Guide and in-app docs name the five card pages, "Got it" on the two
visit cards and the Trends card that also checks the reports step;
architecture lists the new models and targets; the QA list gets the
steps 5 and 8 of the guided walk, the Free padlock variant and the
Trends click.
Resolves#432
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The two visit steps get their card: /categories explains keywords and
points at "All keywords" (never "Reset"), /reports explains the hub and
points at the Trends card, with a padlock variant in the Free edition.
"Got it" acknowledges the step; opening the highlighted Trends card does
too (D17), since following the ring leaves the page.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Guide and docs.gettingStarted / docs.dashboard keys (FR, EN): eight
steps, the two visit steps in the quick-start list, "Got it", 3/8 and
7/7, a finished path finds them done.
- ADR 0022: acknowledgeStep and initializeVisitSteps do not notify (§2),
ACKNOWLEDGE among the local actions (§4), the one-time initialization
admitted like rule D9 (§5).
- architecture.md: eight rows, the amended "never by a click" invariant,
the two new preferences, the provider's new action and initializer.
- QA list: step numbers moved to eight, five cases for the visit steps.
- CHANGELOG: Added entry in both files.
Resolves#425
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A mutation that moved "categories" after "manualCategory" in the
constant survived: the rows come from computeChecklist, the constant
only orders the skipped and acknowledged steps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two visit steps join the getting-started path: "Browse categories and
keywords" (5th) and "Browse the reports" (8th). Nothing in the database
counts them: they are done by "Got it" on their card, stored as
onboarding_acknowledged_<key> in user_preferences (no migration).
- computeChecklist: eight rows; the visit rows are locked until the
profile's visit steps are initialized, so a finished path does not
flash a "Next step" before they are marked done.
- initializeVisitSteps: once per profile, after the first successful
load with the Budget entitlement known; acknowledges both steps when
every other step is done or skipped, then sets the flag. No
transaction (D16): idempotent writes, the flag last.
- Reducer: ACKNOWLEDGE wins over an older snapshot like SKIP but keeps
the card latch, so the card says "Done." and offers the next step.
- StepCard: "Got it" before "Skip this step" when onAcknowledge is
given; OnboardingStepCard passes it for a visit step.
- Banner counter, card rank and end of path wait for
budgetReady && visitStepsInitialized; the list shows the visit rows
"pending" until then.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The page help and the guide (Markdown and in-app) described a Budget split
button removed in v0.2.5, an Annual-column warning that can no longer
appear, report tabs and four report cards (the hub has five, no tabs), a
"Category Zoom" whose displayed title is "Category Analysis", and a "+"
icon on transactions (the keyword button is a tag icon, shown on
categorized rows only; the right-click menu works on every row).
Resolves#431
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The two older QA cases that quoted the previous card sentences now quote
the new ones.
Resolves#430
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both Transactions sentences name the context menu's "Add as keyword" and
carry the advice on the word and the boxes (D18); both Budget sentences say
only last-level rows take an amount and what the Annual column does. The
labels go through StepCardText.labels, quoted in the locale files (D11).
The keyword dialog is unchanged; a test pins that it marks the transactions
as categorized by hand, so the step still checks off through it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The source name typed during configuration was saved, but the next import
looks the source and its imported files up by the subfolder name, so a
renamed source looked new: format guessed again, every file pre-checked,
and a second source created under the folder name.
The name field becomes a read-only SourceNameDisplay showing the subfolder
name with a hint, and executeImport writes selectedSource.folder_name on
create and update, whatever config.name holds, behind a selectedSource
guard. The import card sentence no longer asks to check the source name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scan_import_folder now returns ImportFolderScan { sources, root_files }:
the CSV/TXT files at the root, which no source reads, are named in a
warning on the Import page (above the source list, or under the "no
source" empty state). One action sets sources and root files, RESET keeps
both. Names stay on screen, never in the Journal.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Import empty state now shows an "Expected structure" box in both of its
variants (no folder, folder without sources): the parent folder to choose,
one subfolder per account, and files at its root ignored, as
scan_import_folder reads them. New ImportFolderStructure component, a nested
list with decorative icons.
The getting-started cards ask for the parent folder and one subfolder per
account instead of one per bank. The preview card says that an expense must
count among the outflows and names "Flip the signs" next to "Check
duplicates": WIZARD_ROWS rows now carry every label their sentence names.
Every Import card sentence quotes the controls it names, in French and in
English, pinned by a test that reads the locale files.
