@lssm-tech/integration.job-sources
Governed connectors for La Bonne Alternance, French apprenticeship sites, Ashby, Flatchr, Greenhouse, iCIMS, Jobaffinity, Lever, Recruitee, SmartRecruiters, Taleez, Teamtailor, WeRecruit, Workable, Workday, and Indeed France.
The package exposes an executable capability matrix. supported means the deterministic connector path exists; live qualification remains separate. La Bonne Alternance commercial use and Indeed France are permission-blocked by default.
Direct-first acquisition routing
Connectors may resolve an immutable JobAcquisitionPlanV1 for each exact installation request. createRoutedJobSourceAcquisitionV1 always executes its priority-zero direct-fetch route first. It escalates only explicit
retryable_transport or render_required failures to a later source-qualified connection. Rate limits, authentication, policy, budget, source restrictions, cancellation, unknown failures, and ambiguous effects stop routing. Provider rotation is therefore never an access-control or rate-limit bypass.
createWebAccessJobSourceRouteExecutorV1 binds those routes to the existing Bright Data, Browserbase, Massive, Firecrawl, Oxylabs, Reworkd, and Apify adapters. Reworkd requires an exact qualified scraping-group dataset profile. Apify runs one qualified Actor, polls the same run within fixed bounds, and reads its resulting dataset without replaying the Actor. Authority resolution must return the plan's exact policy and budget references before an adapter is called. Every attempt, provider-reported request/cost usage, and selected route receives a deterministic replay receipt. An indirect provider 404 or 410 can never become job-closure evidence; closure remains bound to a direct detail recheck. The Web Access bridge also requires an injected acquisition rate gate, and the exact source rate policy, installation revision, and opaque credential reference participate in the request digest.
InMemoryJobSourceAcquisitionRateGateV1 is the conservative reference for a single worker process. It keys spacing and concurrency by installation or egress profile exactly as the source policy requires, honors cancellation, and returns a terminal rate_limited classification rather than rotating providers. It is not a substitute for a shared durable gate when multiple worker replicas serve the same source or egress identity.
jobSourceCapabilityMatrixV1 is the canonical 17-source × 7-capability truth table and never promotes deterministic fixtures into live receipts. Documented API adapters, Lever cursor ingestion, generic structured-page parsing, tenant-specific browser flows, and permission-blocked sources are reported separately; a generic parser fixture does not qualify a source-specific connector.
renderJobSourceCapabilityMatrixMarkdownV1 feeds the repository-level generated job intelligence matrix; bun run osint:job-matrices:check prevents drift.
jobSourceLiveQualificationPlanV1 and
verifyJobSourceLiveQualificationReceiptsV1 provide the separate exact-SHA live evidence plane. Partial receipt sets project passed, failed, and blocked cells without mutating deterministic truth or promoting permission-blocked sources. bun run live:lever exercises the public Lever installation through the hardened direct-fetch adapter and production connector. It requires an explicit live confirmation, complete scope, release SHA, policy authorization and expiry, site, and region. Public discovery, listing, detail, cursor, revalidation, and application discovery can pass; closure remains blocked until an observed live transition supplies real closure evidence. The command also requires the claimed release SHA to equal the current git rev-parse HEAD so a syntactically valid but unrelated digest cannot become qualification evidence.
bun run live:public-source exposes the same ceremony for any named connector: an exact source key and canonical installation URL select the existing connector, while the common runner derives discovery, listing, detail, incremental, revalidation, closure, and application receipts from the executable matrix. Missing credentials, cursors, permission, application URLs, or observed closure transitions stay explicit blockers. SmartRecruiters detail normalization accepts the vendor's live postingUrl and applyUrl shape as well as list-response ref links.
bun run live:public-batch consumes a strict, secret-free V1 manifest and produces one deduplicated receipt set and summary. The repository qualification manifest lives at
docs/evidence/osint/job-source-public-live-manifest-v1.json. Unknown fields, duplicate sources, credential-bearing URLs, permission-blocked sources, or unpaired closure evidence are rejected before network execution. Every live command also requires a clean worktree in addition to exact Git HEAD equality.
An optional live closure probe requires both a bounded external listing ID and an immutable prior-publication evidence reference. It can pass only when the same source returns a current 404 or 410 with closure evidence. An absent prior reference, ambiguous transport outcome, still-open listing, or unrelated status cannot promote closure.
