The Agent Skills Directory

[PROMPT_INJECTION]: The skill provides documentation for prompt-based hooks that interpolate untrusted data (e.g., $TOOL_INPUT, $TOOL_RESULT) directly into LLM prompts for validation purposes. This creates an indirect prompt injection surface where malicious data could potentially influence the validation decision.
Ingestion points: SKILL.md and multiple reference files (references/migration.md, references/patterns.md) describe interpolation of tool data into prompts.
Boundary markers: Examples lack explicit isolation markers (like XML tags) around untrusted variables.
Capability inventory: Hooks can influence the execution of file writes and shell commands.
Sanitization: Guidance focuses on the logic of the prompt rather than sanitization of the interpolated content.
[COMMAND_EXECUTION]: The scripts/test-hook.sh utility facilitates local testing of hook scripts by executing them with sample inputs via a bash subshell. This is an intended development feature for verifying script logic before deployment.
[EXTERNAL_DOWNLOADS]: Documented patterns in references/advanced.md provide examples for interacting with external services via curl (for Slack notifications) and psql (for database logging). These are presented as architectural integration patterns for workflow automation.
[SAFE]: The skill includes a dedicated linter (scripts/hook-linter.sh) that checks for common security pitfalls in shell scripts, such as unquoted variables that could lead to command injection and the use of hardcoded absolute paths.
[SAFE]: Provides numerous security-oriented examples, including logic for preventing path traversal, blocking access to system directories (/etc, /sys), and detecting sensitive files like .env or credentials.

Hook Development