As of v1.236.0 the scanner ships 94 adapters — 89 EN-region + 5 RU.

Keyword prefixes (v1.227.5). title_filter / content_filter entries default to case-insensitive substring matching, so a bare negative intern also rejects "International Product Manager". Two opt-in prefixes narrow one entry without changing that default: word:intern matches the whole word only, and stem:agent must start a word and may continue (separating "Agentforce" from "Reagents"). Boundaries are Unicode-aware, so they behave the same in accented and Cyrillic titles. A bare word: / stem: is treated as a typo and matches nothing — as a negative, an empty pattern would veto an entire scan. web-ui ignored both prefixes until v1.227.5, matching them as literal text. All scanning is in-process and zero-token: adapters hit public job-board APIs/feeds directly over HTTPS; no LLM is involved and no external script is spawned.

Adding an EN board touches two registries:

1. Source meta registry — server/lib/sources/<slug>.mjs with an export const meta = { value, label, region, configKey? } block. server/lib/sources/registry.mjs readdirSync-scans the folder at boot and dynamically import()s every *.mjs (P-14, v1.69.0), so dropping a valid file auto-registers it (malformed meta → skipped with one console.warn). This drives GET /api/scan/sources and the #/scan dropdown.
2. Fetch-walk registry — server/lib/portals/adapters/<slug>.mjs plus its entry in server/lib/portals/registry.mjs::ALL_ADAPTERS (asserted === 89 with the exact sorted id list by tests/adapter-registry.test.mjs). Adapters implement matches()/buildEndpoint() (string-or-null contract) and the fetch/parse walk over tracked_companies: in portals.yml.

RU sources additionally need a RU_DISPATCH row in server/lib/ru-scanner.mjs, keyed by the meta's configKey.

Kind: board-wide = one global feed searched by keyword · per-tenant = per-company slug/board (driven by tracked_companies:) · single-company = one fixed employer's careers backend.

Querying the results (v1.228.2). GET /api/scan-results?q= matches search terms independently, combined with AND — менеджер продукта also finds "Менеджер по развитию продукта", and word order does not matter. It was a single whole-phrase includes(q) until v1.228.2, which found only the exact adjacency: продакт менеджер returned 32 rows where 162 carried both words. That is worth knowing when a search looks thin — job titles vary in wording far more than expected, so a short distinctive stem (продукт, product) beats a two-word phrase.

Telegram is the one source configured outside tracked_companies, because a channel is not an employer — the scanner filters that list through detectApi(), and a channel has no careers URL to detect. expandTelegramChannels() in server/lib/en-scanner.mjs expands the block into ordinary adapter-selected entries, so quarantine, filters and dedup all run on one scan path rather than a parallel one.

telegram_channels:
  enabled: true
  max_posts: 100          # default cap per channel (hard cap 300)
  channels:
    - { name: "PHP jobs", channel: rabotaphp }
    - { name: "Salary PM", channel: salary_pm, max_posts: 50 }

channel: accepts rabotaphp, @rabotaphp, https://t.me/rabotaphp, or a link to a single post. Each channel is read from its public web preview at https://t.me/s/<channel> — plain server-rendered HTML, no bot token and no API key. (Telegram publishes no RSS, and the Bot API cannot read a channel your bot does not administer, so neither obvious route works.)

Three behaviours are deliberate and should not be "fixed":

• A channel post is prose, not a job record. No structured fields exist. Separating real postings from ads and digests is your title_filter's job — the same one already tuned for other sources. These are leads, not listings.
• The company is never guessed. With no explicit label the row is attributed to the channel (@rabotaphp); a wrong employer would enter the tracker as fact.
• An empty parse is an error. t.me answers a private or missing channel with a redirect; returning zero rows would read as a quiet day and hide a typo in the config forever.

Parsing anchors on the three things a Telegram redesign is least likely to move: data-post="<channel>/<id>" (which both windows the page and carries the post's only stable URL), the message-text container, and the ISO <time datetime>.

Dispatched by ru-scanner.mjs::RU_DISPATCH and configured under russian_portals: in portals.yml (sources: ["hh", "habr", "trudvsem", "getmatch", "geekjob"] is the default).

The full walkthrough lives in the in-app help guide §17 ("How to add a new job-portal source"). Summary:

1. Create server/lib/sources/<slug>.mjs exporting meta = { value, label, region: 'en'|'ru', configKey? } and a search<Name>(query, opts) returning normalized job objects: { id, title, company, url, salary, location, isRemote, workplaceType, relocates, date, snippet, source }. It auto-registers at boot — no registry edit.