The Flatchr adapter follows the documented public
/company/{slug}.json
feed and preserves nested vacancy identifiers and application URLs. The Workable adapter uses the vendor-documented public account endpoint
/api/accounts/{subdomain}?details=true. The documented Flatchr detail endpoint is used for direct revalidation and explicit 404/410 closure evidence. Flatchr cursor paths remain gated because the public feed exposes no qualified cursor or change token.
When Flatchr omits the optional apply_url, the connector derives the source-owned careers detail URL from the qualified vacancy ID. This preserves a public application entry without inventing an external destination or relaxing the canonical host boundary.
Authenticated Taleez and
Teamtailor listing/detail adapters bind bounded pagination, source-specific payloads, and opaque credential references. Their deterministic fixtures do not replace a credentialed live receipt. The managed worker requires an explicitly injected credential-aware acquisition adapter and never copies the credential reference into a generic provider payload.
JobAffinity uses the same credential boundary for its Basic-auth REST feed. Pagination accepts only the API-provided, same-origin /restapi/v1/job?pagecursor=... link; arbitrary cursor URLs and cross-origin resource links are rejected.
Taleez hasMore/page state, Teamtailor page[after], Lever skip, SmartRecruiters offset/limit totals, JobAffinity _next, and credentialed Workable paging.next are extracted into durable checkpoints. Unverified cursor paths remain gated. Flatchr and Workable bind source-specific detail URLs, direct revalidation, explicit closure evidence, and response-schema drift comparison. Workable's public account feed remains the anonymous default; its SPI pagination and detail routes require an opaque credential reference. JobAffinity, Taleez, and Teamtailor application discovery accepts only source-provided, canonical-host application URLs.
Credential-free Workable detail revalidation uses the vendor-advertised
/jobs/view/{shortcode}.md representation linked from each public job page. The bounded Markdown parser accepts only the H1 title and source-owned Apply link; credentialed installations continue to use SPI detail routes.
Every source has an executable conservative rate policy forwarded to the acquisition port. Policies bound concurrency, request spacing, retryable statuses, and Retry-After handling. The JobAffinity policy preserves its documented 45 requests/minute per-IP ceiling and Workable preserves 10 requests per 10 seconds.
The France-first direct HTML wave includes public ISCOD and CmonAlternance listing cards plus direct detail rechecks. Both parsers bind only canonical source hosts and qualified detail paths, decode stable public fields, and reject cross-origin links. Incremental ingestion remains gated because neither public page currently exposes a qualified cursor or change token.
ISCOD discovery tracks the current
/offres-emploi-en-alternance/ listing path. A failed live discovery request marks every dependent supported capability failed while retaining closure as a separate transition blocker; downstream cells are never left looking merely unevaluated after their prerequisite failed.
CmonAlternance discovery uses the server-rendered /public/ latest-offers surface. The JavaScript /public/OFFRES search page is Algolia-backed and does not fabricate listing cards during direct HTML acquisition; full-index access requires a separately qualified search adapter.
Documented public-board adapters normalize vendor-specific listing shapes for
Ashby,
Greenhouse,
Lever,
Recruitee, and SmartRecruiters. Lever global/EU routing and numeric pagination cursors are explicit. Other ATS connectors use qualified structured-page/JSON-LD acquisition and remain subject to their matrix and live-receipt state.
Greenhouse accepts both the legacy boards.greenhouse.io and current
job-boards.greenhouse.io public job hosts while keeping acquisition on
boards-api.greenhouse.io. The public absolute_url is the qualified hosted application entry and never widens the accepted host set.
Recruitee list responses expose offers[], while current detail responses wrap the same public job under offer. The connector flattens that source-specific envelope before normalization and preserves careers_apply_url as the hosted application entry.
iCIMS
and Workday use tenant-bound public career pages with source-specific structured-data fixtures, safe same-origin detail routing, and browser fallback. WeRecruit uses the same public-page path; its customer API
remains account-gated because WeRecruit supplies API documentation only on request. None of these three sources advertises a qualified public change token here, so incremental ingestion remains explicitly gated while successful scans and direct detail rechecks qualify revalidation and closure evidence.
Current WeRecruit career pages expose a source-owned window.allOffers JSON payload for listings and HTML-encoded application/ld+json on detail pages. The direct parser consumes those public artifacts first; browser fallback remains available when an installation omits them.
