processCompositionAudio's per-element loop races each track's async chain
concurrently via Promise.all, which rejects as soon as the first element
fails without waiting for siblings still in flight. A VstChainProcessingError
from one track therefore let another track's VST sidecar subprocess (up to
BOUNCE_TIMEOUT_MS = 10 minutes) keep running after the finally block deleted
workDir out from under it.
Thread an internal AbortController's signal through every element's chain
(extract/prepare/applyVstChainToWav), aborting it in a catch around
Promise.all before workDir is removed, and abort it on the caller's own
external signal too. applyVstChainToWav now accepts { signal } and kills its
spawned sidecar on abort, following runFfmpeg's existing spawn-abort
convention.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A missing VST plugin or chain file was caught by the per-element try/catch
shared with every soft-degrading failure mode (missing source asset,
download failure, etc): the track was silently dropped and MixResult.success
stayed true, so a render with a missing plugin completed successfully with
no visible error anywhere.
Introduce VstChainProcessingError to mark VST-specific failures and let the
per-element catch rethrow it instead of swallowing it into the warnings
list. The rethrow escapes the Promise.all and rejects
processCompositionAudio itself, which (traced through audioStage.ts's
uncaught await and observability.ts's observeRenderStage, which rethrows)
surfaces as a real render failure through the CLI's existing error-handling
path — the same one every other hard pipeline failure already uses. Every
non-VST failure mode is untouched.
Fixed the two "hard-fails" tests to assert the promise actually rejects
(rejects.toThrow) instead of only checking tracksProcessed/error on an
still-resolved, still-successful result — the previous assertions would
have passed even with the bug present.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Spawns the hyperframes-vst sidecar (packages/vst-host) to bounce a
track's dry WAV through its VST chain before the volume-envelope bake
in processCompositionAudio. A missing plugin or sidecar failure is a
hard failure for that track (named plugin + track id), never a silent
fallback to unprocessed audio.
An unrelated client sending {"cmd":"transport","action":"pause"} could
cancel another connection's active playback task, the same class of bug
the disconnect and play-takeover paths already guard against via
self._play_owner. pause now only cancels/clears the play task when the
sending connection is the current owner; otherwise it is a no-op.
Adds a regression test mirroring the existing disconnect-ownership test.
Track which websocket connection owns the running play task so an
unrelated client disconnecting no longer cancels another client's
active playback. Broaden _dispatch's exception handling so any error
(not just PluginMissingError) — e.g. a KeyError from a command
referencing a not-yet-loaded/unloaded/unknown trackId — replies with
{"event":"error","code":"bad_command"} instead of crashing the
connection with a 1011 close.
Adds two TDD regression tests reproducing both bugs against the
pre-fix code before implementing the fixes.
Adds server.py (VstServer) implementing the full control protocol over one
websocket: load/unload-chain, set-param, get-state, scan, open/close-editor,
and transport (play/pause/seek) with realtime-paced binary PCM streaming via
Task 4's TrackStream/wire-frame codec. Wires up `hyperframes-vst serve
[--port N]` in __main__.py. Adds pytest-asyncio dev dependency and
asyncio_mode = "auto" pytest config for the new async websocket tests.
Moves assetResolution.ts from packages/lint into @hyperframes/parsers behind
a ./asset-resolution subpath export, mirroring the ./ff-binaries shape, and
points lint's two importers at it. Pure move: no behavior change. Unblocks
studio-server consuming the same helpers without a dependency cycle (the CLI
already depends on lint, so importing lint from studio-server is not an option).
The cli, engine, and lint packages each carried their own copy of the
ffmpeg/ffprobe lookup, annotated fallow-ignore code-duplication, and the
copies had drifted: the engine copy handled Windows PATHEXT and executed
which/where without a shell but lacked the Homebrew-dirs fallback for
GUI-spawned processes; the cli copy had the opposite. One resolver in
@hyperframes/parsers (the dependency-graph bottom) now carries the union
of both hardenings, and all three packages delegate to it. Every
consumer gets strictly more robust resolution; env-override semantics
per call site are preserved via configuredMustExist.
A composition variable mirrored as --<slug> for a mounted sub-composition
(default or an explicit data-variable-values value) previously overrode any
same-named custom property the document already authored elsewhere (e.g. a
:root theme token), since the mirroring had no "already defined" guard —
unlike the two other emission paths, which already skip re-emitting when the
name collides with an authored definition. Extend that guard to the
sub-composition mount path in both the compiler (htmlBundler.ts) and the
runtime loader (compositionLoader.ts / getVariables.ts), so an authored
definition always wins; render-time --variables overrides still always win.
