Skip to content

Automation

Matthew Feroz edited this page Sep 30, 2026 · 3 revisions

Automation

This page distinguishes implemented automation from sprint work. Documentation and prototype publication does not deploy the application or run paid evaluations.

Continuous integration

.github/workflows/ci.yml runs on main pushes, pull requests, and manual dispatch. It installs dependencies using the frozen Bun lockfile and the project's configured Bun version.

Job Observable checks
CLI on Node 22 Type check, CLI tests, build/pack, installed-artifact browser smoke; retain the tarball
Connected evaluation end to end Harbor 0.23.0, Docker, live daemon/supervisor and local Convex; Oracle execution produces child and combined reports without model inference
Verify Lint, TypeScript/Vite build, server/domain/chart/Convex/export tests, evaluation UI integration, first-smoke onboarding, browser behavior

.github/workflows/release-checks.yml is a separate manual release gate. It checks installed setup on Ubuntu x64 and ARM64; its compatibility job covers Node 24, packaged CLI viewing, deployment build modes, and the older Docker fixture runner. Those jobs are not claimed to run on every pull request.

Deployment and release

Vercel uses the configured build script; a configured Convex deployment key allows a matching backend deployment and frontend URL injection. CLI publication and worker upgrades are explicit release operations. A local source change does not upgrade an installed worker or publish a package.

The presentation renderer uses a pinned Chromium snapshot in Vercel Sandbox. Operators update that snapshot when renderer assets change. Hosted export jobs persist immutable inputs, store completed private files, and stop rendering resources. This is separate from evaluation compute.

Evaluation automation

  1. Pairing, a health check, and the Oracle setup check establish worker readiness.
  2. Advertised profile/task digests constrain browser selections to locally approved execution.
  3. Backend claims coordinate dispatch; a detached supervisor snapshots tasks and launches Harbor.
  4. Bounded task telemetry updates the UI; upload failure does not block final reporting or cancellation.
  5. Normalization uploads sanitized child results; terminal experiment cells trigger a combined private report.
  6. Shared chart recipes support presentation edits and exports. The owner explicitly publishes or revokes a revision.

Execution-error retries are configurable and differ from verifier failure. A reconnection is not permission to start a duplicate paid run. Runtime/resource settings do not enforce a dollar budget.

Sprint follow-up

Sprint 1 scopes documentation usability validation, reconnect/cancel validation, report-sharing validation, and reproducibility evidence into achievable work with explicit completion criteria. Larger infrastructure work and paid benchmark studies are outside this sprint.

CI workflow, release checks, connected smoke, CLI release guide, and hosted exports.

Clone this wiki locally