mirror of https://github.com/heygen-com/hyperframes.git synced 2026-08-31 02:41:44 +00:00

feat(cli): search the catalog by meaning, on this machine (#3089 )

* feat(cli): search the catalog by meaning, in three named tiers

Browsing the registry means matching names and tags, which fails whenever the
author's wording differs from yours. "make the pace feel faster" finds nothing
when the move is described as "velocity-driven blur". This ranks by meaning
instead.

Three tiers, and the command always says which one answered:

words shared vocabulary, free, offline, no account
on-device bge-small, free, offline, one opt-in download
hosted Gemini, free for signed-in HeyGen users

The tier is stated because a quietly worse answer looks exactly like a good
one. --json carries it as a token alongside dropped, shown, total and
top_score, so an agent reads provenance as data rather than matching English
that is written to be reworded.

Two consents, asked once each, and never conflated. Sending a query is a
privacy question, so the prompt says the query is sent. Downloading a model is
a disk and bandwidth question, so that prompt talks about size. Neither fires
without a terminal: an unattended run sends nothing and downloads nothing
unless a flag records that a person agreed.

The catalog is derived from registry-item.json rather than from a separate
document, so the set that is ranked and the set that can be installed are the
same object by construction. Only the on-device vectors are committed; the
hosted vectors are nine megabytes and belong on the server.

top_score is reported and never acted on. A "nothing matched" threshold looked
clean on long briefs and collapsed on the short queries people type: "a logo
appears" scores 0.6181 and keyboard mash scores 0.6417, so any cut that catches
the noise rejects the real query. The measurement is in the evals directory
rather than in this branch.

Not covered here. The published recall figures were measured against a separate
hand-written document, not against registry text, so they should not be quoted
for this catalog until re-measured. The offline tier needs a normal install: a
single-file build cannot load the native ONNX runtime, which the command now
reports instead of silently degrading. And the drop-detection path has never
been observed firing outside its author's tests.

* fix(cli): make this branch pass the repo's own gates

Three things `bun run lint` and `fallow audit --base origin/main` rejected.
CI runs both, so none of this branch would have gone green. Found by running
them, not by reading the diff.

process.exit in catalog.ts, twice: an invalid --type and a cancelled picker.
check:cli-process-ownership reserves that for cli.ts, and the rule is not
cosmetic — process.exit tears the process down where it stands, so anything
cli.ts has queued to run on the way out is dropped. finishCommand throws a
CliResultSignal that cli.ts turns into the exit code, which is what init.ts
already does for a cancelled prompt.

Three exports with no consumers. normalize keeps its body and loses its export;
localEmbedder is the only caller. modelsDirectory goes entirely, having no
caller inside its file or out. The WordPieceConfig re-export goes, and with it
the import it existed to forward: the type is exported from wordpiece.ts, where
its consumers already take it from.

Complexity. prepareOnDeviceTier is lifted out of run(), which took run from 64
cyclomatic and CRAP 948 to 54 and 684. That block is one decision — can the
offline tier run, and if not, why not — and its only product is a list of
warnings, so it reads and tests as a unit, which it could not do inline.

The rest is suppressed rather than refactored, each with its reason on the line
above. Finishing run() means extracting its three output paths, and that is a
refactor of a command this branch already changes for other reasons: a separate
initiative, not something to absorb here. Every suppression says what shape the
function has and why; a bare marker on a function nobody can justify is how a
threshold stops meaning anything.

Verified: `bun run lint` exits 0, fallow reports no issues across 27 changed
files, and 2540 CLI tests pass.

* feat(cli): ship the local search tiers only, drop the hosted one

Search now has two tiers, both local: shared-vocabulary word matching, and the
opt-in on-device model. The hosted tier, which sent the query to a HeyGen
endpoint and ranked it with a hosted model, is removed.

This is a scope decision, not a defect. The endpoint works and its own change is
reviewed and green; it is simply not what we want to ship first. Landing local
only means the feature has no backend dependency, no auth requirement, and
nothing leaves the machine unless someone opts into downloading a model.

Gone: registry/smartSearch.ts and its test, the --smart and --no-smart flags,
the outcome plumbing through the command, the remote branch of applySearch, the
remote tier, and the hosted-only JSON fields (ranking, catalog_version,
top_score). Also the smartSearchEnabled consent field in telemetry config, which
was the persisted storage behind the hosted consent and would otherwise have
been left as dead configuration surface.

Kept exactly as they were: both local tiers, the --on-device and --yes flags,
the download consent prompt, and the runtime check that happens before the
download rather than after it. The --json envelope still reports query, tier,
tier_detail, shown, total, dropped, warnings and results, so an agent can still
tell which tier answered and why. tierToken now distinguishes on-device from
words.

Verified: lint exits 0, fallow reports no issues, 2522 CLI tests pass, and the
command was exercised directly. A query answers on the on-device tier where the
model is installed and falls back to word matching where it is not, reporting
that fallback in warnings rather than silently. An unknown --type still exits 1
with a readable message, and --smart is now rejected as an unknown flag.

