| name | litestar-build |
| description | Auto-activate for uv build, hatch build, PyApp, PYAPP_*, wheel assets, GitHub release matrices, cargo-zigbuild, or python-build-standalone. Not for runtime deployment. |
litestar-build
Build-side packaging patterns for Litestar applications: how to produce a self-contained wheel that embeds the Vite/Bun frontend, how to wrap that wheel in a PyApp onefile binary, and how to wire the whole pipeline into GitHub Actions CI and releases.
This skill is the counterpart to litestar-deployment — build is about producing artifacts, deployment is about running them.
The Core Idea: One Wheel, Self-Contained
A Litestar wheel is the single source of truth for a release. It contains:
- Python code (
src/py/app/ or app/)
- SQL migrations, Jinja templates, INI configs
- The built Vite/Bun frontend bundle (JS, CSS, HTML, images)
- Email templates rendered from React/MJX to static HTML
Once produced, that wheel can be:
pip installed into a container (litestar-deployment).
- Wrapped in a PyApp binary (
dist/<app>, dist/app-x86_64-linux-gnu) for zero-dep distribution.
- Uploaded to PyPI or a private index.
All three paths assume the wheel is already complete — no bun run build happens at deploy/install time.
Why bundle assets into the wheel (and not serve from a CDN)
| Property | Bundled wheel | External CDN |
|---|
| Deploy artifacts | 1 (.whl or binary) | 2+ (wheel + CDN upload) |
| Version alignment | Atomic — API and UI lock-step | Easy to skew; rollback is painful |
| PyApp onefile | Required — the binary embeds the wheel | Not possible — binary can't fetch CDN URLs at install time |
| Offline/air-gapped | Works | Doesn't |
| Dev server startup | Instant (files on disk next to package) | Fine |
| Frontend-only deploys | Rebuild + redeploy wheel | Push to CDN only |
For most Litestar apps that ship as a product (CLIs, internal tools, enterprise installers), bundled-in-wheel is correct. Projects like litestar-fullstack-inertia and litestar-fullstack all bundle.
Why litestar-vite configs look the way they do in reference apps
This is the piece most developers miss. The Vite/litestar-vite configs in the reference apps are deliberately set up so the Vite output lands inside the Python package directory — because that's what makes the wheel pick them up automatically.
litestar-fullstack (src/js/web/vite.config.ts):
export default defineConfig({
build: {
outDir: path.resolve(__dirname, "../../py/app/server/static/web"),
emptyOutDir: true,
},
plugins: [
litestar({
bundleDir: path.resolve(__dirname, "../../py/app/server/static/web"),
hotFile: path.resolve(__dirname, "../../py/app/server/static/web/hot"),
}),
],
})
litestar-fullstack-inertia — the litestar-vite plugin resolves bundle_dir relative to the project root, and Python settings point it at a package-internal path:
return ViteConfig(
paths=PathConfig(
root=BASE_DIR.parent,
bundle_dir=Path("app/domain/web/public"),
resource_dir=Path("resources"),
),
)
Advanced reference pattern — same approach: Vite and the offline-report build write to src/py/<app>/server/public/ and src/py/<app>/domain/web/static/reports/offline/, both under the package root.
Contrast with a naïve vite build that writes to ./dist/ at the repo root: those files are outside the package directory listed in [tool.hatch.build.targets.wheel] packages = [...], so Hatchling silently drops them. The wheel ships without a frontend.
Rule: Vite's outDir and litestar-vite's bundle_dir must point inside one of the Python packages that Hatchling is told to include. Everything else flows from that.
Quick Reference
Canonical Makefile Build Graph
Every Litestar app with bundled assets has some variant of this:
.PHONY: install build-assets build-wheel build-onefile
install: ## Install Python + JS deps
@uv sync --all-groups
@cd src/js/web && bun install --frozen-lockfile
build-assets: ## Build frontend into the Python package
@uv run app assets install
@uv run app assets build
build-wheel: build-assets ## Self-contained Python wheel
@uv build --wheel
build-onefile: build-wheel ## Single-file PyApp binary
@./tools/scripts/build-onefile-package.sh
The dependency chain is load-bearing: build-onefile depends on build-wheel, which depends on build-assets. Running them out of order produces an empty or broken artifact.
The two-variant story
Real projects have multiple JS build outputs that all need to land in the wheel:
js-build-all: js-build-web js-build-offline-report
build-wheel: generate-licenses build-templates js-build-all
@uv build --wheel
Each js-build-* target emits into a distinct subdirectory of the Python package (src/py/<app>/server/public, src/py/<app>/domain/web/static/reports/offline, etc.). Because they're all inside the package, a single uv build --wheel captures everything.
Workflow
Step 1: Point Vite output inside the Python package
Open vite.config.ts. Set build.outDir to an absolute path inside your Python package (src/py/<pkg>/... or <pkg>/...). Set litestar({ bundleDir, hotFile }) to the same path. Do not let Vite default to ./dist/.
