* fix(skills): reverse hands billing commands to the person instead of running them
The Snyk audit on skills.sh still fails reverse-any-video on E004 because the skill has the agent run
`billing subscribe` and `billing add-credits` itself. media-use, which hands the person the billing command
instead, passes the same audit.
reverse now does the same: on Free it gives the upgrade link and the subscribe command for the person to run,
and mid-run it gives one add-credits command for the run's total. It waits until they say they're done, then
starts again or relaunches the refused steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HebQXQjEiSeVmhvTbJvSAw
* test(skills): pin reverse's billing handoff per payment section
start.md's opening said the agent runs every command, which contradicts the billing sections below it.
It now excludes `billing` and says the person opens and pays on the upgrade and add-on credits pages.
The test now checks the handoff itself. Both payment sections must give the person the command and
forbid running billing. Neither may tell the agent to run, execute or launch a billing command, and
SKILL.md must keep its prohibition. Each of those mutations, and main's start.md, fails the test.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HebQXQjEiSeVmhvTbJvSAw
* fix(skills): reverse hands over one add-credits for the run when an amount is unknown
"If a refusal prints no amount, give plain `billing add-credits` for that one" let the agent hand over a
purchase for each step with no amount, on top of the aggregate one. Now, if any refusal prints no amount, the
total is unknown and the person gets one plain add-credits for the whole run, never one per step.
The test now requires that sentence and rejects a per-step fallback in the out-of-credits section, even when
the aggregate sentence is still there.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HebQXQjEiSeVmhvTbJvSAw
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* test(engine): blur browser tests fail on CI when their browser cannot run
On CI a missing Chrome, a missing core runtime build or a Chrome without drawElementImage
fails the suite's beforeAll instead of skipping every case. Off CI the suites still skip.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): every inside-the-trail case proves blur drew, not only held-card
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): the liveness test watches held-card through its slide-out
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): stateful-callback joins the multi-worker blur render, split across five suites
The render differential's suite moves into __test_utils__/blurRender.ts and five files run it
over one or two fixtures each, so none nears the flake gate's 120 s cap. Every fixture that
moves must now blur more than insideMaxMae, not only held-card, and CI fails instead of
skipping when ffmpeg is missing.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): each blur render file imports its runner so the test classification reads it
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(core): video starts a second early and catches up hidden, so the picture meets the cut
* fix(core): a video that plays from the start of its file also starts early
* fix(core): a from-file-start clip leads by half the lead hold so a fast start is not paused
* fix(core): steering for hidden video is its own function; trim comments and format
* chore(player): raise the packed size budget for the picture pre-roll
* chore: release v0.9.17
* chore(producer): raise the unpacked size budget for windowed video frames
The new modules and this version bump put the unpacked package 49 bytes
past the 1% tolerance. No tests or source maps are in the tarball.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(producer): render long videos with windowed source frames, behind HF_VIDEO_FRAME_WINDOW
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(producer): windowed calibration sees real frames, and the late full extraction reports itself
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(producer): windowed frames match a full extraction, frame for frame
Renders ten compositions twice on the segmented route, with and without
HF_VIDEO_FRAME_WINDOW, and compares FFmpeg framemd5 of the two outputs: a
mid-clip start, NTSC 24000/1001, a loop crossing a slice edge, a playback-rate
ramp, a held tail, two clips of one file, a VFR source beside a CFR one, two
workers, auto workers (capture calibration runs), and motion blur, which
declines the window. Each pair must also report the same capture mode.
A separate test samples segment-frames during a four-segment, two-worker
render and holds its peak under workers x segment frames x the largest frame
file x 1.2, and checks the folder is empty once the render closes.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): windowed frames test finds its own work dir and asserts what it claims
Find the render's work dir with resolveRenderWorkDirPrefix and the owner record's pid, so
the peak test works on Linux and never samples or deletes a concurrent render's dir.
Assert the measured worker count, calibration, and the exact up-front frame count of the
VFR video. Retry a dead Chrome only with HF_TEST_RETRY_BROWSER_STALL=1, and fail a case
whose windowed render alone needed it.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(producer): windowed segments honour keep-segments and cleanup never fails a render
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(producer): trim the comments the windowed frames route added
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(producer): the VFR-beside-CFR extraction test moves to the integration lane
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(producer): raise the package size budget for windowed video frames
The new segment frame modules add about 9 KB to the packed producer package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(producer): windowed frames header says what the real runs measured
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): every published skill reports its own run
`hyperframes events --skill=<name>` exists so a skill can say it ran, but
only reverse-any-video and figma asked the agent to call it, so the other
skills' use was invisible. Each of the nine published skills now opens with
a one-line usage beacon naming itself. The line says what it sends (the
skill's name only) and that the telemetry opt-out applies. The router's
`init` line also passes `--skill=hyperframes`, so renders of new projects
carry the skill.
check-core-skills fails a core skill that lacks its own beacon; each core
skill's size budget grows by 192 B for the line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): checkout links name the CLI and the skill that sent them
Only `billing subscribe` marked its pricing link `from=cli`, and the
terminal's Press-Enter Pro checkout sent no return page, so a purchase made
from the terminal looked like one made on the site. Now:
- `pageUrl` adds `utm_source=cli` and `skill`: the one named, else the
project's own `authoringSkill` from hyperframes.json.
- `billing` takes `--skill`. The terminal Pro checkout returns to the plans
page marked `from=cli&skill=...`, and retries without the return page if
the server does not allow it.
- Pricing links the server hands back (reverse's upgrade link, a refused
media call's link) gain `from=cli` and the skill when they lack them.
- The Pro gates name their skill: reverse -> reverse-any-video, templates ->
style-sampling, paid media calls -> media-use, in the CLI's printed fix and
in those skills' billing commands.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): report a refused checkout return page, and say where --skill goes
When the server refuses the terminal checkout's return page, the CLI still
sells Pro without it, but that purchase loses its CLI origin. A
cli_billing_checkout_return_refused event now makes that visible.
