Developing #

Run prek install once; it wires both the pre-commit and the commit-msg stage, and until someone runs it a commit here is checked by nobody. The pre-commit stage runs cargo fmt, cargo clippy and cargo doc across the whole workspace, so against a cold target/ it is minutes, and cargo doc is wired to docs/*.md as well as to the Rust sources, so a documentation-only commit pays for it too. Against a warm one it is seconds.

Keeping it seconds is what scripts/lint.sh is for: lint by hand with that rather than a bare cargo clippy. Cargo caches a check against the flags it was given, so the two are separate cached results over the same source, and alternating between them re-checks every crate in the workspace each way.

The file size hook counts the lines in each staged file: over 5000 it refuses the commit, and over 2000 it prints the file and lets the commit through. Both numbers are at the top of scripts/check-file-size.py.

Two hooks call a tool rustup does not ship, so install each once: cargo install cargo-deny and cargo install --locked --features cli cargo-about. The first judges the dependency graph against deny.toml; the second writes THIRD-PARTY-NOTICES.txt, the licence text the image carries beside the binary. A dependency change that moves the graph fails the notices hook until scripts/gen-notices.sh is run and its result staged.

The commit-msg stage checks that the subject is a Conventional Commit and stamps the Change-Id trailer that stacked pull requests are matched by; that trailer is what a --no-verify commit gives up along with the checks, and a commit made without one has to be rewritten to get one.

The test suite runs from scripts/ci.sh [<base>]: it runs the hooks over every file rather than a staged set, replays the commit-msg stage over every commit in <base>..HEAD, checks that brand images are up to date via scripts/build-brand.sh --check, and then runs cargo test --workspace --all-features --no-fail-fast. A branch is checked when a person runs it.

The runner #

.tangled/workflows/ci.yml runs scripts/ci.sh on a push to main and on each pull request round. Tangled runs pipelines on a spindle, and a spindle is self-hosted: it is a service somebody runs, registered to their own account and attached to this repository from its settings page. Until one is attached to this repository, the workflow is a file and a person runs the script.

The workflow carries three things a laptop already has: npm ci in both site/ and policy-site/, because scripts/build-site.sh refuses without site/node_modules; wasm-bindgen-cli at the version Cargo.lock resolved, because it and the wasm-bindgen crate talk a private ABI to each other; and cargo-about at the version the workflow names, which has to be the one that last wrote THIRD-PARTY-NOTICES.txt, because another version writes different notices and the notices hook fails. Its compiler is the one rust-toolchain.toml pins, as a laptop's is. It also raises CARGO_BUILD_JOBS: .cargo/config.toml holds cargo to two jobs for a machine that is also running an editor and several agents, and a runner is that machine's opposite.

A spindle ends a workflow after SPINDLE_NIXERY_PIPELINES_WORKFLOW_TIMEOUT, five minutes by default, and this one runs far longer; whoever runs the spindle sets it.