Flagged Stat

Single source of truth for the headline stat: the share of SCANNED tools that carry a high or critical finding. The homepage, the Index and the State of Agent Security report all read this, computed server-side over the launch corpus + community rows.

Every count is an exclusive bucket (see _severity_bucket), so a tool is counted once no matter how many findings it has, and the denominator is only tools that actually received a verdict. x402 endpoints are compliance probes, not code scans, and are excluded entirely.

Response:

• scanned_total: tools with a verdict (the denominator). Alias: scanned.
• critical: tools with >=1 critical finding.
• high_only: tools with >=1 high finding and no critical.
• flagged: critical + high_only. Every flagged tool counted exactly once.
• clean: scanned tools with no high or critical finding.
• skipped: catalog rows with no verdict (fetch skipped / scan error). NOT in the denominator. total = scanned_total + skipped.
• pct: flagged / scanned_total, whole percent (None if nothing scanned); flagged_pct and critical_pct carry one decimal.
• by_surface: the same keys per surface (mcp, npm, pypi, openclaw, github, ...).

pct, flagged and scanned are kept for existing callers. Before 2026-10-01 they double-counted tools with both a critical and a high finding and divided by every catalog row including never-scanned ones; they are now the honest figures.

Agreement guarantee: critical, high_only and skipped (overall and per surface) equal the total of GET /public/scan-catalog?severity=critical|high|skipped. The list's severity=clean additionally requires a score >= 80, so it is a subset of clean here.