Troubleshooting

You see	Try
Empty table, or a couple of rows that look invented	Open the Markdown node’s output. If it’s nearly empty, switch to Everything on the page
Only the first page was read	Raise Follow links; check your patterns match the page’s links; check the fetcher output for JS paging
Run fails with a semantic / 404 model error	Remove prompt from the extraction node; use extraction_instruction
“A resource isn’t available”	The model or store isn’t shared with you or failed its check; pick another
Save error ending “Reload to get the latest version”	Someone (or your agent) saved after you opened it; reload
A node is red after Validate	Read the panel message; it names the option
A failed node	The run panel shows its message. Include it when reporting a problem
Crawl wanders the site	Add url_pattern filters; lower depth and max_pages

Capture a useful failure report

Record the workflow id and version, run id, first failed node, returned error, and a small redacted sample of its input/output. Include whether a run was resumed or rerun. Never attach access tokens or connection strings.

Missing records after a successful run

Check fetch coverage, Markdown length, extraction output, foreach_result.failed and skipped, Combine unkeyed, and storage counts in that order. A completed run can still reflect a page cap or missing source data. See execution and debugging.