dev.sase.tr and sase.tr both run NODE_ENV=production (and ship the same
production Vite build), so every Sentry event — api, worker, and browser — was
tagged environment=production, making dev errors indistinguishable from prod.
Resolve the environment from the canonical prod host instead: COOLIFY_FQDN /
BETTER_AUTH_URL on the server (the same signal isCatalogBackfillEnabled uses),
window.location.hostname on the web. So dev now tags as "staging" and stays
filterable. Explicit SENTRY_ENVIRONMENT / VITE_SENTRY_ENVIRONMENT still win.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The existing CAPI only sent CompleteRegistration (signup). For a B2B funnel where
trials are cheap (~25 TRY) but paid is rare (~1% of trials), the highest-value
signal Meta can optimize on is the realized-revenue Purchase. Add
MetaCapiService.sendPurchase and fire it from activateSubscription — the shared
chokepoint for BOTH Stripe (webhook) and EFT/manual activation — so all paid
revenue is sent regardless of method. Hashed-email Advanced Matching (no browser
fbp/fbc in the webhook); event_id = purchase_<subscriptionId> dedupes a browser
Purchase. Awaited so it ships before the short request returns; fail-open.
This is the "teach Meta to find payers, not end-users" lever from adsOpt.md Phase 0.
Still gated on activating CAPI in prod (merge + META_CAPI_PIXEL_ID/ACCESS_TOKEN env).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The subscription-expiry cron filtered status="active" only, so trials
never transitioned out of "trial" after end_date. Every access gate
keys off status, so trials past end_date kept full product access for
free (revenue leak) and inflated the active-trial count. On prod: 105
stuck trials, 3528 brand grants still live.
- expiry processor now sweeps status IN (active, trial) past end_date
(lt() still skips NULL end_date, so perpetual subs are untouched)
- add "trial" to SubscriptionStatus union — it was used in the DB and
code but missing from the type (both subscription.ts and user.ts)
Proven read-only on prod: old WHERE caught 0, fixed catches 105.
Revocation uses the existing set-expired + delete-userBrands path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mock @sentry/nestjs and assert that a getCategoryWithParts loadError drives the
redis dedup key + Sentry.captureMessage("…drill-load-error…"), so the silent-
failure reporting can't regress unnoticed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The catalog failures that hurt UX — a drill/parts fetch that fails into an empty
"couldn't load" panel, or a decoded vehicle whose category tree comes back empty
("model var ama parça yok") — all return HTTP 200 with a degraded body. Nothing
throws, so the global Sentry exception filter never sees them and they go
unnoticed (serkan's complaint was exactly this class). Report them explicitly.
- new common/catalog-degradation.ts: reportCatalogDegradation(kind, ctx),
fingerprinted by kind+source+brand so each failure mode collapses into one
countable Sentry issue (e.g. "drill-load-error · pl24/Ford — N events, M users").
- categories.service: capture on getCategoryWithParts loadError and on an empty
getCategoryTree, Redis-deduped to <=1 event/hour per category/vehicle so a
broken catalog can't flood the stream; telemetry never throws into the request.
tsc + biome clean, categories suite 10/10.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
payment_success / payment_failed / subscription_activated are captured inside
the Stripe webhook handler and activateSubscription — short requests that return
immediately. posthog-node's fire-and-forget flush was abandoned before the send
completed, so these events were written to the DB but never reached PostHog
(DB had 4 completed Stripe payments in 30d; PostHog had 1 payment_success and 0
payment_failed). payment_initiated, fired in a normal user request, landed fine —
which is what isolated the cause to the webhook/short-request context.
Add PostHogService.flush() and await it at the end of handleWebhook and after the
subscription_activated capture in activateSubscription (the shared Stripe+EFT
chokepoint). Restores server-side paid-conversion visibility so trial→paid ROI is
measurable in PostHog instead of only the DB.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
isLeafLinkPath flagged EVERY pcat node as a leaf (any pcat: linkPath), so the
backfill worker fetched a mid-group folder's "parts" — a 1-level drill that
only revealed its sub-groups — and never recursed. Deep pcat trees (whose top
level is ~100% folders) were left one level shallow: no parts seeded, so the
new tree part-counts stayed 0 and the cross-tree catalog search (DB-only) found
nothing on a freshly decoded vehicle. Exactly why serkan's Ford Mondeo looked
like an empty catalog.
