The `metadata` JSON part in the manga-create and chapter-upload handlers was
read with `Field::bytes()`, buffering the whole part before any check — a client
could send a multi-hundred-MB metadata blob (up to the 200 MiB request limit) as
one allocation. Image parts were already streamed with a cap; metadata was the
last unbounded read.
Cap it at a shared upload::MAX_METADATA_BYTES (64 KiB) via the read_capped
streaming loop (413 on overflow), and remove the now-unused read_field_bytes
helper so the unbounded path can't return.
Tests: manga + chapter oversized-metadata parts are rejected with 413.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
`get` (fs::read → EISDIR) and `get_stream` (File::open succeeds on a Unix
directory, then streams garbage that fails mid-read) both mishandled a key
resolving to a directory. Mirror `size`'s existing guard: `get` maps
IsADirectory to NotFound; `get_stream` checks metadata.is_file() after open.
Test: get_on_directory_is_not_found.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A `%` or `_` in a search term silently acted as a LIKE wildcard rather than
matching literally (`50%` matched everything, `a_b` matched `axb`). Not
injection — terms are bound — but a search-correctness bug across every
user-facing ILIKE site.
- repo::escape_like: shared pub(crate) escaper (promoted from page_tag), unit-tested.
- Trigram-entangled sites (manga, tag, author) append a separate escaped param
used only by the ILIKE branch; the trigram/similarity branches keep the raw term.
- Pattern-built sites (admin manga list, admin users, analysis coverage + history,
crawler search incl. JSONB payload title) escape inside format!("%{}%", ..) and
pair each ILIKE with ESCAPE '\'.
Tests: escape_like unit tests + integration tests on public manga search, author
autocomplete, and admin user search proving `_` matches literally.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Document responses carried no CSP, X-Frame-Options, Referrer-Policy, or
Permissions-Policy — leaving clickjacking and zero script-injection defense
in depth (security audit T4).
- svelte.config.js: enable kit.csp hash mode. SvelteKit hashes its own inline
hydration scripts; the inline theme <script> in app.html isn't part of
%sveltekit.head%, so its sha256 is pinned by hand (csp-config.js) and added
to script-src alongside object-src 'none', base-uri 'self', frame-ancestors
'none'. Styles stay unconstrained (Svelte emits dynamic inline style= attrs).
- hooks.server.ts: applySecurityHeaders() sets X-Frame-Options: DENY,
Referrer-Policy: strict-origin-when-cross-origin, X-Content-Type-Options:
nosniff, and a Permissions-Policy locking down unused features, only when the
response hasn't already set them.
- src/csp-theme-hash.test.ts: drift guard against editing the theme script
without updating THEME_SCRIPT_HASH.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
listAllChapters used `page.total ?? all.length` as the loop bound. The
/chapters endpoint always returns `total: null`, so a full 200-row first
page broke the loop immediately — the reader dead-ended past chapter 200.
Stop on a short page instead; only trust `total` when actually reported.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The network-error wrapping added in the error-boundary work also caught
deliberate AbortController cancellations and turned them into a network_error
ApiError, breaking callers that distinguish "cancelled" from "failed" (the
cancellable admin fetches). Rethrow DOMException AbortError unchanged; only
genuine network failures become ApiError(status 0).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Overlays didn't lock body scroll, so the page behind a modal/sheet scrolled
under it. Add a ref-counted scroll lock (so stacked overlays don't prematurely
release it) and hold it via an effect in Modal and Sheet while open, restoring
on close or unmount. Focus-trapping was already handled.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
After a partial upload failure (manga created, a chapter failed), re-submitting
re-ran createManga and spawned a duplicate manga. Track the created manga id:
a retry skips creation, re-sends only the not-yet-done chapters, disables the
now-saved title, relabels the button "Retry failed chapters", and shows a note.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drag-dropping page images then clicking a nav link silently discarded the
whole form. Add a reusable guardUnsavedChanges helper (SvelteKit beforeNavigate
confirm + native beforeunload) and wire it into the manga upload, manga edit,
and chapter upload forms. The dirty check returns false while submitting so the
post-save redirect isn't prompted.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both lists fetched a single capped page (100 / 200) with no pager, so entries
past the cap were silently unreachable. Add a reusable LoadMore component that
seeds from the streamed first page (preserving the auth gate + inline error
handling) and fetches further offset-based pages on demand, and wire it into
the bookmarks and collections pages.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The reader fetched only the first 200 chapters, so a deep chapter (#250) fell
outside the window and its prev/next chevrons resolved to null — a dead end —
and the chapter dropdown was truncated. Add listAllChapters, which pages
through the API's 200-row cap, and use it in the reader loader.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Chapters/Mangas aggregation views accept an OCR ?text= filter and the
page search intersects tags, but the client dropped both: any text query