The --skill help said the skill is recorded with the purchase. It is sent
with the pricing link; the purchase itself does not store it yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): every published skill reports its own run
`hyperframes events --skill=<name>` exists so a skill can say it ran, but
only reverse-any-video and figma asked the agent to call it, so the other
skills' use was invisible. Each of the nine published skills now opens with
a one-line usage beacon naming itself. The line says what it sends (the
skill's name only) and that the telemetry opt-out applies. The router's
`init` line also passes `--skill=hyperframes`, so renders of new projects
carry the skill.
check-core-skills fails a core skill that lacks its own beacon; each core
skill's size budget grows by 192 B for the line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(skills): say the usage beacon goes out with the CLI's usual telemetry
The beacon line said it sends only the skill's name. The events command uses
the CLI's normal telemetry, which also carries machine details, the install
id and, when signed in, the account. The line now says so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(producer): a wedged effect scan or screenshot-session open trips the 60 s stall watchdog
The sequential streaming watchdog raced only the capture call. Session init
(where the fast-capture effect scan runs) and opening the screenshot session
for effect frames waited on CDP's protocol timeout instead, which scales with
output area up to 30 minutes, so a starved page hung a render for ~25 minutes
before the watchdog saw the next frame. The effect scan now gives up when a
chunk outlasts its 120 s budget, opening the effect-frame session gets its own
no-progress window, and session init gets a 300 s one; each trip retries on a
fresh session like any other stall.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): one local render at a time per machine; a second waits and says for whom
Renders started side by side from different sessions starve each other until
frames stop landing. `hyperframes render` now takes a machine-wide slot (a lock
file in the HyperFrames cache directory, held by pid) after its preflight and
holds it for the render or the whole batch. A second render waits and prints
the pid and project it is waiting behind; `--no-wait` fails at once instead.
A holder whose process is gone is taken over, and HYPERFRAMES_RENDER_LOCK=off
restores side-by-side renders. `--workers` is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(lint): flag fonts a cloud render cannot resolve, including the font shorthand
Lint read only `font-family`, so `font: 500 16px "PingFang SC", sans-serif`
passed while the renderer, which reads the shorthand too, drew the installed
Mac font locally and the cloud refused it. The shorthand now counts for
font_family_without_font_face (its parser moves to @hyperframes/parsers so
core and lint share one), and a new font_face_local_only warning names an
@font-face whose only sources are local(): it renders here and falls back in
the cloud. The rule's hint and the format reference stop recommending local()
for OS fonts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): a cloud render that refuses a font says which project files use it
The cloud's "fonts that could not be resolved: PingFang SC" named no
composition. For a render submitted from a local project, the failure now
lists each refused family with the project files that name it, and says how
to ship the font so the cloud can use it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): keep comment shares and exports within the gates, and split the single-render notes out of executeHeldRender
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(parsers): read a spaced line-height slash and var() before the size in the font shorthand
`font: 16px / 1.5 Inter` was read as the family `/ 1.5 Inter`, and
`font: var(--weight) 16px Inter` as `16px Inter`, because a var() counted as
the size. The parser now follows the shorthand grammar: optional style,
variant, weight and width keywords (or var()), then a length, percentage,
keyword or calc() size, then a `/ line-height` in any spacing. When no size can
be told apart from a var(), it says so: lint skips that declaration instead of
reporting a missing font, and the renderer keeps its best guess.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): say the render slot is one per user account, not per machine
The lock lives in the account's HyperFrames cache folder, so renders by two
user accounts never queued behind each other while the messages, flag help and
docs said "this machine". They now say "per user account" and that
HYPERFRAMES_CACHE_DIR set to a shared folder shares the slot across accounts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): reclaim a render lock or takeover file left half-written
A lock or `.break` file that exists but cannot be parsed (a crash mid-write
where hard links are unavailable) was never reclaimed: a waiting render waited
forever and --no-wait refused forever. Once such a file is older than any real
write (5 s by mtime) it counts as abandoned: the lock goes through the same
`.break` takeover as a dead holder's, and a takeover file is removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(producer): opening the capture session counts against the setup stall budget
The sequential streaming attempt opened its own capture session (Chrome launch
and new page) before the setup watchdog and outside the attempt's cleanup, so a
wedged launch still waited for the protocol timeout. Opening now shares the
300 s setup window with session initialisation, a trip fails the attempt like
any other sequential stall, and a session that finishes opening after its
attempt gave up is closed instead of leaking a browser.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(parsers): keep the shorthand result type module-private
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): reclaim an unreadable render lock only when its writer is provably gone
Where hard links are unavailable the lock is created by an exclusive open and
then written, so a contender paused between the two leaves an unreadable lock
while it is still alive. The 5 s mtime grace took such a lock over, and both
processes then held the slot. Each contender now writes an owner marker under
a name only it can take (pid plus random suffix, holding its process identity)
before it creates the lock, keeps it while it holds the slot, and removes it
otherwise. An unreadable lock or `.break` is reclaimed only when no other
contender's marker belongs to a live process; markers of dead processes, or of
a reused pid, are deleted. No reclaim is decided by time alone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor: split acquireRenderLock and the streaming-stage session stub under the complexity gate
acquireRenderLock hands contender setup and a single marked claim to their own
functions, and the test stub for createCaptureSession splits its held-open
cases from the fake session it builds, so neither exceeds Fallow's cognitive
or CRAP thresholds. Behaviour is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(ci): raise the CLI and engine size budgets for the watchdog, lock and font-lint additions
Main's CLI unpacked size was already ~0.99% over its budget; the render lock and
owner markers push it past the 1% tolerance. The CLI unpacked budget takes the same
value as #7169 and #7170 (25,159,107) so the PRs merge cleanly. The engine's packed
size grows by about 4.7 KB for the effect-scan and session-open watchdogs; its packed
budget moves to 5% above the measured 467,097, as the budget file's headroom rule sets it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(parsers): a var() size glued to its line height is the shorthand's size
`font: 700 var(--size)/1.1 "Inter"` read as no family at all: the token was
tested whole against var(), and the slash after it hid that it was the size,
so the renderer injected no face for Inter. A var() followed by `/ line-height`
(glued or spaced) can only be the size, and is now read as one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(parsers): split the shorthand size search out of parseFontShorthand
Keeps parseFontShorthand under Fallow's cognitive-complexity gate; behaviour is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): render lock never removes another render's takeover, knows a reused pid, works in a shared folder and on Windows
- A `.break` takeover file is removed only by the render that wrote it. One
whose writer is gone stops the render with the file to delete, instead of
being removed on an earlier reading, which could remove a live `.break`
written since and let two renders hold the slot. The takeover re-reads the
lock under its own `.break` right before replacing it.
