Firefox Add-ons approved the first version. The owner reported it on
2026-11-21. AMO's public API shows the add-on public, version 1.3.0,
Firefox 140 or later, slug "bookmarkmagic". Both the slug URL and the
gecko-id URL the About link uses answer 200.
- README: the Firefox install row links the listing, with an AMO version
badge beside the Chrome Web Store and Edge ones.
- docs/00: the Type line, the 1.3.0 status (live on all three stores), §9
(gecko id, slug, listing), §10 item 21 done and item 24 done. Item 23's
automation state was re-checked and is unchanged: firefox-add-ons still
does not exist.
- docs/13 §8 and store/listing.*.md: the licence to pick is "GNU General
Public License v3.0 or later". AMO offers "v3.0 only" beside it, and that
is what the listing ended up with (GPL-3.0-only, per the API) — the docs
only said "v3.0". Item 21 keeps it open, along with the empty support site.
- docs/15 and CLAUDE.md: Firefox Add-ons live; 1.3.0 is on all three.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Chrome Web Store approved 1.3.0 on 2026-11-19. Its listing shows version
1.3.0 ("Updated: November 19, 2026", 4 users) — the first update since 1.2.0,
carrying 1.2.1 as well. Firefox Add-ons is the one store still reviewing.
- docs/00: the 1.2.0, 1.2.1 and 1.3.0 status paragraphs. §10 item 20 is done,
but the 11 §5 upgrade pass and 13 §1b step 10 still have no recorded
result. Item 23 records the store automation as configured today: all
three *_AUTOPUBLISH variables are true, so a pushed v* tag is a store
submission. chrome-web-store and edge-add-ons have their secrets and the
v* tag rule but no required reviewer. firefox-add-ons does not exist yet.
- docs/15: CWS live, Firefox Add-ons in review.
- CLAUDE.md: the banner, and a carry-forward note in Phase 10: a tag is an
upload to users' browsers, so it needs the owner's go.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owner submitted v1.3.0 to all three stores on 2026-11-18, and Edge
Add-ons approved it on 2026-11-19: the first Edge version, public, five
languages. The Chrome Web Store update and the first Firefox Add-ons
version are still in review.
- README: the Edge row of the install table links the listing, and an
Edge Add-ons version badge (a shields dynamic badge over the store's
public getproductdetailsbycrxid JSON) sits beside the Chrome Web Store
one.
- docs/00: the Type line and v1.3.0 status; §9 gains the Edge item id
(CRX ID), Product ID, Store ID and listing; §10 items 20-22 and 24
record what was submitted, what is live and what is still open (the
11 §5 passes have no recorded result; sync-on Edge is still unverified).
The 1.2.1 paragraph says it never went up on its own.
- docs/13 §9: the About link matches the CRX ID, not the Product ID; done
for 1.3.0.
- docs/15: v1.2.1 submission skipped; v1.3.0 submitted; v1.3.0 live on
Edge; Chrome Web Store and Firefox Add-ons in review.
- CLAUDE.md: the banner and the Phase 10 exit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BookmarkMagic went live on Edge Add-ons on 2026-11-19. Its "Store listing"
link chooses the listing by runtime.id, and only the Chrome Web Store's id
was known, so an install from Edge Add-ons was sent to the README's install
section.
storeUrlFor now looks the id up in a map of each Chromium store's item id
to its listing. Edge's item id is Partner Center's "CRX ID", which is what
an install reports; the Product ID is a different number, the one the
EDGE_PRODUCT_ID secret takes. Any other id (an unpacked build) still goes
to the install section.
Ships with the next release (CHANGELOG [Unreleased]); docs/06 §3.5 says
so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Partner Center warned on the zh-CN description: "Avoid referencing other
browsers in your extension description such as Firefox". Edge's developer
policy (1.1.2) says a listing must not reference other browsers, and the
first Edge copy was the Chrome Web Store description with two bullets
swapped: the export bullet still listed seven browsers, and the swapped
import bullet itself named Chrome, Brave and Firefox.
- store/listing.*.md: Firefox Add-ons and Edge Add-ons get a heading
each. The Edge section is a complete description, pasted whole, that
names no browser but Edge: the profile "Bookmarks" file is the one "Edge
and similar browsers keep", HTML export is "the standard bookmark file
every major browser can import". Every other line matches the Chrome
Web Store copy; the Firefox copy is unchanged.
- tests/unit/store-listing.test.ts: one Edge description per locale, 250
to 5,000 characters, no other browser's name; and the same pattern must
find the browsers the Chrome Web Store copy lists, so it cannot pass by
matching nothing.
- docs/08 §10, docs/13 §9, docs/11 §4, store/README.md: the rule, where it
is tested, and policy 1.5.2's "should" about the privacy policy, left
as it is unless review raises it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The hand sync every Dependabot bump needs: docs/01 §2 names the version
the lockfile holds, and Dependabot only moves the lockfile.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v1.3.0 is tagged (signed, on 80c7190) and released: three packages and
SHA256SUMS.txt, each checksum verified after download, and both manifests
passing check-manifest as downloaded. Nothing is submitted to a store yet.
- CLAUDE.md: the banner names 1.3.0 and what it is; the doc index points
at docs/08's Firefox and Edge sections; Phase 10 gets its exit and two
things to carry forward (no Firefox in CI; every change ships in two
packages).
- docs/00: the Type line says Firefox; a v1.3.0 status paragraph; doc
suite 1.13.
- docs/15: "v1.3.0 released" with the checksums and sizes, and
"v1.3.0 submitted" pending for the owner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Firefox and Microsoft Edge Add-ons, and maintenance mode. Permissions are
unchanged; nothing re-asks for consent (LEGAL_VERSION stays 1).
- package.json and the lockfile's root at 1.3.0; CHANGELOG [1.3.0] with
Added / Changed / Fixed / Packages.
- docs/00 §10 items 19-26: what only the owner can do for this release
(the store uploads, the first AMO and Edge submissions by hand, the store
environments, real-browser QA, native review).
- docs/15: Phase 10 in the progress table; 1.2.1 may be skipped on the
store, since 1.3.0 contains it.
- docs/13 §8: the sources archive rebuilt every file of the Firefox
package byte for byte. docs/08 §9: Options can open a second Manager tab
on Firefox.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An adversarial review of the Firefox port found:
- #edit treated a Firefox separator as an item in three places: a folder's
count badge, whether a folder can be expanded, and whether it can be
sorted. A folder of separators and one bookmark offered Sort A-Z.
- Replace's confirmation counted every writable root, so a signed-in
Chrome's account roots, which Replace never empties, were counted among
what it would delete. In every release since 1.1.0; it now counts the
replaceable roots only.
- The store jobs checked their files against the SHA256SUMS.txt that came
in the same artifact, with --ignore-missing. Each file is now checked by
name against the published Release's checksums.
- Qodana's publish job ran only after a successful scan, so a scan failing
on findings never published them.
- getRoots' position fallback had no test. It now has two, with unknown
ids and a Bookmarks Menu listed first.
Docs follow: the two browser modules and the warning codes (02), where the
landed index is compared (05), the separator markup and folder count (06),
what the Firefox suites are (11), the release staging and environments
limited to v* tags (12, 00), and which loose entries reach the Bookmarks
Menu on Firefox (03, 04, README).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- BUILD.md: rebuilding the release packages, written for Firefox Add-ons'
source review (Node.js 26 stated outright: the sources archive drops
.nvmrc), and useful to anyone checking a store build against the source.
- README: a browser/store/requirement table (Edge and Firefox listings not
live yet, and said so), the Firefox package on each Release is for
checking, not installing; the privacy section names the same CSP in every
browser; limits gain Firefox's separators, address and title lengths.
- SECURITY.md: the latest version on each of the three stores is supported.
- ACCESSIBILITY.md: Firefox and Edge are best effort and untested with
assistive technology; separators are drawn but not announced, and Alt+↑/↓
past one leaves the spoken position unchanged — a known limitation.
- store/listing.*.md: the AMO and Edge blocks — the two bullets that named
Chrome rewritten for browsers in general, one more limitation for Firefox,
and each store's fields. Translations best-effort.
- CHANGELOG: the 1.3.0 entry under [Unreleased] until the bump.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- ci.yml: the Firefox package is built and checked beside the Chrome one.
- release.yml: three release files — the Chromium package (Chrome Web Store
and Edge Add-ons), the Firefox package and AMO's sources archive — each
staged by exact name so a missing one fails the run, both manifests checked
against the tag, SHA256SUMS over all three.
- One store job per store (docs/12 §3), each off until its repo variable is
true, each in its own GitHub Environment holding only its own secrets, each
checking the downloaded file against SHA256SUMS before uploading. They
install tools/publish — publish-browser-extension 6.2.0 and its three
dependencies, from their own lockfile, scripts off — instead of the ~320
packages of the dev toolchain, which the single old job installed next to
four Chrome secrets.
- The old job could not have worked: it sent CWS API v1.1 credentials, and
v1.1 was shut down on 2026-10-15. store-chrome uses v2 (publisher id and a
service account) and sets CHROME_API_VERSION explicitly. store-firefox is
listed and desktop-only; store-edge sends the Chromium package.
- Dependabot watches tools/publish; the licence gate reads its lockfile too.
- qodana.yaml: clean since the first run's findings were handled, so any new
one now fails that (non-required) workflow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What made Firefox more than a build target (docs/02 §4, 03, 04, 05 §10):
- Roots. getRoots (now browser/roots.ts) tries Firefox's GUIDs after
folderType and the Chromium ids, and never takes the Bookmarks Menu by
position — Firefox lists it first, and the old fallback took it for the
bar, so a Replace emptied the Menu and the toolbar and restored into the
wrong roots. The Menu is a fourth destination Replace also empties.
- The Menu in the file model: root "menu" (BM JSON, additive); HTML writes
its entries loose at the top level, as Firefox's own export does; on
Firefox loose top-level entries are routed into the Menu, elsewhere a menu
folder stays a folder under Other. New-folder mode is unchanged.
- Separators come back as { type: "separator", url: "data:" }. Recognised by
type (browser/live-tree.ts): left out of exports and backups, of every
count, of the dedupe index and the duplicate finder, never a search match;
in #edit they are aria-hidden lines that take no focus or command, with
posinset counting items only, Alt+↑/↓ stepping past one as one place, and
Sort A→Z sorting each run between them. The Replace confirmation says how
many will go and that nothing brings them back; files' separators are
counted (SEPARATORS_SKIPPED).
- Move index: Chromium counts a rightward same-folder move before removal,
Firefox after. The adapter sends each its own number and the reply's index
is checked — a mismatch re-reads the tree. Before this, every rightward drag
on Firefox landed one place too far and Alt+↓ moved two.
- Firefox refuses addresses over 65 536 characters, mid-write: dropped before
planning on Firefox (URL_TOO_LONG); nothing changes on Chromium.
- Strings true in every browser: the date note says "the browser", the About
link is "Store listing" and goes to the listing the copy came from (AMO by
gecko id, the Chrome Web Store by item id, else the README's install
section); the smart-bookmark note drops "dead link". Diagnostics names the
browser. All five locales (ja/zh-CN/ro best-effort).
bookmarks.ts was over the 300-line target, so its pure half moved to
live-tree.ts and root resolution to roots.ts before anything was added.
Tests run the root, Replace, export and #edit paths over a Firefox flavour of
the fake (docs/11 §4). Also fixes the last Qodana finding, a redundant escape
in the i18n interpolation regex.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
wxt.config.ts now builds for the target browser (docs/08 §1, §9):
- `manifestVersion: 3` for every target — WXT built Firefox as MV2;
- Firefox gets browser_specific_settings.gecko — the permanent id
bookmarkmagic@poli0981.dev, strict_min_version 140.0 (the first that
accepts data_collection_permissions, which AMO requires of new add-ons),
and data_collection_permissions { required: ["none"] } — and no
minimum_chrome_version; Chrome keeps its floor of 136;
- Vite targets firefox140 for that package;
- the sources archive AMO reviews leaves out coverage, captures, store copy,
docs, tests and tools.
Permissions, CSP, pages and icons are identical in both packages.
check-manifest.mjs takes --browser chrome|firefox, mapped through a fixed
table to .output/<target>-mv3 (still no path from the command line), and
pins each package's exact key set and floor; the gecko block is compared key
order aside. New npm scripts: build:firefox, zip:firefox; verify:full builds
and checks both packages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Docs first, as CLAUDE.md requires; the code follows in this branch.
- 08: the manifest per target (two packages from one source, MV3 for
both); Firefox's gecko id, data-collection declaration, floor 140 and
source-code submission; Edge takes the Chrome package unchanged.
- 02 §4, 03 §1/§3, 04, 05 §10: Firefox's GUID roots and fourth root, the
Bookmarks Menu (`root: "menu"`, unmarked entries routed into it on
Firefox); separators recognised by type, kept out of the file model and
shown as untouchable lines in #edit; the move index sent by each browser's
own rule and checked against the reply; addresses Firefox refuses dropped
before planning; separators a file holds counted.
- 06/07: the About link follows the install; three strings that blamed
Chrome for what every browser does are reworded; new keys listed.
- 11: a Firefox flavour of the fake, and a manual pass on real Firefox and
Edge (including Edge signed in with sync) before the first submissions.
