Animated media volume (GSAP/JS fades) dropped the audio track entirely for dense
fades. The 60 Hz timeline probe emits 100-300 keyframes for a multi-second fade,
which were folded into an FFmpeg `volume` expression nesting one `if(lt(t,...))`
per keyframe. Past ~95 nested levels (build-dependent, lower on some Linux ffmpeg
builds) the expression overflows FFmpeg's evaluator, fails filter-graph init,
fails the whole mix, and the muxer omits audio — so a `data-volume="0"` fade-in
rendered with no audio at all (follow-up to #1066; this is why #1064's own
scenario regressed once the fade was dense enough).
Apply volume automation as sample-accurate gain, layered so audio is never lost:
1. Primary: bake the envelope into the prepared PCM samples in-process
(audioVolumeEnvelope.ts). The track WAV is always pcm_s16le/48k/stereo;
multiply its samples by the interpolated envelope and atomically rename the
result into place, then mix at unity. No expression, no keyframe ceiling,
exact at every sample, and the downstream ffmpeg amix/AAC encode is untouched
so golden baselines only change where a fade is applied. The RIFF parser
scans chunks order-independently and accepts only 16-bit PCM, falling back
otherwise. The output is written to a random-named sibling and renamed, so a
crash can't leave a truncated WAV and there's no predictable-path write.
2. Fallback: RDP-bounded ffmpeg `volume` expression (0.5% tolerance, capped at
32 segments) for the rare case a WAV is not 16-bit PCM. 0.5% keeps the
rendered envelope within ~0.2 dB of the source curve.
3. Backstop: if an automated mix still fails, retry once at base volume and
surface the degradation rather than dropping the track.
This mirrors how OSS NLEs render automation (sample-level gain): MoviePy,
Kdenlive/Shotcut (MLT), Remotion.
Verified end-to-end: a 297-keyframe fade that rendered with no audio now bakes
all 297 keyframes sample-accurately. Adds unit tests for sample-accurate gain,
track-start offset, base/tail holds, thousands of keyframes, order-independent
chunk parsing, and format rejection, plus mixer regression tests for bounded
nesting and the base-volume backstop.
Follow-up to #1115. Makes the Design-panel editor recognise every target
shape real compositions use. The panel stays behind STUDIO_GSAP_PANEL_ENABLED
(default off) — no flag change here.
- Array targets: tl.to([a, b], {...}) resolves to a CSS group selector
(".a, .b"). The source array is never rewritten — the joined string is for
display/matching only; edits still touch just the vars object.
- Chained calls: tl.to(a, ...).to(b, ...) — the matcher now walks the member
chain to its timeline root, so every link is captured (previously only the
first). Deletion is chain-aware: it splices out the single targeted link and
re-points the chain instead of dropping the whole statement.
- gsap.utils.toArray("sel") resolves like querySelectorAll, inline or via a
variable binding.
- Lexical scoping: element-variable resolution is now per-scope (walks the
enclosing function/program chain) instead of a flat map. Fixes silent
wrong-resolution when two IIFEs reuse a variable name, and unlocks
multi-scene files. (Addresses review: flat-binding-scope.)
- forEach/map callback params (items.forEach(el => tl.to(el, …))) and items[i]
indexing resolve to the collection's selector, so loop-generated tweens are
editable.
- Panel matching: an element matches a tween when its id/selector is any member
of a comma-group target, so either element of an array/toArray tween surfaces
the shared animation.
- Review items: mutation parse failures now console.warn instead of swallowing
silently; buildTweenStatementCode no longer emits duration on `set`; the
id-only serialize-side filter is renamed getAnimationsForElementId to
disambiguate from the panel's id-or-selector matcher; added fromTo round-trip
and variable-target overlap-lint tests.
Genuinely runtime-only targets (template-literal selectors, unbounded loops)
still skip gracefully — they can't be resolved or matched statically.
The Design-panel GSAP editor only recognized tweens written as
tl.to(".selector", {...}) with inline string-literal targets, in a
contiguous block, with no interleaved setup. Every scaffolded
composition instead targets tweens through element variables
(const kicker = root.querySelector(".kicker"); tl.to(kicker, {...})),
wraps the script in an IIFE, and interleaves gsap.set() calls — so the
parser returned zero animations and the panel was inert.
Three coordinated fixes make it work end to end:
- Parser read: resolve querySelector / querySelectorAll / getElementById
variable targets (and inline lookup calls) back to their CSS selector,
so variable-targeted tweens are recognized.
- Parser write: replace the full re-serialize (preamble + tweens +
postamble) with in-place recast AST mutation. Edits now touch only the
targeted tween's vars/position node and reprint, preserving every
surrounding statement — gsap.set calls, element declarations, the IIFE
wrapper, comments and formatting. Previously the first edit would
discard all of that.
- Linter: build overlap/clip windows directly from the parser's
structured animations instead of a regex walk paired positionally with
the parsed list. The old pairing skipped variable targets and would
drift once the parser started returning them. Removes the now-dead
regex meta helpers.
- studio-api: extractGsapScriptBlock now searches inside <template>
content (sub-compositions wrap markup + the GSAP script in a template,
which linkedom's querySelectorAll doesn't descend into), and the
frontend matches tweens to the selected element by id OR selector
rather than id only (class-targeted elements have no id).
Verified end to end against a real 10-scene project: all compositions
now parse (previously 0), the panel populates editable tween cards, and
property/duration/ease edits round-trip while leaving the rest of the
script byte-for-byte intact.
* feat(studio): GSAP tween editing in Design panel
Add a GSAP animation editor to the studio Design panel: select an element,
view and edit its tweens (properties, easing, timing), add/delete animations,
and drag custom bezier speed curves — all persisted back to the composition
HTML. Gated behind VITE_STUDIO_ENABLE_GSAP_PANEL.
Parsing of existing GSAP source now uses a recast + Babel AST parser instead of
regex, giving scope resolution, stable tween IDs, and round-trip preservation of
extras and unresolved raw values.
recast compiles to CommonJS that calls require("fs"), which breaks browser and
Vite SSR bundles. To contain it, @hyperframes/core is split into an isomorphic
layer and a Node-only AST layer:
- gsapSerialize.ts holds the recast-free helpers (serialization, keyframe
conversion, validation, shared types). htmlParser.ts is now fully isomorphic.