Step 2: Choose a Hatchling bundling strategy
force-include (inertia): List the built-asset directory explicitly under [tool.hatch.build.targets.wheel.force-include]. Built assets stay .gitignored. Explicit, auditable.
ignore-vcs = true (SPA): Tell Hatchling to ignore .gitignore. All package files ship. Simpler; requires discipline to keep dev junk out of package dirs.
See references/wheel-assets.md for full config.
Step 3: Wire Makefile targets
Create install, build-assets, build-wheel. Make the wheel target depend on the asset target. Add any secondary generators (build-templates, generate-licenses) as additional wheel prerequisites.
Step 4: Add PyApp (if shipping a binary)
Decide which flavor:
- Simple: add
[tool.hatch.build.targets.binary] to pyproject.toml and run uv run hatch build --target binary. Good when end-users have PyPI access. See pyapp-simple.md.
- Advanced: write a
tools/bundler.py that pre-installs deps into a python-build-standalone archive, patches PyApp's src/app.rs for a custom install dir, then runs cargo zigbuild. Good for air-gapped distribution or bespoke install locations. See pyapp-advanced.md.
Step 5: Add GitHub Actions CI
Start with a reusable test.yml that accepts python-version + coverage inputs. Call it from ci.yml across a matrix. Use astral-sh/setup-uv@v7 and oven-sh/setup-bun@v2. See github-ci.md.
For larger projects, factor setup-python and setup-node into .github/actions/ composite actions.
Step 6: Add release workflow
Trigger on v* tags. Run the test matrix first. Then build the wheel once. Then build PyApp onefiles in a per-target matrix (x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu, Apple, Windows). Upload to gh release create. See github-release.md.
Guardrails
- Vite/bun output must land inside a Python package directory. Otherwise Hatchling drops it. Set
build.outDir and litestar({ bundleDir }) to an absolute path under src/py/<pkg>/ or <pkg>/.
uv build runs last. Assets, licenses, templates, OpenAPI TypeGen all run before uv build --wheel. Hatchling can't build Vite itself.
- Pick one bundling strategy.
force-include or ignore-vcs = true, not both. Mixing them causes duplicate-file warnings and unpredictable wheel contents.
- PyApp envs are build-time, not runtime.
PYAPP_PROJECT_NAME, PYAPP_PYTHON_VERSION, PYAPP_DISTRIBUTION_EMBED are consumed when cargo build compiles PyApp — not when the resulting binary runs. Setting them at runtime does nothing.
- PyApp version upgrades touch multiple files.
pyproject.toml, build-onefile-package.sh, .github/workflows/release.yml, tools/bundler.py. See upgrading.md.
cargo-zigbuild for portable glibc. Plain cargo build on a modern Linux runner produces binaries that fail on older distros (glibc too new). Use cargo zigbuild --target x86_64-unknown-linux-gnu.2.17 to link against glibc 2.17 (CentOS 7-era). Required for broad compatibility.
- Static-link native deps in PyApp. Set
BZIP2_SYS_STATIC=1 and LZMA_API_STATIC=1 before cargo zigbuild, or patch Cargo.toml to add features = ["static"]. Otherwise the onefile fails to load on systems without matching libbz2.so / liblzma.so.
- Pin
uv and bun versions in CI. Use exact pinned versions (e.g., UV_VERSION=0.11.6 and BUN_INSTALL_VERSION=bun-v1.3.12). Drift in either breaks reproducible builds.
- Create placeholder asset dirs in CI. Hatchling's wheel target fails if
app/domain/web/public or src/py/app/server/static/web doesn't exist at wheel-build time. CI jobs that don't build the frontend (lint, mypy, pyright) still need mkdir -p <asset-dir> before uv sync.
- Never commit built frontend output. Keep
bundle_dir paths in .gitignore. CI rebuilds them on every run. Reason: JS builds are non-deterministic across machines and cause noisy diffs.
- Coverage on one Python version only. Multiple versions uploading the same
coverage.xml silently stomp each other. Pin it to one version in your matrix (if: matrix.python-version == '3.12').
- Disk cleanup on self-hosted runners. GitHub's
ubuntu-latest has ~30GB free; building wheels + PyApp + Docker images can blow past that. Aggressive cleanup before the build job is routine.
Validation Checkpoint
Before claiming "the wheel builds":
Before claiming "the PyApp binary works":
Before claiming "CI works":
Example Projects
Everything in this skill is distilled from three production projects. Read these for the full picture:
- litestar-fullstack-inertia — monolithic
app/ layout, Inertia.js + React 19, force-include bundling, hatch build --target binary for 4-platform PyApp.
- litestar-fullstack — nested
src/py/app/ + src/js/web/ layout, React + TanStack Router SPA, ignore-vcs = true bundling, React Email templates.
Official References
Cross-References
- litestar-deployment — runtime deployment (Dockerfiles, K8s, Railway, Cloud Run, systemd) that consumes the artifacts this skill produces
- litestar-vite — Vite plugin config (asset pipeline details, TypeGen, HMR)
- litestar-granian — Granian ASGI server (what the wheel's entry-point starts)
- litestar settings — env-driven
@dataclass settings that work both in-wheel and as a PyApp binary
Shared Styleguide Baseline