- 12/13: three store jobs, each in its own environment with only its own
secrets and a four-package install; CWS API v2 (v1.1 died 2026-10-15);
AMO and Edge runbooks for the first, manual uploads.
- 09/14 and legal/: three stores named; a clarification, not a
LEGAL_VERSION bump.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first scan found 27 issues and no defect:
- two regular expressions that could be simpler, fixed: the sanitizer's
[\p{Extended_Pictographic}] class and wxt.config.ts's split(/[/]/). A third
is in src/ and waits for the 1.3.0 change, so this branch still builds
byte-for-byte what v1.2.1 shipped;
- 24 that contradict a rule of this repository, switched off in
qodana.yaml with the reason beside each: JSDoc completeness (doc comments
here are prose, docs/10 §6), bitwise operators in the PNG CRC-32 of
gen-icons.mjs, and the deliberately doubled dark palette in tokens.css.
`publish` failed: the qodana-report artifact holds a zip whose entries have
absolute names, so a glob for the SARIF found nothing. It now extracts the
one entry /qodana.sarif.json by name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recorded now rather than with the port, because the maintenance
announcement names it (docs/15, docs/00 §7). Settled with the owner: gecko
id bookmarkmagic@poli0981.dev, desktop only (Firefox for Android has no
bookmarks API), MV3 everywhere, and Firefox separators shown, protected and
warned about rather than stored in a file format.
The roadmap is pruned to what MAINTENANCE.md allows: scheduled snapshots
need the alarms permission and leave it; the permission rule joins the never
list; the fflate zip export, the duplicate auto-merge assistant and
bookmarklet notes are not planned; the old v1.3 "resilience" becomes v1.4.
MAINTENANCE.md no longer lists "a dependency advisory" as a release trigger
on its own: with no runtime dependencies, an advisory against the build
tools needs a release only if it changed what ships.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
qodana.yml runs qodana-js (Ultimate, QODANA_TOKEN) on pushes to main, pull
requests, a weekly schedule and by hand, in two jobs (docs/12 §2.7):
- `scan` installs the dev dependencies with scripts off — qodana-js resolves
types through them — and holds only a read-only GITHUB_TOKEN; PR comments
and annotations are off, the SARIF report is kept as an artifact;
- `publish` installs nothing and uploads that SARIF to code scanning
(category qodana), so the findings sit beside CodeQL's.
The action is pinned to the commit of v2026.2.2 (the newest tag is younger
than the 7-day cooldown). Sharing a runner with dependency code is recorded
as an exception to the Sonar rule: the token can only upload reports.
triage.yml, from review of the first draft:
- comment and close first, label last — the label is what a re-run reads as
done, so it must not exist before the work;
- an `enhancement` issue with no answer to the form is closed only when the
owner added the label; otherwise it goes to needs-triage;
- checkout and setup-node pinned to commits, no cache: this workflow runs
them with a write token on events anyone can trigger.
README gains the Qodana badge.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MAINTENANCE.md is the policy: dependency updates on a schedule, high and
critical advisories fixed as soon as a fix exists, major platform upgrades,
bug/UI/UX/accessibility/translation fixes, and features only when they are
on docs/15's roadmap or genuinely needed.
- The feature form asks one required question — on the roadmap, needed for
a fix, or something new — and links the policy and Discussions › Ideas.
- scripts/triage.mjs decides what happens, as a pure, unit-tested function;
.github/workflows/triage.yml carries it out with `gh` alone. Something new
is labelled maintenance-mode, commented and closed as not planned; the two
in-policy answers get needs-triage; an outside `feat…` pull request is
closed the same way. Owner, members, collaborators, bots and `roadmap` are
exempt.
- pull_request_target checks out the base branch only, runs nothing from the
pull request, and never interpolates a title or body into the shell.
- README gains a status badge and a "Project status" section; CONTRIBUTING,
the PR template and the issue chooser say what is accepted.
- docs: 00 status, 11 inventory, 12 §2.6 and §5 (labels), 13 §6, 15 (a
pointer to the policy, not a copy, and the decision-log rows); CLAUDE.md
Phase 10.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scripts/check-licenses.mjs reads package-lock.json, where npm records each
package's own `license` field, and fails on anything outside two lists
(docs/09 §3.2):
- what the build compiles into the extension (svelte, clsx, wxt,
@wxt-dev/browser) must be permissive and GPL-3.0-compatible;
- everything else is build tooling that never ships, so file-level copyleft
(MPL-2.0, lightningcss) is also accepted; AGPL, source-available terms and
missing or custom licences are not.
SPDX expressions are evaluated, nothing is installed or fetched, and the
tests call the functions directly rather than spawning the script. It runs
in `verify`, CI's quality job and the release build.
The shipped list moves out of wxt.config.ts into
scripts/credited-packages.mjs, so the build gate, the notices test and the
licence gate read one list instead of two copies and a regex.
legal/THIRD_PARTY_NOTICES.md said the tooling was all MIT/Apache-2.0/ISC; the
lockfile also has BSD, 0BSD, MIT-0, BlueOak, CC0 and MPL-2.0. Corrected — a
description of tooling that never ships, so no LEGAL_VERSION bump (14 §3.1).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dependabot's #39-#41 moved the packages and left the rest behind, as
docs/15 (2026-11-10) says it always does:
- biome.json and docs/10 name the schema of Biome 2.6.6;
- docs/01 records Biome 2.6.6, knip 6.44.2 and jsdom ^31.0;
- docs/12 §2.5 names the scan action's new pin, v8.4.1;
- docs/11 §6 gains a jsdom 31 column for the 100k parse (median of three
runs, 2 484 ms) instead of overwriting the jsdom 29 figure.
docs/09 §3.1: npm audit now reports 14 high findings with one root cause,
picomatch GHSA-fc5q-39xc-vjpf, for which no fixed version exists. It is the
toolchain's glob engine, matching patterns this repository writes, and
never ships (--omit=dev stays 0). The deepmerge row goes: GitHub withdrew
that advisory on 2026-11-16.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed tag v1.2.1 on d61318a (PR #37, merged at the reviewed head). The
Release asset verifies against SHA256SUMS.txt (4975b5df…ca13254, 106 290
bytes); its manifest differs from 1.2.0's in `version` alone and the file
inventory is the same. `announce` and the store publish were skipped, as
their variables are unset.
Still the owner's: the docs/11 §5 pass with the 1.2.0 -> 1.2.1 upgrade over
an existing profile, a real Windows contrast theme and a screen-reader pass
over Alt+Up/Down, and the store upload.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
stores/edit-live.ts was untested — it is the block that used to sit in
EditTab's onMount, where nothing measured it. As a .ts module it counts, and
SonarCloud's new-code gate failed at 73.6 % (80 % required).
edit-live.test.ts drives it with wxt/browser mocked to hand-fired event
emitters and a hand-fired frame: one patch per frame, a parentless created
node ignored, nothing patched during Chrome's own import and one re-read
after it, a re-read on a wholesale reorder, and a teardown that detaches every
listener and drops what was queued. connectLive takes the batcher's scheduler
as an optional argument for that.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Accessibility. An accessibility statement and its intake, Alt+Up/Down to
reorder from the keyboard, Windows contrast themes, and the focus and
target-size fixes an audit of src/ found. Permissions unchanged;
LEGAL_VERSION unchanged.
The built manifest differs from the v1.2.0 release asset in `version` only;
the file inventory is identical.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Two independent reviews, of the code and of the documentation against it.
Code:
- A first click on a row button or a chevron jumped the tree to the top:
handing focus to the tree with no row focused made it pick the first one.
The clicked control's row becomes the focused one, and the tree only picks
the first row for a focus that came from the keyboard — which also stops a
plain click on a row in a scrolled tree from landing on the wrong row.
- Three forced-colours rules lost to more specific ones: the selected theme
lost its label on hover, About's status-bar entry had no right edge, and a
drop into the focused folder showed nothing.
- Under a search, Alt+Up/Down announced the position among all siblings while
the row's aria-posinset counts the ones on screen; it announces the latter.
Moves run one at a time, and the live region is emptied before each
message so a repeated text is spoken.
- The preview's "Show more" was still 22 px; the ring on "Show earlier / Show
more" and on the "Move to…" folders was clipped; placeholders used the
browser's grey, which no contrast test saw.
- EditTab's script was over the 150-line limit: its live-sync wiring moves to
stores/edit-live.ts.
Documentation:
- ACCESSIBILITY.md promised more than the code does in nine places ("every
clickable target is 24 px", "nothing times out", "focus is always a 2-pixel
ring", four operating systems "supported"). Rewritten to what the source
supports; the untested list and the known limitations grew.
- docs/05, 06, 10, 11, 13, 15 and CLAUDE.md corrected where they named tests
that did not exist, miscounted, or were overtaken by the fixes above.
tests: a managed node, hidden siblings under a search and an in-flight move
for nudge; the statement's key table is checked against tree-keyboard.ts.
capture/about.png shows the new About link.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Svelte's a11y_* checks are compiler warnings, and a warning never failed
anything: `npm run check` now runs svelte-check with --fail-on-warnings. It is
the only lint that reads the markup (Biome skips .svelte). The suite is at
zero warnings; the two deliberate exceptions carry a svelte-ignore.
scripts/screenshots.mjs built its sample tree without parentId, which Chrome
gives every node but the root. In the harness nothing could be placed beside
a sibling and the toolbar's Sort stayed disabled on a bookmark — found while
driving Alt+Up/Down through it.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The tree's keys are shown nowhere in the UI, and someone who meets a barrier
has no route from inside the product to reporting it. About now links
ACCESSIBILITY.md beside the changelog — the file, not an address: no email
appears on an in-product surface (docs/06 §3.5).
links.test.ts pins the URL and the file's existence, as it does for the
changelog. The label is new in all five locales; ja, zh-CN and ro are
best-effort like the rest of those files.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Placing an item before or after a sibling took a drag: "Move to…" only
appends, and Sort A→Z was the one keyboard reorder (WCAG 2.1.1; backlog since
the 1.1.0 review). Alt+Up and Alt+Down now move the focused row one place
among the siblings on screen.
No second write path: nudgeTarget picks the neighbour, and the drag's own
planDrop and moveTo do the rest, so a key and a drag cannot disagree about
what is legal — nothing happens at either end of a folder, on a permanent
root, on a managed node or while a bulk command runs. A polite live region
speaks "<title>: position n of m": aria-posinset changes too, but a screen
reader does not re-read the active descendant for an attribute change.
Focus no longer falls to <body>:
- a route change that takes the focused control with it ("Open in Edit" on
the import report) moves focus to <main> — only when it was lost, and not
while the page loads;
- after "keep the first of each", to the panel's close button;
- after a confirmed toolbar Sort A→Z, whose button is disabled while it runs,
to the tree.
EditTab's dialogs move to EditDialogs.svelte to keep the tab under 300 lines.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Found by an audit of src/ after a keyboard-only and Narrator pass over 1.2.0
found nothing; verified in a headless Chrome driven over CDP against the
harness, with forced-colors emulated.
- The theme control's ring was invisible. It is inset to stay inside the
group's border, and roving focus only ever lands on the selected segment,
which is filled with --accent: accent on accent. It is --accent-fg now, 4px
in, so the fill shows on both sides of it.
- Contrast themes (forced-colors) remove every fill, shadow and background
image. The tree's focused row and drop lines, the selected theme, the
current tab and status-bar entry, the selected mode and format cards, the
progress fill and the drop zone's dashed edge each get a system-colour
equivalent. A border or outline wherever the state has a label: Chrome
paints a backplate behind forced text, and a label on a Highlight fill
became a box that hid it.
- Every pointer target is at least 24 x 24 px: the row buttons (22), the
three disclosure chevrons (16, now one Disclosure component with a grown
hit area), the toast and dialog dismiss buttons, the popup's gear, the
picker's "Show more".
- A click on a row button or the chevron hands focus back to the tree.
Chrome focuses a tabindex="-1" button on click and the tree ignores keys
aimed at a row's own control, so one click switched the arrow keys off;
"Move to…" from the row button now returns focus to the tree as well.
tests: focus-ring.test.ts fails on an inset ring over an accent fill;
forced-colors.test.ts holds the list of components that must carry the block.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
ACCESSIBILITY.md at the repository root, after opensource.guide's
accessibility practices and the accessibility-toolkit template: WCAG 2.2 AA
as a goal and not a conformance claim, supported environments, the tree's
full key table (shown nowhere in the UI), what was tested and with what
(keyboard only and Narrator with Chrome on Windows 11), what was not, the
known limitations, and how a report is handled by a one-person project.
