Jobich Gap Analysis

This is a behavioural study, not a code reference. Jobich is the proven Swiss job-aggregation reference; we have no licence to its code. Every observation below is described in our vocabulary (Adapter, Observation, Env, Method, family, source, binding) so we can act on it without re-opening Jobich source files.

Coverage gap

Direct Adapters: jobs-ch, jobup-ch, ostjob-ch, zentraljob-ch, nzz-jobs. The three Yousty-platform boards share one factory in adapters/yousty.ts.

Families we need (priority-ranked)

Drawn from Jobich's actual coverage, ranked by volume per Adapter:

Tier A — multi-tenant ATS families (one Adapter factory, many sources):

1. Workday — ~27 enterprise tenants (Roche, Novartis, Julius Baer, Logitech, …). Tenant URL pattern; JSON-paginated.
2. Greenhouse — ~20+ tenants (Stripe, Figma, On, Frontify, Proton, …). Open JSON API per company slug — the cleanest family to implement.
3. SmartRecruiters — ~8+ tenants (Bosch, Vitol, Swissquote, Roland Berger, Syngenta). JSON API.
4. Lever — ~10+ tenants (ANYbotics, Wingtra, Sonar, Neural Concept, …). JSON API.
5. SuccessFactors (HTML) — ~20 tenants (ILO, ITU, ICRC, Rolex, Octapharma, SAP, Nestle, Sonova, Zurich, Helvetia, …). HTML scrape, JSON-LD on detail.
6. SuccessFactors (SPA / RMK) — Pictet, Breitling, PostFinance. JS-rendered; __NEXT_DATA__ cache parsing avoids Browserbase on most.
7. Oracle HCM — IOM, UBP, EFG, JP Morgan, ELCA. JSON.
8. Phenom — large enterprise SPAs; embedded-state parsing first, JS fallback.
9. Personio, Teamtailor (RSS), Ashby, Hibob, Workable, Recruitee — each adds 5–10 tenants.
10. Refline, Prospective, Softgarden, Umantis — Swiss-specific ATS, smaller fanout but high CH relevance.

Tier B — search-driven general boards (single source per family, high hostility):

• LinkedIn (public listings only — auth-walled detail pages are out of scope by contract).
• jobup.ch, Indeed, Xing, jobwatch, Stellenanzeiger, Adzuna, Reed.

Tier C — Swiss employer career pages, RSS feeds, niche boards:

• Migros, Axpo, Alpiq, BKW, SBB, Sika, CERN, ETH Zurich, EPFL, Swiss TPH, PSI, EMPA, Eawag, myScience.
• RSS family: Climeworks, Deloitte CH, Leonteq, myScience — one generic RSS Adapter.
• HotelCareer, HotelJob, EuroPharmajobs, Reliefweb (UN/NGOs), SwissDevJobs, ICTCareer.

End state: every Jobich-covered source has a corresponding (family, source, binding) entry in JobCache.

Patterns to adopt

1. Sources-as-config (not sources-as-code)

Jobich's biggest scaling lever: adding a new Workday tenant is a YAML edit, not a Python module. JobCache's equivalent: a TypeScript per-source config consumed by a family factory that returns N Adapter instances, each with a distinct id (greenhouse:stripe, greenhouse:figma, …) and pinned source_id. The Yousty factory in adapters/yousty.ts is the existing precedent for this shape; new family adapters use the same factory pattern and emit Observations directly.

2. JSON-LD first, HTML last

Many ATS families expose JobPosting JSON-LD on detail pages. Parse JSON-LD before falling back to selector-based HTML. adapters/schema-org-jobposting.ts already wraps json-ld.ts and emits a full Observation; every JSON-LD-capable direct Adapter should delegate observe() to it (jobs-ch does already).

3. Embedded JSON before headless

Many SPA boards hydrate from <script id="__NEXT_DATA__">, <script type="application/json">, or window.__APOLLO_STATE__. Parsing those is faster, more deterministic, and free of Browserbase quota. jobs-ch already parses window.__INIT__ for listing discovery — same pattern works for SF SPA, Phenom, Xing. Reserve JS-rendering capability (Browserbase / Firecrawl via the Router) for sites that genuinely need it.

4. Query × location fanning for search-driven boards

For general boards (LinkedIn, Indeed, jobs.ch), discovery loops over a known set of role-keyword queries (~40 EN + ~20 DE) × Swiss city/canton list. This is how Jobich gets broad coverage without a sitemap. The query and city lists belong in shared catalog files consumed by every Adapter that needs them, not inside each Adapter.

