[chilcano]

Brings the PR skill scan up to date with the canonical workflow in zama-ai/ai-security.

• Scanner 2.3.0, and advanced-2 on Haiku 5.5 instead of Opus 4.6, at a fraction of the cost.
• A critical or high finding in the parts a PR changes fails the check (--fail-on-severity high).
• One summary covers the whole run.
• scripts/skill_scan_budget.py is new and scripts/skill_scan_summary.py is replaced, both with the uv script header.
• No paths filter: the check runs on every PR.

Known gap: each scan unit is skills/<name>, so plugin components at the repo root (hooks/, .mcp.json, .lsp.json, commands/, agents/) are not scanned. There are none today. The fix will come from the canonical workflow.

[jfrery]

Please include repository-root components in plugin scanning. The plugin installs from ./, but changes to root .mcp.json or hooks/hooks.json select zero scans and pass. Add a root-component scan unit and a root-only change test to complete the promised coverage.

This is an inherited coverage gap, reproduced with local fixtures. Live scanner execution remains unverified.

[chilcano]

Agreed, the gap is real: the plugin installs from ./, but each scan unit is skills/<name>, so a root hooks/, .mcp.json, .lsp.json, commands/ or agents/ is in no unit.

The repo has none of these today, so nothing unscanned ships now. The fix belongs in the canonical workflow in zama-ai/ai-security, because this copy must stay identical apart from PLUGINS_DIR. It is tracked in the backport issue there, with a test for a change only at the root. I corrected the PR description, which claimed full coverage.