Intake: an "accessibility" issue form and label, accessibility@poli0981.dev
in the issue chooser and the README's contact table, an interface section in
CONTRIBUTING.md and in the PR checklist.
tests/unit/accessibility-statement.test.ts pins the routes and the key
table, since nothing in the build reads the file.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
v1.2.0 is on the store (listing observed 2026-11-10: version 1.2.0, updated
November 7, five languages); CLAUDE.md, docs/00 and docs/15 still said it had
not been submitted.
docs/06 §5 now specifies what an audit of src/ found the 1.2.0 keyboard and
Narrator pass could not: a ring inside a filled control takes the fill's text
colour, forced-colours equivalents for every state shown by a fill or a
shadow, Alt+Up/Down as the keyboard's before/after (through planDrop, docs/05
§10), three more places focus fell to <body>, row buttons handing focus back
to the tree, and 24 px targets.
docs/11 §5 gains the checks no script runs: a screen reader, a contrast
theme, 200 % zoom, and "ACCESSIBILITY.md is still true". docs/12 §4 says an
accessibility fix is a patch. docs/15 schedules what is not closed.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Dependabot bumped @biomejs/biome to 2.6.4 and left the `$schema` URL in
biome.json at 2.6.2; `vitest` stayed at ^5.0.5 beside a ^5.1.0 coverage
plugin that docs/01 says must match it.
docs/01 §2 now states what is installed (WXT 0.21.5, Svelte 5.57.3,
svelte-check 4.7.8, Knip 6.44.1, Vitest 5.1.0) and docs/12 the Sonar scan
action's pin (v8.4.0).
WXT 0.21.4 -> 0.21.5 passed the built-manifest diff docs/12 §2.3 requires:
manifest.json byte-identical to the v1.2.0 release asset, same file
inventory, CSS identical, +33 bytes in one chunk from Svelte 5.57.3.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Signed tag on e2f36bc; the Release zip (105 284 bytes) passes
`sha256sum -c`, and its manifest differs from 1.1.0's only in `version`,
with _locales/zh_CN and _locales/ro added. Store submission, the upgrade
pass and the two new listing languages stay with the owner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The template still asked for "all three" locales; it now points at
LOCALES and the Romanian `few` form (docs/07 §5).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Simplified Chinese and Romanian (docs/07). Permissions, the Chrome
floor and LEGAL_VERSION are unchanged; the built manifest differs from
1.1.0 only in `version`, plus the two new _locales directories.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Translations (an independent native-style review of every key):
- ro export.csvNote said the bookmarks-bar folder itself is lost on CSV
export; only *which* folder is the bar is lost.
- ro: counts that are plain numbers use a label form ("Omise — deja în
acest browser: {n}") so 1 never reads as a plural; "din Chrome" for
the brand; mode names quoted exactly; the tagline, appDesc and listing
summary as a noun phrase instead of an informal imperative; the
backup promise in replaceWarning made explicit; GPL wording.
- zh-CN: quoted mode names match the real labels (全部替换); 一半 no
longer claims 50%; one name for Chrome's Bookmarks file; 您 only where
it earns its place; the English-only note no longer implies
translations exist.
- Uncertain strings carry TODO(review-zh/ro) for the native pass.
Code:
- Each language option carries its own `lang`, so endonyms are voiced
and drawn in their language (WCAG 3.1.2; Chinese vs Japanese glyphs).
- resolveLocale ignores -u-/-x- subtags; the docs/07 table states that
an explicit script wins and other zh regions get zh-CN.
- The listings' LANGUAGES line is built from LOCALES, and format tests
derive their locale lists, so a new locale cannot slip past them.
- screenshots.mjs back under 300 lines: one job list for --shoot.
- Unneeded casts removed from dictionaries.test.ts; docs/11 describes
the popup shots and what the sweep cannot cover.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- store/listing.zh_CN.md and store/listing.ro.md (best-effort, marked
TODO(review-zh) / TODO(review-ro)); every listing and the README now
name all five languages.
- The translation issue form offers 简体中文 and Română; CONTRIBUTING
explains the plural forms and the comma-below rule, and points at
docs/07 §6.
- scripts/screenshots.mjs photographs every locale, and the popup per
theme and language.
- tests/unit/i18n/touch-points.test.ts holds the files that cannot import
LOCALES (harness, check-manifest, the form, the listings) to the list;
the cedilla scan covers the Romanian listing too.
- Comments that said "three locales" or "VI/JA" now say what they mean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- ro.ts, ro-notices.ts and ro-import.ts, written from Chrome's own ro
terminology (marcaj, Bara de marcaje, dosar) in the polite plural, and
marked TODO(review-ro) until a native pass (docs/07 §4).
- The first locale with three plural forms: every plural key carries
one / few / other ("1 marcaj", "5 marcaje", "20 de marcaje"); count-less
labels repeat the bare noun (docs/07 §5).
- Tests: comma-below ș ț only (a cedilla anywhere in the Romanian files
fails, comments included), "de" only in `other`, selectPluralForm at
the CLDR boundaries (0, 1, 2, 19, 20, 101, 120), tPlural end to end,
and ro-RO collation for Sort A→Z.
- _locales/ro/messages.json; check-manifest requires it.
- Română joins every dictionary's endonym list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- zh-CN.ts, zh-CN-notices.ts and zh-CN-import.ts, written from Chrome's
own zh-CN terminology (书签, 书签栏, 文件夹, 导入, 导出) and marked
TODO(review-zh) until a native pass (docs/07 §4).
- _locales/zh_CN/messages.json; check-manifest requires it.
- resolveLocale maps zh, zh-CN, zh-SG and zh-Hans* to zh-CN. zh-TW,
zh-HK, zh-MO and zh-Hant* get English, matching Chrome's own manifest
fallback (docs/07 §1-2).
- <html lang="zh-CN"> makes Chrome draw Simplified glyphs rather than
Japanese ones; Sort A→Z collates by pinyin.
- 简体中文 joins every dictionary's endonym list.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Groundwork for zh-CN and ro, with no new language yet (docs/07 §2, §5):
- resolve-locale.ts exports LOCALES; Locale is derived from it. dicts,
TAGS, the storage allow-list and the Settings choices all follow it, the
first two as Record<Locale, ...> so a missing table is a compile error.
- Dict is Localized<typeof dict>: plural branches may carry `few`.
selectPluralForm keeps `few`; requiredPluralForms says which forms a
locale's own Intl.PluralRules need.
- dictionaries.test.ts checks plural shape per locale, guards against
locales that need zero/two/many, holds the endonyms equal, and pins
the Dict type with @ts-expect-error cases.
- manifest.test.ts maps locales to _locales directories, rejects strays
and caps appDesc at 132 characters; options, links and storage tests
iterate LOCALES.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
v1.1.0 cleared CWS review (listing shows 1.1.0, 6 users, observed
2026-11-06). The dashboard's install figures make Simplified Chinese and
Romanian the next minor, so the spec says the new thing before the code:
- 07: one LOCALES list, the zh-CN/ro resolution table (Traditional
Chinese falls back to English, matching Chrome's manifest fallback),
a third plural form for Romanian, tone and glossary per language,
the best-effort review policy, and a new §6 "Adding a locale".
- 12 §4: a new UI language is a minor; translation fixes stay patches.
- 15: v1.2 is languages; the old v1.2 items move to v1.3.
- 00, 02, 03, 06, 08, 11, 13, 14 and CLAUDE.md follow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- docs/15: Phase 7 done (#26–#29); a "v1.1.0 released" row with the signed
tag on 9040f84, the checksum-verified zip and its manifest diff against
1.0.1 (version, Chrome floor, CSP — nothing else); a pending "submitted"
row; a decision-log row on why the tag moved.
- docs/00: the header says released, not in preparation; §10 item 12 lists
what is still human-only for the store; doc suite 1.9.
- docs/13 §7: the one case where re-pointing a tag is right — a run that
failed before any Release existed.
- CLAUDE.md: phases 0–7 complete; the Phase 7 exit as it stands.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first tag run of the split release workflow (v1.1.0) failed in "Gate +
package": actions/upload-artifact skips hidden paths by default, `.output`
is one, so `.output/*-chrome.zip` and `.output/SHA256SUMS.txt` matched
nothing and `if-no-files-found: error` stopped the job. The publish job was
skipped, so no Release or asset was created.
The build job now copies both files into a plain `dist/` and uploads that.
The artifact layout the publish and store jobs download is unchanged, and
`include-hidden-files` stays off, so nothing else under `.output` can be
swept into a release.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A second adversarial review, of the review-fix commits themselves, found
six problems. All are fixed here.
- Focus after a delete: a delete from the duplicate panel, or from the trash
button of a row that did not have focus, left keyboard focus on <body>.
The dialog cannot hand focus back to a button that went with what it
deleted. `confirmDelete` now says where the delete was asked for: the
tree gets focus back, and the panel moves it to the next copy's button,
else to its close button. Checking for <body> was not enough: after a
dialog closes, focus sits on its button until the browser's next
rendering update, so a control in a closed dialog, or a detached one,
now counts as lost too. Confirmed in the harness for all three paths.
- Import step 04: confirming the safety backup removed both of its buttons
and dropped focus. A phase change now hands focus back to the heading.
- WebRTC: both gates matched `\bRTCPeerConnection\b`, which no word
boundary lets match `webkitRTCPeerConnection`, an alias Chrome still
ships. Prefixed aliases now count. The evasion comment in guard.mjs no
longer claims to cover every way of hiding a member name.
- New folder: Chrome fires onCreated before create() resolves, so the live
event could insert the folder first and newFolder then inserted a second
row with the same id. It now inserts only when the folder is not already
there, with a test that loses the race on purpose.
- Screenshot harness: `wxt build` empties .output first, and re-listing it
threw inside the request handler, which ended the process. It now answers
404 or 503 and recovers once the build is back.
- The duplicate summary uses one wrapper key instead of two fragments glued
together in markup (docs/07 §2); Japanese gets its own separator again.
- Docs: the popup's tabId and `tabs.getCurrent` (docs/02 §6, 08 §2), the
new-folder race (03 §3), focus after deletes and import phases (06 §5).
SonarCloud's scan job now also receives the generated .wxt/, so
tsconfig.json's `extends` and the `@/` alias resolve (12 §2.5).
- Tests for edit/tree-keys.ts, which had none, and for the focus helpers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud's quality gate failed PR #27 on two findings in new code.
- S2083 (blocker): the screenshot harness built the file path from the
request and then checked it stayed inside the build. The check held, but
taint analysis cannot see a prefix test, so the harness now lists the
build once and serves only files from that index; a request can name a
file only by looking it up. It re-lists once on a miss, so a rebuild
while it runs still works. Probed: valid files 200; `..`, `%2e%2e`,
`%2F` and `%5C` traversals 404; a malformed escape 400.
- S8786: `^\s*\d*\s*$` in the CSV sanitizer let two `\s*` compete for the
same blanks. `^\d*$` on the trimmed value means the same and is linear.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- docs/03 §3: one bulk command at a time, focus that stays on a rendered
row, the search that remembers closed folders, the new-folder path.
- docs/05 §9: the 500-row window around the focus replaces the growing
prefix; measured numbers in place of the old "not done".
- docs/06: rename focus, the pinned failure toast, step-heading focus and
the import's live region, dialog naming, the new component names.
- docs/02 layout, docs/07 plurals (eighteen since v1.1.0), docs/08 §3 CSP
with default-src 'none' and the wider bundle gate, docs/09 §7 typed
diagnostics, docs/12 the Sonar job split and the ops build job's plain
`npm ci`, docs/15 decision rows and the review's backlog.
- README and PRIVACY say what the CSP refuses instead of "any request".
- capture/: the README set regenerated by `screenshots.mjs --shoot` from
neutral sample data, lowercase names as the harness writes them; the
1.0.0 shots are gone. store/screenshots/: 1280×800 dark candidates.
- CHANGELOG [1.1.0] Fixed; CLAUDE.md size figures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Findings 5–9 of the security review of 1.1.0. None broke the zero-network
or permission promises of the shipped build; each made a claim stronger
than the code behind it.
- Diagnostics: every DiagEvent string field is now a union of fixed values
(import phases, FileFormat, the ErrorKeys describeError returns, the two
edit actions, the three storage issues), and error names pass an
allowlist, logged as 'other' otherwise. `@ts-expect-error` cases keep a
title from compiling instead of trusting review.
- guard.mjs: optional-chained computed access on globals, `Reflect`, and
member names assembled from strings or templates (`el['inner' + 'HTML']`)
all got past it.
- check-manifest.mjs: the bundle is scanned for the same identifiers as the
source, as bare words — `fetch.call`, aliases, RTCPeerConnection and
importScripts shipped unseen — and every src/href in the built pages must
exist inside the package.
- CSP: `default-src 'none'` closes everything unlisted. Every directive the
pages use is listed, so nothing changes for them; the harness pages load
with no violation.
- screenshots.mjs: listens on 127.0.0.1 only, checks `OUT + sep` rather than
a bare prefix, decodes paths safely, and runs Chrome asynchronously — the
synchronous exec blocked the very server Chrome was waiting on, so
`--shoot` hung forever.
- SonarCloud: coverage runs in a job with no secret; the scan job installs
nothing and receives lcov as an artifact. A job's secrets are readable by
every step on its runner.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The adversarial review of the 1.1.0 UI found fifteen problems that the unit
tests and svelte-check could not see; all but one were reproduced in the
built harness first.
- Edit tree: F2 and New folder now focus and select the rename box (Svelte's
autofocus only acts when nothing else has focus, which was never the case
here), and Enter or Esc returns focus to the tree. A folder created during
a search is inserted at once and every folder above it is opened as the
user's own, so clearing the search no longer hides it and locks the keys.
A focused row that leaves the screen hands focus to its nearest visible
folder (edit/focus-repair.ts), and a rename whose row is hidden is
dropped. A folder closed during a search stays closed until the text
changes. Right arrow on an empty folder does nothing.
