browser-check

Files (1)

• SKILL.md

SKILL.md (first lines)

# Browser check

For every commit, follow docs/WPT.md: run the entire pinned reftest selection into a new
ignored directory with tools/wpt/run.py --update-score, review its full result,
then regenerate tmp/site and include the deterministic score and page. The
runner rebuilds the Rust adapter; do not substitute an old executable. Preserve
the exact corpus and manually reviewed pass/support ratchet when scoring engine
changes. XML, fonts, scripting and other unsupported prerequisites stay visible
in the denominator. Never count harness controls or readable fallback as WPT
passes. Pages reruns --check-score on the exact commit, including docs-only work.
This headless Browser lane does not replace native/public-release acceptance.
Every PASS page/reference must have an empty rendering-diagnostic list and zero
omissions as well as all existing replay/resource/viewport/pixel guards. A
nonempty diagnostic is a conservative FAIL, not a post-render UNSUPPORTED escape
or permission for a case-specific exemption. Opaque host/artifact failures stay
ERROR; completed=false or PENDING evidence cannot update the tracked score.
Corpus/runner expansions must precede scored engine changes. Preserve the
immutable original 24-test manifest/ratchet as well as the expanded required
pass/support sets. Freeze a source-derived roster before running the browser;
never filter by observed failures. Retain full selected/PENDING totals on an
interrupted run, but never accept PENDING as a completed score. Check the
documented process, input and evidence bounds and all runtime helper hashes.
The report stores each unique page once; inspect its lossless RGB/PNG/log and
typed failure record before accepting new passes. No source navigation or
script-proposed navigation may escape the local test-server lane.

For WPT orchestration changes, follow the separately reviewed two-lane policy
in docs/WPT_EXPANSION.md. Keep a one-lane control and compare complete runs
using the same binary/font/source: all case identities/statuses, every rendered
page's raw RGB hash and original relations, and per-page HTTP request multisets.
Aggregate counts alone cannot establish equivalence. Preserve shared run,
evidence/report/request limits, single-flight references and bounded queues.
Unknown process/fixture cleanup stops admission; it cannot free another renderer
slot. Keep partial PENDING evidence honest and verify exact-commit cold CI and
actual public report delivery after normal publication.

For JSPLAN engine research, also read `experiments/jsplan/README.md` and
`docs/jsplan/RESULTS.md`. Keep the experiment's workspace/lockfile separate from
production. Build its release executable with Rust 1.91.1, run protocol and
runner tests, verify the exact dependency audit, then run the pinned-input
comparison and `tools/jsplan/check_report.py`. The check includes known failures:
do not shrink the denominator or update expectations automatically. Use the
supervised empty-environment pipe protocol with the shared production isolation;
never initialize an experimental engine in the browser parent. Record debug and
release results separately. GNU time measurements exclude the Python launcher's
inherited memory high-water state; distinguish peak process RSS from live heap.
Research-only tooling is not a browser backend or a framework compatibility claim.
Any actual browser adoption still requires the standing release/installer gates.

Production Boa page execution uses the separate process-contained profile in
docs/BOA.md. Run modern, boa_pages, boa_page_events and boa_worker tests in debug
and release, including real restricted-child allocation/job/opcode failures.
Preserve the old interpreter's tests with --features legacy-test-engine and its
explicit --legacy-page-tests fixture lane. Never package that feature or infer
production Boa compatibility from original-engine fixtures. Packaged production
acceptance uses /script-boa and the unchanged /script-events through native input
and external CDP, plus the installed-binary worker/session smoke checks.
For typed external classic source or performance/cache changes, also build normal
journey_server/cdp_journey examples and run tools/page-services-smoke.py against
the packaged and public-installed executable. It owns its loopback/display and
checks external startup order, retained event behavior, worker accounting and
actual CLI cache/failure paths. Preserve the original inline-script journeys.
Run external_scripts regressions in debug and release; never encode trusted
source metadata in page HTML. Run inspector-smoke for Performance shortcut/tab,
compact/scaled and light/dark acceptance, and inspect its saved screenshots.
WeakRef/removed-listener tests must use actual engine checkpoint cleanup, not
test-only ClearKeptObjects. Distinguish requested System allocation counters,
GC reachability and RSS; parser/native/regex/GC work still relies on hard process
containment where cooperative hooks are absent. This is not full JSPLAN P1/P4.

