Jobstitch: free job sources merged into one clean feed

A job feed built from several sources goes bad quietly. The same opening shows up three times, old postings rank as fresh because a missing date became "now", and a source that lost its API key returns nothing for weeks while the UI looks normal. Jobstitch pulls listings from free official APIs and public ATS boards into one normalized, deduped feed, with a health record per source so a dead source is marked down instead of hidden.

What it does

• Terms-first sources: Careerjet, Jooble, Greenhouse, Lever, Ashby, Himalayas, Jobicy and RemoteOK, each with a terms card (endpoint, checked date, attribution, link and cache rules). Portal scraping is excluded.
• One adapter contract: every source returns the same NormalizedJob shape or a typed error such as missing_env or rate_limited.
• Honest normalization: unknown dates, salaries and currencies stay null; ids come from the source, never random.
• Cross-source dedupe: a key from company, title, country and city; the direct ATS row wins the apply URL.
• Ghost job expiry: hides expired, older than 30 days (tunable) and closed rows, and down-ranks reposts.
• Health and degraded mode: a source_runs table, alerts on failure or a drop in valid rows, and "1 source unavailable" in the UI instead of mock jobs.

How it works

1. Pick sources by licence; write each terms card in a sources.ts registry.
2. Build one adapter per source, then normalize and dedupe.
3. Back off: timeouts, Retry-After, backoff with jitter, max 3 tries.
4. Store and schedule: upsert with source_url and fetched_at; batch on cron, live per search.
5. Record health every run and degrade in the open.

Frameworks & tools

The official APIs above, with Supabase, Vercel cron and Next.js route handlers as defaults, not requirements.

Why this beats prompting it yourself

A quick prompt writes a scraper and a catch-all that returns demo data. This skill encodes per-source rules (Careerjet needs the real user IP and agent, Jooble's free key is 500 lifetime requests, Jobicy shows only 7 days so absence is not closure) and treats a missing key as an outage.

Use cases

• Job boards and aggregators merging free sources.
• Matching apps that need stored, dated job rows.
• Adding a new source with tests.

Known limitations

No scraping of job portals or social networks, no paid data feeds, and not for one-off research datasets. Source terms and endpoints change, so the registry dates must be re-checked when a source starts failing.