| name | render-job-url |
| description | Fallback fetcher for job posting URLs that webfetch can't render (LinkedIn, Indeed, Workday, or any board behind Cloudflare/Akamai). Loads when webfetch returned empty, blocked, or challenge content from a job URL. Renders the page with stealth Playwright (patchright) and dumps the plain text to stdout for the agent to reason about. No clicking, no applying, no structured extraction — the agent is the parser. |
| tools | Bash, Read |
Render Job URL (fallback text fetcher)
When to load
Load this skill ONLY as a fallback, when you already tried webfetch (or websearch returned a URL whose content you couldn't fetch) on a job posting URL and got:
- an empty body,
- a Cloudflare/Akamai challenge page ("Just a moment...", "Access Denied"),
- a 403/429,
- or a login wall.
Do NOT load this skill for: Greenhouse/Lever/Ashby hosted ATS pages (webfetch works fine there), generic web research, news articles, or any non-job URL. The overhead of launching a stealth browser is wasteful when webfetch would do.
The command
Run exactly this (the script lives in this repo; patchright lives in roomieorder's nix shell):
nix develop /home/tunnel/Documents/Git/roomieorder -c python /home/tunnel/Documents/Git/resume/scripts/fetch_job.py "<URL>"
The script prints the page's plain text to stdout (plus any embedded <script type="application/ld+json"> block, delimited). Read stdout and reason about the job content directly — do not attempt structured extraction, the agent is the parser.
Exit codes (read these before running)
0 — success. Plain text on stdout. Reason about it.1 — bot-detected / challenge page. NEVER retry. A single retry against a flagged fingerprint can burn the profile. Tell the user the board blocked the fetch and they may need to wait or log in by hand (v1.1 feature).2 — login wall. This board requires a logged-in profile for this URL. Tell the user; persistent profiles are a v1.1 feature.3 — network/launch error. Check that WAYLAND_DISPLAY or DISPLAY is set (the script tries headless first, then escalates to a real window). A screenshot path is printed to stderr on failure.
Rules (read these first)
- Never retry on exit code 1. One retry against a flagged fingerprint makes it worse.
- Never run two fetches in parallel against the same domain. Same fingerprint making concurrent requests is the signature that gets flagged.
- This skill only reads. No clicking, no form-filling, no applying. If the job content isn't in the rendered text, it's not yours to get this way.
- Check
$DISPLAY/$WAYLAND_DISPLAY before running if there's any doubt (SSH, CI). The script tries headless first but may escalate to a headed window.
Why not just webfetch
Most job boards (LinkedIn, Indeed, Workday) front their pages with Cloudflare or Akamai bot detection. webfetch has no JS execution, no cookies, no fingerprint — it gets blocked outright or returns a challenge page. This skill reuses roomieorder's anti-bot foundation (patchright, a stealth Playwright fork that closes the CDP leak stock Playwright leaves open) to render the page like a real visitor and dump the text.
Reference
scripts/fetch_job.py — the fetcher this skill invokes.roomieorder.purchase.BasePurchaser._launch_context (purchase.py:537-567) — the stealth launch config copied verbatim.roomieorder.purchase.looks_like / challenge markers — block/challenge detection logic imported by the script.