- parseGsapScript and the script-mutation helpers live in gsapParser.ts,
reachable only via the @hyperframes/core/gsap-parser subpath, loaded
server-side by the studio-api mutation routes and the linter via dynamic
import (recast stays external under SSR).
- The barrel and the gsap-constants subpath are recast-free, so studio browser
bundles never trace recast.
Adds AST parser unit + stress coverage and e2e helpers for the panel.
* fix(lint): await async lintHyperframeHtml in all callers
lintHyperframeHtml became async (gsap rules use dynamic import)
but lintProject and check-hyperframe-static weren't awaiting it,
causing typecheck failures and runtime crashes in CI.
Also wire LintRule type in gsap rules to fix fallow unused-type
finding, and suppress render.ts exported-for-tests symbols.
The runtime had a maxTimelineDurationSeconds field defaulting to 1800
(30 minutes) that clamped the TransportClock duration. Any seek beyond
this cap was silently clamped, so GSAP tweens starting past ~1700s
never received their totalTime() call and stayed at their pre-tween
state (e.g. opacity:0).
The data-duration attribute is the authored source of truth. The loop-
inflation guard (timelineLooksLoopInflated) already handles the infinite
repeat:-1 case this cap was meant to protect against.
Closes#1107
* docs: document feedback collection — cadence, data, opt-out
Adds guides/feedback.mdx covering: when CLI and Studio prompts
appear (render cadence, session cadence), what data is collected
(PostHog survey fields, doctor_summary shape), what is not
collected, the hyperframes feedback command for manual/agent
submission, agent runtime detection and structured hint,
config file fields, and all opt-out paths (HYPERFRAMES_NO_TELEMETRY,
DO_NOT_TRACK, CI guard, --quiet).
Also adds hyperframes feedback command entry to packages/cli.mdx
(Utilities tab, alongside telemetry) and registers guides/feedback
in the docs.json nav.
— Magi
* docs(feedback): fix cadence, agent env vars, docker gate, telemetry scope, why-we-ask
- Cadence: 1st/16th/31st (not 15th/30th/45th) per actual code
- Agent vars: CLAUDECODE/CLAUDE_CODE_ENTRYPOINT, CODEX_THREAD_ID/CODEX_CI,
TERM_PROGRAM=cursor, Copilot value checks; add Hermes/openclaw/Pi
- Remove docker gate claim (non-TTY only, not docker-specific)
- Telemetry disable only suppresses CLI prompt, not Studio bar
- Add why-we-ask opening section
- Remove Studio 'skip' action (CLI-only); fix 'counter resets' phrasing
- Fix 'values never read' — Cursor and Copilot do value comparisons
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(feedback): remove invented Studio opt-out flags; document localStorage workaround
VITE_HYPERFRAMES_FEEDBACK_INTERVAL=0 falls through to default (n > 0 guard).
VITE_HYPERFRAMES_FEEDBACK feature flag doesn't exist. Bar is mounted
unconditionally. Document the localStorage key workaround instead and
note that a proper flag is a follow-up to hf#1101.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(feedback): fix localStorage workaround — only lastPromptedAt needs to be large
Setting both keys to the same value just delays 10 sessions before the bar
reappears. Setting only lastPromptedAt to 9999999 keeps count - lastAt
negative indefinitely.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(studio): add VITE_HYPERFRAMES_NO_FEEDBACK build-time disable flag
Sets isFeedbackDisabled() guard in shouldShowFeedback() — when
VITE_HYPERFRAMES_NO_FEEDBACK=1, bar never shows regardless of session count.
Updates docs to document the flag and remove the localStorage workaround.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Both flags were silently broken via the same root cause: citty parses
`--no-FOO` as a negation of the base flag `FOO`, so a flag literally
named "no-confirm" gets routed as `args.confirm=false` (not
`args["no-confirm"]=true`), and same for "no-wait".
Surfaced during the end-to-end smoke test on the just-merged stack:
- `cloud delete <id> --no-confirm` was hitting "Confirmation required"
and exiting 1 without calling the API.
- `cloud render --no-wait` was running the full poll + download flow
instead of submitting and exiting with the render_id.
Renamed the arg keys to `confirm` (default true) and `wait` (default
true) so citty's built-in negation handles the user-facing flags
correctly. Flag names stay the same; only the runtime arg keys change.
Live-tested both: delete now removes the render and a subsequent get
404s; --no-wait now returns just {render_id, status: "queued"} and
exits.
Note: a third instance of the same pattern exists in commands/add.ts
(`--no-clipboard`) and is also latently broken. Out of scope for this
fix; should be addressed alongside any audit of the CLI's interactive-
vs-noninteractive defaults.
* feat(cli): vendor initial hyperframes cloud client codegen
Generated by experiment-framework/scripts/generate_hyperframes_cli_client.py
(see heygen-com/experiment-framework#37896). Sets up the baseline for the
sync workflow to diff against on future spec changes.
The follow-up PR adds the orchestration layer (zip + upload + poll +
download) and the user-facing 'hyperframes cloud render/list/get/delete'
commands on top of this generated client.
The fallow ignore pattern is necessary because the generated request()
method is intentionally a single switch that handles all 5 endpoints
in one place; refactoring it here would just be re-introduced on the
next codegen run.
* chore(cli): regenerate cloud client with mimeType parameter on multipart uploads
Adds optional mimeType arg to uploadAsset (and any future multipart
endpoints). Without it, FormData sends application/octet-stream which
is correct for the documented media surface (png/jpeg/mp4/etc.) but
ambiguous for the private-beta zip uploads the cloud render flow uses.
Callers that pass `mimeType: "application/zip"` tag the multipart
part with the right Content-Type so downstream proxies, WAFs, and any
future server-side change that keys off the part MIME (instead of the
current magic-byte detection) all see the intended type.
Addresses review feedback on heygen-com/experiment-framework#37896.
Generated by scripts/generate_hyperframes_cli_client.py with the
matching update to the multipart emit path.
