pr-to-video

pr-to-video - dispatch entry

Input is a GitHub pull request (a code change), supplied as a PR URL, an <owner>/<repo>#<N> ref, or "this PR" while a repo with an open PR is checked out. Output is a code-change explainer: what shipped, why, and how it works — rendered from the diff/commits as before-after, diff-highlight, file-tree, and impact scenes. Default length up to ~3 min (sweet spot ~30-90s); a genuinely longer or exhaustive every-file walkthrough (5 min+) is a different register → /general-video. There is no website scrape and no headless Chrome for ingest — ingest is the gh CLI. The shipped style preset is always claude (warm editorial; signature navy code window).

Confirm the route before Step 0. This skill explains a GitHub pull request (a code change read via gh). If the input is a marketing / product site → /product-launch-video; a general website → /website-to-video; a topic / article with no PR → /faceless-explainer; a whole-repo tour or multi-PR release → /general-video. Out of scope: live / at-render-time data — PR facts are read once at author time and baked in. Handed a non-PR input, or unsure? Read /hyperframes first.

This workflow owns only the PR-specific front (ingest + story-design); every phase marked shared reuses the engine copied from faceless-explainer unchanged (it lives under this skill's own scripts/ + agents/ + phases/, so <SKILL_DIR> resolves to pr-to-video).

All artifacts go to PROJECT_DIR = videos/<project-name>/ (created in Step 0); all paths below are relative to it. Dispatch is harness-portable: before the first subagent dispatch, read <SKILL_DIR>/../hyperframes-core/references/subagent-dispatch.md once — it maps the dispatch verbs (parallel fan-out / background / wait) to your harness's primitives; a concurrency cap below N means waves of the cap size, never fewer workers. This file is a binding runbook, not background reading: execute the steps in order and produce every phase artifact with its designated script or agent role — do not substitute a freestyle pipeline, and do not skip a pause step because the request seems clear. A step you cannot perform → stop and report.

Phase	Execution	Primary artifact	Detailed flow
init	Bash	hyperframes.json	Step 0
ingest (own)	Bash (gh CLI + ingest.mjs + fetch-people-avatars.mjs, NO agent, NO scrape)	capture/pr.json + diff.patch + extracted/{tokens.json,visible-text.txt,people.json} + public/avatars/	Step 1
design-system (shared)	Bash (no agent, deterministic claude)	design-system/design.html + chunks/	Step 1b
story-design (own)	subagent	narrator_scripts.json	agents/story-design.md
audio (shared)	audio.mjs in Bash	audio_meta.json	phases/audio/guide.md
visual-design (shared)	subagent	section_plan.md	agents/visual-design.md
prep (shared)	prep.mjs in Bash	group_spec.json	scripts/prep.mjs
captions (shared, det.)	captions.mjs group -> captions.mjs html in Bash (no subagent)	caption_groups.json + compositions/captions.html	scripts/captions.mjs
scenes (shared)	N x subagent (parallel)	compositions/scene_*.html or compositions/group_w*.html	agents/hyperframes-scene.md
finalize (shared)	Bash prelude (wait-bgm + assemble + inject/verify-transitions + hoist-videos + sfx-verify + preflight) -> finalize subagent (fix brief findings in place + one lean contact-sheet look + render)	renders/video.mp4	Step 7 / agents/hyperframes-finalize.md

Prerequisites

macOS Apple Silicon or Linux x64. System tools: brew install python@3.11 node ffmpeg (use Homebrew Python, not /usr/bin/python3, or pip install is blocked by PEP 668); then npx hyperframes doctor once (downloads Chrome — needed for snapshot/render, not for ingest). The rendered overlap gate (scripts/check-overlap.mjs, run in worker self-checks and preflight) reuses that same cached Chrome — it never downloads a browser; its only dep is the puppeteer-core npm module, ensured once before scene fan-out (Step 5.5, --ensure-deps, ~5s, no full puppeteer install). CLIs: gh (GitHub CLI, authenticated — gh auth status must pass) and hyperframes. Optional cloud keys (else local fallbacks) — inject in Step 0.5:

Key / requirement	Used for	Default / fallback
gh auth status OK	Reading the PR (public or private)	required — fail fast with the auth hint
HEYGEN_API_KEY (or hyperframes auth login)	TTS (cloud, word-level timestamps)	voice: auto (first English starfish voice; override --voice)
ELEVENLABS_API_KEY	TTS (cloud; needs pip install elevenlabs)	voice 21m00Tcm4TlvDq8ikWAM (Rachel)
neither, and not logged in	TTS	local Kokoro, voice am_michael (non-English: pass --voice)
GEMINI_API_KEY / GOOGLE_API_KEY (aliases)	Lyria BGM	unset -> local MusicGen (first run downloads ~300 MB)

