Source confirms an automatic, project-scoped Claude skill installation during postinstall, with skip switches but no affirmative consent check. The inspected behavior supports a capability warning; the reported binary entrypoints remain unresolved, so a clean verdict is not established.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
Installing the package when the hook’s skip conditions do not apply.
Impact
Claude Code may load the added or updated project skills; the automatic agent configuration change warrants a warning.
Mechanism
The postinstall script reads skill files shipped in the package and writes them into the project’s .claude/skills directory, installing missing files or newer versions.
Evidence
package.jsonscripts/postinstall.mjs.claude/skills/<skill>/SKILL.md