SKILL.md
5,500 tokens · o200k_base · 22,268 bytes
Source excerpt starting at line 1.---name: hunk-launch-videodescription: Produces Hunk videos by driving the real TUI headlessly in a PTY, compositing captioned 1080p frames in Chromium, and encoding with ffmpeg. Use for feature demos, workflow explainers, announcements, launch videos, and full-release roundups.--- # Hunk video pipeline Source-checkout only: the pipeline lives in `scripts/launch-video/`, which never ships to npm.It can be used for pull-request demos as well as maintainer release videos. Unix-only — thecapture scripts exec `/bin/bash`. Generates product videos where every terminal frame is the real Hunk TUI —no screen recording, no mockups. Three stages: ```textcapture.ts (bun) drive Hunk over a PTY, snap styled keyframes to PNGcompose.mjs (node + Playwright) render each PNG on a 1920x1080 HTML stage in Chromiumffmpeg encode the composited PNGs at 30fps to mp4/webm``` Playwright controls headless Chromium: for each planned frame it loads theterminal PNG onto the HTML stage, applies the window chrome, cards, captions,and transition state, then screenshots the completed stage back to PNG. ffmpegsequences those composited screenshots into the final videos. The generic machinery (PTY driving, keyframe rendering, storyboard planning,Chromium compositing, the stage template) is the `@hunk/term-video` workspacepackage in `packages/term-video/`; `scripts/launch-video/` holds only Hunk'sscenes, captions, and cards on top of it. ## Choosing a recipe The pipeline is not release-specific. Choose the editorial scope, then use thesame capture → composite → encode stages: - **Single feature:** a short demonstration of one capability or workflow. Use the single-feature recipe below and capture only the required scene.- **Full release:** a multi-feature roundup based on a release's changelog or highlights. Use the full-release recipe and update the canonical storyboard.- **Custom video:** author any set of scenes and `SHOTS` for tutorials, comparisons, announcements, or workflow explainers; follow the scene and storyboard rules below. ## Creating a video Expect ~3–6 min for capture and ~2–4 min for compose — run both with a longtimeout (or in the background); each logs per-snap / per-shot progress.`compose.mjs` needs node ≥ 18 on PATH (bun alone is not enough). ```sh# 0. dependencies (tuistory is a devDependency; ghostty-opentui arrives transitively)bun install # if a postinstall hook fails in a sandbox, retry with --ignore-scripts # 1. capture the current storyboard's default scenes. Output defaults to# <repo>/.video-work/ regardless of cwd (pass a path argument to override),# but the PROCESS must run from the repo root — see gotchas.bun run scripts/launch-video/capture.ts # 2. one-time portable compositor setup. Playwright installs a Chromium build# that exactly matches its browser driver (see gotchas to reuse a system or# sandbox browser instead).printf '{"name":"hunk-video-work","private":true}\n' > .video-work/package.jsoncd .video-workbun add playwright playwright-corebunx playwright install chromiumcd .. # 3. composite the storyboardnode scripts/launch-video/compose.mjs .video-work # 4. encodecd .video-workffmpeg -y -f concat -safe 0 -i concat.txt -vf "format=yuv420p" -r 30 \ -c:v libx264 -preset slow -crf 18 -movflags +faststart launch.mp4ffmpeg -y -f concat -safe 0 -i concat.txt -vf "format=yuv420p" -r 30 \ -c:v libvpx-vp9 -b:v 0 -crf 32 -row-mt 1 launch.webm``` Override the default scene set to iterate on one or more scenes: ```shSCENES=review bun run scripts/launch-video/capture.ts # comma-separated scene names``` Scene names are the `wants("...")` guards in `capture.ts`'s `main()`. `SCENES=`replaces the canonical default scene set; `compose.mjs` preflights that everyframe its `SHOTS` table references exists in `frames/` and fails fast listingany missing ones, so a full composite still needs every referenced scenecaptured at least once. ## Single-feature recipe For a short test, demo, or one-feature announcement, capture and composite onlythe scene for that feature: 1. Pick one user-visible capability and find its scene name in the `wants("...")