Use the captured hasSubgroups flag: a pcat node is a parts leaf only when it is
NOT a known parent group. Folders now queue a children job and the existing
recursion (queueCategoryJob → processChildren → getChildren) drills to full
depth (MAX_DEPTH=5), seeding parts at every leaf. The rolling backfill rescan
warms existing shallow vehicles wave by wave; new decodes warm deep via the
reactive schedulePrefetch. Rate limiter + PCAT_PACE + backlog guard keep it
bounded (all env-tunable to ramp).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
pcat/emex vehicles decode into a tree whose top level is ~100% folders with
zero parts (6746/9458 root nodes are parents, 0 direct leaves across 477
vehicles). Users — serkan filazi among them — saw an opaque wall of identical
folder cards, assumed the parts feature was missing, and bounced without ever
drilling (his session: 3 view-mode toggles, 0 category clicks; his data was
2-3 clicks deep and fully reachable).
Two changes, no upstream/proxy cost:
- getCategoryTree now rolls up a subtree partCount per node (one grouped parts
read + JS post-order sum). The grid already renders "N parça", so folders
now read "Fren sistemi · 142 parça" and steer drilling toward the branches
that actually contain parts; empty branches stay quiet. Counts reflect
drilled data and grow as branches are visited.
- A one-line drill hint on a "cold" tree (nothing drilled yet → no counts) so
first-visit users know parts live inside the subcategories instead of reading
the folder wall as a broken/empty catalog. Drops away once any branch is hit.
Verified: subtree roll-up unit test + existing suite (10/10), tsc + biome clean
across api/web/shared.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The Meta Pixel (fbevents 2.9.x) dispatches ALL events from a Web Worker created
off a blob: URL. With no worker-src directive, the browser fell back to
script-src (which lacks blob:), so the worker was CSP-blocked and NO pixel
events fired — PageView, CompleteRegistration, everything were 0 in a live
headless test (even a manual fbq('track','PageView') emitted nothing).
connect-src/img-src already allowed facebook.com, so this was the sole
remaining client-side blocker. Adds worker-src 'self' blob: (also unblocks
Remotion's blob worker on the landing demos).
Note: CAPI (server-side) already covers the signup conversion regardless; this
restores the browser pixel for PageView + retargeting/audience events.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
VIN-decoded Peugeot/Citroën scope categories ("mekanik", "kaporta", …)
carry json-vin-main-groups.action / json-vin-illustrations.action link
paths, but getCategoryWithPartsInner's isPsaParent gate only recognised
the catalog-browse paths (psa:: / json-illustrations.action). So every
VIN PSA parent fell through to the leaf path, fetched no parts, and
rendered an empty panel — the largest current "0 parça" cluster
(~20-30 empty parts_panel_viewed/day across all Peugeot/Citroën models;
serkan filazi's complaint among them).
The drill already works end-to-end (getChildren → fetchSubGroupsByPath →
fetchVinMainGroups/fetchVinIllustrations → vin-image-board parts); only
the gate was out of sync with PL24Service.isPsaVin*Path. Verified live on
prod: scope "mekanik" → 6 main-groups → "motor" → 10 illustrations →
12 OEM parts + schema + 15 hotspots.
Also surface a retryable loadError (matching the pl24/emex group-node
branches) when a PSA parent drill comes back empty, instead of a
misleading empty grid. +2 regression tests pin the dispatch to the
PL24Service.isPsaVin* helpers so it can't silently drift again.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The browser pixel under-counts signups badly: ~96% of paid traffic is mobile
in-app browsers where iOS ITP / ad-blockers drop client events, and the OAuth
path never fired it reliably. Meta recorded ~0 registrations for a 7.5K-spend
campaign while PostHog saw 98 facebook signups — so Meta could neither optimize
toward nor attribute signups, which is the main driver of the low signup rate.
This adds a server-side CAPI CompleteRegistration:
- MetaCapiService + @Global module. Fail-open: no-ops unless META_CAPI_PIXEL_ID
+ META_CAPI_ACCESS_TOKEN are set; never throws (signup must not break).
SHA-256 hashed email + fbp/fbc/IP/UA.
- Fired from the better-auth user.create.after hook for ALL signups (reliable,
covers Google OAuth which the browser pixel missed entirely).
- A session-gated POST /analytics/meta/complete-registration endpoint adds
fbp/fbc/IP/UA (ad-click attribution) for the email path.
- The browser pixel now passes a shared event_id (signup_<userId>); the
premature Google client-pixel fire (fired on click, before completion) is
removed.
- All sources dedupe via event_id=signup_<userId>.
