[atinux]

An alternative to #946 for comparison. Written by an AI agent (Amp) on behalf of @atinux.

Each nuxt dev start runs tsc to generate hub/db/schema.d.mts, even when the schema has not changed. This PR caches the declarations instead of moving tsc to the background:

• After a full build, schema.d.mts is copied to <buildDir>/cache/nuxthub/db/. A hash of its inputs is stored next to it. Nuxt keeps cache/ when it clears the build directory on dev and prepare.

• The hash covers:

  • every local file the JS and declaration builds read, including type-only imports such as a $type<Meta>() interface
  • the versions of the packages the builds import (for example drizzle-orm)
  • the aliases the schema imports actually use
  • the TypeScript version

  Only the used aliases are hashed, because Nuxt aliases such as #app-manifest change on every prepare/build.

• When the hash matches, only schema.mjs is built and the cached declarations are copied back. Otherwise the build runs as before.

The change is contained in buildDatabaseSchema, so dev, prepare, build, nuxt db and the schema watcher use the same path. There is no background work and no extra CPU, and a schema edit refreshes the types as fast as before.

Numbers

Playground (SQLite, 3 schema files), Node 24, 4-CPU Amp orb. Each round deletes playground/.nuxt, then starts nuxi dev playground twice:

• The first start rebuilds everything. This PR has no cache yet, so it runs tsc like main.
• The second start reuses .nuxt, which is the usual restart. This PR hits the cache.

The rounds alternate between the three branches. Each value is the median of 5 rounds.

	main	#946	this PR
Second start
types phase	2.95 s	1.11 s	1.13 s
"Ready in"	8.49 s	7.15 s	6.46 s
First HTML	18.3 s	16.5 s	15.7 s
CPU to first HTML	27.3 s	27.2 s	22.5 s
First start (no .nuxt)
types phase	3.52 s	1.22 s	3.12 s
"Ready in"	10.8 s	8.74 s	9.47 s
First HTML	23.4 s	20.0 s	19.2 s
CPU to first HTML	33.5 s	31.6 s	28.5 s

"First HTML" is measured from process start until GET / returns 200. "CPU" is user plus system time of the nuxi process tree at that moment. The orb is noisy: single runs of the same branch differ by up to about 30%, so read the medians as trends.

On the second start, #946 and this PR cut the types phase by the same amount. #946 still runs tsc in the background during startup, while this PR skips it. On the first start, #946 is ahead on "Ready in" because this PR still runs tsc in the foreground.

buildDatabaseSchema alone, on a small schema: 1.5 s without the cache, about 7 ms with it, and 0.7 s to rebuild after a change to a file imported only for its types.

Checks

• New test: the cached types are reused, and a change to a file imported only for its types invalidates them. The test fails if that file is not tracked.
• pnpm lint, vue-tsc --noEmit and pnpm test pass locally (167 passed, 3 skipped).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
nuxthub	Ready	Preview	Oct 27, 2026 8:50am UTC

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	playground	7cd0640	Commit Preview URL Branch Preview URL	Oct 27 2026, 08:50 AM

[pkg-pr-new]

npm i https://pkg.pr.new/@nuxthub/core@947

commit: 7cd0640