- One bulk command at a time: sort and keep-first now lock every other
write in the session itself, not just the toolbar, and only the command
that set `busy` clears it.
- #edit mounts a 500-row window around the focus (edit/row-window.ts) with
"Show earlier" / "Show more": End on 20k rows mounted all of them.
- Failures get a pinned toast slot; later toasts still show and announce.
- Import: the preview's new/dup badges show again (`$state.raw` — a deep
proxy broke the WeakMap lookups, already broken in 1.0.x), each step's
heading takes focus, and one live region announces the outcome.
- Dialogs name their close button "Dismiss" and are described by their
body, so the counts are read with the title.
- 375 px: the edit toolbar wraps; titles keep their room before the domain.
- Wording: untitled items named by their address in dialogs; plurals for
show-more and the duplicate summary; a diagnostics copy failure says so;
the Move dialog's filter is labelled as a folder filter; the Japanese
notices spell ブラウザ and フォルダ like the rest of the dictionary.
- The logo pulse stops after two cycles (WCAG 2.2.2), and the popup
switches only the Manager tab it brings forward.
- Back under the size limits: EditSession is three layers (edit-view,
edit-commands, edit-session); RowActions, RenameInput, ImportProgress,
ImportAnnouncer and StatusBar are their own components; the tree's key
handling lives in edit/tree-keys.ts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the v1.1.0 adversarial review (data integrity):
- Replace re-reads the tree once the backup is on disk. If anything changed
while the save dialog was open (or while the user checked Downloads on the
fallback path), the picker path writes the fresh snapshot to the same file
and checks again (up to 3 times); the fallback path stops with BACKUP_STALE.
Before, a bookmark added meanwhile inside an existing folder was deleted with
no copy anywhere.
- Roots resolve by folderType among this device's roots; a signed-in account's
own roots (Chrome 134+, syncing) are never emptied or written. Replace backs
up and clears exactly the three destinations.
- An empty Mobile root, which desktop Chrome leaves out of getTree, is found
with bookmarks.get('3'), so Mobile content stops landing under Other. The
fake now hides it the same way.
- Plans report writeDepth; past MAX_WRITE_DEPTH (90) the import is refused with
TOO_DEEP before any backup — Chrome re-reads its own bookmarks file with a
200-level JSON nesting limit, two levels per folder.
- The suggested backup name carries seconds, and a picked file that never
received a backup is removed again.
- Edit tab: a stale "created" event for an id the view already holds is
dropped, and the duplicate finder never pairs a node with itself — together
they made "Keep first" delete the only copy. Sort re-plans from the folder's
children at confirm time.
- Netscape: an unescaped <h3> inside a bookmark title no longer turns the
bookmark into an empty folder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From the v1.1.0 adversarial review (security) and the docs pass:
- clsx (MIT) ships too — the Svelte runtime imports it — and wxt's browser
module as well as @wxt-dev/browser. All four are now credited in About,
PRIVACY, README and both notice files, and wxt.config.ts fails the build when
code from any uncredited package would ship; tests/unit/third-party.test.ts
keeps the notices in step with that list.
- Sanitizer, CSV: default-deny. A title with an unquoted comma — the row that
makes the extension report CSV_ROW_MISMATCH — shifted a file: path and free
text into columns no rule rewrote, and the host-only check passed them.
- Sanitizer, HTML: unknown attributes (Firefox POST_DATA, any DATA-*) are
rewritten; only dates, root markers and the charset line are kept.
- Sanitizer: a value-level leak check over CSV fields, attribute values and
JSON strings, and a refusal that prints counts only — it used to print the
leaked words and ask for them in a public issue.
- BM_ENV_ERROR_CODES joins the per-locale message check, with the four fixed
keys describeError maps; the popup no longer logs crashes nothing can read.
- docs 00/02/03/05/06/07/08/09/11/13 describe v1.1.0 (diagnostics moved to
09 §7; 13 §3 points at store/listing.* instead of a drifting copy); tokens
and gen-icons comments corrected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
No LEGAL_VERSION bump: every change clarifies, corrects or removes an
obligation (docs/14 §3.1, new "contact details → No" row and a table of what
version 1 means).
- All four documents: version/date header and a copyright line.
- EULA clause 3 becomes a non-binding note — GPL-3.0 §10 forbids adding
restrictions; clause 6 gives legal@ and dmca@.
- DISCLAIMER describes Replace as it now behaves, and adds the name/fork
notice with takedown@ and dmca@.
- PRIVACY describes the in-memory diagnostics log, the Chrome Web Store's own
role, and that opening a bookmark is the browser's request, not ours;
contact privacy@.
- THIRD_PARTY_NOTICES lists the Svelte runtime and WXT's browser shim (MIT),
which the build always compiled in; the notices now ship as
THIRD_PARTY_NOTICES.txt and check:manifest requires the file.
- security@ (SECURITY.md), code@ (CoC, issue chooser), README Contact table;
contact@ retired. links.test.ts now scans all of src/ for an address.
- README: new formats and features, Sonar/CWS badges, verify:full, the
screenshot harness. Store listings EN/VI/JA follow.
- docs/15 decision rows, next direction and backlog; CLAUDE.md Phase 7, the
TS 7 / TS 6 alias rule and --ignore-scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- About no longer claims "no runtime libraries at all": the build compiles in
the Svelte runtime and WXT's one-line browser shim, both MIT (sC2.1).
- The backup save dialog's file-type label was hardcoded English.
- VI: a browser tab is "tab", an in-app section is "thẻ", consistently. JA:
the katakana middle dot instead of "·".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- browser/diagnostics.ts: an in-memory ring buffer of 200 allowlisted events —
import phases, failure codes (logged where every failure is described, by
code, never by detail text), bulk edits, storage read failures, uncaught
errors by name only. Page-relative times, no storage, no network. About →
"Copy diagnostics" copies it with the version, locale, theme and Chrome major.
- browser/write-lock.ts: Web Locks, ifAvailable — a second Manager tab trying
to import, sort or delete duplicates while another writes is refused at once
with WRITE_LOCKED, before anything is touched. The import takes the lock after
opening the save picker, so the click's activation still holds.
- The settings store reads the UI language through browser/ (uM15); progress
reports the first create, so a slow start no longer looks like a hang (uL9).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every surface moves to the v1.1 design system: a header with the brand mark,
a route crumb and icon tabs with a sliding indicator; an IDE-style status bar
(local only · bookmarks, storage · version · licence · Settings · About);
numbered mono step labels on #import and #export; readout stats; mode and
format cards; marching-ants drop zone; striped progress; slide-in toasts; the
popup as launcher tiles over a live readout. Emoji glyphs are gone in favour of
the self-drawn icon set; all motion sits behind prefers-reduced-motion.
Behaviour fixed on the way:
- The Start button counts what will really be imported (a preview plan), and
Replace asks first, with the exact numbers it will delete and write (T4).
- The report counts bookmarks and folders separately, what was skipped and
reused, and where it all went — after a cancel or failure too (uM10, uM11).
- The backup hint names the file as the user saved it, not as suggested.
- Failure toasts stay until dismissed; export says "download started" and
keeps a failure's technical detail on screen.
- The popup routes an open Manager over a BroadcastChannel (refused while an
import writes), and no longer counts the three permanent roots as folders.
- Read-only previews and the export picker drop the role="tree" they never
implemented; both render 500 children at a time; "Everything" shows
indeterminate on a partial selection; the drop zone stops flickering.
- Per-route document titles; the status bar stays at the bottom of the window.
- The Move-to dialog no longer renders every folder while closed.
The Import tab's strings move to <lang>-import.ts so each locale file stays
under the size limit. scripts/screenshots.mjs serves the built extension with
a fake chrome and the production CSP, for visual QA and store screenshots.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
EditTab's state and commands move to stores/edit-session.svelte.ts (an
injectable bookmarks API, tested against the in-memory fake); the component is
markup and wiring, 481 → 290 lines. New:
- Sort A→Z per folder (edit/sort.ts): folders first, natural, case- and
accent-insensitive, stable; planned first so the confirmation states how many
items move; moves only what is out of place, with indices valid for
chrome.bookmarks.move; a refusal part-way says how far it got and re-reads.
- Drag to reorder: before/after/into zones (edit/move-target.ts planDrop, which
converts between Chrome's pre-removal indices and onMoved's post-removal
ones); dragend clears a cancelled drag.
- Live sync batched per frame, re-read once on bursts and after Chrome's own
import or a wholesale reorder; handlers isolated so one exception cannot stop
the tree following the browser. Duplicates are scanned only while shown.
- Bounded rendering: rows mount 500 at a time with "Show more", growing to keep
the focused row in the DOM; explicit aria-setsize/posinset.
- Move-to is a real modal (shared Modal: focus in, Esc, focus returned, no
cancel-after-confirm) with a filter; the menu key / Shift+F10 opens it, Ctrl+C
copies a link, row buttons leave the tab order; loading state, focus follows a
delete, a new folder clears a search that would hide it, resyncs keep
expansion and focus, renames are trimmed and Esc never commits.
- Duplicate rows show the folder each copy lives in.
Also the v1.1 design foundation these use: new tokens (sunken/subtle surfaces,
--border-strong ≥ 3:1, mono stack, motion/shadow/z tokens, color-scheme,
accent-color, faint grid), a self-drawn SVG icon set with motion, and
tests/unit/styles/motion.test.ts, which fails any animation or moving
transition outside prefers-reduced-motion: no-preference.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
normalizeUrl trimmed the last character of a root URL to drop its slash, but
`x/?` and `x/#` serialize with the delimiter last, so the `?` was cut instead
and `x/?` never matched `x/`. Empty query and fragment are now removed first.
The default-port code was dead: the URL parser already drops default ports.
The export folder picker looked up each row's check state by walking the tree
from the roots - O(n) per visible row. checkStates builds the whole map in one
pass, once per selection change.
The seeded round-trip now asserts no warnings per tree as docs/11 §3 says.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Import Chrome's profile `Bookmarks` file and Firefox's JSON backup directly.
Each reader is held to a parity rule against its browser's own HTML exporter,
tested on Chromium/Firefox-faithful fixture pairs. One entry point,
parseBookmarkFile, sniffs by content (JSON first, parsed exactly once, routed
by shape); the picker loses its extension filter, which hid the extensionless
`Bookmarks` file. Firefox's internal root names are localized by the caller.
Merge finally reuses the folders a user already has, level by level, appending
new items (core/merge.ts) - 1.0.x planned it exactly like Replace. Root folders
are unwrapped by marker (core/route.ts), so restoring a backup no longer nests
Other and Mobile one level deeper each time, 1.0.x backups included. Replace
never dedupes, plans before the backup is taken, and refuses a backup the
import could not read back (BACKUP_EXCEEDS_LIMITS) before anything is written,
downloaded or deleted. The save dialog's late rejection is marked handled.
Also found: the seeded round-trip suite had compared 200 empty trees since
Phase 1 (xorshift32 seeded with 1-200 draws near zero first). makeRng now
scrambles the seed and every seeded suite checks its corpus is non-trivial.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Several shapes made v1.0.x import a smaller tree than the file held, with no
warning at all:
- a loose bookmark before the wrapper list (no </DT>) swallowed the whole
list; with an explicit </DT> the bookmark "consumed" the next list;
- an unescaped <b> in one title re-opened around every later entry
(dt > b > a, dl > b > dt) and hid them from a direct-children walk;
- a root list wrapped in <div>/<table> was invisible;
- PERSONAL_TOOLBAR_FOLDER="false" made a folder the bar.
Every reader now classifies addresses before anything is written (core/
href.ts): blank -> MISSING_URL, unparseable -> INVALID_URL (chrome.bookmarks.
create rejects those mid-write, after a Replace has already cleared the
tree), Firefox place: -> PLACE_QUERIES_SKIPPED in browser-written formats
(kept in BM JSON and CSV, our restore formats). A DT that still yields
nothing is counted as UNREADABLE_ENTRIES instead of vanishing.
Also:
- empty titles stay empty (the URL fallback ruined favicon-only bars);
- the HTML node cap is counted on the raw text before DOMParser builds a
DOM; the depth cap no longer rejects an empty list one level past it;
- root markers for Other and Mobile (UNFILED_/MOBILE_BOOKMARKS_FOLDER in
HTML, "root" in BM JSON, version stays 1), and a 1.0.x full-tree backup
is recognised by shape and marked on read;
- BM JSON dates go through the shared normalizer; url "" (which Chrome
makes a folder) is skipped;
- CSV: an unclosed quote is CSV_UNCLOSED_QUOTE with its line instead of one
cell eating the file; formula cells get one reversible guard quote; sep=
lines and padding rows are skipped; folder lookup is a Map, not a scan;
- Markdown escapes every CommonMark-active character and collapses line
breaks in titles (raw HTML used to pass through);
- shared WarningBag/NodeBudget/title rule, an iterative walkTree, an
isBookmark type guard, a CSV dialect module, chromeTimeToSeconds and
microsToSeconds for the JSON formats that follow;
- locales split warnings/errors into <lang>-notices.ts to stay under the
size limit; every warning and error code must have a message (test).