* fix(cli): count only moves this registry cannot install as dropped

The dropped count was computed against the list left after the user's own
--type and --tag filters, so every move the user excluded was reported as one
the registry is missing. Filtering made the number go up: the same query
reported 277 unfiltered and 302 with --type block.

The count exists so a caller can tell "nothing matched your words" apart from
"the ranker suggested things this project cannot install". Conflating it with
user filtering destroys exactly that signal, and worse, genuine index skew and a
self-inflicted filter printed a byte-identical line with opposite remedies --
one means refresh the shelf, the other means drop a flag, and refreshing does
nothing.

Now counted against the registry rather than the filtered view. The manifest is
already fetched whole and narrowed in memory, so keeping the unnarrowed name set
costs no extra request, and item loading still runs only on the filtered subset.

Verified against ground truth rather than by eye: the vector artifact holds 411
names, the registry holds 168 installable items, and 134 of those names exist in
both, so 277 are genuinely uninstallable. The count now reads 277 unfiltered,
277 under --type block, 277 under --type component and 277 under --tag, and the
skew it reports is real -- the artifact predates dropping the UI primitives and
still ranks moves that are no longer on the shelf.

Reported by Vance Ingalls, who also noted this closes an item the status doc
listed as unverified. Two earlier sweeps could not make the count fire because
neither combined a filter with a query.

Tests pin the three cases: a genuinely absent name counts, a filter-excluded
name does not, and a fully installable ranking reports zero.

* fix(cli): tell the user when meaning search cannot see the catalog

The on-device index was fetched once and never revalidated: the only
freshness check was two existsSync calls. A move added after that fetch was
invisible to meaning search permanently, not down-ranked but absent from the
candidate set. The registry manifest on the same command carries a 24h TTL,
so the two halves of one feature disagreed about staleness.

The dropped count reported over-coverage only, names the index has that the
registry lacks. Under-coverage was never computed, so the harmless direction
was instrumented and the costly one was silent. Reproduced with an index
truncated to 120 of 168 moves: dropped read 0, perfect health, while 48
moves were unreachable.

Counts under-coverage from the name list the artifact already carries, so no
extra request. Warns only when non-zero, and names the remedy.

The remedy had to be made true: --on-device could not refresh a stale index
because hasLocalVectors short-circuited the fetch. That flag now refetches
when the index is absent or no longer covering.

Two defects the reproduction surfaced. A failed refresh reported the tier
unavailable while the old vectors were still on disk and still ranking. And
the fetch wrote its two files one at a time, so failing between them paired
a new name list with an old matrix, a hard load error rather than stale
data. It now writes both or neither, which matters more once refresh runs on
staleness.

top_score returns, scoped to the on-device tier and set to the score of the
best result actually shown rather than the ranking head, which can describe
a row the caller never received.

Also: scripts/ is now typechecked. It never was, which is how a build script
that crashes after the paid embedding call, and two scripts whose imports do
not resolve at all, went unnoticed. 43 errors fixed, no suppressions.

And the docs stop describing a --smart hosted tier that was deleted, an
item that does not exist, and a registry refresh that cannot fix a stale
vector index.

* ci: fail when the search index stops covering the registry

The catalog vector artifact is regenerated by hand. Nothing in CI, in
package.json or in a hook rebuilds it, because embedding needs the 32 MB
model. So adding a registry item silently makes it invisible to meaning
search until someone remembers to regenerate.

The failure is asymmetric, which is what makes it easy to miss. Removing an
item is self-healing: the ranker still scores the dead vector, then filters
the name before display, so a user is never offered something they cannot
install. Adding one is not: the item is absent from the candidate set
entirely, not ranked low.

Comparing the two name lists needs neither the model nor a network call, so
the gate runs in seconds. CI checks rather than fixes, for the same reason
it cannot regenerate.

Scoped to blocks and components. Examples are starter projects a user
scaffolds, never something catalog ranks, and the artifact carries no vector
for them, so demanding one would keep this gate permanently red and it would
be ignored within a week.

Verified in both directions rather than assumed: adding an unindexed item
exits 1 and names it, restoring the registry exits 0.

* fix(catalog): rebuild the search index from the registry

build-local-vectors.ts read registry/catalog-artifact/catalog.json, a file no script in this repo writes and which is not committed, so the documented regeneration command failed on a missing path. That is why the index could drift from the registry with nothing to run to fix it.

It now reads registry/blocks/* and registry/components/* through catalogFromRegistry, the existing helper that already produced the right shape but had no caller. Rebuilding reproduces the shipped 168 rows byte for byte.

A lefthook catalog-index command regenerates and re-stages both artifact files whenever a staged registry-item.json changes, mirroring the skills-manifest pattern, so adding or removing an item keeps the index in sync without anyone remembering to. Verified end to end: staging a new item took the artifact 168 to 169 rows and staged it in 0.80s.

* fix(cli): refuse a half-downloaded vector cache

The two artifact files have to agree on how many rows there are, and until now nothing checked that before writing them. A truncated or wrong-model response landed in the cache and only failed at load, on every later search, until someone cleared it by hand. The pair is now checked first and refused as a unit, and the cache is created 0o700 with 0o600 files rather than inheriting the umask of a directory the caller may have pointed anywhere.