* feat(cli): add hyperframes cloud render/list/get/delete commands
Hand-rolled orchestration layer on top of the auto-generated cloud
client (vendored in the previous PR):
- cloud render <dir>: zip via createPublishArchive → upload to
/v3/assets → submit /v3/hyperframes/renders → poll
/v3/hyperframes/renders/{id} every 10s (max 60min) → stream the
signed video_url to disk.
- cloud render --no-wait: submit and exit with the render_id.
- cloud render --asset-id / --url: skip zip+upload and use a
pre-uploaded asset or public HTTPS zip.
- cloud render --variables / --variables-file: same UX as the local
render command; variables are validated against
data-composition-variables only when there's a local project.
- cloud list / cloud get / cloud delete: thin wrappers around the
matching client methods, with cursor-pagination support on list.
Auth comes from the existing cli/src/auth/ chain via cloud/auth.ts —
no new credential store, no new env var. The cloud client receives a
getAuthHeaders() callback that re-resolves credentials on every
request, so OAuth refreshes mid-poll are picked up automatically.
Also extracts a parent-scoped path lookup in help.ts so 'cloud render
--help' surfaces the right examples instead of falling through to the
top-level 'render' command's examples.
* fix(cli): address 15 code-review findings on cloud commands
Correctness fixes
- delete: require --no-confirm when stdin isn't a TTY OR --json is
passed; previously both silently auto-bypassed the irreversible-
delete prompt. Explicit decline now exits 2 (distinct from API/system
errors which still exit 1).
- render: mutex check now counts the positional dir alongside
--asset-id / --url; `cloud render ./foo --asset-id X` now errors
instead of silently dropping the dir.
- render: docstring updated — only --no-wait short-circuits the poll
loop; --callback-url is independent (webhook fires either way).
- render: removed dead try/catch around resolveProject (it calls
process.exit, never throws). resolveVariablesAndValidateIfLocal also
takes the resolved project source instead of re-parsing args.
- render: createPublishArchive errors now surface via errorBox instead
of bubbling a raw stack trace past citty.
- help: loadExamples now only catches ERR_MODULE_NOT_FOUND; real load
errors (syntax error, broken import) propagate so a broken
cloud/render.ts no longer silently shows the local render command's
examples. Also skips the parent-scoped lookup when parentName is the
root command ("hyperframes").
- list: fetchAll gained a 50-page safety cap + duplicate-cursor
detection so a buggy backend serving the same next_token on a loop
can't OOM the CLI.
- download: drain await now listens for error / close / abort so a
failing write stream (ENOSPC, AbortSignal) rejects promptly instead
of hanging forever. Partial files are unlinked on any error so the
caller never observes a truncated MP4. content-length is verified
against the actual byte count.
- poll: default sleep is abort-aware so Ctrl+C feels immediate instead
of waiting out the full interval.
- pollWithProgress: ANSI carriage-return redraws now gated on
process.stdout.isTTY — non-TTY runs (CI, file redirects) emit one
line per status transition instead of polluting the log with
literal escape codes.
Cloud client: 401-retry-with-refresh
- createCloudClient now wraps the generated client with a Proxy that
catches HyperframesApiError(status=401), force-refreshes the OAuth
token via forceRefreshCredentials, and retries the call exactly
once. Mirrors AuthClient's onUnauthenticatedRefresh so server-side
revocations and clock-skew rejections recover automatically.
- auth.ts gained forceRefreshCredentials() and now updates expires_at
on the refreshed credential it returns (fixed stale-expiry race).
Shared helpers
- cloud/errors.ts: reportApiError(stage, err, opts) is the single
error-funnel. ERROR_CODE_HINTS now applies to every subverb — fixes
hyperframes_render_not_found being unreachable from get/delete and
cuts ~70 LOC of duplicated try/catch/instanceof from render/list/
get/delete.
- cloud/parsing.ts: parseIntFlag / parseNumericFlag / parseEnumFlag
strict-mode parsers reject trailing garbage that Number.parseInt
silently accepts.
- cloud/ansi.ts: stripAnsi / visibleLength / padEndVisible — covers
ESC + 24-bit truecolor (c.accent palette) instead of the previous
regex which undercounted overhead and missed truecolor.
JSON-output consistency + _meta envelope
- Every cloud subverb's --json output now goes through withMeta(...)
so it carries the standard _meta envelope documented in cli.mdx.
- Single-render outputs use {render: detail} across get, delete,
render-no-wait, render-failed, and render-success. list uses
{renders: [...], has_more, next_token?}. delete adds deleted: true.
Tests
- 25 new tests across ansi.test.ts, parsing.test.ts, plus truncation
+ abort-cleanup tests for download.test.ts.
- 589 / 589 total CLI tests pass.
* fix(cli): address Vai's review on cloud commands
- render: pass mimeType: "application/zip" to uploadAsset so the
multipart Content-Type is correct (was application/octet-stream).
Server currently magic-byte-detects from file bytes so this is
belt-and-suspenders today, but any downstream proxy / WAF / future
server change that keys off the part MIME now sees the intended
type instead of relying on detection.
- render: poll error path now surfaces "Resume with: hyperframes
cloud get <renderId>" via reportApiError's new `suggestion`
option, matching the PollTimeoutError handler. The server-side
render keeps running through a transient 5xx; the user just
needs the right command to pick it back up.
- list: fetchAll now errorBox-exits on the malformed
{has_more: true, next_token: null} shape instead of silently
returning a truncated list (matching the duplicate-cursor guard).
- download: closeFile now listens for 'error' on the write stream
in addition to the end() callback, so a late ENOSPC during flush
doesn't leak an unhandled error onto the stream and resolves the
finally promptly.
- errors: reportApiError accepts an optional `suggestion` that's
used as the errorBox third line when no code-specific hint
matches — gives callers a place to surface always-actionable
recovery context.
- docs(cli): document --idempotency-key as the safe-retry mechanism
for the upload step. The 401-retry Proxy replays POST requests
on a stale token; without an idempotency key, the upload may
land twice. A UUID per logical render is the recommended pattern.
* feat(cli): vendor initial hyperframes cloud client codegen
Generated by experiment-framework/scripts/generate_hyperframes_cli_client.py
(see heygen-com/experiment-framework#37896). Sets up the baseline for the
sync workflow to diff against on future spec changes.
