectedKeyframes to timelineEditingHelpers.ts, to keep
StudioApp / the timeline hook / Timeline under the studio 600-LOC cap.
U3: scoped stacking row order in the timeline
U4: vertical drag commits z-index via the shared reorder commit
U8: restore keyframe beat-snap on retime
The storyboard view's empty state was a single "no storyboard yet" line. Add a
copy-to-clipboard prompt box (a ready-to-paste handoff prompt carrying the
canonical STORYBOARD.md frontmatter + per-frame format) so users can hand it to
their coding agent, plus a faded skeleton of the contact-sheet grid so landing
on an empty board previews what a filled one looks like instead of a dead end.
Crop is now part of the element selection instead of a separate mode. Selecting
a croppable element shows edge handles just outside each side and, once cropped,
the full content with the cropped-away area dimmed plus a center reposition
handle to pan the crop window. Dragging the body moves the element, edge handles
crop, the center handle pans; corners stay free for the resize handle. Removes
the crop-mode toggle (toolbar + property-panel buttons), the cropMode/
cropAvailable player-store state, and the double-click-to-crop gesture. The
clip-path inset model is unchanged.
Sub-composition <head> styles targeting html/body/:root (width/height/
overflow/background) were injected into the parent document unscoped by both
the Studio runtime mount (compositionLoader) and the render-time inliner
(inlineSubCompositions/htmlBundler). scopeCssToComposition deliberately passed
html/body/:root through unchanged, so a sub-composition smaller than the root
clobbered the host <body> dimensions and its overflow:hidden clipped the
composite to the last sub-comp's size. Only the top-left element painted;
everything else (and framework-owned video positioned outside that box) was
clipped away.
Add a scopeRootSelectors option to scopeCssToComposition that remaps
html/body/:root to the composition's own box, and enable it everywhere
sub-composition styles are scoped. The universal selector stays untouched.
Top-level composition scoping is unchanged (it legitimately owns the document).
Covered by new compositionScoping tests.
* feat(media-use): color grading — grade/lut resolve, smart-grade, grade-compare CLI
Add color grading to media-use as first-class resolve types plus a faithful
comparison command. All local, offline, deterministic — no model, no GPU.
- resolve -t grade / -t lut: produce a data-color-grading block (or a frozen
.cube). Look cascade: core preset (no file) -> bundled .cube library ->
parametric buildCube. Emitted .cube is Rec.709 and validated against core's
colorLuts constraints (LUT_3D_SIZE <= 64) before it is frozen.
- smart grade (grade --for <media>): ffmpeg signalstats -> adjust suggestion
(exposure / contrast / white balance), surfaced with the measured evidence on
stderr as a starting point; never auto-applied.
- hyperframes grade-compare: renders N candidate grades onto a reference frame
through the real runtime shader into one labeled comparison PNG, so an agent
picks a look without opening Studio. Prepends an "original" baseline cell by
default (--no-baseline to omit). Shares the headless-capture pipeline with
snapshot via capture/captureCompositionFrame.
- media-use SKILL: proactive "media opportunity pass" guidance (grounded
signal -> offer, ask once, surface don't mutate).
Verified: media-use 116/116, grade-compare 7/7, snapshot 9/9, lint + format
clean, full build green, comparison renders end to end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* test(cli): narrow grade-compare baseline assertion off unknown-typed grading
Assert the whole cell via toEqual instead of reaching into .grading.preset /
.grading.lut on the unknown-typed field, keeping the test typecheck-clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* feat(media-use): agent-authored LUTs via --params + validate --from cube; never-read-.cube guardrail
- resolve -t lut / -t grade --params '<json>': build a parametric .cube from
explicit params (bypassing the intent cascade), validate, and freeze in one
step. --intent becomes the optional description. Lets an agent commit a look
it computed itself.
- --from <file.cube> now validates the ingested LUT for lut/grade types and
rejects an invalid/oversized cube (no partial write) — the escape hatch for a
LUT the agent generated with its own code.
- SKILL.md: hard rule to never read a .cube body into context (~size^3 lines,
zero legible signal) — inspect via grade-compare (see it) or cube-validate
(ok/size), read the manifest description for meaning; plus both authoring
paths and the parametric-vs-film-stock ceiling note.