Activate by setting META_CAPI_PIXEL_ID + META_CAPI_ACCESS_TOKEN (Events Manager)
in the api env; META_CAPI_TEST_EVENT_CODE routes to Test Events for verification.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
So a non-prod env can hold the (public) project key and evaluate flags - kill
switches, rollout - WITHOUT shipping analytics events to the shared prod
PostHog project. The posthog-node client (which powers both capture and local
flag eval) now initializes whenever the project key is present; capture() is
gated on POSTHOG_CAPTURE_ENABLED (default on; set "false" on dev). Wired into
the api + worker compose blocks.
This lets dev.sase.tr exercise the kill switches end-to-end while keeping the
prod analytics project clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add server-side PostHog feature-flag evaluation to the API and wire three
high-leverage uses. The flags live in PostHog (project 127747), dormant and
fail-open, so this commit is a no-op until POSTHOG_PERSONAL_API_KEY is set and
a switch is flipped.
Phase 1 - upstream kill switches:
PostHogService gains local flag evaluation (onlyEvaluateLocally +
sendFeatureFlagEvents:false -> zero decode latency, no VIN leak) and
isSourceLive(). Guards at each source's decode entry (parts-catalogs, emex,
pl24 whole-source + per-brand via LEGACY_ARCH_SOURCE_TAG) let a flailing
upstream be disabled from the PostHog UI in ~5s instead of a
code-fix -> dev -> prod redeploy. Fail-open: any unresolved flag keeps the
source live, so a PostHog outage can never black out decoding.
Phase 2 - guarded rollout primitive:
isEnabled()/variant() with VIN bucketing, ready to ramp a new decode/parser
path 0->100% behind a decode-*-v2 flag (recipe in feature-flags-strategy.md).
Phase 4 - remote-config ops tuning:
prefetch-worker reads cfg-backfill-tuning to retune backfill
batchSize/maxBacklog/businessHoursOnly live; malformed/missing -> the
compiled-in constants.
POSTHOG_PERSONAL_API_KEY wired into the api + worker compose blocks (empty ->
flags inert, no added latency). Tests updated for the new constructor params.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two visible defects when a user picks a Ford model (variant selector step):
1. **N identical buttons.** Ford's `modelFamilyToModelList` gives every sub-catCode
the same caption — just the family name. parseFordCatCodesForFamily returned
`[{code:CBV,name:Kuga},{code:CBS,name:Kuga},{code:CTD,name:Kuga}]` for Kuga
2012-2020. The UI rendered three indistinguishable "Kuga" pills with no way
for the user to tell them apart. Same on Galaxy ("Galaxy"/"Galaxy"), every
multi-catCode family.
Fix: detect duplicate baseNames per family and graft a disambiguator —
year first (friendliest), catCode as a last resort. Result:
"Kuga (2013-2016)" / "Kuga (2016-2020)" / "Kuga (CTD)" — or, when year is
present even for unique entries, always include it for a uniform look.
2. **Section title was lying.** UI label was `catalog.fordVariant.modelYear`
= "Model Yılı" ("Model Year") but for Ford LEGACY these are catCodes
(generation identifiers), not years. Rename the key value to "Varyant" /
"Variant" — semantically correct for both Ford catCodes and Volvo years
(Volvo flows through the same selector and its codes ARE years; "Varyant"
covers both).
Cleared the affected Redis cache (`pl24:ford:config:*`) on dev so the next
hit fetches fresh disambiguated data; prod cache will roll over on its own
TTL after promote.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
These three brands rendered parts with NO schema illustration (and no hotspots):
schema_pics=0 vs 56/29/19 parts on dev. Root cause: extractPsaImageTicketUrl only
read imageViewerParamsUrl from the id="jsinitparams" data-params attribute (where
Ford/PSA/Opel/Volvo put it). Hyundai/Kia/Nissan's jsinitparams carries only a
localization dict ({"commonTxt":{"ok":"TAMAM"}}); their imageViewerParamsUrl lives
in a separate <script> JSON blob → extractor returned null → image pipeline skipped.
Fix: when the jsinitparams-attribute parse yields no URL, fall back to a whole-HTML
scan for "imageViewerParamsUrl":"…" (decoding &). Brand-agnostic and additive;
Ford/PSA/Opel/Volvo unchanged (attribute path still wins). Proven live via the de
client: Hyundai 600x820/3 hotspots, Kia 600x820/11, Nissan 1024x560/13 — all download OK.
Also: drop a pre-existing noParameterAssign in fetchP4Page (derive `account` const
from `accountParam` instead of reassigning the param). +3 extractor unit tests.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fiat catalog browse returned 0 models because the generic single-endpoint P5
model-list flow can't express Fiat's two-level hierarchy and the drill defaulted
to the tr account (Fiat is licensed only on de-708171).
