Test Lumen

Outcome

Each completion claim is backed by a test at the real ownership boundary rather than a narrower proxy.

Prerequisites

Bootstrap the repository and start Docker. Identify the normal, failure, restart, duplicate, cross-workspace, cancellation, budget, secret, artifact, and recovery paths that apply to the change.

Steps

1. Run the owning unit or contract test while iterating.
2. Run integration tests against migrated PostgreSQL and real filesystem artifact bytes.
3. Use deterministic providers for ordinary CI; gate live providers separately.
4. Use Playwright only for actual browser behavior, including keyboard, accessibility, viewport, motion, console, network, and screenshot state.
5. Run chaos and recovery tests at the process/provider boundary being claimed.
6. Run pnpm check before integration.

The workbench browser gate compares deterministic light and dark screenshots with the committed baselines in apps/web/visual-baselines. Regenerate them only after reviewing an intentional visual change: run the gate once with --update-visual-baselines, inspect both PNGs, then rerun without the flag so the comparison itself is the saved proof. A missing baseline or a difference above the committed threshold fails the gate.

That gate drives the production artifact through loading, empty, success, retryable failures, offline/reconnect, settings ambiguity and conflict, invalid deep links, command palette, contextual inspector, Research activity sheet, and the structured @ reference composer. It checks keyboard navigation and focus return, 44px coarse-pointer targets, 320px reflow, dark appearance, forced colors, and reduced motion. The zoom case changes Chrome's actual default zoom to 200%, verifies the resulting CSS layout viewport and responsive shell, then restores the browser setting; a device-scale override is not accepted as zoom evidence. Every exercised state must have no unexpected console or network errors and no automated accessibility violations.

Static presentation tests keep Hugeicons as the workbench's single general-purpose icon set and reject scale transforms in interface controls. Motion evidence checks the computed styles of panels, tooltips, the composer, disclosures, sheets, and navigator labels under reduced motion rather than assuming a media query is sufficient.

Verify

Inspect what each test asserts. A green command is insufficient if its fixtures never exercise the named requirement. Save exact commands, environment, results, and honest limits in execution evidence.

Recover

When a test fails, preserve the first failure, reproduce narrowly, and fix the owning path. Do not weaken a gate, skip the case, or declare a delegate's report to be proof.

Next task

Apply the design system.