- The lock record keeps its holder's process start time, so a pid that now
belongs to another process (an OOM-killed render in a container with a
persisted cache) no longer holds the slot forever.
- The `locks` folder is created writable for every account, so accounts
sharing HYPERFRAMES_CACHE_DIR take turns instead of the second failing with
EACCES; a folder this account cannot write in says how to fix it, and
another account's file that cannot be removed is left, not fatal.
- Windows: no PowerShell start-time lookup per check (pid liveness only), a
file another handle holds (EPERM/EBUSY/EACCES) reads as held, a lock that
vanished is retried once instead of in a tight loop, and a wait stops when
its signal aborts. The tests end every acquire they start before removing
their folder, and the paused contender writes through its open descriptor.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(ci): raise the lint packed size budget for the font-shorthand rule
With main's audio-attribute rule (#7169) merged in, the font-shorthand and
local()-only font findings take @hyperframes/lint's packed tarball to
181925 bytes, over the 179338 budget. The new budget is that measurement
plus the usual 5% headroom.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* docs(catalog): every card on screen plays a small looping video instead of a live player
* docs(catalog): the preview generator also writes each card's small loop and poster
* fix(catalog): end ffprobe options before the card file, as the argv contract requires
* fix(catalog): a pause control for the card loops; none play in a hidden tab or under 25% visible
* test(engine): pixel invariants for capture blur against blur off
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): fixtures for every capture blur failure class seen so far
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): nested canvas fixture reproduces #6975, sharp-tile fixtures have no moving times
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): capture blur must match blur off outside the trail and the shutter average inside it
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): differential tests pin known bugs by symptom, history test checks the reference frame
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): a playing preview keeps changing with blur on
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): a multi-worker blur render must match blur off outside the trail
Renders five blur fixtures through the real producer as a PNG sequence with 4
workers (contiguous blocks, three of them starting mid-film, the #6967
condition): blur off twice, blur on once. Outside the trail of everything that
moved, blur on must equal blur off in every frame with two neighbours. The
callbacks and scroll-list known differences are pinned by symptom.
Import route: a relative import of blurDifferential.ts, fixtures.ts and
thresholds.ts from packages/engine/src/services/__fixtures__. The producer
tsconfig excludes test files, so I type-checked the file with a temporary
tsconfig (rootDir widened); no error. No package subpath was needed.
RENDER_OUTSIDE_MAX_MAE: render noise is 0 (blur off, twice, bit-identical), the
rewind mutation measured 111.085 when caught, so the 1.0 floor applies.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): render callbacks with workers that start between its calls
With 4 workers the blocks started at 0, 15, 30, 45, each on a call, so the
rewind mutation showed in a render only 3 times in 5. Seven workers over 60
frames start at 0, 9, 18, 27, 36, 45 and 54, two of them strictly between the
calls at frames 16 and 30; the mutation then fails callbacks in 8 of 8 runs
(worst MAE 97.02), the clean build in none.
A new case checks the partition with distributeFrames. The #list and #label
rects now come from the fixture's markup, and MEASURED_RENDER_REWIND_MUTATION
is 97.02.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): blur pins bound the symptom, liveness detector reports blindness, fixtures time out
Pinned known bugs now require every bad tile inside the pinned region and a worst
MAE within half to double of the measured one. tileTouches, Rect and the band check
move into blurDifferential.ts with a dpr 2 unit test. The blank-frame census counts
canvases it cannot read and asserts none, and names both readings when no blank frame
is seen. No console.info on passing runs. Fixtures cases get a 30 s timeout.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): scroll-list render case bounds the race, pins share the engine helpers
The scroll-list loss depends on timing per worker block, so the case now requires
frames 2..27 clean and any later bad tile inside #list, without requiring the loss.
The callbacks pin keeps requiring its symptom, bounds it to the label and to half to
double of its measured size, and its message names the entry to delete. Wording says
multi-worker, not 4-worker.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* build(engine): keep the blur test helpers out of the published tarball
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): print the worst outside tile of each arm to explain the CI text diff
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test: both arms draw grayscale text so host subpixel settings are not read as blur
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(producer): run the multi-worker blur render alone so it cannot skew other Chrome tests
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test: re-pin the producer lane runner after splitting out the heavy blur render
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): read the blank-frame pin from a wrap observer, not an animation frame poll
The pin sampled in requestAnimationFrame, so it could miss the gap between the
wrap and the first draw. A MutationObserver now reads the output canvas the
moment the root is wrapped, which is deterministic, and every fixture must
report a wrapped root.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* test(engine): drop the blurs field no fixture could set to false
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* ci: give the flake gate 60 minutes, it budgets runs by file count not time
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* fix(lint): flag a data-automation or data-fx-chain the render cannot read
A cue whose automation lane has no "version": 1 passed `hyperframes check`
and only failed at render: locally with "Unsupported automation version:
undefined", in the cloud with no cause at all. The new error
`audio_attribute_unreadable` reads each data-automation and data-fx-chain
envelope (JSON object, version 1, a lanes or nodes array, a target and
points per lane, a type per node) and names the element.
Lint cannot import the renderer's parsers from @hyperframes/core, which
depends on it, so core gains a parity test that fails if the two disagree
on any envelope case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(lint): an empty data-automation or data-fx-chain is ignored, as the mixer does
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(lint): split the audio attribute shape check and raise the CLI size budget
audioAttributeShapeError splits its envelope and per-entry checks to stay under
the cognitive-complexity gate, and media.ts drops the added rule label comment.