The follow-up PR adds the orchestration layer (zip + upload + poll +
download) and the user-facing 'hyperframes cloud render/list/get/delete'
commands on top of this generated client.
The fallow ignore pattern is necessary because the generated request()
method is intentionally a single switch that handles all 5 endpoints
in one place; refactoring it here would just be re-introduced on the
next codegen run.
* chore(cli): regenerate cloud client with mimeType parameter on multipart uploads
Adds optional mimeType arg to uploadAsset (and any future multipart
endpoints). Without it, FormData sends application/octet-stream which
is correct for the documented media surface (png/jpeg/mp4/etc.) but
ambiguous for the private-beta zip uploads the cloud render flow uses.
Callers that pass `mimeType: "application/zip"` tag the multipart
part with the right Content-Type so downstream proxies, WAFs, and any
future server-side change that keys off the part MIME (instead of the
current magic-byte detection) all see the intended type.
Addresses review feedback on heygen-com/experiment-framework#37896.
Generated by scripts/generate_hyperframes_cli_client.py with the
matching update to the multipart emit path.
* feat(cli): prompt for render satisfaction after successful renders
* feat: add text feedback, doctor context, and Studio render feedback UI
* feat(studio): replace render feedback with session-based Studio experience bar
Move the feedback prompt out of RenderQueueItem (where it triggered every
5th render) into a standalone StudioFeedbackBar mounted at the bottom of
the preview area. The new bar is session-gated (shows after the 5th studio
session), auto-dismisses after 20s, and respects a 30-day cooldown once
dismissed or submitted. Renames telemetry to trackStudioFeedback with a
"studio_experience" survey ID to reflect the broader scope.
* feat(studio): attach browser doctor summary to feedback events
* fix(studio): use recurring interval for feedback instead of one-time cooldown
* fix(cli): skip feedback prompt when an agent runtime is detected
* feat(cli): add hyperframes feedback command and agent render hint
- New `hyperframes feedback --rating <1-5> --comment "..."` command
for submitting anonymous render satisfaction feedback via telemetry.
- When an AI agent runtime is detected after a render, print a dimmed
hint to stdout so the agent can optionally call the command instead
of silently skipping the readline prompt.
- Export getDoctorSummary from telemetry/feedback.ts to share the
system-info collector between the interactive prompt and the CLI command.
- Register the command in cli.ts and help.ts under Settings.
* fix(studio): align feedback interval to every 15 sessions
* fix: show CLI feedback on first render, Studio every 10 sessions
* feat: add env flags to disable feedback prompts
* feat: env flags to configure feedback prompt frequency
* fix: address review — agent hint reachability, cadence gate, session debounce, deprecated API
## What
Adds OAuth 2.0 + PKCE login as the default for `hyperframes auth login`,
plus refresh-token + 401 auto-retry + `auth refresh`. Stacks on top of
PR #1081 (the API-key + shared store work).
- `hyperframes auth login` (no flags) — opens the user's browser to
`/v1/oauth/authorize`, captures the code on an ephemeral
`127.0.0.1:<port>/oauth/callback`, exchanges it for tokens with
PKCE S256, and persists. `--api-key` opts back into the legacy
long-lived-key path from PR #1081.
- `hyperframes auth refresh` — force-refresh the OAuth access token
using the stored refresh_token. Mostly useful for testing the path.
- `hyperframes auth logout` — best-effort revokes via
`POST /v1/oauth/revoke` (RFC 7009) before wiping local state.
- `AuthClient` now refreshes-and-retries once on a 401 when the
caller wires `onUnauthenticatedRefresh`. `auth status` wires it.
Internals added in `packages/cli/src/auth/`:
- `pkce.ts` — RFC 7636 code_verifier + S256 code_challenge.
- `loopback.ts` — ephemeral 127.0.0.1 HTTP server; state validation,
120s timeout, styled success/error page.
- `browser.ts` — wraps `open` with a `BROWSER=none` /
`HF_NO_BROWSER=1` fallback that prints the URL.
- `oauth.ts` — `startAuthorizationCodeFlow`, `refreshTokens`,
`revokeTokens`, `requireOAuthConfigured`, `parseTokenResponse`.
## Why
This is the foundation OAuth flow that lets free-tier users authenticate
without managing a long-lived key. Refresh + auto-retry means CLI
commands keep working past the access_token lifetime without bugging
the user.
The OAuth client_id (`q2A2QRSke2LrFTPJhoDbHtXh`) is the one James
created in the `oauth2_client` table. Baked in as a build-time default;
override via `HYPERFRAMES_OAUTH_CLIENT_ID` for dev/test.
## How
- Public client: PKCE only, no `client_secret`. Backend already
requires PKCE (`movio/logic/oauth2.py:638`).
- Loopback port is ephemeral (`server.listen(0)`) — the backend
wildcards localhost ports for public clients
(`movio/model/oauth2.py:check_redirect_uri`), so the registered
redirect URI's port is a placeholder.
- State parameter is generated per-flow + validated on callback to
prevent CSRF.
- Token-response parsing is permissive on `expires_in` type (some
servers return it as a string) but strict on `access_token` presence.
- 401 retry happens at the `AuthClient.fetchUser` layer, not the
command layer — so future endpoints inherit it for free.
- `persistOAuth` merges into the existing store (preserves co-located
`api_key`). `auth login` (API-key path) does the symmetric thing.
## Test plan
- [x] 80 unit tests, all green. `vitest run src/auth/`.
- [x] PKCE: verifier within 43-128 chars, challenge = SHA-256, S256
method, distinct outputs each call.
- [x] Loopback: state mismatch / IdP error / missing-code / timeout /
404 non-callback paths all rejected; success path captures `code`.
- [x] OAuth: `refreshTokens` posts correct body, persists, throws
`REFRESH_FAILED` on 400/401 and `API_ERROR` on 5xx. Existing
api_key preserved on refresh.
- [x] AuthClient: 401 retries with refreshed bearer on OAuth, does
NOT retry for api_key, returns 401 if refresh hook fails.
- [x] `bunx oxlint` / `bunx oxfmt --check` / `bunx tsc` clean.
- [x] `bunx fallow audit --base origin/main --fail-on-issues` — only
inherited `help.ts:showUsage` finding (from main, not this PR).