Flow

Step 0.0 - Confirm the brief (ALWAYS ask one round, then build)

Before Step 0, always pause and ask the brief in one message, then wait for the user — never skip this, even for a request that looks complete. Lead with a recommended default for each field and pre-fill anything the user already gave (confirm it rather than re-asking blindly): the angle (changelog / feature reveal / fix / refactor — default: infer from the PR), the audience (developers vs general users — default: developers), length (default ~60-90s), and — if /hyperframes didn't set them — aspect (default 16:9) and language. Style is always claude. Proceed to Step 0 only after the user replies; a "go" / "use the defaults" is a valid reply that accepts every default.

Step 0 - Initialize the video project

cwd is the agent workspace root (e.g. /tmp/pr-video-...). Write all video artifacts under PROJECT_DIR = videos/<project-name>/.

<project-name>: use the directory the user gave (e.g. Use ./videos/retry-pr), else a short kebab-case name derived from the PR (<repo>-pr-<N>, e.g. widgets-pr-1187). Not the workspace basename or a timestamp.

Only when $PROJECT_DIR/hyperframes.json is absent:

PROJECT_DIR="${PR_VIDEO_DIR:-videos/<project-name>}"
mkdir -p "$(dirname "$PROJECT_DIR")"
npx hyperframes init "$PROJECT_DIR" --non-interactive --skip-skills --example=blank

hyperframes init drops a generic AGENTS.md / CLAUDE.md into $PROJECT_DIR; leave them in place — they are agent scaffolding for whoever opens the finished project later.

Constraints: never run hyperframes init / generate AGENTS.md / CLAUDE.md in the workspace root; never nest another hyperframes/ inside PROJECT_DIR; every Bash command (master + subagents) is a (cd "$PROJECT_DIR" && ...) subshell — never bare cd.

Step 0.5 - API key guidance

Skip if $PROJECT_DIR/.env exists or context.log is non-empty (= not the first run). Otherwise first detect what's available (HeyGen TTS on if $HEYGEN_API_KEY / $HYPERFRAMES_API_KEY set or ~/.heygen/credentials exists from hyperframes auth login; ElevenLabs / Gemini only if their env keys set), then always pause and offer the menu — wait for the user; do not proceed on your own even when a workable config is detected (the user may want to add a key like Gemini). State what's detected, then: paste keys (→ Write $PROJECT_DIR/.env, one KEY=value per line, overwrite same-name) / "go" (proceed with what's configured — env, .env, or hyperframes auth login) / "skip" (proceed with local fallbacks for anything unconfigured). Then proceed to Step 1.

Step 1 - Ingest (Bash, NO agent, NO scrape)

Resolve the PR ref and pull structured facts with gh, then fold them into the synthetic capture package the shared backend expects (mirrors faceless-explainer's no-scrape scaffold). gh runs here, in the orchestrator, so auth / not-found / private-repo errors surface with gh's own stderr; ingest.mjs is a pure offline transform.

# PR ref: a full URL, "<owner>/<repo>#<N>", or "<N>" inside a checked-out repo.
PR="<url | owner/repo#N | N>"

# Fail fast if gh is not authenticated.
gh auth status || { echo "gh not authenticated — run: gh auth login"; exit 1; }

(cd "$PROJECT_DIR" && mkdir -p capture/extracted capture/assets)
(cd "$PROJECT_DIR" && gh pr view "$PR" \
  --json number,title,body,author,url,baseRefName,headRefName,commits,files,additions,deletions,changedFiles,labels,reviews,latestReviews,comments,assignees,reviewDecision,mergedBy \
  > capture/pr.json)
(cd "$PROJECT_DIR" && gh pr diff "$PR" > capture/diff.patch)

# Fold pr.json + diff.patch into tokens.json (colors:[] → claude native palette) +
# visible-text.txt (the narrative brief) + people.json (PR author + commit authors w/ counts +
# reviewers / commenters / assignees, bot-filtered + deduped, each with a GitHub avatar URL).
# (The PR `author` is only the opener; commit authors from commits[].authors[] are tracked too.)
# ingest is OFFLINE.
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/ingest.mjs \
  --pr-json ./capture/pr.json --diff ./capture/diff.patch --out-dir ./capture/extracted)

# Network step (the people front's only one — ingest stays offline): download each
# contributor's GitHub avatar to public/avatars/<login>.png for an optional credits /
# shipped-by close. Best-effort — a missing avatar or offline run never blocks (exit 0).
# Avatars + that close are the ONE place pr-to-video relaxes the faceless default.
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/fetch-people-avatars.mjs \
  --people ./capture/extracted/people.json)

Validation:

[ -s "$PROJECT_DIR/capture/pr.json" ] && \
[ -s "$PROJECT_DIR/capture/diff.patch" ] && \
[ -s "$PROJECT_DIR/capture/extracted/tokens.json" ] && \
[ -s "$PROJECT_DIR/capture/extracted/visible-text.txt" ] && \
[ -s "$PROJECT_DIR/capture/extracted/people.json" ] && \
[ -d "$PROJECT_DIR/capture/assets" ] && echo ok || echo missing
# public/avatars/ is best-effort — its absence is NOT a failure (no avatars resolved / offline).