5. Multi-stage post-extract enrichment

Jobich runs canton/region inference, language detection, industry classification, work-mode/employment-type normalization, role/function mapping, and salary parsing as separate post-scrape passes. JobCache's shared Observation contract supports this: Adapter Observations carry only what the source page directly provides; later enrichment passes (Codex side, separate task class) attach additional annotations and observed fields against the same ad_id. Adapters must not collapse the two into one pass.

6. Two-level deduplication

URL x language ad identity is necessary but not sufficient — the same role lands on multiple boards with different URLs. Jobich computes a semantic hash from (title.lower().alnum, company.lower().alnum) to cluster cross-source duplicates. JobCache uses adIdFromUrlLanguage(url, language) for ad identity and roleIdFromUrl(url) for the URL-family role above those ads. Cross-source semantic role reduction is deferred async work over consistent title + organization Observation fields; Adapters' job is to emit those fields in a normalized form.

7. Staleness tracking via observation windows

scraped_at (first-seen) + last_seen_at (most-recent) + 60-day stale threshold = jobs that disappear get marked inactive without losing history. In our model, Adapters emit observed_at correctly on each field; the observation application path (Codex) flips an ad's lifecycle cell when no fresh Observations land within the window. Adapter-side concern: emit observed_at.

8. Per-source kill switch

A DISABLED_SCRAPERS env var + a JSON file lets ops disable a broken source without redeploy. JobCache equivalent: binding-level "paused" flag on the catalog row, surfaced through ops/qa. Adapter-side: nothing — the runtime simply stops leasing tasks for paused bindings.

9. Defunct-source tracking

Companies migrate ATS (~12 documented Workday departures alone). The per-family source list should mark defunct sources with a date + reason instead of deleting them, so the quarterly re-audit knows what to re-check.

10. Non-CH location blacklist

Many international boards leak DE/AT/FR roles into CH searches. A city-name blacklist filters at discovery time, cheaper than per-job geocoding. Belongs in a shared locations catalog consumed by every search-driven Adapter.

11. Incremental scrape decoupling — already better

Jobich's "scrape-to-JSON-then-upload" pattern decouples network conditions from DB writes. JobCache's Observation-submission path already does this differently and better: Adapters emit Observations; submitObservation writes the DB; if the broker is down, Observations queue at the device. Do not re-implement Jobich's two-phase pattern.

Patterns to reject

1. No bare time.sleep rate limiter

Jobich uses a flat 1.5s-per-domain sleep. scrapers/router.ts already implements per-(scraper, host) spacing, 429 cooldown with exponential backoff (30s → 30min), and a 6-attempt ceiling. Keep that for hostile-source Adapters (Adapters with method: "headless-proxy" whose Env.fetch is runtime-routed through the cascade); never regress to flat sleeps.

2. No flat vacancy/posting shape

Jobich stores everything as a single dict (title, company, location, url, …) and dedups by URL. JobCache: every public ad is one Observation anchored to one ad_id, with fields stored as Record<FieldKey, ObservedField> and evidence refs proving every claim. Adapters MUST emit Observations; the old Adapter/posting projection is gone.

3. No proxy rotation

Jobich has none. JobCache's Router cascade (Raw → Browserbase → Firecrawl) ladders on capability (JS, anti-bot family) rather than on IP-rotation alone. Adapters that need JS / anti-bot evasion declare method: "headless-proxy" and the device runtime injects a cascade-backed Env.fetch. Proxies are only justified if a specific source proves they're the missing piece.

4. No raw HTML in hot SQL

Jobich keeps only the cleaned description text. Our contract demands online S3-compatible cold evidence for retained raw HTML/JSON via object_refs; hot rows store the content hash + object key. Adapters never write the blob themselves — they declare the object ref the device runtime produced.

Where JobCache is already ahead

• Router cascade with cheapest-first execution (scrapers/router.ts): capability-gated, quota-aware via available(), host-spaced, 429-backed off. No Jobich equivalent.
• Browserbase tier inside the cascade: 60 browser-minutes/month flat free tier absorbs hostile-board JS rendering at zero per-call cost; Firecrawl picks up the tail.
• S3-compatible cold evidence: replaceable raw storage by object_refs; Jobich keeps none, so it can't re-extract or audit.
• One canonical submitted shape: Observation from @cv/jobcache/contract, validated by ObservationSchema. Jobich's dict-shape is impossible to validate.
• Frozen URL x language identity: adIdFromUrlLanguage(url, language) from @cv/jobcache/ids — every caller shares the same ad identity contract, while roleIdFromUrl(url) groups language-specific ads under the URL-family role.
• Cockroach observation path: Observation → submitObservation → current current ad payload cells, device trust, and follow-up verification tasks. Jobich writes directly to Postgres; no explicit device trust loop.

Net gap summary

Closing this gap is the work the slice plan schedules. See adapter-slice-plan for ordering and adapter-code-inventory for what is kept, replaced, or retired along the way.