- fetchVehicleList: dispatch fiatp_parts/fiatt_parts to new fetchFiatVehicleList,
which expands modelOverview (34 families) → models?modelFamily=N (model codes +
year ranges) into flat catalog vehicles whose catalogPath is the maingroups
endpoint. Verified live: 34 families → 123 models, end-to-end drill to parts+image.
- fetchMainGroups / fetchP5Restrictions: account-aware (resolveAccount → de + DE
proxy for Fiat; tr unchanged for every other P5 brand) so browse maingroups no
longer hit the tr demo/empty page.
- web: case-insensitive "/maingroup" gate so Fiat (lowercase /mdl/maingroups,
already a maingroups endpoint) skips the empty restriction selector and loads
categories directly.
- formatFiatYear: "(2016,2020)" → "2016-2020"; +unit tests.
Subgroups/parts/images already resolve account→de for Fiat (unchanged).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Finding: /p5fiat/extern/vehicle/models IS the endpoint (de-licensed, demo:false) but
returns 'Cannot parse null string' regardless of manufacturer/make/brand param — needs
an undiscoverable param from the P5 SPA session (launchCatalog.do flow). Requires a real
PL24 web-UI network capture to resolve. Turkish Tofaş NM4/Egea is not in this catalog
anyway, so catalog-browse value is European-Fiat-only. Kept the generic P5 endpoint
self-discovery fallback. Fiat VIN-decode (European) is unaffected and still works.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Ford parts are licensed only on the de account. The de drill chain is:
group.action (engine/transmission restriction nodes, HTML) → json-main-group.action
(maingroups JSON) → json-sub-group.action (subgroups JSON) → image-board.action (parts).
The existing json-sub-group handler only knew the Nissan shape ({caption,code,url});
Ford de subgroups are {caption,id,subheader,jsonUrl:null,url:null} — leaves carry no
URL. Add parseFordSubGroupsJson: skip subheader rows, and for each leaf construct the
illustration URL (subGroupId = mainGroupId+subheaderId, bomId = subGroupId+leafId →
image-board.action?bomId=…), which drills to parts via the existing image-board path.
Route Ford (fordp/fordt) to de alongside Hyundai/Kia/Nissan.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tested Ford & Polestar on de: Ford returns a real variant tree but drilling can't
reach parts (needs dedicated Ford variant-drill logic); Polestar returns 0
categories (data gap). Revert them to tr; keep the verified Hyundai/Kia/Nissan
de-routing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Verified on dev: the tr account serves NOT_LOGGED_IN_DEMO for Hyundai/Kia/Nissan
parts (browse-only), while the de (account 2) subscription is licensed and returns
real data (de: 266 Hyundai / 169 Kia models, non-demo, ~88/64 drill rows). Route
every P4 fetch for these brands to de in fetchP4Page (+ resolveAccount), and ensure
the de service token is authorized. Existing tr-coded catalog_vehicles must be
re-seeded so they carry de catalog codes (ops step on deploy).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Confirmed via dev probes: the de account is non-demo (licensed) for Hyundai
parts but does NOT have the EUR/TR-region catalogs our vehicles use, so the
fallback returned non-demo-but-empty and just doubled upstream load. Root cause
of empty Hyundai/Kia/Nissan parts is PL24 licensing/region (tr account not
licensed for these brands' parts) — a commercial issue, not code-fixable.