If gh errors (auth / not found / private), report the exact stderr and stop — do not fabricate PR contents. If ingest.mjs exits 1, read its stderr (usually a malformed pr.json), fix, rerun (deterministic, finishes instantly). fetch-people-avatars.mjs always exits 0; if avatars are missing, story-design simply has no credits scene to author.

Step 1b - Design system (Bash, NO agent, deterministic — SHARED)

Three deterministic commands produce a fully-styled design.html + chunks against the synthetic input, with the claude preset (its code-window / number-lockup / stat-card components are the PR visual vocabulary):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/phases/design-system/scripts/build-design.mjs ./design-system --no-emit --style claude)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/phases/design-system/scripts/build-design.mjs ./design-system --style claude)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/phases/design-system/scripts/emit-chunks.mjs ./design-system)

Validation:

[ -s "$PROJECT_DIR/design-system/inference.json" ] && \
[ -s "$PROJECT_DIR/design-system/design.html" ] && \
[ -s "$PROJECT_DIR/design-system/chunks/index.json" ] && echo ok || echo missing

If any is missing, read the build-design / emit-chunks stderr, fix the invocation, and rerun (deterministic, finishes in seconds).

Step 2 - Story-design (subagent) — OWN

Dispatch one subagent. prompt = full contents of agents/story-design.md + the ## Dispatch context below, passed through verbatim:

SKILL_DIR: <absolute path>
PROJECT_DIR: <video project root>
Schema validator: <SKILL_DIR>/scripts/validate-narrator.mjs
PR facts: ./capture/pr.json                          # title / body / commits / files / +/- stats — read first
Diff: ./capture/diff.patch                           # the actual change — pull 2-4 representative hunks
Brief: ./capture/extracted/visible-text.txt          # the assembled narrative brief
People: ./capture/extracted/people.json              # contributors (PR author + commit authors w/ commitCount + reviewers/commenters) + avatarFile; avatars in public/avatars/ — optional credits close
Design DNA: ./design-system/inference.json           # Read site_dna once to set register (soft hint only)
Orientation: <landscape | portrait | square>        # From the Step 0.0 aspect (16:9→landscape, 9:16→portrait, 1:1→square; default landscape). Emit VERBATIM as the top-level `orientation` field — dictated, not a choice; sets the canvas (portrait→1080×1920) for the whole pipeline.
Script style: concise, dev-facing — 1-2 sentences/scene, <=20 words; name the change, the why, the impact

The agent picks a PR archetype for narrativeArchetype (changelog / feature-reveal / fix-explainer / refactor-walkthrough, or "<outer> with <inner>"), echoes the dispatched orientation as a top-level field (Step 5 prep → canvas size), and emits narrator_scripts.json (it runs the validator before returning). continuity drives worker grouping: continue = same worker as the previous scene (cap=3); break = new worker; scene 1 is always break. intent / sharedMotif are soft hints. assetCandidates is [] on essentially every scene (faceless) — the one exception is an optional credits / shipped-by close that may reference the contributor avatars in public/avatars/<login>.png (from people.json).

Step 3 - Audio — SHARED

After narrator_scripts.json exists:

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/audio.mjs \
  --narrator-scripts ./narrator_scripts.json \
  --hyperframes . \
  --out ./audio_meta.json \
  --lyria-recipe <SKILL_DIR>/phases/audio/lyria-recipe.py)

BGM generation runs detached in the background. Backend selection (audio.mjs Step 5b): cloud Lyria is used only when a GEMINI_API_KEY/GOOGLE_API_KEY is set, the --lyria-recipe exists, AND import google.genai actually succeeds — if the key is set but the package is missing, audio.mjs tries to pip install google-genai on demand. When Lyria can't run, it falls back to local MusicGen (facebook/musicgen-small via transformers, no key; deps auto-installed in the background, parallel with TTS). BGM is only skipped entirely when neither backend can be made to run (e.g. no network for pip). It never blocks the render. Flags + BGM mechanics: top of audio.mjs.