Verified: media-use 116/116, lint + format clean; smokes — --params builds a
valid frozen cube, grade --params returns a lut block, bad JSON and an oversized
--from cube are both rejected with no stray file.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* fix(cli): grade-compare validates referenced LUTs, warns on no-op cells, caps candidates
Bug-bash follow-ups — grade-compare silently accepted bad input:
- Validate LUT *content*, not just existence: each referenced .cube is parsed
with core's parseCubeLut (now exported from @hyperframes/core) and rejected
with a per-cell error ("LUT for \"<label>\" is not a valid .cube: ..."). A
file that exists but isn't a valid cube no longer renders a silent no-op cell.
- Warn on inactive cells: a grading that normalizes to inactive (e.g. a
malformed {lut:12345}) emits a stderr warning naming the cell; the
auto-prepended "original" baseline is intentionally inactive and stays silent.
stdout remains valid JSON.
- Cap candidates at 16 (excluding baseline): over-cap input renders the first N
and reports {truncated:true, total:M} on stdout + a stderr note — no silent
drop, no unbounded giant sheet.
Verified: grade-compare 10/10; non-cube LUT → clear error; {lut:12345} → warning
+ ok; 20 cells → cells=17 truncated total=20; valid runs unchanged. Lint/format
clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* feat(cli): general `hyperframes compare` visual-variant primitive
Generalize grade-compare's "render N variants → one labeled sheet → the agent
looks and picks" loop into a standalone command that works on ANY variation
(font, layout, motion, grade, whole compositions) — the tool never needs to
know what differs.
- `hyperframes compare <path...> [--at <sec>] [--labels a,b,c] [--out] [--cols]
[--json]`: renders each agent-authored composition variant through the real
runtime (captureCompositionFrame) and stitches one labeled comparison sheet +
JSON ({ok, sheet, rendered, variants, truncated?/total?}). 2+ paths required;
caps at 16 with loud truncation. It presents, it does not judge — choosing is
the caller's job.
- Factored the shared "render a labeled set → contact sheet" path so compare,
grade-compare, and snapshot all sit on it (no duplication). grade-compare is
now the first color-specific specialization of this primitive.
- New pathArgs util + contactSheet test; hyperframes-cli SKILL documents compare
as the agent's "see your own renders and choose" primitive.
Verified: 26/26 across compare + grade-compare + snapshot + contactSheet (no
regressions); compare renders 3 variants into one visibly-distinct labeled
sheet; 2+-path error path clean; lint/format clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* fix(ci): green the skills CI — skip ffmpeg tests when absent, oxfmt markdown
The "Test: skills" CI job runs bare `node --test` with no ffmpeg on PATH (by
design — skills tests are meant to be node-builtin-only). The grade-analyzer +
smart-grade tests shell to ffmpeg and were failing there with ENOENT. Guard
them to skip when ffmpeg isn't on PATH; they still run locally / where it is.
Also oxfmt README.md + hyperframes/media-use SKILL.md (the whole-repo
`oxfmt --check .` Format job caught markdown left unformatted by the rebase
conflict resolution).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* fix(ci): skip core-conformance test when tsx is unavailable
The "Test: skills" CI job installs no deps, so the normalizeHfColorGrading
conformance test (which imports core's TS via `node --import tsx`) failed there.
Guard it to skip when tsx can't resolve; runs locally / in the deps-installed
Test job. Completes the skills-CI greening (the ffmpeg guards handled the rest).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* fix(cli): escape grade-compare src double-quotes (CodeQL XSS) + Windows-safe compare test
- grade-compare built `<img src="...">` (double-quoted) with the single-quote
escaper, leaving `"` unescaped — a `"` in the frame path could break out
(CodeQL: incomplete HTML attribute sanitization). Use escapeXml for src.