Four new fixtures, golden serializer suites for HTML and BM JSON (docs/11
named one; neither existed), and docs/02, 04, 05, 07, 09, 11, 15 updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud's second pass still flagged S8707: the directory to check came
from argv and flowed into file reads, and a startsWith() allowlist did not
count as sanitizing it. The CLI now only ever checks the repository's own
.output/chrome-mv3; the logic is an exported checkExtension(dir) (typed by
check-manifest.d.mts) that the tests call directly, so no command-line path
reaches a file read at all.
Path resolution moved inside the entry point: under Vitest's jsdom
environment import.meta.url is not a file: URL, so resolving it at module
load threw on import. The two Node-script suites now run in the node
environment.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first CI-based SonarCloud scan failed the quality gate on its new-code
security rating, with two findings in check-manifest.mjs:
- S8707: it read whatever directory the CLI argument named. It now accepts
only paths inside this repository or the OS temp directory (where its
tests build fake extensions), and says so when it refuses.
- S8689: its success line echoed values read from the file. It now prints
constants — the manifest has just been proven equal to them.
And four S8786 regexes in guard.mjs plus one in the sanitizer, where a \s*
beside an optional quote beside another \s* could backtrack: each is now
written so no two quantifiers compete for the same characters, and the
markup rule is two linear steps (find the tag, then look inside it). The
sanitizer's JSON tokenizer no longer reassigns its loop counter (S2310).
While here: with no default-src, style-src was open, so a remote stylesheet
could still load. The CSP now adds `style-src 'self' 'unsafe-inline'`
(Svelte sets style attributes); check-manifest pins the new string.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
sanitize-bookmarks.mjs accepted any JSON object and rewrote only "url" and
"title". Chrome's profile Bookmarks file keeps titles under "name" (plus
meta_info and a base64 sync_metadata blob), so it came out with every title
intact, passed the host-only leak check, and was written as "sanitized" —
the exact file a reporter whose import failed would attach.
- JSON dialects are recognised by shape (BookmarkMagic, Chrome profile,
Firefox backup); any other JSON is refused. Rewriting is default-deny: a
small tokenizer tracks the key of every string value and replaces all of
them except structural keys (types, ids, timestamps, root GUIDs), by span,
so layout and key order stay byte-identical.
- The backstop now reads the output's TEXT: for HTML it strips markup and
refuses on any word that is not a placeholder — independent of the
rewriting rules, unlike the old host-only check whose test shared its
blind spots. <DD class=…> notes, file: ICON_URIs and <H1>/<TITLE>
owner names are rewritten.
- Attribute values decode in one pass (CodeQL js/double-escaping), <DD>
bodies are found in linear time (Sonar S8786), the input is never
overwritten, size is checked before reading, and bytes are bytes.
- The 627-line file is split into scripts/sanitize/ modules; the
user-facing entry stays plain .mjs for reporters on any Node.
gen-fixture.mjs validates its flags (whole numbers >= 1, values required, no
unknown flags) and its loop visibly advances (Sonar S2189 blocker).
gen-icons.mjs says what is true: it is the icons' source of truth.
Every script now carries // @ts-check and tsconfig.scripts.json runs them
through the TS 7 tsc in `npm run check`; docs/10's size rule names .mjs.
Regenerated icons are byte-identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The "no network access" claim rested on one grep gate. An audit ran its
exact patterns against 28 bypasses and every one got through — an alias
(`const go = fetch`), computed global access, a call split over two lines,
`innerHTML +=`, `{ @html }`, remote url() in CSS — and a checkout path with
a space or diacritic made it scan nothing and print "clean". Now:
1. The production manifest declares an extension-page CSP with
`connect-src 'none'` (plus closed img/font/media/frame/form sources), so
Chrome itself refuses every fetch, XHR, WebSocket, EventSource and beacon.
No default-src, so Svelte styles are untouched; dev builds omit it so
the HMR websocket works. docs/08 used to say "never add a CSP" because
MV3 forbids loosening — tightening was always allowed.
2. check:manifest scans every shipped .js for network APIs (dependency code
included), pins the exact CSP and the Chrome floor, checks that every
file the manifest and pages reference exists, resolves paths from the
script instead of the cwd, and sets exitCode instead of exit().
Vite's modulepreload polyfill — the one network call in the bundle — is
disabled; every Chrome at the floor supports modulepreload natively.
3. guard.mjs matches whole files (so \s* spans newlines), bans the API
identifiers outright, bans computed access on globals, scans .html and
.css, has no test-file exemption, uses fileURLToPath, and fails when it
scanned nothing. docs/08 §3 points at the file instead of restating it;
the restated copy had already drifted.
Both scripts finally have tests (48 + 18 cases, every audited bypass
included). One src comment named the API in prose and is reworded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud ran Automatic Analysis: no coverage, and a scan of the repo rather
than of what CI builds. sonarcloud.yml replaces it — full-history checkout,
the coverage run, then the scan with the package version as the new-code
baseline. It is its own workflow so a Sonar outage or gate opinion can never
turn CI red, the scan action is pinned to a commit because it holds
SONAR_TOKEN, the token reaches only the two steps that use it, and runs
without secrets (Dependabot, forks) skip with a notice instead of failing.
Coverage now measures every src/**/*.ts module and emits lcov, but the CI
gate is unchanged: a glob threshold applies the docs/11 §1 numbers to
src/lib/core/** only (verified to still fail a run below them).
Hardening, from SonarCloud's findings and the scripts/config audit:
- npm ci --ignore-scripts everywhere, with `wxt prepare` run explicitly;
- checkouts drop persisted credentials;
- release.yml is split: gates and zip run with a read-only token; only a
job that installs nothing and runs no repo code creates the Release; the
store submit uses the local wxt binary, never `npx` with secrets in scope;
- reusables get named secrets instead of `inherit`. The announce job could
never have worked: this repo stores RELEASES_WEBHOOK/REPO_WEBHOOK, the
reusable reads DISCORD_*, and inherit passed neither;
- CI runs `npm audit --omit=dev`, which docs/09 already claimed it did;
- a non-required Windows smoke job (guard + unit tests).
serializeMarkdown's trailing-newline trim used /\n+$/, which backtracks
quadratically over a long newline run that is not at the end (S8786); it is
now a single backward scan, with a 20k-newline regression test.
docs/12 no longer copies the workflow YAML — the copies drifted three times,
and under docs-win a stale copy invites reverting a pinned action.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every dev dependency moves to the newest release that has been public for
at least 7 days, and that rule is now enforced rather than remembered:
.npmrc sets min-release-age=7 and Dependabot gets a 7-day cooldown on both
ecosystems. `npm ci` was verified to ignore the age rule for versions
already in the lockfile, so it only shapes what gets locked.
vitest / coverage-v8 4.1.10 -> 5.0.5 (suite passed unchanged)
@biomejs/biome 2.5.5 -> 2.6.2 (exact)
svelte 5.56.8 -> 5.57.2
vite-plugin-svelte 7.2.0 -> 7.3.1
svelte-check 4.7.4 -> 4.7.7
jsdom 30.0.1 -> 30.1.3 (dom-environment canary first)
knip 6.29.0 -> 6.42.1
wxt 0.21.3 -> 0.21.4 (+ transitive vite 8.3.5)
npm audit: 13 (7 high) -> 3 low, all one deepmerge advisory with no fixed
release, inside vite-plugin-svelte's dev-time option merging; documented in
docs/09 §3.1. --omit=dev stays 0. This supersedes Dependabot #19-#25.
The six-entry overrides block is deleted: none of those packages is in the
tree any more (WXT 0.21 dropped web-ext-run, Vite 8 bundles with rolldown),
and a stale override would silently pin a package the day it came back.
WXT got the built-manifest diff the 2026-08-03 decision requires: identical
keys and values apart from the deliberate minimum_chrome_version, identical
file inventory. Biome 2.6 began formatting Markdown and YAML, so both are
excluded rather than letting a formatter reflow hand-written docs.
dependabot.yml now keeps WXT out of the routine group and ignores majors of
the TypeScript pair and @types/node.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Node 26 entered Active LTS on 2026-10-28 and Node 24 went to maintenance on
2026-10-20, so the toolchain moves with them: .nvmrc, engines and @types/node
all say 26.
TypeScript 7.0 is GA but still has no programmatic API, and svelte-check
peers `typescript ^5 || ^6`. So this uses Microsoft's documented
side-by-side aliases instead of choosing one:
"@typescript/native": "npm:typescript@~7.0.2" -> tsc (7.0.2)
"typescript": "npm:@typescript/typescript6@~6.0.2" -> the 6.0.3 API
`npm run check` therefore type-checks .svelte files with TS 6 (svelte-check)
and every .ts file with both. `tsc --version` is 7.0.2; `tsc6` is 6.0.3.
tsconfig targets ES2025 with lib ES2025, which is only honest if the oldest
Chrome allowed to install has every ES2025 built-in. That is Chrome 136
(RegExp.escape), so minimum_chrome_version goes 120 -> 136 and Vite's
build.target follows the same CHROME_FLOOR constant. manifest.test.ts pins
all three so they cannot drift apart. Users below 136 keep what they have
installed; they just stop receiving updates.
Built output: the manifest changes only minimum_chrome_version, the JS chunks
are byte-identical apart from a CSS hash, and the CSS loses one
-webkit-user-select prefix Chrome 136 no longer needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The listing has shown version 1.0.1 since it cleared review; the suite still
said "in review, not yet approved" in docs/00, docs/15 and CLAUDE.md, the
exact mirror of the week it spent claiming review after 1.0.0 was live.
Recorded only what was observed (2026-10-28: listing shows 1.0.1, 5 users).
The approval date and the 13 §1b step-10 clean-profile check of the
published build stay marked as unrecorded rather than inferred. CLAUDE.md's
EditTab figures were the v1.0.0 numbers; they are now the current ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Windows clone with core.autocrlf=true checked every text file out with
CRLF, and Biome's formatter (which expects LF) then failed `npm run lint`
on all 114 files before a single line had been edited. The two scripts
with a `#!/usr/bin/env node` line were also unusable from WSL.
`eol=lf` fixes the checkout, not the repository: the index already held
LF everywhere (`git add --renormalize .` stages nothing else), and
tests/fixtures/** stays -text so the CRLF records the CSV fixtures exist
to exercise are untouched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owner tagged, released and uploaded v1.0.1 to the dashboard on 2026-08-03.
Recorded as **submitted, in review** — deliberately not as published.
That distinction is the whole reason Phase 6 existed. This suite spent a week
asserting "in review" after the listing had already gone public, and the
opposite error is the worse half of the same mistake: telling a future reader
that users have a fix they have not received yet. Users are on 1.0.0 until the
review clears.
docs/15's progress table gains two rows — released (signed tag, checksum-verified
zip, announce and store-publish both correctly skipped for want of their repo
variables) and submitted. docs/00 §10 gains item 11: what to write down when it
clears, and the pointer to §7 if it is rejected instead, because a rejection and
its resolution are the only memory this project would have of what reviewers
actually object to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Applying the §5 checklist item exposed two errors in it.
The required check is `Build + package / build`, not `Build + package`. A job
that calls a reusable workflow publishes its check as `<caller job> / <reusable
job>`, so requiring the caller's name alone matches nothing that will ever be
reported — every PR waits forever and `main` becomes permanently unmergeable.
The doc was confidently specific about this ("the job names — not the workflow
name") and confidently wrong, which is the worst combination.
"Require review" is set to zero approvals, deliberately. It is correct advice
for a team and a lockout for one maintainer, who cannot approve their own pull
request. Zero still forces every change through a PR and through the checks;
it only drops a second pair of eyes that does not exist. Admins are not
enforced, for the same reason: with a single maintainer, an escape hatch from
your own rules is what separates a safety net from a locked door.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first CI run on this branch failed `npm run check` with 18 errors —
"Cannot find name 'node:fs'" — including in `tests/unit/styles/contrast.test.ts`,
which this branch never touched. So it was not the new test: the package has
always used Node types without declaring them.
Two things hid it. WXT 0.20 supplied `@types/node` transitively and 0.21 does
not, and this worktree sits under the parent checkout, so TypeScript walked up
to `E:\bookmarkmagic\node_modules\@types\node` and resolved them anyway — even
after a correct `npm ci` in the worktree itself. A green local type-check in a
worktree is therefore weaker evidence than it looks, which is now written down
in CLAUDE.md next to the existing worktree gotcha.
Types only: erased at build, nothing ships, so hard rule 2 — zero *runtime*
dependencies — is untouched, and the built manifest and bundle are unchanged.
docs/01 §2 records it with the reason, since that table is the pinned-version
spec rather than a summary of it.
This is the failure the tag run would otherwise have hit, and the reason a red
CI is worth reading before merging rather than after tagging.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Exactly the two files a patch touches — package.json's version and CHANGELOG.md,
section plus the link references at the bottom. wxt.config.ts declares no
version key, so it must not appear in this diff, and it does not; the built
manifest reads 1.0.1 and `npm run zip` emits bookmarkmagic-1.0.1-chrome.zip.
`npm run verify:full` green: lint, svelte-check + tsc, knip, the grep gate, 694
tests, core coverage 97.4%/92.1%, the build, and check:manifest against the
built artifact. The built manifest is byte-identical to the pre-phase baseline —
same permissions, same key set — across ten commits and a WXT major-slot bump.