• exit 0 -> voice + transcribe complete (BGM may still be rendering; audio_meta.json records bgm_log / bgm_pid), continue.
• exit 1 -> zero scenes produced voice; report and stop.

Step 4 - Visual-design (subagent) — SHARED

After design-system/chunks/index.json, narrator_scripts.json, and audio_meta.json exist, concatenate all inputs into one dispatch packet (contracts first, static references middle, work items last):

# Dispatch packets live in $PROJECT_DIR/.dispatch/ (transient; safe to delete after the run).
# NEVER use a fixed /tmp path: it persists across runs/projects, so a failed write silently
# reuses another project's stale packet and contaminates every worker.
mkdir -p "$PROJECT_DIR/.dispatch"
DP="$PROJECT_DIR/.dispatch/vd-dispatch.txt"
{
  echo "## Design chunks"
  (cd "$PROJECT_DIR" && cat design-system/chunks/index.json \
    design-system/chunks/composition-hints.md design-system/chunks/voice.md \
    design-system/chunks/tokens.css design-system/chunks/easings.js 2>/dev/null)
  echo "## Effects catalog";  cat <SKILL_DIR>/phases/visual-design/effects-catalog.md
  echo "## Design rules";     cat <SKILL_DIR>/phases/visual-design/rules/{typography,color-system,composition,motion-language}.md
  echo "## SFX library";      cat <SKILL_DIR>/assets/sfx/manifest.json
  echo "## Narrator scripts"; (cd "$PROJECT_DIR" && cat narrator_scripts.json)
  echo "## Audio meta";       (cd "$PROJECT_DIR" && cat audio_meta.json 2>/dev/null)   # Optional; overrides Duration if drift >10%
} > "$DP"
# Guard: a partially-failed build must fail LOUDLY here, not downstream in the subagent
grep -q '^## Narrator scripts' "$DP" || { echo "FATAL: vd-dispatch.txt incomplete — rebuild before dispatching"; }

# Captions planning hint (put it in the Captions: line of the dispatch below)
(cd "$PROJECT_DIR" && node -e 'try{const m=require("./audio_meta.json");process.stdout.write(Object.values(m.scenes||{}).some(s=>s.wordsPath)?"enabled":"disabled")}catch{process.stdout.write("enabled")}')

Then dispatch the visual-design subagent. prompt = full contents of agents/visual-design.md + the ## Dispatch context below, verbatim:

SKILL_DIR: <absolute path>
PROJECT_DIR: <video project root>
Schema validator: <SKILL_DIR>/scripts/validate-section.mjs
Canvas: <width>×<height>   # default 1920×1080 (16:9 landscape); 1080×1920 (9:16 portrait) or 1080×1080 (1:1 square) if requested upstream (narrator_scripts.orientation/dimensions). Plan layouts for THIS aspect ratio — see composition.md "Portrait & Square".
Captions: <enabled | disabled>   # Planning hint from the node -e above: enabled => leave the bottom ~17% of canvas height as caption territory in prose
Dispatch packet: <PROJECT_DIR>/.dispatch/vd-dispatch.txt   # Step 0 reads it once for all inputs
Visuals: faceless code-change — every scene is a code-window / before-after split / file-tree / +/- counter / diagram / typography invented from the script + the featured diff hunk. assetCandidates is [] for most or all scenes; plan visuals from the script and diff, not from captured assets.

Output is section_plan.md. The Captions: line is an optimistic hint; the authoritative gate is group_spec.captions_enabled from Step 5.

Step 5 - prep (deterministic script, NO subagent) — SHARED

After section_plan.md exists:

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/prep.mjs \
  --section-plan ./section_plan.md \
  --narrator-scripts ./narrator_scripts.json \
  --audio-meta ./audio_meta.json \
  --rules-dir <SKILL_DIR>/../hyperframes-animation/rules \
  --capture ./capture \
  --design-system ./design-system \
  --hyperframes . \
  --sfx-lib <SKILL_DIR>/assets/sfx \
  --out ./group_spec.json)

Merges all upstream artifacts into group_spec.json (parse section_plan anchors, validate effect/component ids, group by Continuity with cap=3, build visual_clips[] where a multi-scene continue worker becomes one group_wN.html, compute Tier-B transitions[] between different visual clips, copy assets/fonts/SFX). capture/assets/ is empty, so asset-copy is a no-op (faceless). Internal logic: header of prep.mjs.

--audio-meta ./audio_meta.json is what carries each scene's voicePath / wordsPath and the bgm_path into group_spec — and therefore into the assembled index.html. Omitting it (or pointing it at a path whose wavs don't resolve under --hyperframes) silently blanks every voice / caption / BGM track and renders a SILENT, caption-less video while every gate stays green. prep now defaults this flag to ./audio_meta.json and prints a CRITICAL banner when audio_meta lists voiced scenes but none get wired; assemble-index.mjs re-asserts the same guard before render. Keep passing the flag explicitly anyway.