For function timers or parent-task scheduling, read the reviewed increment in
docs/WPT_TESTHARNESS.md as well as BOA.md/PAGE_SESSIONS.md. Run boa_timers and
script_tasks in debug and release, existing modern/Boa/legacy event/projection
and actual-worker tests, strict Chassis task controls and independent no-chrome
embedding. Preserve shared pending/lifetime work, CPU/active/lifetime/wire/child
bounds. Canceled work never refunds cumulative admission. Test real checkpoints
and traced callback/argument release, hostile delay conversion, direct versus
Promise-created nesting, stale/early/replayed wakes, exactly-one projection
charge, native edit preservation and owned-child cleanup. Do not repair roots
with test-only checkpoint cleanup or sleep inside active worker execution.
Run tools/task-services-smoke.py on packaged and public-installed bytes: its
authored timer-created controls must complete the unchanged native and external
CDP cancellation/form/destination assertions. This is not upstream WPT coverage.
The static adapter still does not await arbitrary timers; audit current-cohort
passing script logs for unfinished work before score adoption. A new asynchronous
scoring prerequisite/corpus requires its own frozen contribution, not a silent
change to waits or exceptions while scoring an engine fix. The original timer
increment's bundled Init was not completed Window-testharness readiness; queued
snapshot lifecycle has separate acceptance below and is not full readiness either.

For Window uncaught errors, also read docs/WINDOW_ERRORS.md and run boa_errors,
external_scripts and actual boa_worker/script_tasks controls in debug/release.
Check five-argument onerror/literal-true cancellation, capture-phase stop versus
same-invocation continuation, retained diagnostics, recursive/fatal no-JS paths,
real traced exception release and sticky foreign/unknown/dynamic-source muting.
Parent origin proof covers every redirect and survives cache/defer; DOM/base
attributes cannot supply it. Preserve the current strict modern wire version,
strict legacy v1 bytes and existing resource counters; never guess an old modern
frame's shape. Run tools/error-services-smoke.py on packaged and
public-installed bytes: require its exact expected diagnostic/PARTIAL per
journey plus native/CDP identity, checkpoint and retained form proofs. Do not
relax the old diagnostic-free page/task checkers. These are authored controls,
not rejection notification, exact source attribution or new upstream WPT passes.

For Promise rejection notification, also read docs/PROMISE_REJECTIONS.md. Run
boa_rejections, boa_rejection_resources and real script_tasks in debug/release,
plus core private fault/storage/ID/weak-map/finalizer controls and all old gates.
Check one frozen batch per parent pump, true late handling, shared work admission,
active-candidate re-entry, genuinely weak outstanding membership, and terminal
root/permit/wake cleanup. GC tests must use normal checkpoints, not test-only
kept-object repairs. Muted Reject retains no payload/permit; the conservative
observation-gap notice is sticky across replies, including loss of outstanding
handled observation. Actual notification diagnostics belong to their transaction
and cannot be retracted by cancellation, handling or later clean-looking output.
Run tools/test_rejection_services_smoke.py, then tools/rejection-services-smoke.py
on packaged and public-installed bytes. Its notification-created form requires
native/CDP payload identity, checkpoints, retained input/cancellation/navigation,
exactly one expected diagnostic and zero trap requests. Keep the old
diagnostic-free and Window-error checkers unchanged. This procedure derives from

From pierce403/mgbrowser, Apache-2.0.

More from pierce403/mgbrowser