Help tips, docs.import.steps, the user guide, the QA list and architecture.md
follow; CHANGELOG entry under Changed.
Resolves#422
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
buildBudgetYearRows sorted each top group by depth, so an expanded
subcategory showed at the end of its group instead of under its parent.
It now builds one block per child (lone leaf, or [subtotal, ...children]),
sorts the blocks once ("(direct)" first, then by name, leaves and
sub-groups mixed, as the Budget vs Actual report) and concatenates them;
top groups are sorted by type, sort_order and name.
reorderRows put a leaf that follows a sibling sub-group into that
sub-group's block. A row at the subtotal's depth or shallower now closes
the block, which also fixes "subtotals at the bottom" in the Budget vs
Actual and Compare reports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Written before the code, from spec-decisions-mcp-server.md and the
reviewed spec-plan-mcp-server.md (local, git-ignored). Status Proposed;
to pass Accepted when the spec-mcp-server milestone merges. Indexed in
docs/architecture.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The small badge next to "Fetch price" on a stock rendered the literal
"best-effort" without going through t, so French users saw English.
It now reads balance.priceFetching.bestEffortBadge ("non garanti" in
French). Raised by the review of PR #435 (suggestion 4).
Refs #389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- withoutFolders no longer keeps the first word of a folder that holds
a blank (`C:\Users\Jean Dupont\...\profile_x.db` gave
`Jean profile_x.db`): a path starts at a recognised prefix (drive,
`/`, `~/`, `./`, UNC, `<profile folder>`) and runs to its last
separator, blanks included, until a line break, a double quote, a
backtick, `;`, `|`, a `,` or `:` before a blank, a closing parenthesis
not followed by a separator, or a blank before another path. When
nothing ends it, the text up to the last separator of the line is
dropped rather than a folder let through. A quote before the path is
not part of it, so `'C:\Users\Bob\x.db'` gives `'x.db'`. Relative runs
with no recognised prefix keep the previous fallback. No error line
contains a `/` nor a `\`, as before.
- ADR 0021 section 7 describes that rule and its limit.
- CHANGELOG.md, CHANGELOG.fr.md: no blank line between the two Fixed /
Corrigé bullets.
Refs #395
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- renderCategoryLabel.test.ts passes markedT itself, its assertions in
the marked form (`t(balance.category.cash[defaultValue=cash])`): the
second fake `t` of the repository is gone.
- i18n-mock-guard: a fake `t` is also caught as a binding (`t = (`,
`t = vi.fn(`) or as a `t: vi.fn(` property, and a file that mocks
react-i18next must call `i18nMock(`, comments dropped before the
search. Each new rule has its own case in the guard's tests; the
binding rule found renderCategoryLabel.test.ts before its migration.
- docs/architecture.md: the "only fake t" sentence is scoped to test
files and says the guard only refuses the forms it recognises.
Refs #389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- useOnboardingChecklist.wiring.test.ts: a data event, then a route
change 100 ms later, still gets its deferred re-read (two LOAD_START).
Pins the follower's deps: with `pathname` among them, the cleanup drops
the pending timer and the test goes red (mutation M1 of the review). The
comment that claimed "one subscription only" now points to this test.
- hookHarness.test.ts: a dependency array that shrinks, or grows by
`undefined`, recomputes the memo (mutation M8, length branch of
depsChanged).
- NOT SIMULATED, in hookHarness.ts and docs/architecture.md: the dispatch
of a useReducer keeps the reducer of the first render, where React
applies the latest one.
Refs #409
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- close() on a getDb() handle runs as a "connection close" section and
checks its generation: a handle kept from a previous profile rejects
with DbConnectionClosedError instead of closing the current profile.
- closeUnlocked never rejects: the state is cleared and the generation
bumped before close(), a failing close of the old pool is logged with
logWarn, connectToProfile still opens the next profile and closeDb()
resolves without a connection.
- The blocking alert names a getDb() call in a callback as a possible
cause only: a long restore also waits 15 s.
- The generation check of the transaction handle is documented as a
defence, not reachable today.
- Tests: the "behind the queue" alert (pins the reset of lockHolder),
close() on a stale and on the current handle, a failing close during a
switch and in closeDb().
Refs #408
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- ProfilePreparation.catch_up_backup carries the file name of the copy as
soon as it exists, so the Journal error line of a catch-up that stops
afterwards (history_write_failed, pragma_failed, apply_pending) names it.
- Error lines and the JS fallback message keep only the last segment of a
path (withoutFolders): no error line contains a slash or a backslash.