• exit 0 -> read stdout (scenes / groups / total duration / per-group) and append to context.log.
• exit 1 -> stderr names the failing scene + anchor (usually a malformed anchor or unknown effect/transition id); return to Step 4 and re-dispatch visual-design.

Step 5.5 + Step 6 - Captions (deterministic) + scene worker fan-out — SHARED

Captions: two deterministic scripts (no subagent), after prep exits 0 and before fan-out:

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/captions.mjs group \
  --group-spec ./group_spec.json --hyperframes . \
  --tokens design-system/chunks/tokens.css --out ./caption_groups.json)

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/captions.mjs html \
  --hyperframes . --groups ./caption_groups.json \
  --tokens design-system/chunks/tokens.css \
  --inference design-system/inference.json \
  --out compositions/captions.html)

exit 0 = normal. If either prints captions: skipped (<reason>), skip the whole chain: no captions.html, assemble won't mount track 12. Skin selection / self-check: top of captions.mjs html (the claude preset ships its own caption-skin.html); for offline, pass --skin-file. Do not run npx hyperframes lint on captions.html.

Then ensure the overlap-gate dep once, from the workspace root (NOT inside PROJECT_DIR — the module must land in the workspace node_modules/ where every worker and preflight can resolve it):

node <SKILL_DIR>/scripts/check-overlap.mjs --ensure-deps
# Installs puppeteer-core (module only, no browser download) if not already resolvable; Chrome is
# reused from the hyperframes browser cache. Workers must NOT install it themselves (parallel npm race).

Then read group_spec.json.groups[] for worker count N. Build the shared header once, then per-worker packets (film direction / tokens / easings / voice are identical for every worker):

# Same rule as Step 4: packets go in $PROJECT_DIR/.dispatch/, never a fixed /tmp path
# (a stale /tmp file from a previous project survives a failed write and silently
# poisons every worker with the wrong design system).
mkdir -p "$PROJECT_DIR/.dispatch/scene-dispatch"
# `## Film direction` = the film-level invariants from group_spec.film_direction
# (palette system / motion defaults + budget / ambient system / negative list);
# each scene's creative_brief carries only scene-specific deltas on top of it.
{
  echo "## Film direction"
  (cd "$PROJECT_DIR" && node -p 'JSON.parse(require("fs").readFileSync("group_spec.json","utf8")).film_direction || ""')
  echo "## Tokens / easings / voice"
  (cd "$PROJECT_DIR" && cat design-system/chunks/tokens.css design-system/chunks/easings.js design-system/chunks/voice.md 2>/dev/null)
} > "$PROJECT_DIR/.dispatch/scene-shared.txt"
# Guard BEFORE fan-out: the project's own brand token must be present; a contaminated
# packet here costs a full re-author round across every affected worker.
grep -q -- '--brand-primary' "$PROJECT_DIR/.dispatch/scene-shared.txt" || \
  { echo "FATAL: scene-shared.txt incomplete/stale — rebuild before dispatching workers"; }
# Then per worker: shared header + that worker's Scenes YAML -> $PROJECT_DIR/.dispatch/scene-dispatch/w<N>.txt

Start N scene workers in parallel (concurrent background dispatches; a harness concurrency cap below N means waves of the cap size until every worker has run — never fewer workers). prompt = full contents of agents/hyperframes-scene.md + ## Dispatch context, verbatim. Top-level fields: SKILL_DIR / PROJECT_DIR / Worker ID / Composition width + Composition height (= group_spec.width / group_spec.height) / Captions: <enabled|disabled> (= group_spec.captions_enabled) / Dispatch packet: <PROJECT_DIR>/.dispatch/scene-dispatch/w<N>.txt, plus the shared header body (## Film direction + ## Tokens / easings / voice) + a Scenes: list. Each worker's self-check runs two scoped machine gates before returning — captions.mjs keepout --scene (when captions enabled) and check-overlap.mjs --scene (always) — so layout violations are fixed at the source instead of surfacing at preflight.

