publish-site

Files (1)

• SKILL.md

SKILL.md (first lines)

# Publish site

Standing user policy: feature work includes publishing a new binary release and
verifying that the public installer delivers it, not just pushing main. Use the
next appropriate version, update Cargo.toml/Cargo.lock and docs/RELEASE-vVERSION.md,
refresh README/site, and use the exact-commit gates below. Packaging, notes and
smoke checks must follow that version rather than a hardcoded first release.
Update the installer when its behavior needs to change; its latest-release URLs
already follow each newly published release. Test reinstalls as well as clean
installs. Do not move published tags or silently replace old release artifacts.

Workspace version bumps also change the local mg-butane package consumed by
the excluded experiments/jsplan workspace. Synchronize its Cargo.lock package
version and regenerate its reviewed dependency-audit records, then run the
locked research gates before pushing. Keep third-party versions/features and
frozen outcome expectations unchanged. The production workspace build alone
does not validate this separate lockfile; exact-commit JSPLAN CI must also pass.

Read FEATURES.md and the latest work log. Update descriptive index.html content when direction changes, with no unsupported readiness claims. Compile tools/site.rs into tmp/site and run it to refresh the marked generated section; run tmp/site --check and git diff --check.

For an authorized publication, commit and push, then inspect the Pages workflow
for the intended SHA, including canonical Cloudflare publication. Production now
uses native Cloudflare Git builds on main, consuming that exact validated public
artifact through cloudflare/; GitHub Pages remains the recovery deployment.
Do not develop on the generated codex/cloudflare-static branch or change DNS to
repair a build. Fetch https://mgbrowser.org without bypassing TLS validation,
compare its HTML against index.html and verify wpt-results.json identifies the
intended source commit. HTTP success alone does not establish visual quality;
inspect in a browser when available and state any QA limit.

Keep validation-generated Python bytecode out of source identity: suppress cache
writes in website validation and ignore only the specific generated cache paths.
Require a clean CI checkout before WPT scoring; never ignore real source edits or
accept a dirty public report merely because its counts and workflow are green.

Keep static reftests and Window declared-value parsing as separate score lanes.
Before every commit, run both complete selections with --update-score into new
ignored outputs, inspect full outcomes/resources/cleanup, then run tmp/site.
The Window owner projects only accepted evidence and never edits its manual
ratchet. Check mode must pass in CI on the exact clean source; only then may a
dependent staging job assemble both bounded reports into the public artifact.
Verify wpt-window-results.json separately from wpt-results.json, including its
check-score gate, exact SHA, source fingerprint and full tracked score parity.
Do not add denominators or call authored controls upstream passes. Unknown
partial inventories stay unknown; valid semantic Window timeouts stay non-PASS.
Measure the actual coupled CI critical path against the unchanged native wait,
not merely the maximum allowed build duration or a warm local cache.

Do not publish the repository root as the artifact: the workflow stages only index.html, CNAME, .nojekyll, install.sh, both validated WPT reports and the explicitly listed identity assets. Never include logs, tmp or raw test artifacts. The public target is the apex domain only. Recheck live hosting before any authorized DNS changes. If certificate issuance is pending, report that state and retry within a bounded verification session rather than claiming success. Do not change nameservers or unrelated DNS records to fix a deployment.

For a preview release, verify Rust CI and Pages on the exact main commit before
tagging. The v* workflow packages the locked Rust 1.91.1 binary and tests installation
before publishing a normal GitHub Release. Then verify assets/checksum, the public
installer command in an isolated prefix, version, worker selftest and desktop/icon
files. Preview publication does not complete F-007 or the Google compatibility goal.

Default installer checks with release-smoke.py and public-release-smoke.py must
remove inherited MGBROWSER_NO_AUTO_UPDATE when expecting the enabled marker.
Keep native/worker private opt-outs, explicit opt-out tests and installer
assertions unchanged. Preserve any contaminated failure; correct only the
owner environment and rerun in a fresh prefix, never a reused evidence tree.

After publication, run `python3 tools/public-release-smoke.py VERSION` from the
repo root. It verifies public bytes, the exact advertised curl command, current
version, replacement of an in-use executable, a fresh worker selftest, desktop/icons
and latest links. Existing browser processes must restart to use the new build.
Since v0.2.1, Linux workers re-exec the running inode through /proc/self/exe;
`bash tools/chrome-update-smoke.sh PATH_TO_PACKAGED_BINARY` verifies native About
and an existing scripted CDP journey after replacement of a test-owned binary.
Do not assume this continuity for older releases. For updater changes, also run
the packaged_update_smoke test with MGBROWSER_TEST_RELEASE and `-- --ignored`.
After publication, public_update_smoke with MGBROWSER_TEST_OLD_BINARY and
`-- --ignored` exercises real release download, installation and a no-op recheck
against a temporary copy of an older verified public binary. See docs/UPDATES.md.
This is a manually invoked release gate, not an unattended publication agent.

Since v0.4.0, package only the default Boa-backed binary, without the
legacy-test-engine regression feature. Installed-payload and public smoke checks
must prove modern inline script/Promise DOM execution and retained session
selftests, not just version/isolation. Keep the actual packaged /script-boa and
/script-events native/CDP journeys distinct from the frozen original-engine lane.

Stylo builds require Python 3 for its vendored code generator, but the shipped
browser does not. Regenerate docs/DEPENDENCY_LICENSES.md with
tools/license-inventory.py when the locked graph changes. It preserves upstream
texts/source links and fails on unreviewed omissions; tools/licenses contains
only explicitly reviewed missing-text fallbacks. Keep MPL dependency notices
distinct from the project's Apache-2.0 license. Verify packaged size against
the existing updater download/unpack limits without silently raising them.
The compressed tar.gz must fit the already-installed transport's 8 MiB response
limit as well as the separate 64 MiB unpacked limit. package-release.sh and the
installer smoke test both enforce the compressed ceiling. Do not infer upgrade
compatibility from a successful curl install or an unpack-only updater test.

Release gzip compression uses tools/gzip-release.py with fixed standard-library
zlib settings. Run tools/test_gzip_release.py and the actual packaged installer
and updater checks. Compare unpacked member bytes and executable identity when
reusing native receipts after compression-only changes; an archive-size gain
does not authorize pruning licenses, changing browser compilation/limits or
claiming a cold CI payload fits without measuring it. Keep the ordinary gzip
format and existing checksum/8 MiB enforcement for older installed updaters.

From pierce403/mgbrowser, Apache-2.0.

More from pierce403/mgbrowser