The CLI's unpacked size was already 0.99% over its budget on main (23,958,7xx of
23,722,874); this rule adds about 2.4 KB of bundled code, which crosses the 1%
tolerance. The budget moves to 5% above the measured 23,961,054, as the budget
file's own headroom rule sets it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* style(lint): format the split audio attribute checks
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(ci): raise the lint package's unpacked size budget for the audio attribute rule
The rule and its envelope checks take @hyperframes/lint's unpacked size to 702,230,
just over 1% past its 695,173 budget. The budget moves to 5% above the measured size,
as the budget file's headroom rule sets it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A third-party skill-trust grader (DeltaVerdict D5) fails any skill text
that has the agent modify its own skills, even when disclosed or gated on
the user. Reword the skill lifecycle, plugin, CLI and scaffold text so it
states facts the agent can relay (skills are versioned with the CLI; the
person installs and updates them, and which command does it) instead of
commands for the agent to run. Every agent-run `init` example passes
--skip-skills, since init otherwise installs missing core skills; the
plugin header, plugin-cli and sfx messages are person-addressed; the
person's install commands are marked as theirs; /style-sampling joins
the on-demand lists. Regenerate skills-manifest.json and update the
pinned reverse init regex.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): surface outdated CLI and skills to agents; add hyperframes update
Single commit of #7094 on top of main (no merge commits through the signed
GraphQL path). Includes: 2h update/skills check interval and stale-cache
recheck; `_meta.checkedAt`, `_meta.skillsOutdated`, `_meta.skillsMissing`;
`_meta.updateNotice`, one sentence for the agent to relay to the person, at
most once a day per project (brew/npx-aware command, omitted under CI, the
update-check opt-out and agent plugins); `hyperframes update [--check]
[--yes] [--json]` (CLI then skills at the installed CLI's tag; brew
version read back); doctor names the running binary and its channel; the
hyperframes-cli skill gets one line to pass the notice on, naming no
command. Keeps main's install-target skills check (#7120) in updateSkills.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): update notice names a command that moves this install
- `_meta.updateNotice` (and `update --check`'s skills command) now name a
command that moves the running install: a global npm/pnpm/bun/brew CLI
runs its own `hyperframes update` (or just its manager's command when
only the CLI is behind), since `npx hyperframes` would run a project's
copy or a fresh npx one; a project-local CLI moves with its project's
manager by lockfile (`npm install`/`pnpm add`/`bun add`/`yarn add
hyperframes@<v>`), which `update --yes` now names too instead of
`npx hyperframes@latest`.
- `update`'s confirmation hint states a fact and names no step for the
agent: "Updating is the person's call; `update --check` only reports."
- `upgrade --json`, `upgrade --project --json` and every `skills` JSON
output leave out `_meta.updateNotice` (NO_UPDATE_NOTICE), like `update`.
- dailyNotice: a null/non-object file reads as empty and is repaired; a
future timestamp is stale; writes go through a 0600 tmp file + rename;
withMeta never throws from the notice.
- Docs: --json without --yes exits 2; plugin `update` exits 1
(plugin_managed); full `_meta` sample; per-install notice commands.
`update --check` says "npm registry not reached" instead of "latest".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): a credit limit the person names holds on every paid media call
A limit was only a per-call --max-credits flag that each caller had to remember, so kit tools, voice resolve and transcribe could spend past it. The limit now lives in the project's credit-limit file and runJob, which every paid call goes through, refuses a new job past what is left (unpriced calls included). Voice and avatar video resolved through the heygen CLI are refused under a limit, since they cannot be priced there.
* fix(cli): keep a lost job's reserved credits, and hold the limit on the audio engine's voice
A resumed job refused with a status that says nothing about the job released its reservation while the job may still be running. Only the call that reserved releases, and only for a refusal that rules the job out. The audio engine's HeyGen speech call now refuses under a limit too.
* fix(cli): type withinLimit as returning whether it reserved
* chore: regenerate the skills manifest
* fix(cli): a credit-limit file is a run-time project file, and a lighter runJob catch
* chore(cli): mark runJob's complexity as accepted
* fix(cli): the voice a film uses is listed from the HyperFrames account, never the heygen sign-in
* fix(cli): drop a comment line and an unenforced range from the voices command
* feat(engine): a script-load failure names where the script came from and why it failed
* fix(engine): keep the script-load cause helpers private and within the engine size budget
* fix(cli,skills): say a step's credits are its total, and keep cost out of retries
An out-of-credits refusal named what the step needs, and an agent read
that as the shortfall. It now says the number is the step's total.
After its last try, an agent told the person what was spent and what
another try would cost. The HeyGen reference now says not to.
* fix(cli): plan avatar credits from the HyperFrames price rows
An avatar's credits came from its old API price, about five times what
the plan charges, so --max-credits refused calls that fit and a top-up
asked for too much. They now use the Avatar IV photo row plus the
general avatar video row, 950 credits a minute.
* fix(cli): pin the avatar price test and keep pricing under the comment share
* fix(cli): plan avatar seconds on the Avatar IV row alone
The general avatar row is only checked at admission and billed on
seconds left after Avatar IV, so adding it to every second made a 15 s
script plan at 238 instead of 200 and a 220 limit refuse a call that fits.
* feat(skills): creator templates on the HyperFrames plan
A Pro account asks for a video in a creator's format and the agent pulls that
template on demand. The packs sit in the private bucket behind the plan-gated
bundle route, so a Free account gets a 402 and the upgrade step. A pull costs
no credits; making the film spends them as Reverse does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): make creator-remix the one template skill
SKILL.md holds the catalog, the Pro openings, the pull and the credit
rules. references/remix.md is how to make the episode, from Desktop's
creator-remix with only what the app alone has swapped for the CLI.
The packs now hold only the template, so their hashes change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): the creator-remix skill and its packs
SKILL.md holds the catalog, the Pro openings, the pull and the credit
rules. references/remix.md is how to make the episode, from Desktop's
creator-remix with only what the app alone has swapped for the CLI.
The packs now hold only the template, so their hashes change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(skills): word creator-remix's sign-in steps the way media-use does
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): a creator-remix presenter runs on the plan by default
The person's own HeyGen avatar is recommended, not required: without
one, 小Lin's presenter is made with HeyGen Video 1 from their photo.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): pin template packs whose voice bills the plan first
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): offload media named inside template tags
References in a composition's <template> were never rewritten, because the DOM
query stops at the tag and template.content does not serialize, so the file stayed
in the zip. A stem matched as a plain word and a same-named file elsewhere counted
as a mention too. A 5-minute presented film went from 196 MB to 71 MB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): lipsync, and avatars that speak a finished voice
Video 1 takes audio as a reference and re-paces it, so a shot's lips follow its
own track and drift up to a second from the voice. generate --type lipsync redoes
the mouth to the exact voice. generate --type avatar takes --audio, and --avatar
for the person's own look, so the lips land on the voice from one job.