Auto-generated weekly digest (bun run changelog:weekly), hand-cleaned to
publish quality. Grouped into Features / Fixes / Performance / Docs &
Examples with readable one-liners and commit + PR links.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Packages Jake Moran's changelog-video pipeline (v1, validated end-to-end
by Home on the Jun 23-29 range) as a repo-native skill set that Claude
Code (.claude/skills/) and Codex CLI (.agents/skills/) auto-discover the
moment the repo is opened. No install step; run the skill against a
changelog markdown for a given git range and it produces a lint-clean,
seam-gate-green 1080x1080 MP4 (~45-60s, Annie VO, mock-UI visualizations,
caption rail) end-to-end.
Six skills added byte-identical in both mirror dirs:
- changelog-video (pipeline entry point)
- motion-doctrine (carries seam-stamp.mjs + seam-gate.mjs)
- cut-the-curve, captions-overlay, seam-craft, oversized-cursor
Layout:
- .claude/skills/ - Claude Code project-local auto-discover
- .agents/skills/ - Codex CLI project-local auto-discover (verified via
Magi's clean-home Codex 0.144.3 repro; NOT .codex/skills/)
Fonts, animated background (12 MB), house BGM (5 MB), lexicon, and
align-captions ship inside the skill dirs. .gitattributes routes only
.claude/skills/**/*.{mp4,mp3} + .agents/skills/**/*.{mp4,mp3} through
LFS — narrowly scoped so unrelated Player, Studio, registry, and
marketplace media stay put. HeyGen CLI auth is the one credential the
skill needs; Node >= 22, ffmpeg, and headless Chrome are documented
alongside in both READMEs.
.gitignore: rewrites .claude/ and .agents/ blocks to keep agent-installed
skill hygiene while re-including the six repo-native skill dirs plus
README.md.
CI:
- Extends changes.skills filter to match .claude/skills/**,
.agents/skills/**, scripts/lint-skills.ts, and scripts/check-skill-mirror.mjs.
- New 'Skills: project-native lint + mirror' job runs the extended
lint-skills.ts (schema-driven; required { name, description } + optional
{ license, allowed-tools, metadata }, name pattern check, description
length check) plus a new check-skill-mirror.mjs byte-integrity script
(24 mirrored files must match; README.md deliberately per-CLI).
- Wired into 'bun run lint' locally.
Frontmatter validator:
- Rejects unsupported top-level keys (catches category:-style drift).
- Requires name + description.
- Validates name pattern (^[a-z][a-z0-9-]{0,63}$) and description shape
(non-empty, <=1024 chars).
- Missing frontmatter block itself is a first-class error.
Also strips unsupported top-level 'category:' frontmatter from Jake's
motion-doctrine and cut-the-curve SKILL.mds (both mirrors), rewrites the
TTS invocation from ~/.claude/skills/media-use/... to the tracked
skills/hyperframes-media/scripts/heygen-tts.mjs, swaps npx hyperframes@latest
for the repo-local CLI in the gate step, and fixes a lint issue in Jake's
seam-gate.mjs (ternary-for-side-effect -> if/else).
Validated end-to-end by Home on Jun 23-29 (MP4 posted in C0ACCNHLG3U
thread 1784181166.041319). Independently reviewed R1/R2/R3 by Magi.
Co-authored-by: Jake Moran <jake@heygen.com>
## What changed
- adds a stage-aware review guide for plan, sketch, animation, and final-review states
- makes the review loop explicit: review frames → save feedback → reply in agent chat
- adds copyable agent handoff and approval messages
- keeps frame status agent-owned and presents planned frames before previews exist
- clarifies voiceover ownership, save state, feedback errors, and preview availability
- improves responsive layout and surfaces storyboard warnings without dominating the board
- adds unit coverage for review-stage and handoff-step derivation
## Why
Storyboard reviewers could leave comments, but the UI did not clearly explain when a pass was ready, what to inspect, or how saved feedback reaches the agent. This makes the human-to-agent iteration loop visible while preserving the existing file-based handoff.
## Impact
Users get a clear next action throughout the storyboard lifecycle and can copy the exact message needed to continue the agent workflow. The underlying storyboard and frame-comment formats are unchanged.