- The copy is taken again at every catch-up, written as .tmp, synced and
renamed over a same-name copy; after a successful catch-up only, the
other copies of that profile are pruned (remove_backups keep argument).
- Pins: a pre-existing foreign key violation in the profile without
category 31 (RV4, RV7) and the Journal line that reports it (RT2).
- Ignored test measuring prepare_at on 100 000 transactions at floor 6:
306 ms in release, so STARTUP_TIMEOUT_MS stays 10 s (ADR 0021).
- Guide, docs.profiles.tips, ADR 0021, architecture and changelogs.
Resolves#395
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 10 s startup delay of App.tsx sets the error page while the profile
is still opening; a connection that succeeded afterwards set dbReady but
left the error, so the error page stayed. The startup logic moves to
src/utils/startupConnection.ts (startProfileConnection), which clears the
error on success and is tested with fake timers.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
src/__integration__/i18n-mock-guard.test.ts reads every test file of src
and fails on a useTranslation: or t: ( property (word-bounded) outside
src/test/i18nMock.ts, or on a react-i18next mock that does not go
through i18nMock. docs/architecture.md gets the convention: never a raw
t, always i18nMock.
Resolves#389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 12 test files that already marked their output each carried a copy
of the same fake t. They now use the shared i18nMock; no assertion
changes. LicenseCard keeps its per-test language through a getter.
Refs #389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 15 test files whose fake t returned the bare key (the 12 of the
issue, plus the three category render tests added by #423) now mock
react-i18next with i18nMock. Assertions naming a displayed text are
rewritten in the marked form, t(key[options]), so a text that skipped
t fails them. Components without such an assertion get one:
PriceFetchControl, PriceFetchConsentToggle (rendered through the hook
harness, its loading slot preset), CategoryDetailPanel, SourceList and
SourceCard, and the region label of OnboardingChecklist.
Each component was proved by a mutation removing one t call: every one
turns at least one test red (exit 1).
Refs #389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
src/test/i18nMock.ts holds the one fake t of the suite: markedT renders
t(key) or t(key[a=1,b=2]), and i18nMock({ language?, fixture? }) builds
the react-i18next module for vi.mock. The local echo of
useLicenseMachines.test.ts is replaced by markedT.
Refs #389
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
useOnboardingChecklist.wiring gains three pins the old stubs could not hold:
one loader for the provider's life (refresh and the shared value keep their
identity, no re-read), 'Skip' moving the next step on the next render, and a
change of mode alone reaching the consumers. UsersSettingsPage now commits
its effects, so an effect that read the preference goes red on the
assertion rather than on 'Invalid hook call'. BudgetTable.wiring,
BudgetPage.wiring, AppShell.slot and OnboardingStepCard drop their local
slot copies; OnboardingStepCard plays StrictMode through replayEffects and
runs its effect only on a changed dependency, which now pins location.key.
Refs #409
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
src/test/hookHarness.ts gives the tests that call hooks and components as
plain functions one fake React with React's semantics: state initialized
once with stable setters, the reducer applied with a dispatch spy, refs kept,
memos and callbacks recomputed only on a changed dependency, effects run by
commit() only when a dependency changed (cleanups first), and every hook it
does not cover throws instead of reaching the real React. Each semantic is
proved by hookHarness.test.ts, which also checks that the module imports
React only as a type.
Refs #409
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Strict lock, transaction handle, single section for a profile switch,
connection generation and blocking alert.
Resolves#408
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The global inTransaction flag let every execute/select bypass the FIFO
lock while a withTransaction callback ran, wherever it came from. The
callback now receives its own handle, which runs directly, refuses
close() and throws once the callback settled; any other call, getDb()
from a callback included, queues behind the transaction.
connectToProfile, initializeNewProfileDb and closeDb each run as one
exclusive section (close the old pool, prepare, load), and every open
or close bumps a connection generation: an operation queued for a
closed connection throws "connection closed by profile switch". An
operation waiting more than 15 s for the lock writes one Journal
warning naming the holder and since when.
findOrCreateSecurity now requires its executor: its getDb() fallback
would deadlock if called from a transaction callback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs #423
- parentOptions.ts: say that this list is the only guard today (updateCategory
checks neither depth nor cycle, #393 will), instead of pointing to a
service-side validation that does not exist yet. The height walk is also
bounded by the number of categories, so a missing cycle guard gives a wrong
height instead of an endless loop; a test edits a member of a cycle.