Keeping only the explanatory comment.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
getCategoryTree cached the built tree for 3600s unconditionally, and the
read guard `if (cached)` treats an empty array as a hit. So a single
transient decode/proxy failure (emex via the flaky DataImpulse pool) seeded
0 categories and poisoned the catalog with an empty tree for a full hour,
even after the source recovered. Cache an empty tree for only 60s — it
self-heals on the next request post-recovery while still throttling
re-decode during a genuine outage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hyundai/Kia/Nissan parts-drilling returned a NOT_LOGGED_IN_DEMO page on the tr
account (license-gated), so live drills yielded 0 parts even though the tree
loads. After the tr re-auth retry still demos, try the de account once; if it
serves real (non-demo) content, use it. Tests whether de holds the brand license.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
emex's ssd tokens appear IP-bound: a rotating proxy sends each hop of a
decode→tree→drill→parts flow from a different exit IP, so emex rejects the
ssd and returns empty (the rotating-Floxy verify still seeded 0). Pin the
Floxy exit IP with a sticky session (password suffix
`_session-<id>_lifetime-300`); a rolling session id is reused across the
flow and rotates near expiry or after a transport failure (dead IP → fresh
one). lifetime via EMEX_FLOXY_LIFETIME (0 = rotating).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DataImpulse (74.81.81.81, random port 10001-10099) intermittently throws
connect-timeouts / resets, dropping real emex hits. Add a Floxy residential
fallback (residential.floxy.io:12321): fetchEmexHtml now runs a provider
schedule — 3 DataImpulse attempts (rotating port), then 2 Floxy attempts —
and only falls back on transport errors (a definitive HTTP answer like 404
still stops the schedule). On by default; endpoint/creds overridable via
EMEX_FLOXY_* env. Direct proxy-less last resort still gated by
EMEX_DIRECT_FALLBACK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ford P4 catalog navigation is session-stateful: catCode group.action returns
the full Ford model picker (30 vehicle.action rows), and vehicle.action?modelFamily=X
loops back to the same picker; a bare group.action after select still returns the
picker (parsedGroups=0). Stateless GETs over a rotating proxy can't hold PL24's
server-side model selection → groups are unreachable (explains 0/57 categories_fetched
on prod). Needs session-continuity or browser-based navigation — tracked, not a quick fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
emex catalogs were seeded from the flat QuickGroups taxonomy (~351
unsearchable leaves on one page). Switch the seeder to the Vehicle.aspx
OEM-microfiche tree (top groups -> sub groups -> Unit.aspx leaves), which
also reaches units that have no QuickGroups gid (e.g. the torque
converter, OEM 5205547 on the Fiesta).
- emex-tree.parser.ts: parseVehicleTree / parseUnitLeaves — pure,
GET-based (per-href ssd, no browser/postback). Unit-tested vs fixtures.
- emex.service: fetchVehicleTree + drillVehicleNode; all 3 decode methods
now populate categoryTree; Unit.aspx leaves fetch parts directly.
- categories.service: revive the dormant tree-insert branch with group
markers (linkWid="emex-group" / hasSubgroups), add an emex getChildren
lazy-drill branch + emex-group parent guard, tighten leaf detection,
disambiguate colliding sub-group names.
- prefetch: emex Vehicle.aspx = parent to drill; Unit/QuickDetails = leaf.
No DDL and no frontend changes — the generic tree UI renders it like
pl24/pcat. Existing flat emex categories re-seed lazily once deleted.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosis captured: Ford catCode group.action returns a vehicle-selection
page (30 vehicle.action rows, no group rows) — needs an extra selection
level before groups. Polestar json-model-years returns an HTML landing page,
not JSON — needs a different navigation flow. Both tracked for follow-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
fetchVehicleList fell back to /extern/vehicle/modelfamilies for any backend
not in BACKEND_MODEL_PATH; p5fiat (Fiat) isn't mapped, so it returned 0
models and Fiat seeded nothing (0 catalog_vehicles on dev+prod). When the
primary path yields nothing, try the other known P5 listing endpoints and
use the first that returns models, logging which one worked so it can be
pinned. Only runs on the empty path → mapped backends unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When fetchSubGroupsByPath returns only nav-crumb stubs (e.g. a "Portal"-only
page for a Hyundai BODY node), the post-filter set is empty and drizzle
.values([]) threw → HTTP 500. Compute `unique` unconditionally and only
insert when non-empty; otherwise fall through to the leaf/parts path for a
clean empty node.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
getCategoryWithParts returned persisted child categories without the
nav-crumb filter, so a node whose subgroups were polluted with "Portal"
(e.g. Hyundai BODY → only child "Portal") drilled to a dead junk node.
Drop nav-crumb children, delete them from the DB so the node re-fetches its
real subgroups, and also skip nav-crumbs when persisting freshly-fetched
subgroups. Real subgroups (illustration codes) and parts are reached as
before once the junk is gone.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Nissan's "Repair & Maintenance Information" persists as
`/nissan/nissan_parts/https://eu.nissan.biz/` — the parser prefixes the
external URL with basePath, so the anchored ^https check in isNavCrumbLink
missed it. Match `://` anywhere (non-.action) instead; real PL24 linkPaths
are relative and never contain `://`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The parser filter only stops NEW junk from being persisted; vehicles whose
trees were already fetched still had "Portal"/eu.nissan.biz rows in the DB,
and persistAndBuildLegacyTree/buildCategoryTreeFromDb build from those rows.
Filter dbCategories through isNavCrumbLink before buildTree in both, so the
response self-heals for previously-persisted junk without a DB migration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