2. EN boards: add server/lib/portals/adapters/<slug>.mjs implementing the adapter contract — matches(url) and buildEndpoint(...) (must return a string or null, never an object) — and register it in server/lib/portals/registry.mjs::ALL_ADAPTERS. Update the count assertions in tests/adapter-registry.test.mjs and tests/scan-sources-endpoint.test.mjs.

3. RU sources: add a RU_DISPATCH row in ru-scanner.mjs.

4. SSRF host-pinning (mandatory): every adapter pins its hostname with an anchored regex (e.g. /(^|\.)justjoin\.it$/i — so justjoin.it.evil.com can't spoof) and rejects http:. Adapters with no stable vendor host (e.g. Radancy) validate the endpoint path shape and are selected only via an explicit provider: key in portals.yml. Parity tests must cover: evil host rejected, http: rejected, suffix-spoof rejected, parse happy path, pagination stop condition.

5. Host-controlled ids go through _safe-url.mjs (since v1.234.0): encodeURIComponent throws URIError on a lone UTF-16 surrogate, and a JSON payload can carry one — inside a parse loop that aborts the loop and loses the whole page. Build the job URL with safeEncodeURIComponent(id) and drop just that posting on null. Config-derived values (portals.yml segments, locale codes) keep the strict encode so a config bug fails loud. tests/sources-url-encoding-surrogate.test.mjs enforces it: a new source with a bare encodeURIComponent on a url line fails CI unless it is allowlisted there with a reason.

6. Identity in a URL fragment must be promoted (since v1.236.0): the dedup key (server/lib/url-key.mjs) drops fragments, because almost all of them are presentation-only. If a board puts the posting id ONLY in the hash — MokaHR's …/{tenant}/{id}#/job/{n} is the case in our registry — every job on that tenant collapses to one key and the extras vanish silently (from the pipeline, and from the scanner as "already seen"). promoteKnownFragmentIdentity recognizes #/job/{id} and #/jobs/{id}; a board with a different hash shape needs a case added there and a test, not a change to what the key drops.

7. A locale prefix in a detail path must be keyed on the HOST (since v1.237.0): Jobstreet/SEEK is one adapter over many market hosts, and only the Indonesian ones serve /id/job/<id> — /id/ is a locale prefix, not part of the path. Every other SEEK host (www.seek.com.au, www.seek.co.nz, sg./my.jobstreet.com, hk.jobsdb.com) answers 404 on it. Because those hosts were already in the allowlist, the scan returned their postings and every one pointed at a dead page — validation passes, the link is still broken. ID_LOCALE_HOSTS + jobDetailPath(origin) now decide per host, and an unparseable origin falls through to the common path rather than throwing. Two of our own tests had encoded the bug by asserting the 404 form; when a fix makes an existing assertion fail, check which of the two is wrong before "repairing" the test.

8. A short page is not the end of a list (since v1.237.0): Oracle Cloud filters rows server-side, so it serves fewer rows than the limit in the middle of a board. Measured on American Express: TotalJobsCount 454 over pages of 200 / 199 / 54 — treating the 199 as the end dropped the last 54 postings, 12% of the board. An empty page always ends the walk; when the tenant reports a total, page until offset + PAGE_SIZE >= total; only a tenant that reports no total may stop on a short page. This follows the wider convention that an API "may return fewer results than the number requested … even if not at the end of the collection" (Google AIP-158). Note the consequence: once short pages stop ending the walk, the page cap is the only backstop against a tenant reporting a bogus total, so it needs its own test.

9. Config-driven providers: POST/GraphQL/paginated boards thread opts.company through buildEndpoint; keep buildEndpoint returning a string.

10. Add a parity test suite (tests/sources-parity-*.test.mjs, stubbed transport — no network), bump the registry counts, and fan out docs ×17 per the Release Process.

• Entry point: GET /api/stream/scan?source=ats|regional|both (SSE, one done per phase, final:true on the last).
• dryRun=1 skips writes to data/scan-history.tsv / data/last-scan.json; company=<slug> narrows the ATS walk.
• Results are deduped against data/scan-history.tsv, sanitized (scan-sanitize.mjs), quarantine-filtered (scan-quarantine.mjs), location-filtered, and matched against your target roles / two-pager (◎ badge).
• Client disconnect aborts in-flight upstream fetches (AbortSignal).

Since v1.119.1 the #/scan Source dropdown's offline fallback (FALLBACK_SOURCES in public/js/views/scan.js) is gated by tests/scan-fallback-sources.test.mjs — exact value+label parity with the source registry, so a new adapter can't ship without the client list.