Octocode Scraping

Flow: FRAME → POLICY → ROUTE → FETCH → CORPUS → SEARCH → CITE → RECOVER.

FRAME before the first fetch: fix target URL/domain, goal, depth, and output shape — vague ask → references/user-inputs.md.

Defaults: one public URL, --mode html, omit --provider (keyless cdp→direct), session .octocode/tmp/scrape/{sessionId}, compact stdout. Search corpus before refetch. Live interaction → chrome-devtools on one port, then har-ingest + corpus-run into the same session. Ask before auth, hosted spend, crawl widen, CAPTCHA/MFA, destructive actions. Cite paths + URL metadata — not raw dumps.

Stop when: two same-class failures (report evidence, route tried, sanitized status, next approval); hosted 403 (wrong key or credits gone — status only, no retry); CAPTCHA/MFA, auth wall, or cookie/profile transfer needed; still blocked after one cdp try (ask before --provider scrapingant); personal data, form submits, purchases, sends, deletes, or account changes in scope; the saved corpus already proves the claim (cite it, do not refetch); crawl widen before reports/summary.md is useful. Recovery table: references/failure-recovery.md.