Also moves the plan rows to the 2026-10-30 price book, estimates a voice's credits
so --max-credits can hold it, logs jobs from a folder with no project, and describes
the resolve flags.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): presenters go avatar first, and a voice can be stock
The person's own avatar, then Avatar IV from their photo, then Video 1 with
lipsync. With no clone or recording, a stock voice named in the credits. Pins the
five rebuilt packs. The archive test keeps its ../ under GNU tar with -P.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(cli): fix avatar price and sfx provider tests, trim comments
A wav gives the avatar price test an exact length on every platform. The
sfx tests now expect HeyGen first. Comment share is back under the base.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): read a font shorthand whose size nests parentheses
`font: calc(var(--size) * 1px)/1 "X"` cut at the first closing parenthesis, so a
bad family name failed the render.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): one ledger row per job, and search skips a template's own pictures
A job done at once logs one row, transcribe logs to the project, and resolve no
longer offers a component's file for a search. --reuse and --provider get help text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): pin the tested template packs, and a board needs no voice
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(cli): read the transcribe ledger row as JSON, for Windows paths
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the Big Questions pack without the film length check
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): call them style templates, and the docs example pulls a real slug
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the cleaned template packs
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): drop the stock-presenter ban and the no-spend-before-storyboard rule
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the Paper Trail and Long Story Short packs without the stock-presenter ban
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): rename creator-remix to style-remix, and the five templates to names that say the look
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): style-remix lists the five templates when none is named
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the Finance Explainer pack that varies its motion graphics
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(cli): media-use generate --type cutout takes a shot out of its background on the plan
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the Finance Explainer pack without the provisional variety rule
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): pin the five style packs with variety rules measured on real films
Animated Science gets a worlds-and-boards budget, 3D True Crime a main board and a
brief corner note, 3D Brand caps by measured rate, Finance and Collage set pieces and
more than one look.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(skills): media-use names the on-plan cutout and lipsync commands, not the heygen CLI ones
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): cutout retries dedupe on request_id, uploads carry their real type; style-remix is public
The background-removal route dedupes on the body's request_id and ignores the Idempotency-Key
header, so a retry after a lost answer started a second job: runJob now also sends the key in the
body for such a route. Uploads typed every non-MP4 file as audio, so WebM and MOV inputs were
refused: they now take their type from the extension. Fails without the fix: the new cutout test
in generate.test.ts sees audio/mpeg for WebM and MOV.
style-remix drops metadata.internal and is listed with the other published skills.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(cli): keep the comment share of the touched files at main's
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: count style-remix in the README and CLAUDE.md skill lists
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(skills): rename style-remix to style-sampling
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli): drop --page-motion-blur from cloud render
The flag sent page_motion_blur to the hosted render API, which never
accepted that field and rejects unknown fields. Motion blur is set in
the composition with data-motion-blur, and the runtime turns it off on
its own when it cannot render correctly, so the cloud flag is removed
with its request plumbing, test and docs row.
Local render keeps its own --page-motion-blur flag.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(ci): allow deleting the cloud render page motion blur test
The deletion guard blocks removed files unless they are listed. The
test file only covered the cloud render --page-motion-blur flag that
this branch removes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* docs(skills): reverse asks to top up once for the run's total
* docs(skills): reverse builds the project with init --video
* test(skills): pin reverse's init --video and top-up-once wording
* docs(skills): a refusal with no amount runs a plain top-up
* feat(producer): report browser-computed reference backgrounds
Preserve one initial root/page reference across local, streamed and distributed
render results without guessing from film pixels.
— Jerrai
* fix(producer): preserve plan compatibility and optional inspection
Keep reference metadata outside the v2 manifest hash and read it from the
verified execution-plan artifact. Static planning uses a non-gating page
inspection instead of requiring render-worker capabilities.
— Jerrai
* fix(producer): poll readiness independently of animation frames
Use timed polling while the render clock parks animation frames. Cover readiness
that changes after the first check with a short real-browser regression.
— Jerrai
* fix(producer): budget measured reference metadata runtime growth
Producer adds 2,228 compressed bytes over current main. Tests, fixtures and source
maps are already absent from its dist-only tarball. Allow 7,766 bytes of headroom
above the verified 837,234-byte payload while retaining the unpacked limit.
— Jerrai
* feat(skills): publish reverse-any-video, render it at high quality, and stop on a stuck shot
From Wanbi's Pro test of reverse-any-video (11-01, checked against 0.9.2):
- A standard install never found the skill: it was internal, and /hyperframes did not route to
it. It is now an on-demand skill on every catalog surface, and the routing line names it.
- The cloud render in SKILL.md had no --quality high, so the final rendered at standard.
- A re-shoot that stayed pending had no rule; the run waited four hours. people.md now runs the
command once more after generate's 30-minute stop, then treats the shot as failed.
* feat(cli): generate takes a language for --voice, so a Pro sign-in gets a HeyGen voice
Wanbi's Pro run fell back to Kokoro: --voice needed an id, and the only way to list ids was the
separate heygen CLI and its own login. /v3/voices answers on the HyperFrames sign-in, so:
- --voice <language> picks HeyGen's first public voice in it and names it in the output.
- A missing --voice lists six English voices to pick from instead of a bare error.
- A voice id still goes straight through, with no extra call.
* fix(cli): a voice picked by language stays with its request across re-runs
Rames: `--voice Chinese` re-picked HeyGen's first voice on every run, and the voice is part of the
body and the idempotency key. If the catalog reordered while a job was pending, the same command
could start a second paid job instead of resuming the first.
The pick is now bound to the request (every field, plus the account scope a job bills) in the
project's heygen-jobs.jsonl, or the CLI config dir outside a project, and a re-run reuses it.
Regression: a pending timeout, a reordered catalog, the same command, for voice and avatar; the
second POST keeps the voice, body and key.