` guards. If it does not have a scene, author one using the guidance below. Capture only that scene: ```sh SCENES=review bun run scripts/launch-video/capture.ts ``` 2. Make a scratch compositor beside the canonical one so its imports and repo-relative paths continue to work: ```sh cp scripts/launch-video/compose.mjs scripts/launch-video/compose-one-feature.mjs ``` 3. In the scratch copy, trim `SHOTS` to an opening card, only the selected feature's frames, and an outro card. Rewrite those cards and captions for the scoped cut. Sequence lengths must still match the captured frame names.4. Composite into the same work directory, then run the normal ffmpeg commands with descriptive output names: ```sh node scripts/launch-video/compose-one-feature.mjs .video-work cd .video-work ffmpeg -y -f concat -safe 0 -i concat.txt -vf "format=yuv420p" -r 30 \ -c:v libx264 -preset slow -crf 18 -movflags +faststart hunk-feature-demo.mp4 ffmpeg -y -f concat -safe 0 -i concat.txt -vf "format=yuv420p" -r 30 \ -c:v libvpx-vp9 -b:v 0 -crf 32 -row-mt 1 hunk-feature-demo.webm cd .. rm scripts/launch-video/compose-one-feature.mjs ``` Keep the scratch compositor uncommitted. The canonical `compose.mjs` remainsthe checked-in reference storyboard. If you added a capture scene only to makePR evidence, revert that scene after encoding; retain it only when it is usefulchecked-in demo coverage and belongs to the submitted change. ## Full-release recipe 1. Read the release section in `CHANGELOG.md`. If it has a hand-written **Highlights** list (0.18.0 has one; Changesets does not generate them), use that list as the storyboard. Otherwise distill 4–6 user-visible headlines from the Minor Changes — per-PR entries are too granular to shoot — and confirm the shortlist with the user before capturing.2. Rewrite the canonical storyboard's editorial surface (next section), adding or adjusting capture scenes as needed (see "Authoring scenes").3. Set `SCENES` to every scene referenced by the full storyboard, capture and composite it, then encode both formats using the main workflow above.4. Verify the complete cut (see "Verification") and deliver both files. When `hunk-release` invoked this workflow, return the approved MP4 and WebM to that skill for versioned naming and GitHub user-attachment embedding; do not upload or edit the public release without its confirmation gate. ## Per-video editorial surface The capture machinery is reusable, but the storyboard is editorial content forone video. Rewrite it to match the video's scope. The checked-in reference is asingle-feature Git-history video, not a frozen release artifact: - `compose.mjs`: the whole `SHOTS` table; opening, feature, and outro cards; every caption; camera targets; and callout rectangles. A `NEW` badge is a claim about the video at hand, so drop or move it as features age.- `capture.ts`: `SCENES` selects from a reusable scene library. With no override it captures only the current storyboard's default scene; update that default whenever the canonical storyboard changes. Hunk-side glue (`launchHunk`, `launchHunkShell`, demo repositories, and the keyboard probe) remains reusable. Reusable machinery lives in `@hunk/term-video` (`packages/term-video/`) —extend it there, don't fork it into the scripts: `createKeyframer`,`launchApp`/`launchShell`, `createCommandWrapper`, `typeCommand`,`ensureKeyboardIsLive`, `makeSceneFilter` (`src/capture.ts`); the unit-testedstoryboard planner with the caption/timing semantics (`src/plan.mjs`);`composeStoryboard` with font/Chromium resolution and the missing-keyframepreflight (`src/compose.mjs`); and the stage template (`src/stage.html`). ## Environment gotchas Sandbox-specific bullets are marked; each cost real debugging time. - **Run `capture.ts` with bun from the repo root.** tuistory uses subpath self-imports (`tuistory/pty`) that only resolve inside this repo's `node_modules`; running the script from elsewhere resolves tuistory from bun's global cache and crashes. (Output location is unaffected — it defaults to `<repo>/.video-work/` via `import.meta.url`.)- **ghostty-opentui is a transitive dep** (via tuistory) with an exports map: import `"ghostty-opentui/image"` (not `.../dist/image.js`), resolved relative to tuistory — `@hunk/term-video/capture`'s `createKeyframer` does the `Bun.resolveSync` dance.- **Playwright must match the Chromium it drives.