Testing and Validation

This is the command and coverage reference. The post-edit skill owns validation order and scope; the performance guide owns feature design and cost decisions. CI uses Node 22 with npm ci; use that version to reproduce its toolchain.

Commands and CI

The CI workflow uses one job for master pushes and PRs. Checkout and Node setup happen once. The dependency-free selector decides which steps to run, and npm ci runs at most once. npm downloads are cached only when installation is needed.

Markdown-only changes skip dependency/cache/browser installation and run only the lightweight harness, routing and hook checks. Coverage keeps its exclusions for documentation/harness files, CSS/themes, scripts, webpack, generated types, Playwright tests/configs, developer demos and unrelated metadata. These rules now live in the selector rather than workflow-level paths-ignore, allowing the harness to run even when runtime checks are unnecessary. They control execution, not Jest's collectCoverageFrom or thresholds. Mixed changes select the union of their checks. JS assets remain eligible because editor construction imports them even when their lines are excluded from metrics.

PRs compare the merge base to the PR head, including earlier commits in the PR. Pushes compare before to after, including all commits in that push. Checkout includes full history for local comparisons without GitHub's changed-file API limits. Renames include both old and new paths. Missing history (for example after a force push) selects all checks. The selected flags and comparison reason appear in the GitHub run summary.

After successful dependency installation, selected checks run sequentially. A failed check keeps the job failed while other selected checks still run; cancellation or failed installation stops dependent work. Coverage and browser artifacts are uploaded only if their producing steps actually ran. Independent translation/generation/release tools and the existing native E2E fixture still require task-specific validation when changed.

Run node --test scripts/ci/select-checks.test.cjs scripts/hooks/codex-post-tool-use.test.cjs to validate routing, comparison failure handling, GitHub output formatting and edit hooks without installing packages. Changing the workflow also requires validating step conditions and the single-install invariant. Configuring a workflow is not evidence that a hosted CI run has passed.

The existing playwright.config.js uses port 8088 and runs Firefox for *.firefox.spec.js and the shared security.xss.spec.js corpus. The separate contract configuration serves test/browser/fixture.js on port 8089, includes the real editor CSS, and runs every contract on both browsers. It starts its own server in CI and can reuse a local one.

npx --no-install playwright install --with-deps chromium firefox
npm run test:browser -- geometry.spec.js
npm run test:perf
npx --no-install playwright show-report

Browser results include HTML, JSON, failure screenshots and traces. The CI browser-contracts artifact retains playwright-report/ and test-results/ for seven days. Each performance case attaches JSON samples with browser version, fixture sizes, operation counts, median and p95.

Agent edit hooks

Codex hooks call the edit checker after apply_patch. It reads the patch from tool_input.command, resolves paths against the event's cwd, and checks existing src/**/*.js files in one local ESLint invocation. Added and moved files are included; deleted paths and paths outside this repository are not linted. It does not install packages or apply lint fixes. Node and the project's installed ESLint are required.

Changes to the option schema and Backspace/Delete rules add the owning checklist as hookSpecificOutput.additionalContext. Lint or hook-input failures return blocking feedback; the edit has already happened and is not rolled back. Shell-based file writes are outside this apply_patch hook, so the post-edit workflow still applies.

Codex must trust the project and the current hook definition before running it. Review new or changed definitions with /hooks in the Codex CLI; see the official hook contract. Claude's hooks use its separate tool_input.file_path contract. The Codex regression tests run in the existing CI job before dependency installation and use a fixture ESLint process; they do not replace source lint or establish host trust/execution.

Behavioral contracts

• EventManager tests dispatch real events: exact removal handles/capture options, array targets, host/iframe global listeners, repeated binding/removal, teardown and public-hook cancellation.
• Destroy tests initialize actual editors/plugins and verify delayed history cancellation, listener inactivity after destruction, cleared registries and a surviving second editor. These are resource-lifecycle assertions, not JavaScript heap/garbage-collection measurements.
• Geometry uses native rectangles as an independent oracle: DIV/iframe × LTR/RTL, simultaneous host/editor scroll, selection rectangles, popup alignment, both viewport edges and direction reversal/restoration. jsdom cannot prove these layout contracts.
• Architecture fixtures exercise both sides of each new boundary, including exact allowed ownership edges, root/directory plugin combinations and plugin-barrel bypasses. See Layer Dependency Rules for static-analysis limits.

These browser contracts do not replace native typing/IME tests, mobile/WebKit coverage, multi-root async races, or feature-specific dropdown/controller interaction tests. Run or add the relevant cases when those paths change; see the validation matrix.

Performance baseline and budgets

performance-budgets.json is the versioned baseline. Each scenario uses 50 and 2,000 paragraphs, including nested inline content; DIV and iframe; alignment plugin enabled and disabled; 20 warmup operations, then 20 samples × 50 operations. The selection workload alternates a nested centered paragraph and a right-aligned paragraph, checks the actual caret/toolbar result, then checks repeated unchanged selection and offsets.

Initial local measurements (macOS, Node 24.14.0; Chromium 141.0.7390.37 and Firefox 142.0.1; 2026-09-21) matched:

These counters are enforced; elapsed milliseconds are recorded for investigation, not used as a noisy shared-runner threshold. Use npm run test:perf with the same machine/browser and fixture for timing comparisons. Samples are batch averages; timer resolution can produce zero. Instrumentation and browser-native selection costs affect timing, so this is not a production input-latency benchmark or a claim that enabling a plugin improves performance.

The probes observe the named root APIs and explicit rectangle reads, not every traversal, allocation, layout flush, network request or bundle cost. Test-only wrappers verify their own realm coverage and restore originals in finally; production methods are unchanged. For another hot path, add a representative workload and an output assertion. Change a budget only with measured evidence and a reason for the additional work; never loosen it just to make a regression pass. For layout-latency claims, also collect a browser performance trace.