- AddKeywordDialog.tsx: drop `categoriesLoaded`, which only repeated what the
null initial category already guarantees.
- Tests: the form and the detail panel are pinned to what they hand to
buildParentOptions (editingId, currentParentId), through a pass-through spy.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Refs #423
The review of PR #428 showed that the rule only looked at direct children:
editing a root that carries grandchildren still offered the other roots as
a parent, and saving created a fourth level, which nothing guards against
in updateCategory or moveCategory. The allowed parent depth is now
MAX_PARENT_DEPTH minus the height of the edited sub-tree (0, 1 or 2), so a
category with grandchildren gets no parent at all, only its current one.
Test written first on the reviewer's probe ([1, 10->1, 100->10, 2],
editing 1: [] expected, [2] obtained before the fix).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Fixes#423
The "Parent category" field of CategoryForm and the "Category" field of
AddKeywordDialog were native selects with no typing. Both now use
CategoryCombobox, like the category cell of the Transactions page: typed
letters filter the list, which is ordered hierarchically with indentation
and shows seed categories through their i18n_key.
- CategoryCombobox accepts a minimal structural type (CategoryLike), so
CategoryTreeNode rows pass without conversion; sortHierarchical is
generic over it. Callers are unchanged.
- buildParentOptions (pure, tested) keeps roots and level-1 categories,
leaves out the edited category, and keeps only roots when the edited
category has children. That rule alone also leaves out its descent (a
descendant is never a root), so no separate subtree walk is needed. The
parent currently stored is always kept so the field never looks blank.
CategoryForm gets an editingId prop, passed by CategoryDetailPanel.
- AddKeywordDialog reads the Transactions list (active and inputable,
decision D2) through loadKeywordCategories, keeps the pre-filled
category only when it is in that list (reconcileInitialCategory), and
holds the apply button until the list is in. Its two fields sit above
the scrolling block of matches, so the open list is not clipped by the
scrollport.
- Tests: parentOptions, keywordCategory, and static renders of both
components. Changelogs and the user guide (with its i18n mirror).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Milestone planned-2026-10-10-release-deps merged (PRs #414 to #418,
main 9e8c41b): new position paragraph and decision entry, the oldest
decision (#365) rotated out. Blockers: #419 point 1 before the next tag,
stale mentions of #376, #317, #345, #358 and #320 replaced by their
closure.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-up of PR #417. CLAUDE.md and docs/architecture.md said
nothing re-evaluates `npm audit` in CI. No job fails on it, but the
`npm ci` step of check-frontend audits on every run and prints the count
(run 520: found 0 vulnerabilities; runs 512 and 519: 8), and the
Defenseur reports advisories on its side, which is where #358 came from.
Refs #358
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-ups of PR #414.
- audit.yml points to the concurrency comment of check-rust.yml, whose
measurement covers push and pull_request runs only: say this workflow
runs on schedule and workflow_dispatch, which it does not cover.
- reportService.test.ts said CI runs Node 20, so no node:sqlite; CI runs
Node 24 since #376, and the harness itself is #392's.
- The release skill said /actions/tasks ignores `limit`. Measured
2026-10-12: only without `page` (613 rows); with `page` it pages by 30
or by `limit`.
Refs #376
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-ups of PR #415. The container does have python3 (pulled by
the apt packages of the install step, run 517): awk is kept because mawk
is part of the ubuntu:22.04 base image, which the workflow comment and
the ADR now say. #344 is a PR and the second drift, not the third. The
awk exit-code test is described by what it alone catches (an awk that
fails after printing every fact). The repair procedure gains the two
failures it had no answer for, and the ADR names the canary's limit.
The release commit edits Cargo.lock on main without a PR, so /release
now replays the guard's exact step text right after that edit (step 2),
checked on the stack's lock (exit 0) and on a drifted edge (exit 1).
Refs #345
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Max's decision on PR #416, after its review measured who still read the
key. tauri-plugin-updater looks up "{os}-{arch}-{installer}" first, and
every version that ships the updater (v0.2.1 and later, plugin 2.10.0,
tauri-utils 2.8.2 in each tag's Cargo.lock) carries its bundle-type
marker, so .deb and .rpm installs never reach the generic key. Its last
readers were the AppImage installs v0.4.3 to v0.4.5 (v0.4.2 targets a
route that answers 404, earlier versions read GitHub) and binaries run
outside a package; none of them can install the .deb it served, and the
updater wrote that .deb over the running file. They now get
TargetsNotFound when checking: an error on the Updates card, nothing
installed.