The changelog is written for users rather than for the diff: what changed for
someone who already has this installed, and the fact that they need do nothing.
⚠️ Stops here, deliberately. The tag and the store upload are human steps:
git tag -s v1.0.1 && git push origin v1.0.1
and then docs/13 §1b from step 5 — which begins with the manual QA pass,
including the 1.0.0 → 1.0.1 upgrade over an existing profile that nothing here
can perform.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six lenses over the branch, each finding independently attacked by a verifier
told to refute it. 31 filed, 18 survived, clustering into five real defects —
and the pattern held: every one is in work added earlier in this same branch.
**The sanitizer did not keep its promise (9 findings).** It is the single
privacy control in the intake flow, named as mandatory in SECURITY.md, README,
CONTRIBUTING, the privacy policy and the issue form — and it leaked in nine
distinct ways. Worst: `detectKind` fell back to 'csv' for anything it could not
identify, `sanitizeCsv` then matched no columns, and the input was emitted
verbatim with exit code 0. The users who run this tool are by definition the
ones whose file did not import, so a truncated or hand-excerpted file is the
TYPICAL input — and `tests/fixtures/malformed/no-doctype.html`, already in the
repo, reproduced it. The test could not catch any of this: it returned early
whenever the original failed to parse, so it passed vacuously on precisely the
dangerous inputs.
Nine holes found in one pass is not a signal to patch nine holes. The rules are
regexes over untrusted markup, so "we thought of every shape" is not a claim
anyone can make honestly. So the tool now fails safe:
- unidentified input is REFUSED, matching core detectFormat rather than being
weaker than the parser it fronts;
- the output is checked before it is written — any host that is not one of our
own `*.example` aliases means a rule missed something, and the file is refused
with the leaked host named. A rule this misses now degrades into an annoying
refusal instead of a published browsing history.
The individual holes are fixed too: single-quoted and bare attribute values were
never matched (a straight leak); `spansFrom` located capture groups by searching
the match text, landing on the wrong copy — replaced with the `d` flag's exact
offsets; overlapping spans spliced already-rewritten text using offsets measured
against the original; `[^<]*` made a title containing a raw `<` fail to match
entirely; `<DD>` stopped at the first newline, publishing the rest of a wrapped
note; and replaceSpans rebuilt the whole string per span, O(spans x file).
**The import report pointed users at a backup that was never written.** The
guard was "mode is replace and this is not a BmBackupError", but runImport
resolves the roots BEFORE proving the backup — so NO_WRITABLE_ROOTS, added in
this same branch, reaches the catch with nothing on disk. Now keyed on the
onClearing callback, which is the only signal that proveBackup actually
returned.
**Cancelling a Replace after the clear read as a harmless no-op.** The tree is
gone and only `created` of it rebuilt; "N items created" alone, on a browser
that is now nearly empty, is the wrong story. It now says so and names the
backup.
**Collapsing a folder while a search was active re-opened it instantly.** The
$effect read and wrote the same two `$state` sets, making it its own dependency:
every write re-invalidated it and it re-merged an unchanged `search.expand`.
Reproduced against the real Svelte runtime — `$effect` is stripped under this
repo's vitest config, so no unit test could have seen it. Fixed with `untrack`,
which also removes the self-retrigger loop.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`continue-on-error` is not allowed on a job that calls a reusable workflow.
GitHub answers with "Property continue-on-error is not allowed" and refuses the
**whole file** — a startup_failure with zero jobs and no log, not one failed
job. Confirmed against two GitHub community discussions, one of which is an
open, unanswered feature request asking for the capability.
release.yml's announce job has carried that keyword since 5f4c6e7. That commit
landed *after* the v1.0.0 tag object and no tag has been pushed since, so the
first v1.0.1 tag would have been its first execution — and the release would
have died before running a single job, in exactly the silent shape this project
already hit once with the `contents: write` grant (2c8fba9). The previous commit
in this branch then copied the same mistake into notify-ci-failure.yml.
Both are now gated with `if:` on a repo variable instead. A variable and not a
secret because secrets cannot be referenced in a job-level `if:`, and unset
means skipped — so the release stays green and CI stays quiet until the owner
opts in by setting DISCORD_ANNOUNCE / DISCORD_CI_NOTIFY alongside the webhooks.
Same idiom as the CWS_AUTOPUBLISH gate already in release.yml.
docs/12's samples and its §5 checklist updated in the same commit; docs are
binding here, and leaving the invalid keyword in the spec would have had the
next reader put it straight back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Applied directly rather than merged: both branches predate this phase and would
have reverted `verify:full`.
**wxt ^0.20.27 → ^0.21.3.** A 0.x minor is semver's breaking slot, and WXT is
what emits the shipped manifest — so the acceptance test was a byte diff of the
built artifact, not a green suite. Manifest, file inventory (content hashes
stripped) and bundle size all came out identical, permissions unchanged, zip
name unchanged. It *did* break two things, both dev-only: `wxt/testing` split
into `wxt/testing/fake-browser` and `wxt/testing/vitest-plugin`, and
`tabs.create` is now typed as returning void.
That split is the point of the ritual. `npm run test` passed before the bump and
the *config load* failed after it — while `tests/unit/manifest.test.ts`, the
test that looks like it guards the manifest, would have stayed green whatever
WXT emitted, because it reads wxt.config.ts and a WXT upgrade cannot change that
file. The check-manifest key-set gate added in the previous commit is what makes
the next one CI's problem rather than a manual comparison someone has to
remember to do.
**jsdom ^29.1.1 → ^30.0.1.** Canary first: tests/unit/dom-environment.test.ts
pins the parse5 behaviour the Netscape walk depends on — implied end tags, and
`:scope >` from an element root. It passed, then the parser, fixture and
round-trip suites, then coverage, all unchanged. The 2 869 ms figure in
docs/11 §6 is now labelled as a jsdom 29 number; a DOM major invalidates it
rather than updating it.
**svelte-check ^4.7.3 → ^4.7.4.** A patch.
npm audit stays at 0, and none of the six `overrides` can be dropped yet.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
notify-ci-failure.yml called a reusable that declares DISCORD_CI_WEBHOOK as
`required: true`, which GitHub validates before any of the reusable's own logic
runs. With the secret unset — docs/00 §10.6 lists it as still open — that
workflow failed on *every* CI completion, success included, from the first push
after it landed. The file's own header comment admitted it. This is precisely
the failure mode commit 5f4c6e7 fixed for releases: a run that is red for a
reason nobody acts on teaches everyone to stop reading the colour, and the next
genuine failure goes unnoticed.
release.yml's `gh release create` had no existence check, so any failure after
it — realistically the CWS publish step, once autopublish is on — stranded the
tag: a re-run could not get past that line, and recovery meant deleting the
Release and the tag by hand. v1.0.1 is the first tag to exercise that path at
all, since both the continue-on-error announce fix and tag_override landed
after the v1.0.0 tag object.
check-manifest.mjs now asserts the exact top-level key SET, not just the absence
of ten enumerated reach-widening keys. The forbidden list can only catch what
someone thought to name; this catches a build tool that starts emitting a key
nobody asked for. That matters immediately: WXT is a 0.x dependency, so a
routine minor bump is exactly when it happens, and tests/unit/manifest.test.ts
is structurally blind to it — it reads wxt.config.ts, which a WXT upgrade cannot
change. Verified by injecting a `commands` key into a built manifest and
watching the gate fail.
docs/12's sample YAML updated in the same commit, so the drift this phase just
finished correcting does not immediately start again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/13 §6 counts store ratings and reviews as one of the three post-launch
channels this project has, and there was no way to reach the listing from
inside the extension — the listing URL appeared in exactly one place in the
repo, the README, which a user who installed from the store never reads.
Deliberately no support email alongside it. Publishing an address on an
in-product surface invites people to mail their real bookmarks file, and a
bookmarks export is a browsing history — the exact receipt of personal data
legal/PRIVACY.md and docs/09 §6 are written to bound. contact@poli0981.dev
stays in SECURITY.md and CONTRIBUTING.md, where it arrives next to the
instruction to sanitize first. A test asserts links.ts contains no address, so
the decision has to be made again on purpose rather than by accident.
docs/06 §3.5 sanctioned the fourth link in PR A, before the code added it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/15 recorded this as a known v1 limitation, and with real users it is a
one-star review waiting to happen: `save()` writes the whole settings object
rather than a patch, so a tab opened before the language changed reverted that
change the moment its own theme toggle fired. From the user's side that reads as
"I changed it and it didn't stick" — unfalsifiable, and indistinguishable from
a bug. Accepting the legal terms in one tab likewise left the other padlocked
until reload.
The adoption rule is field equality plus an in-flight guard, and between them
they solve the echo problem and the debounce problem at once:
1. If this context has an unflushed write, ignore the incoming change entirely.
It is about to overwrite that value anyway, and adopting first would clobber
the choice the user just made — in the UI *and* in what gets persisted 200 ms
later. That is last-write-wins, exactly the semantic docs/15 documented, and
it converges because our write fires the same event in the other tab.
2. Otherwise adopt only fields that actually differ. Our own write always
matches what we already hold, so self-echo suppression falls out for free —
no `lastWritten` snapshot to keep in step and go stale, and no dependence on
key order, which a $state spread and coerceSettings do not agree on.
The fake-timer test covering rule 1 is the one that matters; without it the
guard is an unverified assertion about a race.
Legal acceptance is adopted **one way only, toward accepted.** App.svelte
renders the gate in place of the tab body, so an externally-raised gate would
unmount ImportTab mid-write and strand the import it was driving — recreating,
through a third door, the deadlock the routing guard and the store-held
attestation resolver were just written to close. Clearing storage elsewhere
therefore does nothing; a reload re-evaluates it, and a reload has no in-flight
work to strand. Three tests pin that asymmetry so a future simplification that
makes it symmetrical fails loudly.
subscribeStorage coerces through the same functions the read path uses — a
second tab is not a more trustworthy source than disk, it is the same disk — and
swallows a throwing handler, because an exception inside a chrome.* listener
surfaces in chrome://extensions and docs/11 §5 counts that as a failed QA pass.
Verified before writing any of it that WXT's fakeBrowser does implement
storage.onChanged (unlike bookmarks.*), so no injectable event source was
needed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
**#edit and #export showed a raw English Chrome string as the entire message.**
docs/02 §7 requires a localized sentence plus a separate technical block, and
LegalGate carries a comment describing this exact mistake as a defect that was
already fixed once — writing `detail ?? t(key)` looks like a sensible fallback,
but the raw string is almost always present, so the translation becomes dead
code. Three sites still had it: EditTab's load failure, ExportTab's load
failure, and ExportTab's export toast. A Vietnamese or Japanese user got an
untranslated Chrome message as their whole explanation.
One ErrorCallout now owns that contract, so there is a single place it can be
got wrong, and all three tabs go through the shared describeError.
**Clearing the search box rendered the whole tree in one synchronous pass.**
Search-opened ancestors merge into `expanded` (deliberately — a render-time
union made the disclosure button and ArrowLeft dead keys), but nothing ever
un-merged them. A one-character query expands most of a large profile, and
clearing the box then rendered every row the search had opened at once, with no
virtualization behind it; the 150 ms debounce delayed that growth without
bounding it. src/lib/edit/search-expansion.ts tracks search-owned ids
separately, and toggling a folder transfers it to the user so the collapse
cannot steal it back.
**"Keep the first of each" targeted nodes it cannot delete, then gave up
silently.** The duplicate list had no isEditable filter — unlike the
single-delete path — so duplicates inside a policy-managed subtree were offered
for bulk deletion, and the browser rejects every one of them. The loop then
abandoned the rest on the first rejection, leaving a partial delete measured
against a dialog that had promised an exact number, with no way to tell how many
copies remained. deleteEach never throws and reports how far it got.
EditTab was at 481/500 before this, so the ARIA tree container came out into
TreeList.svelte first — 466 now, and the tree pattern lives somewhere it is not
surrounded by loading, searching, mutation and five dialogs. That extraction
immediately caught a real bug the shorthand had hidden: `{ontoggle}` bound
`window.ontoggle`, not the local `toggle`.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three defects on the paths that touch a user's real bookmarks.
**A mid-import rejection used to lose everything needed to recover.**
writeTree awaited create() and let a rejection propagate bare — while the
bookmarks it had already written sat in the user's tree, uncounted. The
neighbouring BmAborted has always carried `done` for exactly this reason.
clearRoots was worse: it is the one path that deletes, and it reported nothing
at all, plus threw BmAborted(0) on cancellation regardless of how many root
children were already gone. Both now throw BmPartialWrite{phase, done}, and the
error state carries the safety backup's filename — which `attesting` knew and
`error` threw away, so the one string that lets someone restore a half-replaced
tree was the one thing not on screen. Only when a backup was actually produced:
on a BmBackupError the file may never have been written, and sending a
frightened user to look for a backup that does not exist is worse than silence.
BmAborted is re-thrown before the wrap, with a test pinning it. Wrapping it
would relabel every user cancellation as a failure and make the `cancelled`
state — and its copy in three languages — unreachable.
**Browser Back during the backup attestation deadlocked the Manager
permanently.** The attestation resolver was a component-local `let`, so Back
unmounted ImportTab, stranded the promise runImport was awaiting, and left
isWriting() true forever: every tab and footer control disabled, beforeunload
prompting, and a remounted tab rendering two inert buttons against a fresh empty
resolver. Closing the tab was the only escape. Phase 4 closed this same hole for
the footer and the tab bar; hashchange was the door left open.
