Problem
The dangerous handoff is a subagent saying "safe to delete/deploy" with no file, diff, test output, or command transcript quoted. The main agent acts on an unverified verdict.
Pattern
Add a check in agentrace/checks.py (sibling: check_url_without_verification, counts evidence signals, fires when absent). Fire risky_verdict (medium) when action verbs (delete, drop, deploy, merge, revoke, force-push) appear with zero evidence tokens (paths, code spans, file:line, quoted output).
Acceptance
New check risky_verdict. _run(result="It is safe to delete the table, go ahead") flagged; same verdict plus a quoted SELECT count(*) / path clean.
Problem
The dangerous handoff is a subagent saying "safe to delete/deploy" with no file, diff, test output, or command transcript quoted. The main agent acts on an unverified verdict.
Pattern
Add a check in
agentrace/checks.py(sibling:check_url_without_verification, counts evidence signals, fires when absent). Firerisky_verdict(medium) when action verbs (delete, drop, deploy, merge, revoke, force-push) appear with zero evidence tokens (paths, code spans,file:line, quoted output).Acceptance
New check
risky_verdict._run(result="It is safe to delete the table, go ahead")flagged; same verdict plus a quotedSELECT count(*)/ path clean.