AGENT SKILL

career-ops-scraper-only

Discover and ingest new career-ops job postings when the user explicitly asks to run the scraper only, with no triage or later stages. Separate from the broader career-ops scan mode.

License 许可
MIT
最近更新
2026-11-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
career-ops-scraper-only
description
Discover and ingest new career-ops job postings when the user explicitly asks to run the scraper only, with no triage or later stages. Separate from the broader career-ops scan mode.

03 SKILL.md 正文

Career-ops scraper-only discovery

Use the existing scraper from the career-ops checkout. This workflow requires no claude-mem service or external model; do not query or start claude-mem. First follow root resolution and shared preflight, including the local-controller data-root check. This selects no broader mode.

Before running

  1. Read the checkout's current AGENTS.md, modes/_custom.md, config/profile.yml, DATA_CONTRACT.md, and local/orchestration/scan-all.sh. Current code and rules take precedence over remembered commands or queue counts.
  2. Inspect git status --short, active scan and batch workers, locks and their owners (including data/pipeline.md.lock), batch claims and .eval-wave. Respect existing work; do not delete locks, rebuild batch input, or start a competing scan.
  3. Inspect data/pipeline.md, data/scan-all-last-success, and data/scan-runs.tsv. Record the pending count and completion marker before running. Do not infer a current result from historical logs.
  4. Read local/README.md, local/pyproject.toml, local/uv.lock, and the effective JOBSPY_SITES. Read the installed version from the lock; use that locked environment and preserve the adapter's metadata/filter contracts. If dependency setup is needed, sync the existing lock from local/ with UV_CACHE_DIR=/tmp/job-agent-uv-cache uv sync --locked; do not upgrade dependencies or restore obsolete provider monkeypatches during a scan.

Run and wait

Run the canonical script from the checkout root:

env -u JOBSPY_HOURS UV_CACHE_DIR=/tmp/job-agent-uv-cache ./local/orchestration/scan-all.sh

The script calculates the window from the last successful full scrape. If its completion marker is absent, it uses the last broad completed scan in data/scan-runs.tsv, or its bootstrap default. Do not calculate routine hours manually. The wrapper also respects an existing explicitly accepted cutoff (data/scan-all-accepted-through.json); never create or advance one merely to hide missing coverage. Clear an inherited JOBSPY_HOURS unless the user explicitly requests an override; only then supply their positive whole-number hours as JOBSPY_HOURS instead of unsetting it.

The script runs ATS discovery, JobSpy, ingestion, and built-in repair. Keep its process/session handle and wait for that same process's final ingest and repair output and exit status. It checks per-process native ATS and JobSpy health receipts before advancing the success marker. Missing/stale receipts, provider errors, an unavailable LinkedIn query, or request/page budgets that truncate coverage leave the marker unchanged. A file or exit code alone is insufficient. If interrupted or a stage fails, report partial progress and preserve written state; do not claim a full scan. Check the completion marker afterward as supporting evidence, never as a substitute for the current process's output. The wrapper has no safe --dry-run/--help preview; inspect its source instead.

Report and stop

Report the window, ATS additions, JobSpy ingested additions, pending queue change, and repair results from this run. State observed coverage limits such as provider errors, NVIDIA page limits, LinkedIn per-query caps, or unrepaired metadata. Do not claim exhaustive coverage. Keep the configured healthy source defaults; probe quarantined Glassdoor only when explicitly requested. Defaults are Indeed and LinkedIn. LinkedIn requests have shared pacing and hard budgets; a 429 pauses that source. Do not add authenticated sessions or proxies, or start a separate recovery run without its requested scope. Existing recovery helpers retain checkpoints when that work is requested. Captured rows remain usable even when the overall window is incomplete. Correlate local/source-health.json and scan receipts with this run; distinguish a provider failure from zero matches or filtering. A partial ATS/provider result remains partial even when the other sources add jobs.

Stop after discovery, ingestion, and the script's built-in repair. Do not run separate hydration, triage, evaluation, packaging, updates, sending, or submissions. The script's Next: footer, the broader scan mode, and next-step.mjs do not authorize advancing to another stage.

04 同仓库更多 Skills