Skip to content

armgit / hypit

Public Imported from GitHub
0

Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one command, and get your 100M views.

docs: rules the voice-clone reconstruction paid for

rponec5db994b924f
diff --git a/.agents/skills/hypit/references/local-author-package.md b/.agents/skills/hypit/references/local-author-package.md
index e128e5fd..9aaf8cce 100644
--- a/.agents/skills/hypit/references/local-author-package.md
+++ b/.agents/skills/hypit/references/local-author-package.md
@@ -112,6 +112,18 @@ nothing.
slots are inputs and each document fills them. Writing the observed content into the package is
the same failure as a missing inner picture, inverted: the frame was seen but the slot was closed.
`media-track:Item` and the `icon` port on `@hypit/ranking` are the shape this takes.
+- **A slot is a graph edge, not a value the Surface can read.** The package's own texture resolves at
+ author time because it is a file the package ships; a slot holds whatever the document generated,
+ which does not exist until the Build runs. A Surface that reaches for it the way it reaches for its
+ own asset fails with the slot's value refusing to resolve during author compilation. Declare the
+ slot as a Fragment input carrying the Artifact type, give the Producer that consumes it a second
+ form for the case where the slot is filled, and read it from `inputs` as the Artifact it is —
+ a blob input arrives as the `BlobRef` itself rather than wrapped inline like every authored value.
+- **The element kind follows the Artifact, not the slot.** A picture slot holds an image once the
+ document has built one and a stand-in *video* frame before it has, because that is what the
+ Playground substitutes for an output nothing has produced yet. Hard-code `kind: "image"` and the
+ component builds for nobody until the whole graph has run, which is the opposite of what a preview
+ is for. Decide from the Artifact's `mediaType` and the same code serves both.
- Never require an input the installing project has no reason to choose. If a Run Source exists only
to produce the component's own texture, the texture is in the wrong place.
- **A slot that is not filled yet costs its own cell and nothing more.** The component's structure —
diff --git a/.agents/skills/hypit/references/playbooks/craft/generated-dependencies.md b/.agents/skills/hypit/references/playbooks/craft/generated-dependencies.md
index 0cda360f..95c1639f 100644
--- a/.agents/skills/hypit/references/playbooks/craft/generated-dependencies.md
+++ b/.agents/skills/hypit/references/playbooks/craft/generated-dependencies.md
@@ -77,6 +77,30 @@ This is why a full-screen board that contains two media boxes is read as *two* m
single flat composition. The component owns its surface; the boxes are inputs, and each one is
filled by generation when the reference shows content in it.
+## An audio Take brings no picture, so its Segment is covered or it plays black
+
+A `speech:Take audio={…}` creates program time and speech and contributes no visual at all. For as
+long as it runs the picture is whatever the peer Tracks put there, and wherever they put nothing the
+Film's own background shows through. A voiceover Segment is therefore an obligation: every frame of
+it belongs to some Item, and the frames nobody claimed are black in the delivery.
+
+Two different holes open, and both look identical on screen:
+
+- **A stretch inside no Selection.** Mark the ranges the B-roll covers and one sentence between two
+ of them belongs to neither, so nothing draws it. The Selections have to *tile* the Segment — each
+ one picking up where the last left off — rather than merely landing in the right places. A line
+ that introduces what comes next usually belongs to the Selection it introduces.
+- **A take shorter than the window it fills.** An Item whose window outlasts its own material runs
+ out partway and leaves the rest empty. Read the model's duration ceiling before deciding: Seedance
+ `mini` stops at 15 seconds, so a longer stretch needs more than one Item rather than one Item asked
+ for a length the model refuses.
+
+Give a silent take a literal duration at or above its window instead of a `SpeechDuration` edge. The
+estimate predicts the words; the window is decided by the audio that was actually produced, and when
+the estimate falls a second short that second is black.
+
+`production-gates.md` measures the delivery for these before it is reported as finished.
+
## A short stretch is not a short take
An authored take stays inside the selected model's declared duration range — read the range from the
diff --git a/.agents/skills/hypit/references/playbooks/craft/production-gates.md b/.agents/skills/hypit/references/playbooks/craft/production-gates.md
index 73a6feff..c31e3456 100644
--- a/.agents/skills/hypit/references/playbooks/craft/production-gates.md
+++ b/.agents/skills/hypit/references/playbooks/craft/production-gates.md
@@ -37,6 +37,20 @@ ended in an error, so it does not feel like a run that produced anything. It pro
and every take that ran before the step that broke, and resubmitting without pinning them buys the
same bytes again. Run `--pin` against the failed Build before touching the cause.
+**Pin what a Provider was paid to make, and nothing the Source computes.** `--pin` emits a line for
+every accepted output, including the ones the Script's own structure produced: `speech.space`,
+`speech.visual`, `timing.map`, the caption plan. Those are free to rebuild and they are measured
+*against the Script as it was*. Paste them back after editing a Segment — renaming one, merging two —
+and the Build pins a map whose anchors describe a program that no longer exists. It does not fail.
+Every Track times itself against those stale anchors and the delivery is quietly wrong, which is the
+one failure this whole gate sequence exists to prevent. Keep the pictures, the takes and the voices;
+let everything downstream of them recompute.
+
+**A Build that failed downstream may still hold the good draw from a model that does not repeat