- [ ] Smoke test against dev API:
`HEYGEN_API_URL=https://api.dev.heygen.com hyperframes auth login`
then `hyperframes auth status` then `hyperframes auth refresh`.
## Out of scope
- Cloud render commands — separate plan.
- PR 4 (heygen-cli read-side JSON support) — independent, ships after.
## What
Introduces the `hyperframes auth` command group + a shared credential
store library that hyperframes-CLI and heygen-cli will both read from.
- `hyperframes auth login --api-key` saves a HeyGen API key to
`~/.heygen/credentials.json` (stdin pipe or hidden-input prompt).
- `hyperframes auth status` resolves the active credential (env vars
→ file) and verifies it against `GET /v3/users/me`, printing
identity + billing.
- `hyperframes auth logout` removes the credential (`--keep-api-key`
drops only the OAuth block).
Internals (`packages/cli/src/auth/`):
- `paths.ts` — `~/.heygen` layout, `HEYGEN_CONFIG_DIR` override.
- `store.ts` — read/write `credentials.json` (file 0600, dir 0700)
with legacy single-line plaintext fallback so existing heygen-cli
users don't lose their session.
- `resolver.ts` — chain: `HEYGEN_API_KEY` → `HYPERFRAMES_API_KEY` →
file (unexpired OAuth wins over api_key).
- `client.ts` — hand-written typed wrapper for `GET /v3/users/me`
(intentionally not OpenAPI codegen — single endpoint).
- `errors.ts` — typed `AuthError` with discriminating `code`.
## Why
This is the foundation for `hyperframes cloud render`. Splitting it
out keeps the cloud-render PR small and lets users sign in today.
The plan originally called for a library-only PR followed by a
commands PR. The `fallow` dead-code gate flagged the library-only
shape as unused exports, so I bundled them — the library and its
first consumers ship together. PR 3 (OAuth PKCE) and PR 4
(heygen-cli read-side JSON support) follow.
## How
- Credential file format: JSON with optional `api_key` + `oauth`
blocks. Both CLIs read it; the resolver picks the freshest valid
credential.
- Auth header selection happens in the HTTP client: OAuth →
`Authorization: Bearer ...`, API key → `x-api-key: ...`.
- `HEYGEN_API_URL` lets dev testing target `api.dev.heygen.com`
without rebuilding.
- The new `auth` command lazy-loads its subverbs (same pattern as
`lambda`).
## Test plan
- [x] Unit tests added (`vitest`) for paths, store, resolver,
client, and errors — 45 tests, all green.
- [x] `bunx tsc --noEmit -p packages/cli/tsconfig.json` clean.
- [x] `bunx oxlint` + `bunx oxfmt --check` clean.
- [x] `bunx fallow audit --base origin/main --fail-on-issues` —
zero new findings.
- [ ] Smoke test against dev API:
`HEYGEN_API_URL=https://api.dev.heygen.com hyperframes auth login --api-key`
then `hyperframes auth status`.
`process.on("beforeExit", ...)` re-fires every time the event loop
drains, and the handler kicks off a fire-and-forget async telemetry
flush — so on a successful command the user sees the
"Update available: …" notice twice (once after the initial drain, again
after the flush settles). Using `process.once` detaches the listener
after first invocation, fixing the double-print and also preventing a
double-flush of telemetry.
Reported during local testing of `auth login`, but the bug affects every
command (any path where `_flush()` schedules work).
When forward playback reaches loopEnd and the loop wraps back to
loopStart, the RAF tick was calling `adapter.seek(loopStart)` without
keepPlaying, then immediately `adapter.play()` to resume. With the
post-3e7b464b wrapTimeline contract (default seek pauses), this means
every loop boundary executes pause→seek→pause→play for GSAP and a
stop/start RAF ticker cycle for the static-seek adapter — purely
unnecessary churn.
Pass { keepPlaying: true } so seek skips the implicit pause; the
follow-up adapter.play() is then a no-op because the underlying
adapter never paused. Adds two tests covering the wrap-around branch
(previously uncovered) and the no-loop terminal path as a regression
guard.
Completes the keepPlaying rollout: #842 introduced the option for A/E
shortcuts, #863 extended it to the runtime player, #1089 aligned the
static-seek adapter, and this applies it to the last internal caller
that explicitly resumes after seek.
Co-authored-by: Carlos Alcaraz <193642530+calcarazgre646@users.noreply.github.com>
* fix(studio): add FFmpeg pre-flight check before starting render
Studio renders now fail fast with a 422 and an actionable FFmpeg
install hint instead of burning through the entire capture pipeline
before hitting "spawn ffmpeg ENOENT" at encode.
* fix(studio): address review — use 503, memoize FFmpeg lookup
* fix(producer): normalize error messages to prevent [object Object] in telemetry
When a render fails and the caught value is a plain object (not an Error
instance), String(error) produces [object Object], masking the real error
in PostHog telemetry (~24 errors/day).
Add normalizeErrorMessage() that tries Error.message, string passthrough,
.message on plain objects, JSON.stringify, and String() as a last resort.
Apply it on the two telemetry-feeding paths: the main render failure
handler (renderOrchestrator.ts:2099) and buildRenderErrorDetails
(cleanup.ts), plus the error classifier isRecoverableParallelCaptureError
so timeout detection works even when the thrown value is a plain object.
* fix: address review — normalize CLI telemetry path, captureCost fallback
* fix: use local normalizeErrorMessage in CLI to avoid cross-package resolution
The Vite test runner can't resolve runtime imports from @hyperframes/producer
since its exports point to dist/. Copy the utility into the CLI package and
import locally instead.
* fix(core): guard timeline method calls for non-conformant objects
User compositions can register timeline-like objects on window.__timeline
where .duration is a number property (not a function) and .pause/.play
may be missing entirely. The runtime player called these unconditionally,
causing ~166 "duration is not a function" and ~38 "pause is not a function"
errors per day.
Add safeNum() and safeVoid() helpers that check typeof before calling,
falling back to reading numbers as properties and silently skipping
missing void methods. Applied consistently across all timeline method
call sites in player.ts.