* fix(cli): a voice pick waits for a known account instead of binding to none
Rames: when /v3/users/me failed, senderScope() was null and the pick was bound to an empty
scope; once the account came back, the same command missed that binding and could pick another
voice and paid key. pickVoice now stops before picking or posting while the account is unknown,
and never reads or writes a blank-account binding.
Regressions, both directions: unknown then known refuses with no POST, then picks once; known
then unknown refuses the re-run with no second POST, then resumes with the same body and key.
Also README: a non-interactive install gets all 8 skills, not 7.
* fix(engine): a scene that has ended gives up its WebGL context so the next scene is not slowed
* build(engine): raise the engine package size budget for the WebGL release module
* chore(engine): shorten the WebGL release comment
* fix(engine): a scene gives up its WebGL context only when the runtime says its window ended
* chore(engine): drop a comment the type already says
* chore(engine): shorten a comment line
* fix(engine): GPU renders use the NVENC preset names the host's ffmpeg lists
* fix(engine): one ffmpeg stdout reader for encoder detection and preset listing
* ci(engine): raise the engine unpacked size budget for the NVENC preset reader
* fix(core): a resized preview video keeps its picture while its new copy loads
Swapping a video to a copy made for its new size emptied it until the copy
decoded, so the preview showed black for a moment. The swap now shows a still
of the frame it had, as the video's poster, until the new copy can draw.
* fix(core): a held preview still is let go when no file loads back
* chore(core): raise the packed size budget for the held preview still
* ci: split the cli test job into two shards so each stays well under its limit
* ci: map the sharded cli test run in the reachability audit
* ci: the reachability audit treats --shard as selection-neutral instead of a pinned mapping
* ci: map the sharded cli test run in the reachability audit (pinned), not a blanket --shard rule
* fix(skills): paid-plan cloud final keeps the composition fps and high quality
Follow-up to #6884 from its review: the cloud command agents copy had
no --fps, so a 24 or 60 fps composition rendered at 30, and cloud's
default `standard` is below the local `delivery` final. Step 5 and
cloud.md now spell out `cloud render --fps <data-fps> --quality high
-o <path>`; the render table uses `-o` to stay within the entry's byte
budget and restores the "Hosted, no local Chrome or FFmpeg" row.
— wanbi
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(skills): pin --fps and --quality high on the paid-plan cloud final
Review on #6886: a later byte-budget trim of hyperframes-cli/SKILL.md or
cloud.md could drop the flags silently. The test asserts both files'
paid-plan command reads `cloud render --fps <data-fps> --quality high -o …`.
— wanbi
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): motion blur preview no longer slows down later in the timeline
With data-motion-blur-events="on", which Studio's Subtle and Strong write,
both blur tiers ran timeline callbacks at every per-frame pose read. Each
frame's reads step back from half a shutter ahead to the frame, so a
callback that keeps state and replays from zero on a backward seek (a
ripple simulation, for example) redid its whole history most frames, and
the cost grew with the playhead.
The mover scan still runs callbacks, so motion set from a render(t)
callback is still found. With events on, it also scans once without them,
and the per-frame reads run callbacks only when the two scans disagree.
The eventful scan runs first, so tweens take their start values after the
callbacks before them and a one-off call fires as often as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): keep the motion blur comments to the invariant
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): motion blur reads run callbacks when a callback moves any box
Comparing which elements each scan finds missed a callback that pans a
mover's ancestor while a tween also changes that ancestor: both scans
found the same movers and the same changing elements, so the reads left
the pan out. The scans now compare every changing element's box at each
step, and the reads run callbacks when any step differs.
Raises @hyperframes/core's unpacked size budget to the measured size plus
the usual 5%: main already sat 0.98% over the old budget, and this change
adds 1,232 bytes, which tipped it past the 1% tolerance.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): silent motion blur reads land back on the frame as its own seek left it
After silent pose reads, the restore seek back to the frame left GSAP in a
different state than the frame's own eventful seek did:
- A step exactly on the frame (a .call() or set) was re-armed, so the next
frame fired it again. Calls fired 2-5 times per crossing, and a toggle
could end inverted.
- In renders, which keep no GSAP structure, landing back on the frame from
later skipped same-time steps' authored order, so the frame showed the
first of two sets instead of the last.
Events-off blur always read silently and had both. This PR moves events-on
roots without callback-driven motion (the Studio default) onto that path,
so both are fixed here.
The host's new holdFrame snapshots every landing mark on the root and scene
timelines before the reads, and its restore seeks back silently, redraws in
authored order where there is no structure, and puts the marks back. Each
call now fires once per frame crossing, as without blur, and same-time sets
land as authored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(core): a held render frame draws its scenes first, so the root lands last
holdFrame redrew the root and then the scenes, the reverse of the frame's
own seek. When a scene and the root tween one property, the frame landed
on the scene's value instead of the root's.
The hold and landing move into the GSAP adapter as holdFrameMarks, which
takes the scenes and the root separately, so the order lives in one place
a real-GSAP test can pin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(cli,skills): say a step's credits are its total, and keep cost out of retries
An out-of-credits refusal named what the step needs, and an agent read
that as the shortfall. It now says the number is the step's total.
After its last try, an agent told the person what was spent and what
another try would cost. The HeyGen reference now says not to.
* fix(cli): plan avatar credits from the HyperFrames price rows
An avatar's credits came from its old API price, about five times what
the plan charges, so --max-credits refused calls that fit and a top-up
asked for too much. They now use the Avatar IV photo row plus the
general avatar video row, 950 credits a minute.
* fix(cli): pin the avatar price test and keep pricing under the comment share
* fix(cli): plan avatar seconds on the Avatar IV row alone
The general avatar row is only checked at admission and billed on
seconds left after Avatar IV, so adding it to every second made a 15 s
script plan at 238 instead of 200 and a 220 limit refuse a call that fits.
