Detect verification-code signups (e.g. "your verification code is
371404") whose only link is a mailto. These cleared the keyword
threshold but were dropped because the detector required an http(s)
candidate link. A code path now raises the flag/badge/banner when a
verification keyword sits next to an OTP-style code; the code is never
extracted or surfaced.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
email-processor parsed input.from twice — once via EmailAddress for the
native-feed base, once via the favicon infra helper extractEmailDomain
just to get the domain. CLAUDE.md forbids reaching across a layer to
parse a domain: parse once and derive both siteBaseUrl() and domain.value
from the EmailAddress VO, removing the infrastructure/favicon-fetcher import.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Delete the `iconBase` local helper (which mishandled display-name form
like `Name <a@b.com>`) and replace it with `EmailAddress.parse(input.from)
?.siteBaseUrl()` — the domain-layer VO that already handles bare and
display-name addresses correctly. Adds TEST C to lock the
display-name + relative-href absolutization fix.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wire extractFeedLinks + detectNativeFeeds into storeEmail so that RSS/Atom/JSON
feed <link> tags in the newsletter HTML are detected and stored per-sender on the
feed metadata.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Return the ranked links directly (string[] | null) instead of an unused
{score, links} wrapper, and drop the redundant hasKeyword helper in favor
of matchesAny(_, KEYWORDS). No behavior change.
Separate the two feed identities so the public read URL never reveals the
inbound address and vice-versa:
- FeedId becomes an opaque high-entropy token (read id + KV key); MailboxId
(noun.noun.NN) owns the inbound address and the untrusted-input boundary
via MailboxId.parse. They map only through the inbound:<mailbox> secondary
index, resolved solely at reception.
- inbound index lifecycle is owned by FeedRepository: written by save/saveConfig,
dropped by removeFromList(Bulk) — symmetric, never mirrored by hand (removes the
manual delete in feed-service + the cron loop, and a silent empty-catch).
- Feed.mailboxId exposes a MailboxId VO (symmetry with Feed.id); the
mailbox@domain shape lives on MailboxId.emailAddress(domain).
- Distinguish mailbox_unknown (no feed claims the address) from feed_not_found
(dangling index) for observability; both forwardable, both 404.
- Drop the redundant EmailParser.extractMailbox pass-through so MailboxId.parse
is the single parse boundary.
Docs (README/INSTALL/CLAUDE.md/landing) and tests updated; 439 tests green,
tsc clean, build dry-run OK.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Batch of four reader-facing improvements (TODO "Compat lecteurs + dedup"):
- JSON Feed at /json/:feedId (feed lib .json1()); all formats cross-link
- OPML export at /admin/opml (admin-protected; the registry lists every
feed URL, so it must not be public)
- Conditional GET on /rss + /atom: strong ETag + Last-Modified, 304 on
If-None-Match/If-Modified-Since, validators shared via http-cache.ts
- Duplicate-send dedup in ingestion: match by Message-ID, fall back to a
SHA-256 of normalized subject+content; a duplicate is a no-op and bumps
the new emails_deduplicated counter (status page + /api/v1/stats)
429 tests green, tsc clean, build dry-run OK. Docs (README/CLAUDE/TODO +
landing cards) updated.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Inline images (referenced by src="cid:…") are now classified at ingest and
kept out of the downloadable attachment lists, RSS/Atom enclosures, and the
API — while still stored in R2 and cleaned up with the email. Fixes the admin
email preview, which injected raw HTML into the data: iframe so cid refs never
resolved; it now rewrites them to absolute /files URLs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Move four DDD tensions on the Feed aggregate to ground:
- #1 The aggregate now holds a domain FeedState (camelCase) instead of the
snake_case FeedConfig DTO; infrastructure/feed-mapper.ts owns the
FeedState<->FeedConfig/FeedListItem translation as the sole snake_case site
outside the HTTP edge.
- #3 Replace the edit() recomputeExpiry control flag with a Lifetime VO:
passing a lifetime recomputes expiry, omitting it preserves the current one
(the dashboard quick-edit path).
- #4 Domain events carry their own feedId; dispatchFeedEvents centralizes the
drain+dispatch in the application layer (no more manual pullEvents at call
sites), keeping infra->application dependency direction intact.
- #6 Rename FeedId.fromTrusted to FeedId.unchecked to make the absence of
revalidation explicit.
Adds Lifetime + feed-mapper round-trip tests. 353 tests green, tsc clean,
wrangler dry-run OK. Docs (CLAUDE.md) synced.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Address five modeling tensions in one pass:
- Encapsulation: the Feed aggregate no longer exposes raw config/metadata
(a shallow Readonly still leaked mutable arrays). It now offers
intention-revealing accessors that return copies, plus
toConfigSnapshot/toMetadataSnapshot for the repository and summary() for
the global registry.
- feeds:list consistency: FeedRepository.save/saveConfig upsert the registry
entry from feed.summary(), so services no longer mirror title/description/
expiry by hand (the old add/updateInList footgun is gone).
- domain/feed.ts: drop the dead applySenderPolicy, internalise resolveExpiresAt
and trimToByteBudget into the aggregate; feed.ts keeps only the shared
isExpired predicate used by the read-model routes.
- Single edit path: remove editDetails; edit(patch, deps) is the sole config
mutation, with a systematic expired guard. Renaming an expired feed now 403s.
- FeedId flows through the application and infrastructure signatures;
fromTrusted/parse happen once at the edge, .value only at the serialisation
boundaries (urls, feed-generator, feed-keys, logs, JSON).
347 tests green, tsc clean, Worker bundle builds.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Light "collect + dispatch" variant: the Feed aggregate records FeedEvents
(FeedCreated, EmailIngested) on the mutations that have consequences, exposed via
pullEvents(). A new application dispatcher (feed-events.applyFeedEvents) maps
those events to their side effects — counters (awaited) plus WebSub pings and
favicon fetches handed to a BackgroundScheduler. This removes the inline,
scattered side effects from the ingest hot path (email-processor) and from
createFeedRecord; the aggregate is now the source of truth for "what happened".
Side effects with no aggregate mutation (rejected email, feed deletion bypassing
the aggregate, bulk admin ops, the cron, unsubscribes-sent) stay imperative by
design — there is no aggregate event for them to ride on.
BackgroundScheduler type moved to infrastructure/worker.ts (shared). CLAUDE.md
updated. 355 tests pass (+4 event tests); tsc --noEmit clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Make the domain stop depending on infrastructure ("imports point inward").
- Point 2: relocate the four KV adapters (FeedRepository, IconRepository,
WebSubSubscriptionRepository, CountersRepository) from domain/ to
infrastructure/, where the logger import is legitimate. The domain now keeps
only the pure key schema (feed-keys.ts), the Feed aggregate and value objects;
it imports nothing outward. Deliberately no hand-rolled 24-method port
interface (YAGNI without DI) — relocation alone fixes the direction.
- Point 6c: EmailParser.extractFeedId now returns a validated FeedId value
object instead of a raw string, so the most untrusted input (an inbound
recipient address) is guarded at the parse boundary and no longer round-trips
through FeedId.fromTrusted in the ingest path.
All import paths updated; CLAUDE.md source layout/KV-schema notes updated.
351 tests pass; tsc --noEmit clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Julien Herr