- compare label test hard-coded POSIX paths that can't match on Windows; assert
the derived labels (the subject); path resolution is covered elsewhere.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* refactor(media-use): generate LUT library from params (drop committed .cube files)
The 3 bundled .cube files were 733 lines each (2,199 total) and were themselves
buildCube output — pure repo bloat. Replace with compact per-look params in
luts/index.json, generated on resolve; add an optional `url` for future scanned
LUTs to be CDN-hosted + downloaded on demand (freezeUrl) instead of committed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* feat(media-use): serve library LUTs from CDN on-demand (static.heygen.ai/luts), params fallback
Looks now carry a CDN `url` (hosted at s3://heygen-public/luts → static.heygen.ai/luts/<id>.cube);
resolve downloads + validates + freezes on demand, like bgm/image. `params` stays
as the deterministic offline fallback (--local-only, or if the download fails), so
resolution is never blocked on the network. Provider prefers url, falls back to params.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
* fix(media-use): address #2041 review — atomic LUT writes, compare telemetry, follow-ups
- Atomic .cube writes: library provider (url + params) and the parametric
generator now write to a .tmp path, validate, then rename, so a crash can
never orphan an invalid .cube at the final path (was validate-after-write).
- track("media_use_resolve") now emits provenance.via (url/params-fallback/params).
- grade-compare + compare: --timeout flag (was hardcoded 5000) and a
media_use_compare event (cells, truncated, total, render_ready_timed_out);
openSettledCompositionPage now surfaces the render-ready timeout.
- compare staging skips node_modules/.git; --for gets an upfront existence check.
- Rec.709 luma comment; HYPERFRAMES_ANALYZE_TIMEOUT_MS override; measured note
uses basename; LUT s3 hosting moved from index.json into luts/README.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`hyperframes transcribe --model large-v3` aborted with "unknown DTW preset
'large-v3'". whisper.cpp's --dtw flag wants a dotted alignment-heads preset
(large.v3), but we passed the hyphenated ggml file stem (large-v3). They
coincide for tiny/base/small/medium(+.en) — why it slipped through — but
diverge for the large-v* family. Map stem -> preset (- to .) so large-v1/v2/v3
(and large-v3-turbo) work; no-op for the others. Also fixes media-use, which
shells to `hyperframes transcribe`.
Claude-Session: https://claude.ai/code/session_01H5k87mPZ4d6yiFwcWSb8Vv
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Address two max-effort code-review findings on PR #2056 not covered by
the earlier review-gap commit:
- captureFrameToBufferPipelined's static-dedup reuse branch never
advanced session.lastEncodeResultFrame, unlike its sibling real-capture
branches. The gap-check window is computed from that watermark, so
every consecutive reuse in a static run rescanned an ever-widening
window instead of just the newest frame — O(n^2) total work over a
long static stretch instead of O(n).
- The "single-threaded, no race" justification on the shared
parallelGuard closure was wrong: the guard has real internal await
points (recapture, PSNR) between reading and writing its
sizes/absFloor/acceptedSmall state, so concurrent workers' calls do
interleave there (confirmed). Replaced with the actual reason it's
safe: absFloor only ratchets down, sizes is append-only and
order-independent for the median, and acceptedSmall's fast path
re-validates by exact byte-equality regardless of which worker wrote
the reference buffer.
Address PR #2056 review feedback:
- Fix totalFrames progress inflation for interleaved tasks — divide
each task's span by its frameStride to match the actual per-worker
frame count (captureFrameRange steps by stride), instead of summing
raw endFrame-startFrame which double(N)-counts interleaved tasks.
- Attach a no-op .catch to each frame's pipelined encodeResult at kick
time so an abandoned promise (loop exits early on abort/error before
draining it) can't surface as an unhandled rejection during teardown.
- Document why the pipelined branch's stride=1 path is validation-only
in production (HF_DE_PARALLEL_STREAM always uses interleaved
distribution) so a future refactor doesn't unknowingly widen it.
- Comment the intentional single shared parallelGuard/parallelStats
across workers (safe single-threaded, better rolling-median signal).
Step 2 of the DE engagement plan: multi-worker drawElement capture through
the streaming encoder, with the full runtime self-verification net riding
along — the confinement rule that kept the parallel clamp in place is now
satisfied on this path. Opt-in via HF_DE_PARALLEL_STREAM=true; default
routing (including the #2026 single-worker inversion) is unchanged.