* fix(core): add observability for non-conformant timeline properties
* fix(studio): compensate GSAP translate when starting manual drag
When an element has an active GSAP transform with translate (x/y),
starting a drag via createManualOffsetDragMember would strip the
GSAP translate from element.style.transform during the probe phase
without accounting for it in the initial offset. This caused the
persisted manual offset to be wrong by exactly the GSAP translate
amount, producing a visible position shift after page reload.
Read the GSAP translate contribution (m41/m42 from the transform
matrix) and fold it into initialOffset before the probe runs. The
offset now compensates for the stripped translate, so the element's
visual position is preserved across the drag start, commit, and
subsequent reloads.
* fix(studio): show visual position in Layout panel and fix save-reload race
PropertyPanel: X/Y fields now display the visual position (manual offset
+ GSAP translate) instead of the raw CSS var offset. Editing a value
reverses the compensation so the correct raw offset is persisted. This
matches what the user sees in the preview during GSAP playback.
persistDomEditOperations: move domEditSaveTimestampRef update before the
patch API call. The server writes the file and emits an SSE file-change
event during the fetch — if the event arrived before the response, the
file watcher would trigger a spurious reloadPreview(), resetting
playback to t=0. Setting the timestamp upfront suppresses that race.
* fix(studio): apply same timestamp race fix to element delete, relocate helper
Move readGsapTranslateFromTransform to manualEditsDom.ts alongside its
sibling stripGsapTranslateFromTransform and re-export through the
manualEdits barrel. PropertyPanel and manualOffsetDrag now import from
the shared location instead of the drag module owning a display concern.
Move domEditSaveTimestampRef update before the remove-element fetch in
handleDomEditElementDelete — same SSE race as persistDomEditOperations.
* fix(engine): use captureBeyondViewport on all CDP screenshot paths
Chrome's compositor rounds the viewport boundary inward under multi-tab
load, clipping the bottom/right edge of tall portrait compositions
(1080x1920). The explicit clip rect already constrains output to exact
composition dimensions, making the viewport-boundary pre-clip from
captureBeyondViewport:false both redundant and unreliable.
Set captureBeyondViewport:true on all three CDP screenshot call sites:
pageScreenshotCapture, captureScreenshotWithAlpha, and captureAlphaPng.
Add portrait-edge-bleed regression test: 1080x1920 grid with bright
magenta bottom rows, rendered with 4 workers. Any compositor clipping
at the bottom edge drops PSNR sharply against the golden baseline.
Closes#1009
* fix(engine): address review feedback on captureBeyondViewport
- Add backref comments on captureScreenshotWithAlpha and captureAlphaPng
pointing to pageScreenshotCapture for the rationale, so the next reader
doesn't treat the flag as unintentional copy-paste
- Note in test meta.json that the static grid fixture covers the
capture-side clipping path but not the video-element compositor surface
timing that produces the t≈37s self-healing in #1009
* test(producer): use video element in portrait-edge-bleed regression test
Replace the static CSS grid with a 1080x1920 portrait video element —
matches the original bug report shape where the compositor surface
allocation timing causes the bottom-edge clipping. The video has a dark
top region and bright magenta bottom 480px, so any viewport clipping at
the bottom edge drops PSNR sharply. Baseline regenerated in Docker with
4 workers.
* feat(core): add probeElementInSource for source-existence checks
* feat(core): add probe-element endpoint for source-existence checks
* feat(studio): gate editing capabilities on source existence
* fix(studio): enrich save_failure telemetry with target details
* feat(studio): async selection resolution with source probe
Make `resolveDomEditSelection` async and wire a `probeSourceElement` call
into the selection path so elements generated by scripts (not present in the
source HTML) are detected early and have all edit capabilities disabled with
a clear reason message ("This element is generated by a script and cannot be
edited visually.").
Part A – core probe logic:
- `domEditingLayers.ts`: `resolveDomEditSelection` is now async; calls
`probeSourceElement` (POST /api/projects/:id/file-mutations/probe-element/:file)
when `projectId` is supplied and the element has a stable id/selector.
`existsInSource: false` flows into `resolveDomEditCapabilities`, which
disables all write capabilities with the appropriate reason.
- `domEditingLayers.ts`: `refreshDomEditSelection` promoted to async.
- `files.ts`: new `probe-element` route; extracted `resolveProjectPath`,
`resolveFileMutationContext`, `writeIfChanged`, and `parseMutationBody`
helpers to eliminate repeated boilerplate across remove/patch/probe handlers.
Part B – caller propagation (all eight consumer sites):
- `useDomSelection.ts`: `buildDomSelectionFromTarget`,
`resolveDomSelectionFromPreviewPoint`,
`buildDomSelectionForTimelineElement`, `handleTimelineElementSelect`,
`refreshDomEditSelectionFromPreview`, and
`refreshDomEditGroupSelectionsFromPreview` all made async; `projectId`
forwarded into `resolveDomEditSelection`.
- `useDomEditCommits.ts`, `useDomEditTextCommits.ts`: updated
`buildDomSelectionFromTarget` parameter type; added `await` at call sites.
- `useDomEditSession.ts`: inner `syncSelectionFromDocument` made async; fire
with `void` to satisfy the surrounding effect.
- `usePreviewInteraction.ts`: `handlePreviewCanvasMouseDown` and
`handlePreviewCanvasPointerMove` made async (React ignores handler return
values, so this is safe).
- `useStudioUrlState.ts`: deferred `buildDomSelectionFromTarget` call
converted to `.then()` chain with `void` prefix so the effect stays sync.
- `LayersPanel.tsx`: `seekToLayer`, `handleSelectLayer`, and
`handleLayerHover` made async.
- `DomEditOverlay.tsx` / `useDomEditOverlayGestures.ts`: `onCanvasPointerMove`
return type widened to `Promise<DomEditSelection | null>`; pointer-down
handler falls back to `hoverSelectionRef.current` (always populated by a
prior hover) instead of awaiting the async move callback inline.
Part C – test and tooling fixes:
- `lefthook.yml`: filesize hook shell loop explicitly skips `*.test.ts/tsx`
files as a guard against a lefthook v2.1.6 bug where `exclude` patterns are
not applied to `{staged_files}` in shell scripts.
- `domEditing.test.ts`: all `it()` blocks calling `resolveDomEditSelection`
made async with `await`.