collapsed every view into a flat page search that discarded the selected tag
and grouping. Add text to AggregateOptions, thread it into
listTaggedChapters/listTaggedMangas, and rework the search loader so a
tag-scoped view runs the aggregation (optionally text-filtered) and the page
search passes the selected tag.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The account page pointed users at a "bot-token list" that didn't exist: there
was no GET endpoint, no UI route, and the client type lacked expiry. Add
GET /v1/auth/tokens (caller-scoped, token_hash never serialised), extend the
auth client with listTokens/expires_at and createToken expiry, and add a
/profile/tokens page to list, create (revealing the raw bearer once), and
revoke tokens. Link the account-page copy to it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A load() that rethrew (e.g. the profile overview) fell through to SvelteKit's
bare default error page, and a connection-refused fetch rejected as a raw
TypeError that blanked the view. Wrap fetch network failures in the client as
ApiError(status 0, network_error) so callers handle them uniformly, and add a
root +error.svelte with a friendly message plus Try again / Go home actions.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The admin checkbox fired an immediate, irreversible privilege change with no
confirmation, and on cancel/failure the optimistic DOM flip stuck while the
server state was unchanged. Add a confirm() gate (like Delete) and make the
checkbox controlled by u.is_admin: onchange reverts the DOM immediately and it
only flips once the server confirms — so a cancelled or failed toggle never
misrepresents the real role.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add /files/{key}?w=<px>: the endpoint decodes JPEG/PNG sources, downscales to
a width snapped to a small allow-list (aspect-preserving, never upscaling),
caches the result under a thumbs/ prefix, and serves it — so cover grids
download ~KB instead of the 1–5 MB original. Cover handlers purge cached
variants when a cover (whose key is reused) is replaced or deleted. Frontend
grids use new thumbUrl/thumbSrcset helpers (MangaCard with responsive srcset).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
?sort=author used a correlated min(lower(a.name)) subquery as the ORDER BY
key, evaluated per filter-matching row before LIMIT. Materialize it into
mangas.sort_author (indexed by mangas_sort_author_idx), maintained by triggers
on manga_authors and backfilled in the migration, so the sort is a plain
indexed column read. Author names are immutable, so only join changes matter.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The content-warning include/exclude filter (and its count) ran a correlated
page_content_warnings->pages->chapters join per candidate manga — O(mangas *
pages) on every filtered list. Add a manga_content_warnings(manga_id, warning)
table holding each manga's deduped warning union, maintained by triggers that
recompute an affected manga's set (delete+reinsert of the tiny five-label set,
robust across page/chapter/manga cascade deletes), backfilled in the migration.
The filter is now a single indexed lookup.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Uploaded chapters have no source_index, and ORDER BY source_index DESC NULLS
LAST sorted every one of them after all crawled chapters regardless of number
— an uploaded chapter 5 landed after crawled chapter 100. Slot each uploaded
chapter by number, just before the crawled chapter with the smallest greater
number, so it interleaves. Crawled rows keep their reversed source-DOM order,
preserving the deliberate placement of non-numeric entries (migration 0021).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Re-adding an already-bookmarked manga returned 409, which the UI surfaced as
a false "Could not add bookmark" toast — collections are idempotent, bookmarks
weren't. On a unique violation, fetch and return the existing bookmark, and
answer 200 (vs 201 for a fresh insert). The frontend toggle already treats a
successful POST as done, so the false toast disappears with no client change.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
try_acquire used .expect on lock(), so a single panic while holding either
limiter mutex would poison it and make every subsequent auth request re-panic
— one blip became a persistent auth outage. Recover the guard with
unwrap_or_else(|e| e.into_inner()) so the limiter keeps serving.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Heartbeat renew failures were only logged, so under DB flakiness a job could
be leased twice concurrently — the original worker kept crawling past a lapsed
lease while another re-leased it. Track consecutive renew failures and, after
MAX_HEARTBEAT_RENEW_FAILURES, signal the worker (via a CancellationToken) to
abandon the in-flight dispatch instead of working a job it no longer owns.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
An idle worker re-issued a row-locking lease query every 1s per worker with
no backoff. Replace the flat sleep with idle_backoff: 1s, 2s, 4s, … capped at
30s, reset to 0 the moment a job is leased, so a quiet daemon stops hammering
the crawler_jobs table.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A job's attempt is incremented at lease time, then the dispatcher acquires
the headless browser. When the browser was down or mid-restart, acquire()
failed and the error was ack_failed — so during an outage the whole pending
backlog was chewed to `dead` within minutes while workers hot-looped.
