The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates a shell script (.claude/hooks/session-start.sh) that executes standard package management commands like npm install. It also performs manual validation by executing this script and running linters/tests.
[PERSISTENCE]: The skill creates a SessionStart hook registered in .claude/settings.json. This is a legitimate platform feature designed to ensure the environment is correctly configured every time a session starts.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external repository files to determine environment requirements, which presents a surface for indirect instructions if the files are malicious.
- Ingestion points: package.json, requirements.txt, Cargo.toml, go.mod, Gemfile, and README.md in SKILL.md.
- Boundary markers: Absent; the agent reads and interprets the content of these files directly.
- Capability inventory: File writing (cat), permission modification (chmod), and shell execution (./.claude/hooks/session-start.sh, npm, pip, etc.) are used in SKILL.md.
- Sanitization: Absent; the skill relies on the agent's reasoning to parse manifest files safely.
[DYNAMIC_EXECUTION]: The skill writes a bash script at runtime using a heredoc (cat > ... << 'EOF') and then executes it to validate the setup. This behavior is the primary intended function of the skill.

session-start-hook