Tests protect privacy, authority, durable Markdown, operation semantics, and complete keyboard access before they optimize coverage percentages. A green command applies only to the exact source, dependencies, environment, profile, and artifact that produced it.
| Purpose | Command |
|---|---|
| Source hygiene | corepack pnpm check:source-artifacts |
| Workspace typecheck | corepack pnpm typecheck |
| Deterministic compiled suite | corepack pnpm test |
| Coverage thresholds | corepack pnpm test:coverage |
| Desktop source contracts | corepack pnpm test:desktop |
| Accessibility source contract | node scripts/check-frontend-accessibility.mjs |
| Build generated public docs | corepack pnpm docs:site |
| Reject public-doc drift | corepack pnpm check:docs-site |
| Browser build + budgets + secret scan | corepack pnpm build:web |
| Supported local browser smoke | corepack pnpm test:desktop-browser |
| Rust authority tests | cargo test --locked --manifest-path apps/desktop/src-tauri/Cargo.toml |
| Packaged desktop build | corepack pnpm build:desktop |
| Frontend evidence manifest | corepack pnpm evidence:frontend |
| Conditional SBOM/checksums | corepack pnpm sbom:frontend |
| Candidate source bundle | corepack pnpm release:frontend:candidate |
| Opt-in provider check | corepack pnpm test:live-provider |
The final V2 orchestrator validation is a separate required state check. It does not substitute for code, browser, packaging, or manual evidence.
- closed runtime schemas and one operation registry across browser, desktop, administrator, provider, and agent audiences;
- Host/Origin, browser bootstrap, cookie/CSRF, rotation, revocation, and project-binding matrices;
- privacy completeness, missing-visibility fail-closed behavior, never-send rules, secret canaries, and agent result projection;
- project-generation A-B-C/A-B-A races, cancellation, polling disposal, idempotency, compatibility, and outcome-unknown reconciliation;
- destructive intent expiry, replay, wrong target/revision, and cancellation;
- write-only provider secrets and destination/DNS/redirect/size policy;
- loading, refresh, partial, stale, offline, refused, failure, and recovery state distinctions;
- dialogs, forms, focus, route recovery, typed route reachability, graph virtualization, bounded keyboard navigation, and structured graph parity;
- semantic tokens, registered contrast pairs, reduced motion, forced colors, formatter ownership, and responsive source constraints; and
- public docs, bundle budgets, forbidden imports/fixtures, source maps, credential names, and emitted secret canaries.
The checked-in CI workflow runs source hygiene, typecheck, deterministic tests, browser builds, Windows Edge smoke, packaged desktop build, Rust tests, and a clean-diff check across its declared jobs. It does not by itself prove the manual assistive/device matrix, live provider compatibility, installer/signing, or rollback rehearsal. Public-doc and accessibility commands must be verified against the final integrated workflow before claiming that CI enforces them.
The Windows CI jobs compile a deliberately inert Rust executable and pass it as the Tauri external-binary fixture. This lets Cargo and Tauri exercise their real command manifest, capability, sidecar naming, copy, cleanup, and packaging boundaries without committing or publishing a binary. The fixture exits immediately and is never a product daemon or release artifact. A releasable Windows application still requires the separately built, approved, hashed daemon sidecar and packaged runtime smoke.
Local automated validation was run on the uncommitted 2026-08-12 working tree with supported Node 24.19.0. The workspace typecheck and 333 tests passed with zero failures and two intentional Windows symlink-safety skips. The production web build and budgets, isolated secret-canary build, fixture reachability, accessibility source contract, generated public docs, source-artifact guard, and headless Edge startup-recovery smoke also passed. Six Rust unit tests and the Tauri compile/package mechanics passed with the explicitly inert fixture. These are local results, not release approval: no real packaged daemon/runtime or signed Windows installer was qualified. The qualification matrix keeps local implementation confidence separate from commit-bound release and device evidence.
- Bind branch, commit, lockfile digest, profile, toolchain, and clean status.
- Review the final changed-file list and production/test boundary.
- Run source hygiene, V2 state validation, typecheck, and deterministic tests.
- Repair only focused failures, then rerun a broad command once if affected.
- Build and check public docs.
- Run accessibility source checks.
- Build browser assets; run budgets, forbidden-reachability, source-map, and secretless emitted-byte scans.
- Generate evidence manifest and SBOM/checksums for the same artifacts.
- Run available browser/Rust/Tauri gates and the declared manual matrix.
- Record unsupported combinations and unexpected skips without upgrading them to pass.
Use disposable stores and synthetic credentials only. Never point tests at a private memory root. See AI provider testing, Frontend qualification matrix, and Frontend accessibility contract.