Add SyncOutcome::BrowserUnavailable: the dispatcher returns it instead of
propagating the acquire error, and the daemon releases the lease (refunding
the attempt, like the cancel/session paths) and backs off, so an outage is
treated as an infrastructure blip rather than per-job failures.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The cover multipart path called Field::bytes(), buffering the entire field
into memory before parse_image checked max_file_bytes — an oversized cover
was fully allocated before rejection. Add read_capped/push_capped, which
reject with 413 as soon as the running total exceeds the cap (the offending
chunk is never copied in), and route both cover handlers and stage_image_part
through it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Commit 134ab54 dropped the safe DNS resolver for any configured proxy, but
that guard is only pointless for SOCKS proxies (where the proxy resolves the
target). On an http(s):// proxy reqwest still resolves the target locally, so
a hostname resolving to a private IP would be reachable. Narrow the exemption
to SOCKS-only via should_attach_safe_resolver.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The CDP Fetch interceptor only paused main-frame Document requests, so a
scraped page's <img>/fetch()/XHR subresources to internal targets
(169.254.169.254, postgres:5432, RFC1918) reached those hosts unguarded.
Drop the ResourceType::Document constraint so every request runs through
is_blocked. Also make open_page fail-closed: if the guard can't install,
close the blank page and return the error instead of navigating unguarded.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ff4ca96 wired SafeResolver (drops any host resolving to a private IP) into
all four crawler/analysis reqwest clients. That broke every proxied fetch:
with a socks5h:// proxy the target is resolved by Tor, so reqwest only ever
resolves the proxy's OWN host — `tor`, on a private Docker IP (172.x) — which
the resolver then rejects ("SOCKS error: failed to create underlying
connection"). 100% of crawler image downloads failed. The vision clients
broke the same way against `mangalord-vision` (latent; analysis was off).
Fix: attach the resolver only on the crawler's direct (unproxied) path, where
it genuinely guards DNS rebinding. Behind a proxy it adds no protection (Tor
can't reach internal ranges and the target is invisible to reqwest) so it's
omitted. Drop it from the vision clients entirely — that endpoint is an
operator-configured internal service that must resolve to a private IP.
Security posture is >= the pre-ff4ca96 baseline (which had no resolver at all).
is_private_ip / retain_public_addrs and their tests are unchanged.
Bump 0.124.13.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A single global token bucket let one attacker at the sustained rate 429
every user's login/register/change-password. Key buckets by client IP: the
SvelteKit proxy now stamps the real peer address onto X-Forwarded-For
(overriding any client-supplied value, anti-spoof), and axum reads it via a
ClientIp extractor — but only when AUTH_TRUSTED_PROXY is set, else it falls
back to the shared bucket (today's behavior). The per-IP map is bounded
(10k IPs, idle buckets pruned) so a spoofed-IP spray can't grow it.
AUTH_TRUSTED_PROXY defaults false (safe for a directly-exposed backend);
compose sets it true since the proxy is the single trusted hop.
Bump to 0.124.12.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The reqwest DNS resolver can't see Chromium navigations, so a scraped page
that redirects the browser to an internal target (302 -> 127.0.0.1:5432, or
a rebinding hostname) was still loaded. Add crawler::intercept: an opt-in CDP
Fetch guard that re-validates every Document navigation/redirect through the
same ensure_public_target + resolved-IP check, failing internal targets.
