Zeabur development loop

Follow the shared skills (install: npx skills add zeabur-oss/skills --skill <name>; skills-lock.json pins them): zeabur-oss-rust (Nix toolchain, workspace, lints, common checks), zeabur-oss-standards (with STANDARDS.md here), zeabur-oss-workflow (commits and pull requests; Zeabur's part in zeabur-workflow). This skill adds what is Zeabur's own.

Crates and binaries

• Each crate has a level in Cargo.toml [workspace.metadata.levels] (table in docs/internal/crates.md). Binaries: zeabur (in-cluster server), zeaburctl (CLI: ui, doctor) and zeabur-installer (runs as root on a Linux server; crates zeabur-installer-*, guide in docs/get-started/quickstart.md and docs/reference/installer*.md).
• zeabur-kube is the only crate that talks to kube-apiserver, and only through clients bound to a person's credential. Never build a client from Zeabur's service account token (the one exception is tools pods in the terminal namespace, zeabur_api::tools), and never impersonate.

Checks (what CI runs, beyond zeabur-oss-rust's)

nix develop -c bash -c 'bun install --frozen-lockfile && node_modules/.bin/zeabur-oss-unsafe-report zeabur zeaburctl zeabur-proxy zeabur-installer'
nix develop -c ./scripts/check-levels.sh
nix develop -c cargo build --release --locked --no-default-features -p zeabur -p zeaburctl

The installer's web interface (crates/zeabur-installer/web) is not in CI: run nix develop -c bash -c 'cd crates/zeabur-installer/web && bun install && bun run fmt:check && bun run lint && bun run typecheck' and nix develop -c scripts/build-installer-ui.sh. The screenshots are kept in git and not taken in CI: after changing a page that web/e2e/shots/docs.ts covers, run nix develop -c scripts/docs-shots.sh on aarch64 Linux, as CI does, and commit docs/images/shots (docs/internal/screenshots.md). The README and cover images take the same script with their set, readme or cover. Cluster tests: see zeabur-cluster-tests. Web interface: see zeabur-ui. Each CI job is nix develop -c scripts/ci/<job>.sh, the same on GitHub and on any Linux machine with Nix; scripts/ci-local.sh [job…] runs several and prints a result line per job (in target/ci-local/*.result), each in its own copy of the checkout but in the machine's own environment; on a VM, snapshot its clean state and restore it after every run (see zeabur-cluster-tests). When a check fails on GitHub, see zeabur-ci.

Lint conventions beyond zeabur-oss-rust

• The installer's output module is out.
• Integration tests wrap their tests in #[cfg(test)] mod tests (clippy tests_outside_test_module).

Comments

• zeabur-oss-standards, section 1, in Rust, TypeScript, CSS and scripts alike: only what the code cannot say; decisions go in docs/internal/.

API

• Handlers live in zeabur-api, described with #[utoipa::path], registered with routes!. After changing one, regenerate the committed description: nix develop -c cargo xtask openapi (also regenerates web/src/api/schema.d.ts).
• Every response item carries its origin (zeabur_model::Origin); errors are zeabur_core::problem::Problem. Gets return the resource itself, lists {items} with every item; event lists take a limit and add totalSize. Every answer is Cache-Control: no-store. A change that breaks a client gets a BREAKING CHANGE: footer.

Tests

• Read docs/internal/testing.md before adding, changing or removing a test: the level each behaviour belongs at, what not to test, the promises that must keep a test, polling and isolation.

Fixtures

• crates/<crate>/fixtures/<area>/<case>/ holds language-neutral cases (format in fixtures/README.md), loaded with zeabur_test_support::load!. A new operation is documented in the README's table first; zeabur-test-support's test fails on an undocumented one. Expected errors are snake_case codes, never message text.