For the worker top-level context, copy from group_spec.json.groups[i]: worker_id, composition_id, composition_file, duration_s, scene_ids; and from the top of group_spec.json: width, height (the worker authors + self-checks the root at these dims — landscape 1920×1080 unless portrait/square was requested upstream). When Captions: enabled, also pass Caption band top y = height − round(height × 0.1667) and Foreground max y = Caption band top y − 20 (landscape → 900 / 880; portrait → 1600 / 1580) — constraint #13 keep-out is computed from these, not hardcoded. Copy every field in the Scenes: list verbatim from group_spec.json.groups[i].scenes[<sid>] (only that worker's 1-3 logical scenes): scene_id / local_start_s / effects / rule_paths / assetCandidates / estimatedDuration_s / voicePath / design_chunks (absolute paths to the whole component library — the worker chooses by visual judgment) / creative_brief. A continue run of 2-3 scenes writes one group_wN.html with true shared DOM across the segments.

assetCandidates is [] for most or all scenes — the worker invents the visual from creative_brief + design chunks (code-window for diffs, before/after, +/- counters); there are no captured assets to place. design_chunks: null (chunks missing) → worker falls back to reading ./design-system/design.html fully; should not happen in the normal path.

After all workers + captions return, run preflight (scans group_spec.visual_clips[]; does NOT check captions.html):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/check-compositions.mjs \
  --hyperframes . \
  --group-spec ./group_spec.json)

• exit 0 -> all compositions pass, continue to Step 7.
• exit 1 -> stderr names the violating scene + rule category; return to Step 6 and re-dispatch the affected worker (do not Edit in the master — fix upstream).

Step 7 - Assembly prelude + preflight gate + finalize — SHARED

After Step 6 exits 0: a deterministic Bash prelude (wait-bgm + assemble + inject/verify-transitions + hoist-videos + sfx-verify + preflight), then one finalize subagent that fixes the brief's findings in place, takes ONE lean contact-sheet look, and renders. Principle: deterministic prelude is all Bash; findings go to finalize (not back to workers); worker re-dispatch is reserved for recomposition. compositions/scene_N.html / group_wN.html are worker source files; editing them edits the source.

(1) BGM wait + assembly (Bash):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/wait-bgm.mjs \
  --audio-meta ./audio_meta.json \
  --hyperframes . \
  --timeout-ms 120000 \
  --interval-ms 2000)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/assemble-index.mjs --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/transitions.mjs inject --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/transitions.mjs verify --group-spec ./group_spec.json --index ./index.html)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/hoist-videos.mjs --group-spec ./group_spec.json --hyperframes .)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/verify-output.mjs sfx --group-spec ./group_spec.json --index ./index.html)
(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/verify-output.mjs audio --hyperframes . --group-spec ./group_spec.json --index ./index.html)

inject only changes the index.html shell data-start/data-duration/data-track-index, never visual roots. hoist-videos reads each scene's poster data-video-src declarations, measures the poster's rendered rect headless, and mounts the real <video class="clip"> at the index.html host root with global timing clamped clear of transitions — the ONLY legal way footage plays, since the runtime never decodes a <video> nested in a scene. Internal logic: header of each script.

• assemble exit 1 -> names a visual composition (root data-duration != group_spec, or file missing) = worker contract break → return to Step 6, re-dispatch that worker, rerun this step.
• inject/verify-transitions exit 1 -> injector bug (prep already validated transitions[]) → report, don't roll back workers.
• hoist-videos exit 1 -> a data-video-src declaration is invalid (missing file / bad numbers / window too small after transition clamping / poster not measurable) — stderr names the scene + declaration; Edit the visual source file (or re-dispatch its worker for a real relayout), then rerun this step. exit 2 -> browser unavailable; run node <SKILL_DIR>/scripts/check-overlap.mjs --ensure-deps from the workspace root, then rerun. exit 0 prints one line per hoisted video (src, global window, track, rect).
• sfx-verify exit 1 -> assembler bug → report.
• verify-output audio exit 1 -> a voice wav / bgm.wav / captions.html exists on disk but was NOT wired into index.html (the silent / caption-less render class). This is an upstream wiring bug — almost always empty group_spec voicePaths because prep ran without --audio-meta. Do NOT render. Re-run Step 5 prep with --audio-meta ./audio_meta.json, then re-run this Step 7(1) chain. ⚠-prefixed lines (BGM / captions intended but never produced on disk) are non-blocking generation gaps — render proceeds.

(2) Preflight gate (Bash):

(cd "$PROJECT_DIR" && node <SKILL_DIR>/scripts/preflight-finalize.mjs --group-spec ./group_spec.json --hyperframes .)