- `DomEditOverlay.test.ts`: mock updated to return `Promise.resolve(selection)`
and `hoverSelection` pre-seeded so pointer-down test works with the new
hover-first path.
- `studioUrlState.test.ts`: `buildDomSelectionFromTarget` mocks wrapped in
`Promise.resolve()`; seek/selection hydration test made async with
`await act(async () => { await Promise.resolve(); })` to flush microtasks.
* feat(cli): add global error handlers for crash telemetry
Register process-level uncaughtException and unhandledRejection handlers
that fire trackCliError so unhandled crashes are captured in telemetry.
Add the trackCliError function to events.ts and re-export it from the
telemetry barrel.
* feat(cli): track per-command success/failure and duration
* test(core): add integration test for JS-created element probe scenario
* fix: address PR review feedback
- uncaughtException handler now calls process.exit(1) after flushing
- cli_command_result uses real exit code from process "exit" event
- drop stack_trace from cli_error (contains filesystem paths)
- skip source probe during hover — only probe on click/selection
- format .fallowrc.jsonc
* fix(cli): restore stack_trace in cli_error telemetry
* fix(cli): use captured module refs in exit handlers instead of dead import()
createStaticSeekPlaybackAdapter.seek now accepts the same options as the
PlaybackAdapter contract and aligns the default-pause semantics with
wrapTimeline (hardened in 3e7b464b). Without keepPlaying the adapter
clears its `playing` flag and cancels the RAF ticker, so on non-GSAP
compositions a scrub during playback no longer leaves the iframe
silently advancing while the public seek wrapper marks isPlaying=false.
Follow-up to #863 review: jrusso called out the type drift and invited
a separate PR; this also closes the asymmetry with wrapTimeline.
Co-authored-by: Carlos Alcaraz <193642530+calcarazgre646@users.noreply.github.com>
* test(producer): add parallel capture regression test
Add a regression fixture that forces workers: 2, ensuring the parallel
capture code path (browser-per-worker in BeginFrame mode) is exercised
in CI. All existing fixtures pin workers: 1, so this is the first test
that would catch a regression in the multi-worker pool isolation fix
from PR #1087.
The composition is 5s @ 30fps (150 frames), which exceeds both
MIN_FRAMES_PER_WORKER * 2 (60) and minParallelFrames (120), so the
parallel coordinator will always split work across workers.
Baseline output/output.mp4 must be generated inside Dockerfile.test
before the fixture can run in CI.
* test(producer): bump parallel capture test to 4 workers
Matches realistic auto-mode worker counts (4-6 on typical machines),
not just the minimum (2) that triggers the bug.
* test(producer): add golden baseline for parallel capture regression
Generated inside Dockerfile.test on amd64 Linux (Docker image
hyperframes-producer:test) to match the CI rendering environment.
* test(producer): address review feedback on parallel capture test
- Add fixture to shard-5 in regression.yml so CI actually runs it
- Reframe description: multi-worker path coverage (frame distribution,
reorder buffer, per-worker browser lifecycle), not GPU-specific crash
guard — SwiftShader CI can't reproduce the hardware compositor race
- Remove dead @keyframes count-up (content doesn't apply to div)
- Remove unused CSS animation reference on .counter
- Regenerate golden baseline with cleaned-up HTML
* fix(producer): replace rAF + CSS keyframes with GSAP in parallel-capture test
The composition used requestAnimationFrame for a frame counter and CSS
@keyframes for animations, which triggered screenshot capture mode
(non-deterministic across workers) and caused 29 PSNR failures in CI.
All animations now use the GSAP timeline, keeping the render in
deterministic BeginFrame mode. Baseline regenerated in Docker.
* fix(engine): disable browser pool for parallel capture in BeginFrame mode
BeginFrame's compositor is process-global — when multiple pages in the
same Chrome instance drive HeadlessExperimental.beginFrame concurrently,
they race the compositor and crash with "Protocol error: Target closed".
Only disable the pool when BeginFrame mode would actually be active
(Linux + headless-shell + not forceScreenshot). Screenshot mode
(macOS/Windows) is unaffected and keeps the pool for memory efficiency.
Also extracts the frame capture loop into captureFrameRange to reduce
function complexity in executeWorkerTask.
* fix(engine): include supersampling in BeginFrame-mode predicate
Match the full capture-mode predicate from createCaptureSession:
DPR > 1 (supersampling) forces screenshot mode, which is pool-safe.
Without this check, supersampled parallel renders on Linux would
unnecessarily launch separate browsers.
Fix the 4th unguarded resolveStartForElement call site in
resolveMediaWindowDurationSeconds that inflated the timeline duration
floor for pip compositions. Extract resolveMediaStartSeconds helper
to consolidate the data-hf-auto-start guard across all call sites.
Renders a 6s composition with a pip video inside a sub-composition
host starting at t=3. Verifies the video is visible during the host
window and not double-offset to t=6. Baseline generated in Docker.
Narrow the raw data-start read to media elements without
data-hf-auto-start (explicitly authored global coordinates). Elements
with auto-injected data-start="0" remain composition-local via the
resolver. Apply consistently across all three consumers:
- visibility loop (init.ts)
- refreshRuntimeMediaCache start/duration (init.ts)
- resolveMediaWindowEndSeconds (timeline.ts)
Add regression test for auto-injected data-start="0" inside a
late-starting host to prove it doesn't regress.
For video and audio elements, data-start is authored in global (composition-root)
time — the same contract used by the render pipeline's discoverMediaFromBrowser,
which reads the raw attribute directly. Previously, the visibility loop called
resolveStartForElement which adds the nearest ancestor composition's global start
on top, causing a double-offset that kept pip-wired media permanently hidden when
the host composition did not start at t=0.
Example: a pip video with data-start="45.40" inside a host composition that also
starts at data-start="45.40" resolved to 90.80, so the video was always hidden
during its actual [45.40, 52.46] window.
Non-media elements (divs, sections, etc.) continue to use the accumulating
resolver because their data-start values are local to their composition.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
`injectVideoFramesBatch` now returns `Promise<string[]>` so the caller can
filter cache entries to videos the page actually painted. The cli-side
snapshot command does not use the return value, but its local `InjectFn`
declared `Promise<void>` which made the `as { injectVideoFramesBatch:
InjectFn }` cast on the dynamic engine import fail typecheck under TS's
"sufficiently overlapping types" rule. Match the engine's actual export
shape.