Gated behind CRAWLER_SSRF_INTERCEPT (default OFF): enabling Fetch is a fragile
critical-path hook and the CDP wiring is not CI-verifiable (no Chromium in
CI). When off, open_page is byte-identical to browser.new_page; when on, a
guard-install failure falls back to an unguarded navigation rather than
wedging the crawl. The pure decision logic (verdict/is_blocked) is unit-tested;
an #[ignore] smoke test covers the CDP path. Validate with a manual crawl
before enabling in production.
Bump to 0.124.11.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The SSRF guard only checked the host string, so an attacker-owned domain
resolving to 169.254.169.254 / 10.x (DNS rebinding, TOCTOU) bypassed it.
Add a reqwest dns::Resolve (SafeResolver) that drops resolved addresses in
private ranges, wired into all four crawler/analysis clients — it fires per
connection so it also covers redirect hops. Also close the is_private_ip
IPv6-embedding gaps (IPv4-compatible ::/96, NAT64 64:ff9b::/96, 6to4
2002::/16 all now unwrap to the embedded IPv4).
Bump to 0.124.10.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The immutable Cache-Control added for the reader preload work was
`public` unconditionally. Under PRIVATE_MODE, /files is auth-gated, so a
shared cache/CDN would store the blob and serve it to unauthenticated
clients, defeating the gate. Emit `private, ...` when private_mode is on,
`public, ...` otherwise.
Found in the security audit; regression from 0.124.3. Bump to 0.124.9.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The new DB_MAX_CONNECTIONS / DB_ACQUIRE_TIMEOUT_SECS knobs were read by the
backend but never added to .env.example or docker-compose.yml, so an operator
setting them in .env would see no effect (caught by the config meta-tests).
Document both in .env.example and wire them into backend.environment.
Bump to 0.124.8.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
GET /mangas recomputed count(*) over FILTER_WHERE (up to five correlated
subqueries) on every page, so pagination cost scaled with catalog size. The
total doesn't change as a caller walks pages, so compute it once at offset 0
and return null thereafter (the envelope already serialises total as
Option<i64>; the frontend types it number|null and the library route ignores
it).
Bump to 0.124.7.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The OCR and vision dispatchers read the whole page image into a Vec via
storage.get() and only then checked max_image_bytes, so the cap couldn't
bound the read. Stream via get_stream() + safety::accumulate_capped so the
read bails as soon as the running total exceeds the cap.
Bump to 0.124.6.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
new_page() opened a tab that was only closed on the happy path; a `?` on
wait_for_nav / content() leaked it (chromiumoxide doesn't close on drop),
so errored fetches accumulated tabs in the shared browser. Add a generic,
unit-tested `close_after(close, body)` helper and wrap the three fetch
sites (chapter content, session probe, source navigate) so the tab is
closed on every exit path — including navigate's content()-read branch,
which previously leaked.
Bump to 0.124.5.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The single Postgres pool (shared by HTTP handlers and the crawler/analysis
daemons) hardcoded max_connections=10 with no acquire_timeout, so under
saturation callers hung on the driver's silent 30s default. Add a DbConfig
(DB_MAX_CONNECTIONS default 20, DB_ACQUIRE_TIMEOUT_SECS default 10, clamped
to >=1) and apply both in PgPoolOptions.
Bump to 0.124.4.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Page and cover blobs are content-addressed by unguessable, immutable keys,
but /files sent no Cache-Control, so the browser refetched on every reader
open — defeating the whole-chapter and next-chapter preloading (and re-
proxying every byte through the node server in prod). Send
`Cache-Control: public, max-age=31536000, immutable`.
Bump to 0.124.3.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Continuous mode rendered every page (up to 2000) with loading="eager",
fetching and decoding the whole chapter up front — enough to OOM a mobile
tab on a long chapter. Eager-load only a leading window ([0..=eagerThrough],
covering the deep-link target + a lead, min 6 pages) and lazy-load the rest;
the browser fetches them as the reader scrolls near. Unloaded lazy pages
reserve height (page width x ~1.4) so they stay distinct scroll targets for
the progress IntersectionObserver, cleared on load.