+itself.** The caption Planner is a language model: the same Source gives a valid plan on one run and
+a plan that drops an Atom on the next. When a Build fails after the Planner accepted, that accepted
+plan is the one that worked — pin it rather than resubmitting into another roll.
+
Apply this checklist to every image:
- identity: correct recurring person/product/location, complete face when required, no drift;
@@ -86,6 +100,12 @@ visual QA.
- Inspect the durable Build and retrieve the final Artifact with `get`.
- Watch and listen to the complete delivery, including the first and last second. Verify dialogue,
captions, overlays, transitions, audio density, claims, and CTA as one program.
+- Measure the delivery for empty picture before reporting it as finished. Watching finds a scene that
+ is wrong; it slides straight past a second of black between two shots, and a program assembled from
+ Selections can hold several. `ffmpeg -i final.mp4 -vf blackdetect=d=0.2:pic_th=0.90 -an -f null -`
+ prints every stretch with its start and duration, and any hit is a hole in the picture rather than
+ an edit — read `generated-dependencies.md` on covering a Segment. Report the delivery only after
+ this comes back empty.
- Preserve accepted Records for deliberate future reuse; never assume a rerun will reuse them.
Use explicit Run Source authoring for every accepted reuse:
diff --git a/.agents/skills/hypit/references/reference-video/final-sources.md b/.agents/skills/hypit/references/reference-video/final-sources.md
index e5fda4d9..0865dc64 100644
--- a/.agents/skills/hypit/references/reference-video/final-sources.md
+++ b/.agents/skills/hypit/references/reference-video/final-sources.md
@@ -26,6 +26,25 @@ any Build is submitted, which is the correct place for the author to find out.
Do not author one source fragment per shot. Do not let check success substitute for unresolved
semantic evidence; return to a narrow `observe_reference` question when necessary.
+## Keep Segment names short
+
+The Spine joins every Segment name into the id its visual Track carries, and the renderer turns that
+id into one filename when it extracts frames. Nine descriptive names — `hook`, `problem`,
+`discovery`, `steps` and the rest — reach that filename three times over and overrun the 255-byte
+limit a path component has, so the Build dies in video extraction with `ENAMETOOLONG` after every
+generation has been paid for. Nothing earlier catches it: the Source is legal, the preview builds,
+and only the delivery render touches the filesystem this way.
+
+Name Segments in two to four characters. It costs nothing — a Segment name is a handle, and the
+Script's meaning lives in the words it holds — and the ceiling is reached sooner than it looks, since
+the joined list appears once per nesting level.
+
+Reach for it late rather than early: a reconstruction whose Segments are already few and short has no
+problem here. The pressure comes from splitting one continuous stretch into several Segments, and
+that is usually worth undoing on its own merits. Consecutive Segments carrying one unbroken
+voiceover are one Segment with Selections inside it, which is both shorter and closer to what the
+reference does.
+
## A take's duration is measured, not estimated
`estimate:Speech` predicts how long a line will take to say. In this route that prediction is the
diff --git a/.agents/skills/hypit/references/reference-video/workflow.md b/.agents/skills/hypit/references/reference-video/workflow.md
index 6bc660d6..a798067d 100644
--- a/.agents/skills/hypit/references/reference-video/workflow.md
+++ b/.agents/skills/hypit/references/reference-video/workflow.md
@@ -84,6 +84,13 @@ selected shots. Each shot produces three observations:
The picture observation also says whether the picture moves and how, separating camera movement from
movement inside the frame. A held still and a moving shot are reconstructed differently.
+Covering content is whatever changes what reaches the eye, not only the things that sit on top with
+an edge. A wash laid over the whole frame to darken it is covering content and gets reported as such;
+so is anything else the observation happens to describe in terms of the picture being altered rather
+than something being added. Read it for what it says rather than matching it against the kinds of
+element you already expect — the observation is the evidence, and the moment a category is treated as
+a list of known items it stops being able to report the one thing nobody thought to list.
+
Each cut produces one `continuity` observation answering both questions under their own headings —
whether the two shots are one continuous camera shot, and whether an overlay continues across the
boundary — and a window of three consecutive shots triggers a three-shot continuity review. The
diff --git a/.agents/skills/hypit/references/runtime.md b/.agents/skills/hypit/references/runtime.md
index 56d111e6..cd12dfd1 100644
--- a/.agents/skills/hypit/references/runtime.md
+++ b/.agents/skills/hypit/references/runtime.md
@@ -56,6 +56,14 @@ Builds remain archived and neither command cancels remote Provider work.
Read an identical repeat failure as this until you have ruled it out. Reasoning about why a correct
fix did not work is how an afternoon goes, and the fix was never loaded.
+- **Stop the Worker between Builds, never during one.** A remote generation belongs to the Provider,
+ and the Worker is the only thing watching it: take the Worker down mid-flight and the request keeps
+ running with nobody to collect it, passes its operation deadline, and fails — taking the whole Build
+ with it, not merely the requests that were in the air. `runtime status` reports what is queued,
+ running and reserved; wait for it to report nothing before stopping anything. This applies to the
+ Profile as much as the code — raising an Endpoint's `defaultConcurrency` needs the same restart, so
+ a change that costs nothing on an idle Runtime costs the whole Build on a busy one. Nothing is lost
+ by waiting: accepted Records are durable, and the requests still in flight are the expensive ones.
## Keep project and package boundaries distinct