cross-corpus-research

Run bounded research across Crawlora public-web, paper, and indexed developer-document sources, retaining source snapshots, evidence, and partial-coverage warnings. Use for focused cited briefs or structured research jobs, not exhaustive web coverage.

Use this skill when a question needs an evidence brief across more than one public source type, a durable research job, or a bounded inventory of one site. Choose the smallest source scope and credit budget that can answer the question.

Access and source readiness

Set CRAWLORA_API_KEY. Use the bundled scripts/crawlora.sh, which sends it in an x-api-key header to https://api.crawlora.net/api/v1; do not place the key in a URL, shell argument, example, or report. Check both HTTP status and the application code. Stop on authentication errors, back off on 429, and retry a transient upstream failure once. Successful calls may consume credits.

Before selecting sources, inspect the live registries and capabilities when the request depends on source availability:

• /research/sources lists paper operations and the current registry revision. Paper keyword search is arXiv-based. Europe PMC adds MED-backed metadata, licensed full-text reads, and one-hop citation operations; it is not a universal paper search index.
• /developer/sources lists indexed documentation sources, supported filters, and index state. /developer/search searches only the index and never starts ingestion or contacts upstream sites. The source set can be narrow (the current ready source is Go documentation), so check before promising coverage.
• /web/capabilities reports which retrieval operations and limits are enabled for the current account.

Run a durable research job

1. Frame one answerable query, the requested date/market, source preferences, and a credit ceiling. Choose scope=web, papers, developer, or mixed; sources can further restrict the registered source IDs. seed_urls, include_domains, and exclude_domains apply only to web or mixed jobs.
2. Submit POST /research/jobs with a 16–128 character idempotency_key, the query, and limits.max_credits. Bound rounds, searches, sources, and timeout for the task. Browser renders are fixed at zero. Reuse the same key and identical request after an uncertain submission; use a new key only for a materially changed investigation. A JSON output requires a strict object schema; otherwise use the report output.
3. Keep the returned job_id. Check /research/jobs/{job_id} for state, progress, warnings, and usage. Read /research/jobs/{job_id}/events with its next_cursor when the audit trail matters. A queued job can wait up to 30 minutes before it times out without charge.
4. Read /research/jobs/{job_id}/results until a terminal state and a validated result are available. ready:false, partial:true, or any source warning must remain visible in the answer. Do not fill missing sources with guesses or describe retained evidence as a completed report.
5. Preserve source_snapshot revisions, result kind, source URLs, source dates, retrieval times, and usage. Search snippets are discovery leads, not citable evidence. A claim is supported only when its full text exactly quotes a retained passage; citation IDs alone do not establish support. Keep conflicting and unverified claims explicit.

sh

scripts/crawlora.sh /research/sources
scripts/crawlora.sh /developer/sources
scripts/crawlora.sh /web/capabilities
scripts/crawlora.sh -X POST /research/jobs '{"idempotency_key":"brief-20261019-001","query":"Compare recent evidence on public EV charging reliability","scope":"mixed","output_format":"report","limits":{"max_credits":40,"rounds":2,"searches":4,"sources":8,"timeout_seconds":180}}'
# Use the returned job ID; this placeholder is for route shape only.
scripts/crawlora.sh /research/jobs/job-example123 | jq '.'
scripts/crawlora.sh /research/jobs/job-example123/results | jq '.'

Read papers and indexed documentation directly

• Search papers with /research/papers/search, then inspect returned stable IDs with /research/papers/metadata. The arXiv search result window is frozen for cursor continuation; preserve the source and coverage warnings.
• Use /research/papers/read with a specific question. Europe PMC full text is returned only when the article license permits it; restricted records remain metadata-only. arXiv PDFs are not fetched. Keep passage locators, license, provenance, and exact source URLs with any extracted evidence.
• /research/papers/related returns one-hop Europe PMC references or citers. It is a bounded indexed graph, not a complete bibliography or a multi-hop literature review. Alias or seed failures and partial lists remain visible.
• For developer docs, call /developer/search only after selecting a source reported ready by /developer/sources. A search hit is indexed content, not a statement that the documentation is current for every software version.

Map or crawl one site

Use /web/search for a bounded public-web discovery sample. Use /web/map when the user needs an approximate URL inventory; it follows public sitemaps and seed-page links, defaults to the seed host, and freezes its cursor snapshot for 30 minutes. It is not a complete site census.

Use /web/crawl for a bounded, resumable page crawl. Supply a stable idempotency_key, the seed URL, and limits.max_credits; set reasonable page, depth, byte, and timeout limits. Poll its status and read results until drained or terminal. Failed/skipped URLs count toward the admitted page limit. Browser rendering is not available on this job path; do not promise JavaScript-rendered coverage. Preserve partial, warnings, stop reason, and accounting.

sh

scripts/crawlora.sh -X POST /web/search '{"q":"site:example.org API rate limit","include_domains":["example.org"],"limit":5}'
scripts/crawlora.sh -X POST /web/map '{"url":"https://example.org/docs/","scope":{"include_paths":["/docs/**"]},"max_urls":100,"page_size":50}'
scripts/crawlora.sh -X POST /web/crawl '{"idempotency_key":"crawl-example-20261019","url":"https://example.org/docs/","scope":{"include_paths":["/docs/**"]},"limits":{"max_credits":20,"max_pages":10,"max_depth":1}}'

Return a concise finding summary with claim-level citations, source and query scope, dates, successful and failed sources, and a clear stopping/coverage statement. Do not imply exhaustive research, browser rendering, future monitoring, or independent verification beyond the retained evidence. Do not publish, contact sources, or create recurring jobs unless requested.