** The portable setup above installs `playwright` and `playwright-core` together, then downloads their matching Chromium build. It works on macOS and Linux and is preferred when bandwidth and browser downloads are available; leave `CHROMIUM_PATH` unset so Playwright uses that managed browser. On Linux, if Chromium reports missing system libraries, run `bunx playwright install-deps chromium` (it may require sudo) before retrying. To reuse an existing system, CI, or sandbox Chromium instead, install the `playwright-core` version provided by that environment and set its executable explicitly: ```sh cd .video-work bun add playwright-core@<matching-version> cd .. CHROMIUM_PATH=/path/to/chromium node scripts/launch-video/compose.mjs .video-work ``` Common executable locations include `$(command -v chromium)` or `$(command -v google-chrome)` on Linux and `/Applications/Google Chrome.app/Contents/MacOS/Google Chrome` on macOS. In an environment with a preinstalled Playwright toolchain, read that toolchain's `package.json` to get the exact driver version. For example, the Anthropic sandbox exposes it through `/opt/pw-browsers/.links/*`: ```sh cat "$(cat /opt/pw-browsers/.links/* | head -1)/package.json" | grep '"version"' # e.g. "1.56.1" -> bun add playwright-core@1.56.1 ``` `compose.mjs` picks its browser as `$CHROMIUM_PATH`, then `/opt/pw-browsers/chromium` when present, then Playwright's managed browser. If a dependency update leaves Playwright asking for an uninstalled browser revision, either rerun `bunx playwright install chromium` in `.video-work/` or set `CHROMIUM_PATH` to a known system browser such as `/usr/bin/chromium`. - **Give `.video-work/` its own `package.json` before `bun add`.** Without one, bun walks up and installs into the repo's `package.json` — revert with `git checkout package.json bun.lock` if that happens.- **Chromium needs `--allow-file-access-from-files`** (already in `compose.mjs`): the stage samples each keyframe through a canvas to color-match the window background, and file:// images taint the canvas without it.- **mp4 needs an ffmpeg with libx264.** Sandbox: `apt-get install ffmpeg` (run `apt-get update` first if packages 404). macOS: `brew install ffmpeg`. Verify `ffmpeg -encoders | grep -E 'libx264|libvpx-vp9'` shows both before encoding — playwright's bundled `ffmpeg-*/ffmpeg-linux` only does VP8/WebM and cannot produce the mp4.- **Caption font**: JetBrains Mono ships inside ghostty-opentui; `findCaptionFont` in `packages/term-video/src/compose.mjs` searches bun's isolated layout (`node_modules/.bun/node_modules/…`) then a hoisted `node_modules/…`. If it still throws `caption font not found`, locate the file with `find node_modules -name jetbrains-mono-nerd.ttf` and pass it as `fontPath` to `composeStoryboard`. ## Authoring scenes (capture.ts) - Shared geometry is 140x32 cells rendered at fontSize 16 / dpr 2 → 2688x1536 PNGs. Keep every scene at this size so all frames fit one window.- Keep time-sensitive fixtures recent and timezone-stable. The canonical history scene derives commit dates from the current UTC day and launches Hunk with `TZ=UTC`, so day groups and relative-age labels remain useful on future runs. Preserve that policy in custom history scenes.- Helpers: `createDemoRepo()` and `launchHunkShell()` are hunk-side glue in the script (git repo built from `examples/2-mini-app-refactor`; interactive bash with a real `hunk` command on PATH and a clean `❯` prompt); `snap`, `typeCommand`, `launchApp`/`launchShell`, and `createCommandWrapper` come from `@hunk/term-video/capture`.- Always `waitForText` on scene-specific content before the first snap, and call `ensureKeyboardIsLive()` before scripted keypresses — the first key after startup can be dropped (real race, the helper toggles `?` to prove keys land).- **Animation = one snap per keypress.** Cursor walks and typing effects are just every `j`/`k`/character captured as its own frame and played back at 0.2–0.3s per frame. Prefer this over sparse keyframes: three stills read as a slideshow, per-press frames read as motion.- `renderTerminalToImage` auto-trims trailing blank rows, so short outputs (CLI scenes) produce short PNGs — the stage handles this by sampling the image's bottom-left pixel and painting the window body to match.- `manifest.json` is a capture-side inventory of the _current run_ only; `compose.mjs` ignores it (frames resolve by name from `SHOTS`), and after a `SCENES=` run it is partial while `frames/` stays cumulative.