Mechanism:
- distributeFramesInterleaved + WorkerTask.frameStride: worker i captures
frames i, i+N, i+2N... — seek-based capture makes stride free and the
ordered writer's reorder window shrinks from totalFrames/N to N (contiguous
chunks serialize workers behind the writer).
- Depth-2 pipelined worker-encode produce in the parallel worker loop (the
same shape as the sequential loop; frame k's in-page encode overlaps
k+stride's produce). HF_DE_PAR_DEBUG=1 traces the first frames per worker.
- Drain guard extracted to createDrainFrameGuard (session-parameterized):
every parallel frame gets the SAME blank-guard + PSNR self-verify as the
sequential drain, against its owning worker's pre-injection ground truth
(all sessions arm identical sample indices from
CaptureOptions.compositionDurationSeconds).
- FrameReorderBuffer.abort(err): a failed worker (e.g. verification error)
rejects all parked and future waiters — without this, peers park forever
in waitForFrame and the pool (which awaits ALL workers before surfacing
errors) deadlocks. Found by the verify-trip test; unit-tested.
- The typed DrawElementVerificationError is preserved past the pool's
error-string flattening so the orchestrator's verify-retry recognizes it.
- Static-dedup stride hazard fixed: lastEncodeResult reuse now requires EVERY
frame in (lastEncodeResultFrame, i] to be predicted-static (sequential
capture reduces to the old has(i) check).
- Workers get separate browser PROCESSES under the flag: pages co-tenant in
one browser starve non-active pages of BeginFrames on the paint-wait path
(measured 86s vs 30s on a 3,245-frame rAF comp).
Validation:
- Happy path W3: verify samples pass across workers (4x inf on the 2,381f
comp), output vs single-worker DE = 59.3dB (encode noise floor) — the
interleave + dedup-stride produce identical pixels.
- Verify-trip (marginal comp + HF_DE_VERIFY_MIN_DB=45): fails at frame 649
(32.2dB < 45), peers abort instead of deadlocking, whole render retries
via parallel screenshot, RENDER_OK in 42.6s.
- Canary suite 7/7 with the flag off (default paths untouched); producer
orchestrator tests 99/99; engine suite 909 passed (14 pre-existing main
failures, stash-A/B verified); reorder-buffer abort unit tests.
Perf note: capture-only parallel speedup measured 1.38x (W2) / 1.52x (W3)
over single-worker DE in the spike; end-to-end numbers on this machine are
currently noisy (separate-browser init overhead + bench load) — clean
benchmarks before any default routing change. The flag stays explicit
opt-in; promoting it into the router replaces the #2026 W=1 pin for the
same cohort.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Address the highest-severity max-effort code-review finding on the now-
merged #2082 (the drawElement Chrome-version-pin fix): findFromHyperframesCache
matched a cached Chrome by browser type only, never comparing its
buildId against CHROME_VERSION. Any machine that already rendered with
an older hyperframes version has an old build (this pin has moved
131 -> 151 -> 152 across releases) sitting in ~/.cache/hyperframes/chrome,
which satisfied the lookup and silently defeated the whole point of
#2082's version bump for exactly the population it was meant to fix —
drawElement's new capability probe would then permanently and silently
fall back to screenshot capture instead of ever fetching a build that
implements canvas.drawElementImage.
Verified directly (not just via review): seeded ~/.cache/hyperframes/chrome
with the old 131 build, confirmed a real render previously kept using it
forever; with this fix it's correctly ignored and 152 is downloaded.
New regression test locks in the buildId mismatch case.
Address max-effort code-review finding on PR #2045 (confirmed, not
addressed by the earlier review-gap commit): the script_failure bail
path skipped the composition-id enumeration entirely, so a render with
multiple sub-compositions sharing a failed script only logged the raw
failed URL(s), never which composition(s) were still waiting on it —
a real observability regression versus the pre-#2045 behavior, which
always logged the missing-id list on any non-ready outcome.
Now enumerate unregistered composition ids unconditionally and log them
alongside whichever reason (script_failure or natural timeout) fired.