Two independent fixes, because one is not enough: startRouting now takes an
injected isBlocked predicate (injected, not imported, so route stays uncoupled
from the import session), and the resolver moved into the store where it
outlives any component. A stale or reset resolver settles **false** — an
unanswered attestation is not consent, and this is the last gate before deleting
every bookmark someone has. The restore-the-hash path has an explicit test for
not recursing; an infinite hashchange loop in a released extension, on the path
a panicking user takes, is not a defect worth discovering from a review.
**Nothing told users their imported bookmarks are re-dated to today.** docs/05
§6 instructs saying so in the import report and docs/00 §8 repeats it as a
success criterion; no string ever did. The README and store listing say it to
prospective users, and the product never said it to actual ones — so someone
migrating a decade-old archive finds every bookmark dated today with no
explanation and no way to tell it from corruption. Now shown before the import
as well as after: telling someone before they commit is worth more.
Also: a profile whose every root is policy-managed now reports
NO_WRITABLE_ROOTS rather than "the browser refused a bookmark operation". The
browser refused nothing; it is an absence, not a refusal, and the old wording
sent people looking for a permission problem that was not there.
describeError moves out of ImportTab into browser/describe-error.ts so #edit and
#export can share it next, with tests asserting the two properties that have
been defects before: it returns a key rather than a resolved sentence, and it
omits `detail` rather than setting it to undefined.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/13 §6 stated as present-tense fact that "GitHub issues — templates route
bug reports with sample files", and docs/12 §5 carried the same as an unticked
box. Neither was true: .github/ held only FUNDING.yml, dependabot.yml and
workflows. Meanwhile the entire v1.0.x roadmap is "fixture + test per report",
which cannot happen if reports arrive without the file that broke.
Five issue forms plus a chooser, blank issues disabled. parser_report.yml is
the one that matters: it requires the sanitized sample, asks which browser
*version* produced the file, and asks explicitly for consent to commit the
sample as a GPL-3.0 fixture — it lives in a public repo forever, and nothing
previously asked.
scripts/sanitize-bookmarks.mjs makes "attach a sanitized sample" something a
reporter can actually do. It works by span replacement rather than by
reparsing, so every byte outside a replaced address or title is provably
untouched: BOM, doctype, line endings, tag and attribute case, quoting,
indentation and every timestamp survive, and so do the properties a parser
branches on — port, path depth, trailing slash, query-parameter count and
order, percent-escapes, HTML entities, and the distinctness of hosts that
dedupe depends on.
Its test asserts the promise rather than the implementation: for every fixture
in the repo, the sanitized file must detect as the same format, parse to the
same stats, warnings and topology — or fail with the same error code — and leak
no original hostname. Writing it immediately found three real defects: a data:
URL containing markup broke the title regex and left that bookmark's title
unsanitized; an & in an HREF came back as a bare &, silently repairing the
escaping of the file being reported; and Firefox's ICON_URI leaked a full
hostname past every other check. The last one is why URL-valued attributes are
now swept by value shape instead of by an enumerated list.
scripts/gen-fixture.mjs is the generator docs/11 §2 named, .gitignore has
always reserved a directory for, and the 2 s parse budget needed in order to be
measurable at all.
CONTRIBUTING.md collects the five-file ritual for turning a user's file into a
regression test, which was previously spread across CLAUDE.md, docs/04 and
docs/11 with no single entry point — including the .gitattributes rule that
stops EOL normalization destroying the bytes a bug depends on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/13 §4 has always said the VI/JA listing text should live in
store/listing.{vi,ja}.md "so listing text is version-controlled". It never did,
which meant every listing edit — one of the three things the v1.0.x roadmap
anticipates — would be made straight into the dashboard with no diff, no
history and no reviewer.
Reconstructed from docs/13 §3, with EN added as the canonical file since it is
the one CWS requires. Marked ⚠️ throughout: these were written after the listing
went live, so they are a copy of the specified copy, not a transcript of what
the dashboard shows. Until someone diffs the two, the dashboard wins.
Summary lengths are recorded per file and verified against the 132-character
dashboard cap — VI has only 7 characters of slack, which is worth knowing
before someone edits it and gets rejected at submission time.
JA carries the same TODO(review-ja) caveat as the in-app dictionary.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/13 had only a first-submission runbook, so nothing described how to ship
v1.0.1 — nor what to do when an update is rejected, when the item is taken
down, or when a tag was pushed and the release run died half-way. All four are
now reachable in a way they were not before the listing went public.
- docs/13: §1 marked historical (and its never-performed step 8 handed to
docs/11 §5). New §1b update runbook — the bump must reach main *before* the
tag, because release.yml compares the built manifest against the tag; the
store artifact is the Release asset, not the CI artifact, because that is
what SHA256SUMS.txt covers. New §7: a rejected update leaves the published
version live, so it is not an emergency; a rejected version string can never
be reused, so the fix ships as the next patch, not as a re-upload; and there
is no rollback in CWS at all — only a higher version carrying the revert.
- docs/12: the sample YAML had drifted twice (setup-node v6 vs the live v7, and
a missing check:manifest in the release job). Docs win here, so reconciling
the wrong way would have downgraded a pinned action and deleted a gate — both
fixed, with a note that this is the one doc where the running pipeline is the
ground truth. §2.4 gains the if:/continue-on-error that stops
notify-ci-failure failing on every green CI run for want of a secret.
§5's checklist now distinguishes verified from unknown instead of reading as
eight things nobody did.
- docs/11: §2's inventory named four fixtures that do not exist and omitted two
that do; each phantom is removed only after naming the test that actually
covers its behaviour. §5 gains the upgrade pass — installing the published
build, updating over it, and proving the legal flag and settings survive —
which is the only failure mode that hits every user at once. §6's parse
budget gains a procedure now that a generator exists to make the file.
- docs/14 §3 stopped embedding drafts of documents users have already accepted
(its copy still carried `Last updated: 2026-⚠️`). It now points at the
committed files and specifies when a change bumps LEGAL_VERSION, since that
re-shows the gate to everybody.
- SECURITY.md: supported versions (latest published only — Chrome auto-updates,
so there is no population to back-port for), an email fallback for reporters
without a GitHub account, and an honest split between "fix prepared in 14
days" and "fix reaching users", which additionally needs store review.
- legal/PRIVACY.md: says what happens to a file you send us. Deliberately not a
LEGAL_VERSION bump — it clarifies handling of data a user chooses to send,
which was never inside "what the Extension collects", and re-gating every
installed user over a clarification is the worse outcome.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The suite froze at the moment of submission and still said "in review" a week
after the listing went public. Since docs win over code here, a stale spec is a
false spec — and one of its "still open" items (the store URL for the README)
had already been closed by ea246aa.
- docs/00: status is published with the listing URL; the store item id and
listing URL join the §9 canonical identifiers table, because the id is also
the CHROME_EXTENSION_ID secret v1.1 automation needs; §8's approval criterion
marked achieved; §10 item 5 resolved and item 6 rewritten — DISCORD_CI_WEBHOOK
is not optional in practice, since notify-ci-failure.yml fails on every CI
completion without it. Two new post-launch items: the JA native review pass
(now shipping to users) and the never-performed post-approval smoke test.
- docs/15: the progress table gains a published row and a Phase 6 row; v1.0.x
stops being speculative and becomes the live workstream, with the intake path
spelled out end to end and a release trigger. Decision log: the Changelog-link
row is marked resolved, the two-tab known limitation superseded, and twelve
rows added for the Phase 6 design — recorded before the code, as the project's
own docs-first rule requires.
- docs/06 §3.5: the Changelog link points at CHANGELOG.md (it has since Phase 5)
and About gains a store listing entry. Also states why no support email
appears in the UI: an in-product address invites unsanitized bookmark files.
- CLAUDE.md: leads with the fact that this is shipped software, and what that
changes. Records EditTab.svelte back at 481 lines rather than the 459 it
claimed, adds a Phase 6 section, and adds the worktree bootstrap gotcha.
Also adds `verify:full`, because hard rule 8 now points at it. `verify` omits
coverage, build and check:manifest, all of which a tag run enforces — so a core
change can pass every documented local step and fail only after the tag is
pushed, when gh release create has already run and cannot be re-run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Status header, the docs/15 progress table and the docs/00 §10 manual list now
say where the project actually is: all six phases done, v1.0.0 tagged and
released, submitted to the Chrome Web Store and in review.
§10 gains two items that were being carried in commit messages and memory
rather than anywhere a reader would find them:
- `npm run dev:firefox` is untested against the dependency overrides. Four of
them sit under web-ext-run and no gate exercises that script. Firefox is a
v1.0 non-goal so it blocks nothing, but nobody should assume it works.
- The 2 s parse budget is unmeasured on a real browser. The recorded 2869 ms is
a jsdom figure and settles nothing either way.
The decision log gains the overrides rationale and why the release
announcement is continue-on-error.
The review-habit note is updated with phase 4 and 5 (19-of-33 and 8-of-20
confirmed), plus the two patterns this build kept demonstrating: the worst
findings were in work added earlier in the same branch, and three defects
reached main that no automated gate could have caught — a missing space, a
glyph that reads as a letter, and a permission grant that failed CI at startup
with no jobs and no log.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The real v1.0.0 run went red. The release job itself succeeded — the GitHub
Release, bookmarkmagic-1.0.0-chrome.zip and SHA256SUMS.txt all published, and
`sha256sum -c` verifies against the downloaded asset. Only `announce` failed,
with "No webhooks configured", because DISCORD_RELEASES_WEBHOOK and
DISCORD_REPO_WEBHOOK are unset — which docs/12 §5 lists as optional.
A red run on a correct release is worse than a missing announcement: it trains
you to stop reading the colour. `continue-on-error: true` on the announce job.
The tag_override fix from the Phase 5 review did work — the job reached the
webhook check, which it could only do after resolving the tag.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six Dependabot alerts (1 critical, 3 high, 1 medium, 1 low) plus six further
`npm audit` highs, all in transitive DEV dependencies whose parents pin them to
vulnerable ranges. That pinning is why Dependabot's own per-package PRs could
not fix any of them — every one of those runs failed. `overrides` is the npm
mechanism for precisely this case.
shell-quote 1.7.3 -> ^1.9.0 wxt > web-ext-run > fx-runner
adm-zip 0.5.18 -> ^0.6.0 wxt > web-ext-run > firefox-profile
tmp 0.2.5 -> ^0.2.6 wxt > web-ext-run
uuid 8.3.2 -> ^11.1.1 wxt > web-ext-run > node-notifier
esbuild 0.27.7 -> ^0.28.1 vite, wxt, unplugin
brace-expansion 1.1.16 -> ^5.0.8 web-ext-run > multimatch > minimatch
The last one is the whole second cluster: `npm audit` reported six highs that
were one brace-expansion DoS propagating through minimatch -> multimatch ->
web-ext-run -> wxt, with fixAvailable:false because multimatch pins minimatch.
`npm audit` now reports 0. It already reported 0 for `--omit=dev`, and still
does: the shipped extension has no runtime dependencies, so none of this was
ever in the artifact.
Correcting something I wrote in the dependabot.yml comment and repeated in a
PR: I said these advisories had "no reachable patched version". That was wrong
— every one of them had a patched version. What they lacked was a path
Dependabot could take on its own.
Re-verified after the overrides: lint, check, knip, coverage, guard, build,
check:manifest and zip all pass, and the built manifest is still exactly
["bookmarks","storage"] at version 1.0.0.
⚠️ Four of the six live only under web-ext-run, the Firefox dev runner behind
`npm run dev:firefox` — a v1.0 non-goal per docs/00 §7, and the one script no
gate exercises. docs/09 §3.1 records that, plus when each override can be
dropped.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two defects visible in the owner's screenshots of the running extension, both
invisible to every gate the repo has:
- The footer rendered "v0.1.0 ·GPL-3.0". Svelte trims trailing whitespace at an
{#if} boundary, so the space after the separator inside the block was
dropped. Separators are `{' · '}` expressions now, which are never trimmed;
verified in the built bundle.
- The compact theme toggle's dark option used U+263E (☾), a thin crescent that
reads as a capital "C" at ~13px — clearly visible in all four screenshots.
Now ☀️/🌙, which also matches the popup's existing ⬆/⬇/🗂 vocabulary. The
accessible name came from aria-label either way, so this is purely visual.
Adds .github/dependabot.yml (docs/12 §5): weekly npm and github-actions, dev
minor+patch grouped into one PR so CI runs once, Conventional Commit prefixes,
and TypeScript majors held because svelte-check needs the TS 6 programmatic API
until 7.1. The comment records what the file cannot do — security updates run
from the repo's alerts regardless, and the ones failing today are transitive
DEV dependencies with no reachable patch, which is why `npm audit --omit=dev`
is still 0.