The key leaves the script, scripts/check-latest-json.jq and
verify-release.sh together; the test compares the three lists. Docs fix
the readers named by #416 (v0.4.2 is not one) and the CHANGELOG entry
says what changes for those AppImage installs.
Refs #320
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-ups of PR #416. The check that release.yml runs before
deleting the published manifest moves from an inline jq program to
scripts/check-latest-json.jq, which release.yml applies with
`jq -e --arg version ... -f`. scripts/build-latest-json.test.mjs now runs
that same filter on every pull request: on the script's output, on a
wrong version, on nine broken manifests and on the shape published up to
v0.17.1. It also compares the key lists of the script, of the filter and
of verify-release.sh, so a list left behind shows before the tag.
Two new cases pin the version argument: 0.18.0 must come out as 0.18.0
(the other fixtures pass 0.17.1, the version the bundle names carry, so a
script that ignored its argument passed them all), and an empty version
must fail. Both mutations survived the review's pass.
No change to what the manifest holds: still four keys.
Refs #320
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
npm audit reported 8 vulnerabilities (5 high, 3 moderate) on the lock
of main: react-router and react-router-dom (one advisory,
GHSA-qwww-vcr4-c8h2), vitest and @vitest/mocker (one advisory,
GHSA-82fw-gwwq-j7x9), browserslist, baseline-browser-mapping, nanoid
and source-map-js. Every fix sits within the range package.json
already declares, so only the lock moves; package.json is untouched.
Lock generated with npm 11.21.0 exactly:
npx --yes npm@11.21.0 update react-router-dom react-router vitest \
@vitest/mocker browserslist baseline-browser-mapping nanoid \
source-map-js --package-lock-only --ignore-scripts
259 entries before and after, 20 changed, each resolved from
registry.npmjs.org with a new sha512 integrity. npm ci, under npm 10
and under npm 11.19.0, leaves the lock as is; npm audit: 0.
Only react-router 7.18.1 -> 7.18.4 changes the built app: a build from
a lock where only react-router moved equals the full build byte for
byte. Since 7.18.3, useNavigate() and <Navigate> reject a navigation
to another origin; the app only navigates to internal paths.
CLAUDE.md and docs/architecture.md no longer describe 2 high
advisories accepted for good: npm audit is at 0 on 2026-10-10, a
dated reading that nothing re-runs in CI.
Resolves#358Resolves#317
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fallback key is also read by the AppImage installs of v0.4.2 to
v0.4.5 (bundled until v0.7.3, published on Forgejo for v0.4.2-v0.4.3
and on the GitHub mirror for v0.4.3-v0.4.5, all reading the registry),
which look up linux-x86_64-appimage, absent, then linux-x86_64. Like a
marker-less binary, they cannot install the .deb it serves: the updater
writes its bytes over the running AppImage. State since v0.4.2, not
changed by this PR; the key stays as decided.
Also quote the error the update card really shows for a refused
package (invalid updater binary format).
Refs #320
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
latest.json only had linux-x86_64, built from the .deb. An .rpm install
looks up linux-x86_64-rpm first, then fell back on linux-x86_64,
downloaded the .deb and refused it (InvalidUpdaterFormat).
- scripts/build-latest-json.sh writes linux-x86_64-deb,
linux-x86_64-rpm, linux-x86_64 (the .deb, as a fallback) and
windows-x86_64 from the release assets. It fails unless it finds
exactly one .deb, one .rpm and one -setup.exe, each with a non-empty
.sig. The JSON comes from jq alone, notes read with --rawfile, file
names encoded with @uri like the upload step.
- scripts/build-latest-json.test.mjs (vitest) runs it under
bash -e -o pipefail on a temporary assets dir; check-frontend
installs jq.
- release.yml calls the script, writes the notes output under a random
delimiter, checks the manifest with jq before the registry DELETE,
and names the .rpm instead of an AppImage in the release notes.
- verify-release.sh requires the four keys, the extension of each URL
and the minisign file: of each signature, and exits 1 on any failure.
- Docs: release skill, qa-update-cycle section 3 (rpm procedure, keys,
reason and removal condition of the fallback), architecture, README,
CHANGELOG, STATE.
Resolves#320
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"Ce sous-arbre" lost its referent once the trigger paragraph stopped
naming a subtree: the sentence now says what is measured, that both past
drifts (#310, #344) came with RustSec advisory fixes. Rule 1's sentence on
a missing block is reworded.
Refs #345
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