Keeps the DOM dense so the scroll-to-?page=N effect, progress observer, and
next-chapter preload trigger are all unchanged. Rewrote reader-preload-window
spec to assert the eager/lazy attribute partition + deep-link scroll landing.
Bump to 0.124.2.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Argon2id hashing (~15-50ms, 19 MiB) ran synchronously inside async
handlers, stalling every task sharing those runtime worker threads under
concurrent auth load. Add spawn_blocking wrappers hash_password_async /
verify_password_async and switch all async call sites; sync primitives
stay for the login timing-equaliser and unit tests. Bump to 0.124.1.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Eager-load every page of the current chapter rather than a bounded
rolling window, so pages are fetched up front (the browser paces them
over its per-origin connection limit, top-to-bottom = reading order)
instead of filling in as they're scrolled onto. This drops the
scrollLeadIdx/eagerThrough machinery in favour of a plain
`loading="eager"`.
The first few pages of the *next* chapter are already warmed when the
reader nears the end of the current one (the next-chapter preloader), so
flipping over stays instant.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Continuous mode eager-loaded only pages 0..=initialIndex; everything
below stayed `loading="lazy"`, so pages fetched and reflowed in as they
scrolled into view. Now a rolling window eager-loads a few pages ahead of
the reading position (PRELOAD_AHEAD) so they arrive — fetched and sized —
while still below the fold, then swap in without pop-in.
The window rolls forward off `scrollLeadIdx`, the top-of-viewport page
measured from live rects and guarded by `scrollY`. It deliberately does
not use the read-progress high-water mark: before images load they are
0-height and pile into the viewport, which would latch that mark to the
last page and eager-load the whole chapter on open. Pages past the window
stay lazy so opening a long chapter doesn't fetch every page at once.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The streamed loaders on bookmarks, collections, library, search, and the
collection detail page swallowed only ApiError into their in-band `error`
field and re-threw anything else. A raw network failure (a non-ApiError
TypeError from fetch) therefore escaped the `{#await}` to SvelteKit's
generic error page. On the collection detail page a content-load failure
showed a misleading "this collection is empty", and on search a failed
results query showed a false "no matches" because the streamed `error`
was never rendered.
Fold non-ApiError failures into the same in-band `error` so every
transient blip renders as an inline message, add a `contentError` branch
to the collection detail grid, and render the search results `error` in
both result views.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The detail loader now streams its six-call bundle instead of blocking
navigation on it. The page component is a thin {#await} wrapper: it shows a
MangaDetailSkeleton (cover + meta + chapter rows) while the bundle loads,
renders the extracted DetailView once resolved, and shows an inline
not-found on a 404 (previously the framework error page). Extracting
DetailView also makes it remount per navigation, so its tag/bookmark/reaction
state re-seeds cleanly on manga-to-manga moves.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The library loader streams its one-shot bundle (also the auth gate), so the
active sub-tab shows a skeleton while it loads — a grid skeleton for the
Collections tab, a row skeleton for Bookmarks/History/Page-tags — instead of
only the global nav bar.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both loaders now stream the whole result (the single fetch is also the auth
gate) so the page shows a skeleton while it loads instead of the global nav
bar only: a ListRowSkeleton on bookmarks and a MangaGridSkeleton on the
collections grid, resolving to the sign-in / error / empty / list branches.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The search loader now awaits only the chip cloud (its auth gate) and streams
the view-specific results, so switching tag / view / sort / text shows a
SearchResultsSkeleton in place of the frozen previous list until the query
resolves.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
submitTag now shows the chip immediately with the typed name and reconciles
it with the server's (normalized) ref on success, rolling back on error —
matching the already-optimistic tag removal instead of blocking on the
round-trip.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both loaders now stream their manga grid instead of blocking on it: the cheap
header renders immediately (and still handles 404/401 in the loader) while a
MangaGridSkeleton stands in until the grid resolves, then swaps in without a
layout shift. Collection detail seeds its optimistic-removal state from the
stream via a guarded effect.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Continue-reading / Recommended shelves were fetched after the catalogue
and rendered only once populated, so they popped in late and shoved the grid
down on every authenticated home visit. Now the shelf fetch fires
concurrently with the catalogue and a reserved ShelfSkeleton holds the space
for logged-in users until it resolves.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