preflight does everything the agent does not need to judge and writes it all into finalize_brief.json: warms a pinned npx hyperframes@<version> cache, runs lint/validate/inspect with that version (inspect runs STRICT — no --tolerance flag, CLI default; by-design transient overflow from 3D morph / tilt / zoom peaks is declared per-element with data-layout-allow-overflow, never absorbed numerically — any re-run of inspect elsewhere must also be plain or verdicts disagree) and captures tails + summary counts, computes the snapshot timeline, runs check-overlap.mjs (the single-rule rendered overlap gate: every scene loaded headless, timeline seeked to 0.4/0.7/0.92 of duration, all non-background paint atoms flattened onto one plane with z-index ignored, pairwise-intersected; persistent overlap = a finding finalize must fix; status: unavailable blocks at exit 2 — the gate never soft-skips), and when captions_enabled runs captions.mjs keepout static check for "foreground lower edge y <= 900" (the bbox math folds in CSS transforms AND margin-top/margin-bottom, so negative-margin-centered cards are measured at their real bbox). Keep-out violations include ready-to-apply Edit strings (edit_old/edit_new) and overlap violations carry both selectors + both rects + the overlap rect — finalize consumes both directly and fixes them in place. Brief fields (preflight_clean / gates_clean / gates.* / bgm.* / overlap.* / caption_keepout.* / anomalies[] / snapshot_times_s[] / npx_prefix / scenes[] / internal_seams[]) and algorithm details are documented at the top of preflight-finalize.mjs. Only contrast and cramped-container remain eye-owned (finalize's one contact-sheet scan); collision / panel-bleed are machine-owned by the overlap gate.

• exit 0 -> dispatch finalize — clean or not. Findings (gate errors / overlap.violations[] / caption_keepout.violations[]) ride in the brief and finalize fixes them in place as its first work step. Do NOT diagnose them yourself, do NOT hand-Edit scene files, do NOT re-dispatch workers for them.
• exit 2 -> ONLY when the overlap gate could not run (overlap.status: "unavailable" — puppeteer-core / Chrome missing). Environment problem with a deterministic remedy: run node <SKILL_DIR>/scripts/check-overlap.mjs --ensure-deps from the workspace root (and npx hyperframes doctor if it names Chrome), then rerun preflight — do not proceed unmeasured.
• exit 1 -> preflight itself crashed (bad invocation / missing group_spec) → fix the invocation.

Worker re-dispatch (Repair Mode) is the EXCEPTION path now, not a preflight branch: it triggers only when finalize STOPs because a scene needs recomposition (content fundamentally wrong / real relayout / animation broken beyond a couple of edits). Then: re-dispatch that scene's owning worker (a continue worker owns its whole group_wN.html and repairs all its logical scenes together) with the full agents/hyperframes-scene.md + the normal dispatch context + a ## Repair context block containing: (a) finalize's verbatim findings for that worker's scene(s) (never paraphrase measurements), (b) npx_prefix copied from finalize_brief.json, (c) Inspect at: <t1,t2,t3> = that scene's midpoint_s + high_risk_extras_s (or start_s + 0.5/0.75/0.9 × duration) from brief.scenes[], (d) Captions: enabled|disabled. Per the contract's Repair Mode section, each worker Edits in place and self-verifies (scoped plain inspect --at + check-overlap.mjs --scene + keepout) before returning — so you (master) do NOT hand-Edit scene files and do NOT re-run the full preflight after each individual fix. When ALL repair workers have returned green, rerun (1)+(2) once and re-dispatch finalize. If the same finding survives two full repair rounds, STOP and surface it to the user instead of looping.

Scan anomalies[] even on exit 0 (loud non-blocking warnings surfaced by preflight; currently rare — read each entry's message and decide whether it changes the dispatch).

(3) Dispatch finalize subagent (fix brief findings in place -> ONE lean contact-sheet look -> render). prompt = full contents of agents/hyperframes-finalize.md + ## Dispatch context:

SKILL_DIR: <absolute path>
PROJECT_DIR: <video project root>
Render quality: high     # Or draft / standard
Finalize brief: <PROJECT_DIR>/finalize_brief.json   # Preflight has already written it; agent reads once for findings + npx_prefix + scene timings
Film direction: |        # = group_spec.film_direction (film-level invariants the briefs assume)
  <verbatim>
Visual clips:            # One line per group_spec.visual_clips[] entry
  - { id, file, kind, worker_id, scene_ids, start_s, duration_s }
Scenes:                  # One line per logical scene, copied verbatim from group_spec.json
  - { scene_id, start_s, estimatedDuration_s, effects: [...], creative_brief: |
      <Phase 3 prose for this scene> }

index.html is already assembled (transitions injected, videos hoisted); all gates have already run. Finalize's flow: fix every brief finding in place first (gate output_tail -> Edit + rerun only that gate; overlap.violations[] -> Edit per the given selectors/rects + scoped check-overlap --scene verify; caption_keepout.violations[] -> apply edit_old/edit_new mechanically), then ONE snapshot call at scene midpoints + group-internal continue-seam mids, one read of the contact sheet (looking only for blank/black panels, cut or unreadable text, crushed interiors, seam jank — escalate single frames only on suspicion), then render + verify-render. No per-frame QA walkthrough. Finalize must never change a visual root data-duration (= visual_clips[].duration_s, fixed upstream; changing it makes assemble fatal — timing is only fixable by returning to Step 6).

