Codetrial Web

The CodeTrial browser boundary

web/ is plain ES modules loaded by the page. There is no bundler, no transpile step and no framework, which is what makes an edit in the checkout visible on reload. Keep it that way: a build step here would turn every browser fix into a build artifact somebody has to remember to regenerate.

How a file reaches the browser

rust-embed compiles web/ into the binary, so a release build serves a complete application with no web/ directory beside it. The fallback is per file rather than wholesale: a file present on disk wins, and one that is missing is filled in from the embedded copy. That is why a partially populated tree serves rather than 404s, and why running the binary from a checkout picks up your edit with nothing set. --web-dir or CODETRIAL_WEB_DIR points that root somewhere else.

The consequence worth remembering: a debug build that reads from disk and a release build that serves embedded bytes can disagree, and only the second is what ships. scripts/test.sh covers this, and CI proves the binary serves with the web tree deleted.

Vendored assets

web/vendor/ holds third-party bytes that are downloaded rather than committed: the LiveKit client, Pyodide, the MediaPipe face detector. They are pinned by a SHA256SUMS beside each of them; scripts/fetch-vendor.sh fetches them and scripts/verify-vendor.sh checks them, both wired into the gate. Never edit a file under web/vendor/, never lint it, and never add a dependency by dropping a file there by hand: the checksum is the whole mechanism. The avatar model is not even vendored, the browser fetches it from a pinned upstream URL against a pinned hash, and the interview falls back to a voice-only panel when it is unreachable.

The data channel is two implementations

Every topic on the LiveKit data channel has a producer in web/lib.js and a consumer in src/agent.rs. Each side has its own tests, and a change that tests only its own side is exactly the shape that once let the integrity hash diverge and silently empty the evidence section of every camera interview while the gate stayed green.

tests/fixtures/*.json are the answer: generated by node scripts/gen-wire-fixtures.mjs from the real producer, consumed by the Rust tests, and held to --check in the gate. So when you change a message:

1. Change the producer in web/lib.js.
2. Regenerate the fixtures and commit them with the change.
3. Change the consumer in src/agent.rs against the new fixture.

Never hand-edit a fixture. Everything in the generator is deterministic, timestamps included, because a fixture that changes on every run cannot be checked.

Generated pages

web/problems/, web/judges/ and the problem cards inside web/index.html are generated from problem-bank/. Edit the bank and rerun the generators; see codetrial-verify.

The lint gate

eslint.config.mjs is narrow on purpose: every rule in it fails only where the code cannot mean what it says, so the gate never becomes something people learn to argue with. no-undef is the reason it exists, since a typo'd identifier in web/interview.js reaches a candidate's browser while the same mistake in src/web/mod.rs never leaves the terminal. Three recommended rules are excluded with the reasoning written above them. Adding an eslint-disable is a stronger claim than it looks: there are none today, and the Rust half carries four #[allow]s in 45k lines, each with its reason on the same line.

Browser behavior gets a Node test in tests/browser/*.test.js and, where it needs a real browser, scripts/browser-check.sh with Playwright and Chromium.