Also lifts the capture setup the two preview generators had drifted into sharing into scripts/preview-capture.ts, and splits the vector builders batching and packing out of main. Both were findings the audit attributed to this branch.

* fix(cli): keep the catalog vitest run with the tests it runs

Restacking took the base package.json wholesale, which dropped the vitest dependency and the scripts/catalog run this PR adds. Both belong here rather than under it.

* fix(cli): stop the declined model download from happening anyway

Answering no to the on-device download offer recorded no and warned, then carried on. The guard below it is localModelConsent() !== false, which the decline had just made false, so it was skipped rather than taken: control reached recordLocalModelConsent(true), overwrote the answer with yes, and fetched the 32 MB model the user had refused. Next run it never asked again.

No test could catch it. The stub pinned localModelStatus to ready, so the prompt never fired, and recordLocalModelConsent was a no-op that recorded nothing.

Two tests now cover the offer, and they need three things the old stubs did not model: the run has to look like a terminal, because off one the command treats --on-device as the consent and never asks; the ONNX probe has to answer true, or an accepted offer returns at the runtime guard before it can download; and the status has to follow the recorded answer, or the second offer later in the run fires as well. Removing the return makes the decline test fail.

* fix(catalog): let someone without the model still add a component

The pre-commit hook rebuilds the search index, and rebuilding needs the 32 MB embedding model. An outside contributor adding a registry item does not have it, so their commit died inside the ONNX loader on an ENOENT naming a path they never set, and the CI gate then told them to run the command that had just crashed.

The model is an opt-in for search, not a build dependency, so nobody is charged for it to contribute. The builder checks first and explains itself, exiting 3 for cannot as distinct from 1 for failed. The hook treats 3 as skip and lets the commit through. The gate now names both paths: regenerate if you have the model, leave it if you do not and a maintainer will.

Verified both ways: with no model the builder explains and the hook exits 0; with the model it still regenerates byte-identically.

* docs: say that anyone can add a registry item, and stop hand-editing a generated file

Two defects, one of them the reason 64 stale entries survived in registry.json.

The checklist told contributors to add their item to registry/registry.json. That file is generated from the item directories, so an entry added by hand survives until the next regeneration and then vanishes, and one left behind for a directory that no longer exists is worse: hyperframes add resolves the name and then fails on missing files. Both CONTRIBUTING.md and the agent-facing skill reference now run the generator instead.

Nothing said contribution was maintainer-only, but nothing said it was not either, and two steps do need assets an outside contributor has no reason to install. Those are now named in a table with what happens if you do not have them, matching how the preview image was already handled. The search index is the new one: the model behind it is a 32 MB opt-in for search, not a build dependency.

* fix(cli): harden on-device catalog search

* fix(cli): refresh stale catalog vectors

* test: create catalog vector temp dirs securely

2026-08-09 22:59:15 -07:00

2.5 KiB

Raw Blame History

Catalog artifact

Two files ship from here, and only these two. The CLI fetches them over HTTP when a user opts into offline catalog search (catalog --query ... --on-device), so they are served from the registry rather than bundled in the package.

File	What it is
`local-vectors.json`	`{ model, dimensions, names }`. `names` is the row order of the binary.
`local-vectors.bin`	Float32, row-major, `names.length * dimensions` values, no header.

Currently 168 rows at 384 dimensions: 258,048 bytes, one row per installable registry item.

Provenance

No digest, revision or build timestamp is recorded in either file, and the runtime only checks that dimensions matches the model it loaded. model is a label the build wrote, not a proof. So the only real provenance check is to rebuild the rows and compare them, which works because both inputs are in this repository:

the text each row was embedded from is itemRetrievalText(registry-item.json) (title, description, tags, joined by newlines, name deliberately excluded), over registry/blocks/* and registry/components/* sorted by name;
the model is the pinned quantized bge-small-en-v1.5 ONNX build that packages/cli/src/registry/localModel.ts downloads.

Re-embedding in batches of 16, the batch size the build uses, reproduces the shipped rows exactly (cosine 1.000000). Batch size matters: the same text embedded alone differs at cosine 0.9969, because padding within a batch changes the quantized result. A rebuild that does not match this way was not built from this registry, or not with this model.

Rebuilding

bun scripts/catalog/build-local-vectors.ts

It reads registry/blocks/* and registry/components/* directly, so the rebuild has no input outside this repository. You rarely need to run it by hand: a lefthook catalog-index pre-commit command regenerates and re-stages both files whenever a staged registry-item.json changes, and CI fails the "Catalog: search index covers the registry" job if the index is ever missing an item.

build-catalog-artifact.ts does not produce these files. It builds the 3072-dimension hosted artifact from an external shelf file, for the hosted search tier that this repository does not ship. Its --shelf, manifest.json and text-embedding-3-large have nothing to do with local-vectors.*.

2.5 KiB Raw Blame History

Catalog artifact

Provenance

Rebuilding

2.5 KiB

Raw Blame History