* fix(cli): the credits link opens the web plan page, which is live
* test(cli): pin the production credits link too
* chore(skills): regenerate the skills manifest after merging main
* docs(cli): keep the person measure's doc comments to four lines
* docs(cli): keep the measure doc comments to four lines
* docs(cli): keep the measure comment lines to 120 characters
---------
Co-authored-by: Miguel Angel Simon Sierra <miguel.sierra@heygen.com>
* fix(cli): exit-time telemetry no longer loses a render's events on Windows
* fix(cli): telemetry smoke stand-in reads the batch file; format the warning
* docs(catalog): put each item page's title, description and Docs text in the server HTML
Mintlify renders snippet components in the browser only, so everything inside
CatalogDetail (title, standfirst, the Docs tab) was missing from the HTML search
engines read: 448 catalog pages served Googlebot no body text. The generator now
writes the H1 and standfirst as page markup above the component and the Docs tab
as page markup after it, shown with that tab (docs/custom.css). The component no
longer draws its own title. Pages patched structurally only; registry drift on 14
pages is left for a regeneration of its own.
* docs(catalog): keep the title's top spacing the component gave it
* fix(docs): catalog titles and descriptions keep markdown markers literal
MDX reads text inside the static <h1> and <p> as inline markdown, so @_blue_sweater_
rendered as italic blue_sweater. mdxText now writes backslash, backtick, * _ [ ] ~ as
character references too, with a unit test; the four affected pages are re-escaped.
* revert: feat(skills): reverse any video (#6754)
Keep the reverse-any-video skill and its CLI billing commands out of main until
web subscriptions are live (/pricing and /account/plan return 404 today).
Re-land together with #6833 once those pages ship.
This reverts commit cb2389cf35.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* chore(scripts): allow the #6754 revert deletions
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(skills): reverse any video on the HyperFrames plan
Reverse rebuilds a video as an editable HyperFrames project and re-shoots its
people with HeyGen. It runs for a person signed in on HyperFrames Pro; Free is
told it is a Pro feature and given the upgrade step.
- skills: reverse-any-video; media-use is the one gate for HeyGen calls
(sign-in, credits, refusals)
- cli: media-use generate (image, video shot, avatar, voice, music) with --plan
pricing and --max-credits; measure, diff and analyze for the rebuild; the
HyperFrames plan in auth status; billing subscribe and top-up wait on the
plans or credits page
- remotion-to-hyperframes compares with `hyperframes diff` instead of its own
shell scripts
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(skills): satisfy CI gates for reverse-any-video
- docs: list diff, measure, analyze and billing in "Find your command" in
docs/packages/cli.mdx, each with its own section.
- skills: bring media-use/SKILL.md under its 5120 B budget by pointing setup,
direct-recipe and treatment detail at the references that already hold it;
the sign-in, credits, --max-credits and billing rules stay in SKILL.md.
- comments: register transcript.json and REVERSE.md as project files for the
citation checker, and drop a history note from face.test.ts.
- comments: cut restating prose or move it into names so every changed file is
at or under its comment share. Allowlist the two deleted remotion scripts in
check-no-main-deletions.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): terminate measure's ffprobe options with --
probeVideo passed the input path straight after its options, so a path
starting with "-" would be read as a flag; the producer's ffprobe argv
contract test failed on it. runFfmpeg's fixed flags move to a constant
so the contract no longer mistakes that ffmpeg wrapper for a probe.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(cli): split diff's video tests into three files under CI's per-file limit
The file took ~131 s on CI against the flake gate's 120 s, and its heaviest test (four
comparisons) hit the 20 s per-test limit. The video tests now live in three files sharing
diff.test-helpers.ts, and that test is two tests of two comparisons each.
analyze reads each part once and checks the bytes it read, instead of a stat then a read
(CodeQL js/file-system-race).
* fix(skills,cli): one credits rule; ask before dropping a step that won't fit
On dev with a 100-credit limit the agent priced a re-shoot with --plan and dropped it without
asking (J5), and in J6 it told the person a price. SKILL.md's Credits rule now covers both:
--plan is for the agent, and when a step won't fit (a --plan over what is left, or a refused
call) the agent says what doesn't fit and asks before dropping or shrinking it.
setup-providers.md's older confirm-before-paying rule points to it.
--plan now runs the same checks as the real call (checkRequest), so it no longer prices a
request that would be refused. The re-shoot recipe in people.md gains the --prompt it needs.
heygen.md: a sign-in also makes images, and --plan's usd is the API charge rounded to $1/60
while its credits come from the list price.
* fix(cli): auth status drops 'balance unavailable' above the plan's credits; sign-in page title
* test(cli): keep diff's video tests inside the flake gate's limits
Under the gate's CPU contention every diff test needing face detection or
optical flow paid opencv.js's 11 MB download and WebAssembly start-up, and
every diffVideos call wrote four worst-moment pictures (about 28 ffmpeg runs)
no assertion read. Tests ran 21-28 s against the 20 s limit.
- loadOpenCvFirst() stages opencv.js in a beforeAll (from this machine's
cache, a temp copy kept between runs, else the network).
- diffVerdicts() runs diffVideos with worst: 0 for tests that only read
verdicts; the worst-moment tests keep the pictures.
- The carried-text test is split into its dirty and clean halves.
Locally with 8 CPU burners on 8 cores: slowest test 7.7 s, files 24-30 s.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): generate --plan refuses what the real call refuses
checkRequest now holds every pre-send check (the --out extension and
--mask without --motion joined the route body and guide checks), and
generate() runs it first, so --plan and the real call share one function.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli): plans link uses from=cli and follows HYPERFRAMES_ENV; no balance notice beside the plan
- pageUrl sets from=cli, the parameter the pricing page reads to send a
CLI buyer back to the terminal.
- The plans and credits pages are on dev.hyperframes.dev with
HYPERFRAMES_ENV=dev, hyperframes.studio otherwise.
- auth status --json omits the 'HyperFrames balance is unavailable' notice
when the plan was read.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(skills): reverse-any-video re-checks sign-in on resume; REVERSE.md holds no prices
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* test(cli): split diff-people's whole-frame reuse test into three
It ran five diff comparisons in one test and took 20.2-20.6 s under the flake
gate's contention, over the 20 s limit. Each piece now runs at most two
comparisons against a shared ledger: copy and burned-in title, footage cut
whole, and cuts that only fed a generation. Under 24 CPU burners locally the
slowest of the three takes 3.0 s, where the combined test took 4.9 s under 8.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli,skills): reverse resume asks after one failed retry; out-of-credits names the step's price
- media-use generate: when HeyGen refuses for credits without saying how many, the refusal
names the CLI's own price for the call (what --plan prints) as "needs about N" and
"billing top-up --needed N", instead of "needs more".
