The Agent Skills Directory

[COMMAND_EXECUTION]: The "SessionStart smoke test" in SKILL.md (Step 1b) parses a command string directly from the .claude/settings.json file and executes it using bash. This pattern allows project-level configuration to trigger arbitrary command execution on the user's system.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes untrusted data from project configuration files (.claude/settings.json, .mcp.json, and .pre-commit-config.yaml) without boundary markers or sanitization, allowing malicious project files to influence the agent's diagnostics.
[DATA_EXFILTRATION]: The skill reads sensitive local files, including global and project-specific Claude Code settings and the plugin registry. These files often contain private environment configurations, paths, and potentially credentials or API keys.
[EXTERNAL_DOWNLOADS]: In Step 1c, the skill suggests that the user manually install the pre-commit package using pip if it is not found during the diagnostic check.
[COMMAND_EXECUTION]: Multiple internal scripts (check-plugins.sh, check-hooks.sh, etc.) execute shell commands and use jq to perform diagnostics or modify the plugin registry, which could be exploited if the inputs are not properly handled.

health-check