Job Posting Watcher

Created with Claude Code to test its capabilities. This watcher checks a private list of company career pages on a schedule, finds new postings since the last run, filters them to ones matching keywords relevant to my skillset (ecology / geospatial / remote sensing), and emails me a digest.

• Free. Runs on GitHub Actions; filtering is keyword-based (no AI API, no billing).
• Private watchlist on a public repo. The company list and run-state never touch the repo — the list is a GitHub Actions secret, the state lives in the Actions cache.

How it works

COMPANIES_YAML secret ─▶ companies.yaml (ephemeral, git-ignored)
   restore seen.json ◀── Actions cache
        └─▶ fetch listings ─▶ diff vs seen ─▶ keyword filter ─▶ email digest (Gmail SMTP)
                                                                     └─▶ save seen.json ─▶ cache

File	Purpose
watcher/fetchers.py	Pull postings from Greenhouse / Lever / Ashby / BambooHR JSON APIs, Rippling (embedded JSON), Gusto boards, or scrape an HTML page.
watcher/store.py	Load/save seen.json; compute which postings are new.
watcher/filter.py	Keyword relevance matching (keywords.yaml).
watcher/notify.py	Build + send the email digest.
watcher/main.py	Orchestrates the pipeline.
keywords.yaml	Include/exclude relevance terms (public).
companies.example.yaml	Dummy format sample (the real list is a secret).

Notes

• Supported platforms: greenhouse, lever, ashby, bamboohr, breezy, rippling, pinpoint, workable, gusto, airtable (public shared-view), notion (public Notion page with an inline jobs database), section (reads items under an "Open positions" header on Webflow-style CMS pages), browser (renders JavaScript-only pages with headless Chromium, then scrapes), and html (best-effort scrape). See companies.example.yaml for the format of each.
• The browser type needs Chromium, installed in CI by the workflow (playwright install --with-deps chromium). Locally: python -m playwright install chromium.
• HTML fallback is best-effort. It only sees server-rendered markup (no JavaScript runs). Titles come from link text, or the URL slug when the text is generic ("Apply"/"View Position"). Fully JavaScript-rendered pages (e.g. Notion) won't yield jobs.
• Workday isn't supported in v1.
• Cache eviction: state lives in the Actions cache, which can evict after ~7 days of no runs. The daily schedule keeps it warm; if it ever evicts, one digest may repeat.
• To upgrade filtering to AI later: watcher/filter.py exposes a single is_relevant(posting, keywords) — swap its body for a model call without changing the rest of the pipeline.