- reverse-any-video: never repeat auth status's plan/credits line; a step an earlier session
gave up on gets one more try, then ask; a command stored in REVERSE.md is a note, rebuilt
from its recipe and checked with --plan, never run as written.
- people.md / heygen.md: a refused or failed job gets one more try, test calls included,
then stop and ask.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* fix(cli,skills): send a failed HeyGen job again as new; resume asks if the limit still applies
- runJob: a key whose job failed or was refused is recorded in the config dir
(heygen-failed-keys); the next run with the same inputs sends key-1, key-2, ... so HeyGen
starts a new job. A finished or still-running job keeps its key and is picked up, so
nothing is paid twice. "fix the request it names, then run again" now holds.
- reverse-any-video: on resume after a run under a credit limit, just ask whether it still
applies, naming no costs or numbers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* refactor(cli,skills): drop what reverse-any-video no longer uses
The skill rebuilds from measure probe, cuts, frames, events and face, and checks with
diff. analyze, diff --explain and eight measure commands (trail, box, ink, rhythm,
objects, track, flow, continuity) have no caller, so they go, with their tests and docs.
Examples now use original.mp4 instead of the old .reverse/source.mp4 evidence folder,
and diff's --segments is described as a file rather than REVERSE.md.
* fix(cli): check no longer copies VP9 and AV1 videos before it starts
Chrome plays both, alpha included, and the proxy is made on demand if a browser ever
asks. Copying remove-background's VP9 cutouts took minutes on projects that have them.
* fix(cli,skills): keep paid retries, credit limits and lip sync honest
Retries keep their key after 401, 403, 408, 409, 425 and 429, which say
nothing about the job, and failed keys are scoped to the API and payer.
Credits come from the plan's video rows instead of list dollars, and
reference pictures past the first two count toward --max-credits.
Lip sync reads sound on its own timestamps, so a delayed track fails
the gate. Look and energy call two near-zero values the same.
--plan refuses a missing input as the real call does. The skills teach
the direct sign-in, its credential path and the billing commands.
* chore(cli): keep the new pricing and energy files under the comment share
* fix(cli): scope failed HeyGen keys to the signed-in account
Failed keys were scoped by the local user block, which is absent for a
host's token and display-only otherwise, so two accounts could share a
scope and one could start a second paid job.
The scope now comes from the account /v3/users/me reports for the token
in use. When that cannot be told, no failure is recorded and the key
stays the same.
* fix(cli): fix a job's sender when it starts
Failure scope was looked up again after polling, so a sign-in during a
job could record one account's failure under another and rotate its
finished key. The credential is now read once per job: it signs the POST
and owns any failure, even if someone else signs in meanwhile.
Prefer username to email, since /v3/users/me returns no user id.
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: James <james.russo@heygen.com>
* ci: one regression shard writes the test image cache, so it stops evicting the rest
* test(ci): each image cache has one writer, shard 0 of a main push
* fix(ci): skills sync steps past any taken ClawHub version, not one hard-coded skill
* fix(ci): skills sync steps past the five hidden ClawHub versions blocking it
* ci: keep the bun cache warm and skip binary downloads in preflight
* test(ci): pin the single cache writer and preflight's skipped downloads
* test(ci): keep the cache-writer scan out of the test body
* perf(aws-lambda): small-chunk cloud renders stitch their output in parallel
When every chunk's payload is under S3's 5 MiB part minimum (typical for short 4K chunks), the assemble step could not copy any part server-side and pulled the whole output through one Lambda's network link. Those range-only parts now go to parallel invocations of the warm function that just rendered the chunks, about 32 MiB each and never more at once than the stack's reserved concurrency allows. The output bytes are unchanged.
* fix(aws-lambda): a throttled stitch helper builds its parts in the assemble instead
Another render can hold the stack's reserved concurrency while this one assembles, so a helper invocation may be refused. The assemble now builds that batch itself, logs it, and counts it in assemble_stages; the helper's error type is kept as the error name. Comments trimmed back to each file's share at the base.
* chore(aws-lambda): raise the package size budget to the measured size plus headroom
The parallel stitch helpers add about 1.7 KB packed and 9.9 KB unpacked; the new budget is the measured size times the checker's own 1.05 headroom.
- .claude/settings.json: the typecheck step piped through grep inside
single quotes, which cmd.exe splits into a broken pipeline, so the hook
denied every commit on Windows. On every OS, `|| true` also made the step
pass regardless of type errors. Test-file errors are now filtered in JS
and any remaining `error TS` line fails the step.
- scripts/check-studio-copy-dashes.mjs: report paths with forward slashes
so EXEMPT entries match on Windows.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* test(studio): record each edit-accuracy shard's file writes and reloads when HF_WRITE_TRACE is set
* style(studio): format the write trace
* fix(core): hash the write trace with sha256
* style(studio): wrap the trace hash helper
* style(studio): format the trace hash helper
* fix(studio): turn the reload log on in the Studio page the trace watches
* ci: a release PR runs a fast release check, not the heavy suites
* ci: keep the release check's header within the comment limit
* ci: shorten the release check's usage line
* ci: trim the release check's comments
* ci: the release check's classifier hands the manifest test to a helper
* fix(core): preview seeks land on the composition's own frame rate
* fix(studio): frame step, frame readout and ruler use the composition's frame rate
* fix(studio): an end trim previews the clip's last frame at the composition's frame rate
* refactor(studio): ruler labels follow the ticks' frame rate and fps parsing splits by form
* style(studio): format the ruler label helper and its test
* chore(inspector): raise the package size budget to the measured size plus headroom
* fix(producer): convert HEIC images to JPEG at compile time
Tiled HEIC photos probe as several video streams, so the media-type
preflight rejected them as image:video, and Chrome cannot decode HEIC.
Sniff the ftyp brands of every <img>, join and rotate the tile grid with
ffmpeg, verify the declared size, and swap in a cached JPEG.
* fix(producer): shorten the HEIC prep comments
* fix(producer): name host tool failures and keep a failed HEIC preview image as authored
Also: HEIC tests run in the integration lane; the primary grid is picked by default disposition.
* chore(producer): raise the package size budget for the HEIC prep