• finalize reports the mp4 (verify-render passed) + gate/snapshot status + files repaired in place -> complete.
• finalize STOP (only when a scene needs full recomposition) -> return to Step 6, re-dispatch that worker, rerun (1)+(2), re-dispatch finalize.

Completion report

Summarize per phase: PR (repo / #N / title), preset (always claude), PR archetype, scene count / total duration, worker grouping, transitions, gate status (lint / validate / inspect (strict) / overlap), hoisted videos (count + tracks), findings fixed in place, lean pass (tiles scanned, escalations), visual files repaired in place, final mp4 path + bytes + duration.

Offer a live preview — never auto-open one. The deliverable is the mp4 above. A browser preview is optional and must not be started until the user asks for it. Do NOT run hyperframes preview / play during any earlier phase: a preview opened mid-run shows half-edited compositions and dies when that phase's own snapshot/render server is torn down. When the user asks, start a long-lived dev server after the render (it serves the final on-disk files and stays up until stopped), then report the actual URL with the real port + project name:

(cd "$PROJECT_DIR" && npx hyperframes preview)   # Studio UI, e.g. http://localhost:3002/#project/<project-name>
# or a lightweight shareable player link instead:
(cd "$PROJECT_DIR" && npx hyperframes play)       # plain http://localhost:<port>

Flags (custom port, external browser) live in the hyperframes-cli skill (references/preview-render.md).

---

Resume table

Read $PROJECT_DIR/context.log and resume from:

State	Continue from
log missing or empty	Full pipeline
capture/pr.json or capture/extracted/visible-text.txt missing	Step 1 (ingest)
ingest done, design-system/inference.json or chunks/index.json missing	Step 1b (three deterministic commands)
chunks/index.json exists, narrator_scripts.json missing	Step 2 (story-design). If the user supplied a final narrator_scripts.json, place it in $PROJECT_DIR/ to skip this state
narrator_scripts.json exists, audio_meta.json missing	Step 3 (audio)
audio_meta.json exists, section_plan.md missing	Step 4 (visual-design)
section_plan.md exists, group_spec.json missing	Step 5 (prep)
group_spec.json exists, any visual_clips[].file missing or caption_groups.json missing	Step 5.5+6 (run captions.mjs group -> html, then dispatch workers for missing clips). Captions-ran criterion = caption_groups.json exists (NOT captions.html)
all visual_clips[].file exist + captions decided, renders/video.mp4 missing	Step 7 (rerun assemble + sfx-verify + preflight, overwriting finalize_brief.json / index.html, then dispatch finalize)
renders/video.mp4 exists	Report completed and stop

Directory shape

./                            # workspace root
├── .claude/skills/
├── node_modules/  package.json
└── videos/<project-name>/    # PROJECT_DIR - HyperFrames project root
    ├── hyperframes.json  context.log
    ├── capture/              # synthetic package (NOT a scrape) — kept for backend layout compatibility
    │   ├── pr.json           # gh pr view --json (now incl. reviews / comments / assignees / reviewDecision)
    │   ├── diff.patch        # gh pr diff (the full change; story-design pulls hunks from here)
    │   ├── extracted/        # tokens.json (synthetic) + visible-text.txt (brief) + people.json (contributors)
    │   └── assets/           # empty (faceless)
    ├── design-system/        # build-design outputs: inference.json / design.html / chunks/ / fonts/
    ├── narrator_scripts.json  audio_meta.json  section_plan.md  group_spec.json
    ├── public/  assets/  compositions/  snapshots/   # public/avatars/<login>.png — contributor avatars
    └── renders/video.mp4

Routing note (for the hyperframes router)

• Input: a GitHub PR — a code change (PR URL, owner/repo#N, or "this PR"). A URL, but a github.com/.../pull/N link, not a product/marketing website.
• Output: code-change explainer, up to ~3 min (sweet spot ~30-90s); 5 min+ exhaustive deep-dives → /general-video.
• Triggers: "make a video about this PR", "turn PR #1187 into a changelog video", "explain what this pull request does as a video", "release-notes video from github.com/org/repo/pull/123", "turn this PR into a video".
• Do NOT use for: a product/marketing website URL (-> /product-launch-video) or a general website to turn into a video (-> /website-to-video); a topic/article/text with no PR (-> /faceless-explainer); adding captions to an existing video (-> /embedded-captions); a whole-repo tour or multi-PR release (no workflow yet -> /general-video).