Two follow-ups to the ancestor-visibility skip in `injectVideoFramesBatch`
and `syncVideoFrameVisibility`.
1. **Mask defence.** Both ancestor-hidden branches previously wrote a plain
`img.style.visibility = "hidden"`. `applyDomLayerMask` writes the
stylesheet rule `#${showId} *{visibility:visible !important}`, and CSS
cascade puts important stylesheet author above non-important inline
author — so a sub-comp host landing in the active layer's `show` set
would revive a stale `__render_frame__` and let it bleed onto the
layer composite. Write the hide via
`style.setProperty("visibility", "hidden", "important")` instead;
important inline beats important stylesheet.
2. **Caller cache hygiene.** `createVideoFrameInjector` unconditionally
wrote `lastInjectedFrameByVideo.set(id, frameIndex)` after calling
`injectVideoFramesBatch`, even for videos the page silently skipped due
to a hidden visual ancestor. On the next call at the same frameIndex —
common with source-fps < output-fps, paused source frames, or
non-frame-aligned host starts — the cache short-circuited the second
inject and the host's first visible frame painted blank because the
replacement `<img>` was never created.
Make `injectVideoFramesBatch` return `string[]` (the subset of ids it
actually painted) and have the caller cache only those. The cli-side
`snapshot.ts` consumer is unaffected: its local `InjectFn` types the
return as `Promise<void>`, which is structurally compatible with
`Promise<string[]>` under TS void-return assignment rules.
Tests: linkedom doesn't preserve `!important` in cssText, so the two new
mask-defence cases spy on the live `<img>`'s `style.setProperty` and assert
the 3-arg call shape. The cache-hygiene case stubs the page-side primitives
via `vi.mock`, drives the hook twice at the same frameIndex with a stubbed
"injected nothing" first response, and verifies the second call still
issues an inject. A counter-test pins the happy-path cache hit so a future
refactor can't trade the skip bug for a never-cache regression.
`isVisualAncestorHidden` was treating any `visibility: hidden` ancestor as a
signal to skip injecting the replacement frame. That's too broad — for plain
`[data-start]` containers, the replacement `<img>`'s explicit
`visibility: visible` correctly overrides the ancestor per CSS spec, and
consumers rely on that to hold the final GSAP-driven frame when an authored
`data-duration` outlives the composition's GSAP timeline (e.g.
`style-9-prod`, where the runtime truncates the host to `visibility: hidden`
after the timeline ends and the replacement frame must paint through).
Restrict the `visibility: hidden` skip to ancestors that carry
`data-composition-src` or `data-composition-file` — the actual sub-composition
hosts this guard was added for. `display: none` keeps the broad behavior:
it takes the whole subtree out of layout and a child override cannot escape.
Update the existing regression suite to mark the host as a sub-composition,
and add two new cases pinning the plain-`[data-start]` behavior: both
`injectVideoFramesBatch` and `syncVideoFrameVisibility` must still produce a
visible replacement `<img>` when the host is `visibility: hidden` but does
not carry a sub-composition attribute.
The screenshotService.test.ts regression-suite comment pointed at the
author's fork branch as backstory. Strip the line so upstream code
doesn't carry a fork-relative reference; the surrounding paragraph
already explains the bug end-to-end without it.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`injectVideoFramesBatch` and `syncVideoFrameVisibility` iterate every
`video[data-start]` whose raw time window covers the current seek.
Inner `<video>` elements inside `[data-composition-src]`
sub-compositions get `data-start="0"` auto-injected by
`compileTimingAttrs` and probed-duration cover the entire timeline,
so they look "active" even when their host has not yet started.
When the runtime then hides the host with `visibility: hidden` (its
out-of-window lifecycle), the inner video inherits hidden via the CSS
cascade — but our injector responded by painting a replacement
`<img class="__render_frame__" style="visibility: visible">` next to
the video. `visibility: visible` on the descendant defeats the parent
`visibility: hidden` cascade, and because the host has not been
morphed by GSAP yet the video's bounding box is its CSS default
(usually full-bleed). The result is one full-bleed frame per inactive
sub-comp painted over whichever moment is *actually* visible — the
overlay symptom the upstream agentic-finecut project saw.
Walk ancestors in both functions; if any has `display: none` or
`visibility: hidden`, skip the inject and hide any stale
`__render_frame__` sibling. The render is now correctly empty for
hidden hosts, which is what the surrounding CSS cascade already
intends.
Tests:
- `screenshotService.test.ts`: cover the new guard for both
visibility:hidden and display:none hosts, both for the fresh-img and
the stale-img paths, plus `syncVideoFrameVisibility` for the case
where the time window calls a video "active" but a hidden ancestor
still requires its frame to stay hidden. Each test fails against
pre-fix `screenshotService.ts`.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
elementRect.left/top from getBoundingClientRect() already reflects GSAP
transforms in viewport coordinates. Subtracting rootRect.left/top
cancels the transform, pinning overlays to the un-animated layout
position. Use elementRect directly so overlays track elements during
scroll (y: -500) and entrance (scale: 0.95) animations.
When GSAP applies transforms (scale, translate) to the root composition
element during playback, rootRect.width/height from getBoundingClientRect()
changes to reflect the transformed size. The overlay scale calculation
(rootScaleX/Y = iframeRect / rootRect) then produces wrong values,
causing overlays to appear at incorrect positions during animated
playback — especially visible during scroll animations (y transform)
and entrance animations (scale transform).
Fix: use the composition's declared data-width/data-height attributes
for scale calculation. These are the canonical dimensions that don't
change with GSAP transforms. Falls back to rootRect dimensions when
the attributes aren't present (non-composition elements).
When outPoint exceeds composition duration, rawLoopEnd > dur makes the
time >= loopEnd branch unreachable after the playhead clamp — the player
ticks forever. Clamp rawLoopEnd to dur in both forward and backward RAF
loops, matching the seek() clamping. Add test for the boundary behavior.
Trim blank lines to satisfy 600-line filesize gate.
Miguel Ángel