- Demo content that must exist: STML notes come from `examples/9-agent-markup-notes` (launch with `--experimental`), extension scenes from `examples/extensions/` loaded via `--extension <path>` (explicit paths skip the repo trust prompt). The pager pipe is `git diff | hunk pager` — bare `hunk` on piped stdin prints help. Sidebar toggle is `s`; comment draft is `c`, save with Ctrl+S (`\x13`). ## Storyboard model (compose.mjs) - `SHOTS` is the whole edit: one entry per shot, `dur` in seconds, played as unique frames + per-frame durations in an ffmpeg concat list (holds cost one frame, so runtime is dominated by transitions, not length).- `capKey` is caption identity: the caption slides in only when `capKey` changes, and continuation shots that share a `capKey` without restating `caption` keep the previous caption on screen. Sequences (walks, typing) are generated with `Array.from` spreads.- **Sequence lengths must match capture loop bounds**: the history capture's `selected = 2..4` loop produces `history-range-2..4`, consumed by a three-entry `Array.from` spread in `SHOTS`. Change one side and the other breaks — the preflight check names any frame that's missing.- `enter: true` fades/scales the surface in — use it for cards and the first terminal shot only.- Terminal shots may set normalized `camera: { x, y, scale }` targets and `highlight: { x, y, width, height, label? }` rectangles. `motion` controls the transition length in seconds. The planner interpolates camera changes and outline geometry, while the stage keeps the camera inside the captured image and paints an animated glow around the source-aligned region. Set the same `cameraKey` or `highlightKey` on related keyframes when their normalized coordinates share one capture geometry. Cross-image camera changes cut to their target unless `cameraKey` explicitly permits a pan, and outlines never leak across unrelated images.- Caption HTML vocabulary: `<span class="badge">NEW</span>` amber pill, `<span class="hl">` amber highlight, `<span class="dim">` muted. Cards use `badge` / `h1`/`h2` / `sub` / `cmds`+`cmd` / `foot` classes from `packages/term-video/src/stage.html`.- Target pacing: money shots hold 3–4s, context shots 2–3s, typing/walk frames 0.2–0.6s; keep the total near 60s. ## Camera and callout composition Treat `camera` and `highlight` as one composition. Both use normalized sourcecoordinates, but the stage projects highlights through the active camera. Arectangle that fits within the source can still lose an edge after zooming. Use this framing process: 1. Start with a full-frame establishing shot, then pan or zoom to one readable subject. Use `motion` around 0.7–0.9s for a deliberate move rather than a cut disguised as animation.2. Fit the target plus its outline inside the camera viewport. Leave about 1% of the visible width on each side and at least one terminal row above and below the meaningful content. If that does not fit, reduce `camera.scale`.3. Put outline edges in actual blank rows. Mathematical padding can move a border onto a neighboring date heading, status row, or metadata line.4. Check both sides of the capture. A zoom that preserves left-aligned subjects can still crop hashes, authors, or status text aligned to the right.5. Give irregular states explicit rectangles. A growing commit range that crosses day separators should use measured per-keyframe heights instead of assuming every added item occupies the same vertical distance.6. Share `cameraKey` or `highlightKey` only when captures use the same source geometry. This allows pans and outline morphs across related captures; unrelated keyframes cut or fade separately. The outline should direct attention, not become the subject. Use one callout ata time, keep labels short, and remove the outline once the viewer has enoughcontext to follow the interaction unaided. ## Command and output legibility Terminal commands and their results must be readable at normal playback size.Never rely on a full-width terminal prompt or output region as the only way theviewer can understand a command scene; text that looks acceptable in a 1080psource frame is often illegible in an embedded player or social feed. Use this default treatment: 1. **Present the exact command in large type.