Version bumped to 1.0.0 with the CHANGELOG section dated, and the icon set
closed in docs/00 §10: the owner adopted the gen-icons.mjs output as final, so
the generator is the source of truth rather than a placeholder. README's
screenshot placeholder is replaced with the four real captures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first real run of ci.yml was a startup_failure — no jobs, no log, just
"this run likely failed because of a workflow file issue".
reusable-chrome-extension.yml declares two mutually exclusive jobs: `build`
(contents: read, `if: !inputs.release`) and `release` (contents: write,
`if: inputs.release`). GitHub validates the caller's grant against EVERY job
the reusable declares, before evaluating any `if:` — so passing `release: false`
does not exempt the caller from the release job's requirement, and the
least-privilege `contents: read` in docs/12 §2.1's sample rejects the whole run.
CodeQL succeeded on the same push because its caller happened to grant a
superset of what its single job needs, which is what isolated the cause.
docs/12 §2.1 updated with the same grant and a warning, since its sample is
what this stub was written from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five lenses filed 20; 8 survived verification.
**release.yml would have failed on the first tag.** The `announce` job called
the ops reusable with no `with:` block, but that reusable resolves
TAG="${INPUT_TAG:-$EVENT_TAG}" where EVENT_TAG is github.event.release.tag_name
-- empty, because this workflow is triggered by `on: push: tags`, not
`on: release`. The job would have exited 1 on "No release tag available" for
every release. Filed independently by three lenses and confirmed against the
reusable's actual source; docs/12 §2.3 had the same omission and is fixed too.
Ironically the comment right above it explains why the workflow deliberately is
not event-driven -- which is exactly what makes tag_override mandatory.
**Nothing checked the BUILT manifest.** The unit test reads wxt.config.ts,
which proves what we ask for, not what WXT emits. Those differ: a dev build of
this same config carries ["bookmarks","storage","tabs","scripting"] and
host_permissions ["http://localhost/*"]. Production is clean, but that is
WXT's behaviour, not ours. scripts/check-manifest.mjs now gates the artifact
and runs in both CI and release; verified by editing a built manifest. The
README sentence claiming a test asserted the built manifest was false and is
corrected -- the same overclaim I had already made once verbally.
**The contrast test I added last commit could not see its own bypass.**
TabBar's `.locked { opacity: 0.65 }` composites --fg-muted against the page and
drops the label to 2.83:1. The test reads token values, so alpha applied on top
is invisible to it. Opacity removed; the padlock and title already carried the
signal. TEXT_PAIRS also widened to check the status colours on --bg, not only
--bg-raised.
**docs/09 T3 was signed off with half its mitigation unbuilt.** The row has
always specified a ⚠ badge for dangerous schemes; only the blocking existed. A
`javascript:` bookmark titled "Your bank" rendered identically to a safe one
and its Open button silently did nothing. isBlockedUrl now drives a badge, and
a refused open raises a toast.
Also: the five row action buttons had glyphs as their accessible names (content
beats title in accname), so voice control could not address them; the manifest
test's literal-harvesting regex would have passed a spread or identifier;
resolveNewFolderParent returned a focused unmodifiable folder's id, which every
create rejects; and clipboard.ts was the only new module with no test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Walking docs/09's threat model against the shipped code turned up two things
that had gone unnoticed because nothing mechanical was watching either.
**The permission list had no test.** "Exactly bookmarks + storage" is
CLAUDE.md's first hard rule, the README's headline privacy claim, the store's
privacy-tab answer and a line in the privacy policy -- and a single added
string in wxt.config.ts would have shipped silently. tests/unit/manifest.test.ts
now asserts the list exactly, plus the absence of host_permissions,
content_scripts, a background worker and eight other reach-widening keys.
Verified to fail by sneaking in 'tabs'. It reads the config as text, with
comments stripped, so the assertion cannot be satisfied by something computed
at build time -- and because the file's own header says "No host_permissions,
no content_scripts", a naive substring check failed on the prose promising the
very thing it checks.
**"Copy URL" was never built.** docs/06 §3.3 lists it in the row menu and
docs/09 T8 describes its mitigation, so the threat model documented a control
for a feature that did not exist. Now implemented in browser/clipboard.ts:
write-only, on an explicit click, never reads the clipboard, and needs no
permission because a transient user activation grants clipboard-write.
Also records the measured pass in docs/09 §2 and the reproducibility finding in
§5: content-identical rebuilds (every zip entry matches by CRC32) but not
byte-identical, because wxt zip records mtimes. The old wording implied more
than is true.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Per docs/12 §2, with every input checked against the ops repo's actual
workflow_call blocks rather than the doc — docs/12 originally described a
template family that had never existed.
Verified before writing: reusable-chrome-extension.yml takes
package-manager/node-version/build-command/source-dir/release and declares NO
secrets block, so passing a named secret to it would be a compile error;
reusable-codeql.yml takes `languages` as a JSON-array string (the older
codeql.yml's singular `language` is the wrong one); notify-ci-failure.yml
declares DISCORD_CI_WEBHOOK required:true; announce-release.yml's secrets are
optional. actions/checkout@v7 and setup-node@v6 are what the ops repo itself
runs.
Three traps the comments now carry, because each costs a run to rediscover:
- leaving package-manager at its 'none' default skips install and build, then
fails on `test -f ./manifest.json`, a file WXT never writes to the repo root;
- reusable workflows do not inherit the caller's permissions, so every caller
repeats its permissions block on the job;
- the release announcement is a job INSIDE release.yml, not an
`on: release: [published]` caller — a Release created by gh under the default
GITHUB_TOKEN emits no event that starts new workflow runs, so the
event-driven version silently never fires.
The quality job is local because the reusable runs no quality gates at all.
All four parse, and `npm run zip` emits .output/bookmarkmagic-0.1.0-chrome.zip,
matching the glob release.yml uses.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
CHANGELOG.md follows Keep a Changelog with the semver policy docs/12 §4
defines, and carries a marker for the human step at first release (rename
[Unreleased] to [1.0.0]).
README leads with what the project is actually for and then spends its longest
section on privacy, because that is the product's reason to exist and every
claim in it is mechanically enforced: the grep gate for network calls, the
manifest assertion for permissions, zero runtime dependencies. It also keeps
docs/13's honesty corner -- favicons, import dates, CSV flattening -- rather
than only listing what works.
src/lib/links.ts now points "Changelog" at blob/main/CHANGELOG.md instead of
the /releases placeholder it used through Phase 4, and the links test asserts
the file exists, the same way it already does for the four legal documents. A
link shipped inside the product must not 404 because someone renamed the
target.
Screenshots and the store URL are marked ⚠️ in the README -- both need a human
and neither blocks the code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
EditTab.svelte was 589 lines against the 500-line hard limit in docs/10 §1 --
the one place in the codebase that broke it. Now 459.
Extracted, in each case because the thing was genuinely cohesive rather than to
hit a number:
- src/lib/edit/edit-node.ts -- toEditNode, flattenTree, isEditable,
resolveNewFolderParent, countDescendants, extraCopyIds. All pure, all
previously inline and therefore untestable. resolveNewFolderParent in
particular encodes a fixed bug (a focused root resolving to the synthetic
root "0", which every create rejects) that nothing was pinning.
- src/lib/browser/open-url.ts -- the docs/09 T3 scheme allowlist, which is the
only thing between a crafted `javascript:` bookmark in someone's import file
and the browser acting on it. It had NO test. It does now, including the
case-shouting dodge and data:/file:/chrome:.
- EditToolbar.svelte and MoveToDialog.svelte -- presentational, with their
styles copied verbatim. This is a move, not a redesign.
The script block is still 279 lines against docs/10 §1's 150-line guidance. The
remainder is the mutation layer (guardEdit and the six operations), which is
coupled to reactive state and would need accessor indirection to move. Adding
that indirection to the app's delete paths immediately before a release is a
worse trade than the line count, so it stays and is recorded in CLAUDE.md
rather than quietly ignored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
docs/06 §5 has said "≥ 4.5:1 in both themes (tokens chosen accordingly)" since
the spec was written. Nothing enforced it, and the tokens were not chosen
accordingly. Measured:
white on dark --accent 3.67:1 (fixed in Phase 4)
white on dark --danger 2.74:1 Button/ConfirmDialog .danger
light --warn on --bg-raised 3.40:1 WarningList summary
light --success " " 3.26:1 TreePreviewNode badge, DuplicatePanel
light --danger " " 4.27:1 DuplicatePanel count
dark --accent " " 4.40:1 links inside raised panels
Adds --danger-fg, which like --accent-fg must be redefined in the dark block in
the OPPOSITE direction: dark mode lightens the colour for contrast against the
page, and that is exactly what breaks white ink sitting on it.
The light status colours are now darker than an eyeball would pick, because all
three are rendered as text on --bg-raised. Adjustments preserve hue and were
solved for numerically rather than guessed; each clears 4.5 with margin.
The real deliverable is tests/unit/styles/contrast.test.ts, which parses
tokens.css and fails on any text pair below 4.5:1 — verified to actually fail
by restoring the old --warn and watching it report 3.40:1. It also pins the two
dark blocks (explicit override and prefers-color-scheme) against each other,
because they already drifted once when --accent-fg was added to only one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Phase 4 is now verified on a real browser, which every earlier phase had been
and this one had not: the gate's block/reach/persist behaviour, EN→VI→JA live
switching with localized numbers and dates, the theme control, settings
surviving an immediate tab close (the path the $state-proxy DataCloneError
would have broken), and the Phase 2–3 flows still working after Phase 4 touched
their tabs and the settings store.
docs/00 §10 item 4 closes: the contact email in legal/PRIVACY.md is filled in
and confirmed live. That was the last thing standing between the repo and the
docs/13 §1 store runbook that is not Phase 5 work.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The plural helper docs/07 §5 asked for now exists, so §5 names it and the two
shape tests that enforce it. The decision log gains the five design choices
the review forced, and the progress note records that 19 of 33 findings
survived verification -- including the first-run dead-end where the gate was
reachable only from the routes it blocked.
Adds one measured defect to the Phase 5 list rather than fixing it silently:
Button's .danger/.success hardcode white, which is 2.75:1 on dark-theme
--danger. Same class of bug as the --accent-fg one Phase 4 fixed, but it needs
new tokens and a visual decision rather than a mechanical change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Last batch from the Phase 4 adversarial review, including one the completeness
critic caught that none of the six lenses looked for.
The num() retrofit was incomplete: FolderPickNode rendered its per-folder count
raw, so a Vietnamese user saw "12345" in the export pick tree while StatsCard
showed "12.345" on the adjacent tab. docs/07 §3, written in this same phase,
enumerates exactly three exemptions and this was on none of them.
ImportTab resolved t('errors.…') inside a catch block and stored the resulting
STRING in the session store. t() is only reactive where it is called, so the
error message froze in whatever language was active when it happened and never
followed a language switch. Phase 4 is what makes that observable -- until now
there was no way to switch language at runtime. The store holds the key now.
docs/07 §5 specifies plural keys as { one, other } objects in every locale with
a helper that selects by locale. No such helper existed and the dictionaries
used "(s)" instead, which produced "1 unreadable date(s) were left blank." --
wrong in more ways than the parenthesis. selectPluralForm uses Intl.PluralRules
(en gives one/other, vi/ja only ever other, so two forms cover every locale we
ship) and tPlural supplies a locale-grouped {n} so callers pass a raw count.
All seven warning codes are plural keys, including the two that interpolate no
count, so WarningList can call tPlural unconditionally instead of maintaining a
list of which codes count things.
Also splits getGeneratorVersion from getAppVersion: the UI hides an unknown
version, but a file field cannot be hidden, and "BookmarkMagic " with a
trailing space is a worse artifact in an export -- and in the forced Replace
safety backup -- than the bare product name.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Five findings from the Phase 4 adversarial review, all of them things a green
build cannot see.
tokens.css redefined --accent in both dark blocks but never --accent-fg, so
dark mode painted white on #8b6dff -- 3.67:1, under the 4.5:1 docs/06 §5
requires. That hit the ThemeToggle's selected chip and every primary button.
Dark ink on the same violet measures 4.84:1.
ThemeToggle's `overflow: hidden` clipped the global :focus-visible ring, which
base.css draws at outline-offset: 2px, i.e. entirely outside the clip region.
The theme control had no keyboard focus indicator at all. Corner rounding is
per-child now and the ring is inset.
ThemeToggle also claimed role="radiogroup"/"radio" while implementing neither
roving tabindex nor arrow keys, so a screen reader announced "radio button,
1 of 3", switched to forms mode, and then the arrows did nothing while all
three stayed separate tab stops -- the inverse of the pattern (WCAG 4.1.2).
Arrow/Home/End now work and only the selected radio is a tab stop. The handler
sits on the radios, not the group: a keydown handler on a non-focusable
container never fires.
Toast built its role="status" region and its text in one DOM insertion. Screen
readers derive announcements from mutations to a region that was ALREADY
present, so success toasts were never announced. Two permanently-mounted
regions now, polite and assertive, with only their text changing -- two rather
than one with a swapped aria-live, because changing politeness at runtime is
not reliably picked up.
applyLocale never touched <html lang>, and both entrypoints ship lang="en", so
a screen reader read Vietnamese and Japanese with an English voice. It also
fixes the toast's dismiss button, which was labelled "Cancel" -- there is
nothing to call off, the operation already finished.
Also routes the last three hardcoded English aria-labels ("Sections",
"Preview" x2) through the dictionary. They date from Phase 2, but full EN/VI/JA
coverage is this phase's exit criterion.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>