## Validation
- `bun run --filter @hyperframes/studio test --run src/components/storyboard/storyboardReviewStage.test.ts` — 13 tests passed
- `bun run --filter @hyperframes/studio typecheck`
- pre-commit: tracked artifacts, lint, formatting, Fallow audit, typecheck
* docs(skills): make the core set the default install on every surface
A field test showed an agent with a real 'make videos' intent installing
all 19 skills: at install time, every surface it could read pointed at the
full set, while the core-eager / workflow-on-demand design only exists
inside hyperframes/SKILL.md - unreadable until after the install decision.
Two traps made full-install the documented default:
- The README Quick Start used 'skills add --yes': skills.sh force-detects
agent environments into non-interactive mode, and a non-interactive run
without --skill installs all 19. Dropping --yes fixes the human path
(the picker opens with nothing pre-selected), but not the agent path.
- marketplace.json listed the full 'hyperframes' bundle first, under the
name an agent installing 'hyperframes' matches; core-skills sat second.
Changes, each behavior verified by an isolated run:
- README Quick Start drops --yes (humans get the picker; verified via a
pty capture that nothing is pre-selected) and points agents and
non-interactive runs at 'npx hyperframes skills update', which from a
clean HOME installs exactly the 8 core skills, refreshes stale ones and
prunes unpublished ones on an existing machine, and is idempotent.
- CLAUDE.md and the docs install guide lead with the same one-liner;
--all is reworded to explicit-request-only at every surface.
- marketplace.json puts core-skills first and both descriptions steer the
default choice; the full entry keeps auto-discovery (no allowlist),
per the skillsManifest core-pin test (56/56 pass).
* docs(skills): close the same install trap in AGENTS.md
Review follow-up on the core-default change: AGENTS.md still carried a
bare 'npx skills add heygen-com/hyperframes' - no --full-depth, and the
same non-interactive-installs-everything trap the README fix closed.
AGENTS.md is the first file Codex/Cursor-family agents read for repo
intent, so it leaked the full-set default to exactly the readers the
core-default policy targets. It now leads with the same core-set
one-liner and policy line as CLAUDE.md.
* feat(skills): probe and bump stale CLI pins during project resume
The entry skill now keeps a resumed project's pinned CLI current instead of
leaving that to a notice nobody acts on. On resuming a project with pinned
scripts, run the read-only probe 'npx hyperframes@latest upgrade --project
. --check'; when it (or the stale-pin stderr notice, or _meta.updateAvailable
from a pinned run) reports the project behind, apply the bump and verify
with 'hyperframes check'. A failed check reverts the bump and keeps the
project on its pinned version, preserving the reproducibility contract the
pin exists for.
The probe matters because the stale-pin notice only exists in >= 0.7.59:
a pinned run of an older CLI prints no warning at all, so a notice-only
trigger never fires for exactly the projects most behind. The probe runs
unpinned, so its behavior never depends on the project's CLI age.
Telemetry: the fleet converges to new releases within about a week via the
background auto-updater and ephemeral npx, but pinned projects form a
persistent stale tail (~10% of weekly actives, e.g. 6.3k users still on
0.6.x three weeks after 0.7.0).
Both skill surfaces now pass an explicit dir ('--project .') because a bare
'--project' followed by another flag consumes that flag as its directory
value and no-ops; the parsing fix is a separate CLI change.
* fix(cli): stop bare --project from eating the next flag as its directory
citty parses --project as a string option, so 'upgrade --project --check'
arrived with project="--check": the dir resolved to a nonexistent path and
the command no-opd with 'No package.json found' while --check was lost.
The documented default-cwd behavior only worked when --project was the
final token — and the trap-prone form is exactly what the scaffolded
template CLAUDE.md instructs.
A leading dash can never be a real directory argument, so resolveProjectArgs
now reclaims the eaten token as the flag the user wrote (--check / --json),
falls back to the current directory, and drops unrelated eaten flags rather
than treating them as paths. Templates and skill references switch to the
explicit-dir form ('--project .'), which behaves correctly on every release
including ones that predate this fix.
* feat(skills): report a successful pin bump in the run summary
Review follow-up on the stale-pin rule: 'hyperframes check' validates
composition structure, not render-output equivalence, so a check-passing
bump can still shift a project's rendered output. The bump stays the right
default for stale projects, but it must not be silent — the summary now
names the old and new version so the user knows the reproducibility
trade was made.