** The easiest treatment is a `cmd` card or an oversized editorial callout immediately before the real terminal result. A typing animation may remain for motion, but it does not replace the large command treatment.2. **Zoom the real output.** Crop or scale the captured terminal frame around the meaningful result so both the command and the important output lines are comfortably readable. A brief full-window establishing shot is fine, but the result's main hold must use the focused view.3. **Keep the evidence real.** Never fabricate terminal output. Large command text may reproduce the exact command as an editorial overlay; any enlarged output must come from the captured PTY frame. If reusable crop/zoom controls are missing, add them to `@hunk/term-video` rather than baking one-off image edits into a storyboard.4. **Trim before shrinking.** Prefer fewer relevant lines and a tighter crop over fitting a long transcript into the frame. Split a workflow across multiple focused shots when one crop cannot keep every important line legible. Treat command legibility as a release gate: if the command or its result cannotbe read when the 1920x1080 video is displayed at 50% size, revise the shot. ## Content accuracy (learned the hard way) - **Verify install commands against reality**, not the README: check `npm view hunkdiff dist-tags`. A prerelease needs `npm i -g hunkdiff@beta`; `brew install hunk` only serves stable (homebrew-core Autobump, lags npm) — omit brew on prerelease cards.- **Label demo extensions as examples.** The triage board, CSS palette, and semver views are `examples/extensions/`, not shipped features — caption them with a dimmed `example:` prefix. The real features are the APIs (sidebars, file views, commands, dialogs).- Window titles are decorative but must not lie: the shell scenes run bash, so keep their titles generic (`shell — …`) rather than naming a shell the capture doesn't launch.- STML requires `--experimental`; say so on the outro card.- The video is silent — never imply audio in the video or its announcement copy. ## Verification and delivery - Eyeball keyframes in `.video-work/frames/` (Read renders PNGs) after capture — especially new scenes — before compositing.- Inspect every command scene at 50% display size. Confirm the exact command is shown in large type and the meaningful real output is zoomed tightly enough to read; a readable caption does not compensate for unreadable terminal text.- Composite and encode the MP4 first. Inspect it before spending time on the slower VP9/WebM encode; produce both final formats only after the visual pass.- After encoding, return to the repository root, set `VIDEO` to the produced MP4, and inspect a settled frame for each callout, a mid-animation point, each new scene, and the outro: ```sh VIDEO=.video-work/hunk-feature-demo.mp4 # or .video-work/launch.mp4 ffmpeg -y -ss 2 -i "$VIDEO" -frames:v 1 -update 1 .video-work/check.png ffprobe -v error -show_entries format=duration,size \ -show_entries stream=codec_name,width,height,r_frame_rate \ -of default=noprint_wrappers=1 "$VIDEO" ``` Captions must persist through extracted animation frames. Every callout must show all four borders, keep its label clear of terminal text, leave visible padding around its subject, and avoid clipping right-aligned metadata. - Compare the storyboard planner's reported duration with `ffprobe`; they should agree within one frame. The concat manifest declares the PNG input rate so 30 fps animation timestamps stay exact. Investigate larger drift before delivery. - Outputs stay under `.video-work/`: the full-release recipe creates `launch.mp4`/`launch.webm`, while the single-feature recipe above creates `hunk-feature-demo.mp4`/`hunk-feature-demo.webm`. `.video-work/` is gitignored — never commit the video or its frames. For a standalone video request, send both files to the user directly (mp4: social/Slack; webm: web embeds), report duration and file sizes, and flag if the mp4 exceeds ~10 MB (Slack) or ~15 MB (X). When invoked by `hunk-release`, hand both files back to that workflow instead; it owns versioned filenames, GitHub's user-attachment limit, public embedding, and inline-player verification. Copy them elsewhere only if the user names a destination.
Discovery context
Discovered by repository scan. No exact path reference found in the snapshot’s root AGENTS.md.