The Agent Skills Directory

[SAFE]: The skill's primary function is to enhance security by preventing destructive Git operations such as hard resets or forced pushes. It uses a local script to filter commands based on a hardcoded list of dangerous patterns.
[PROMPT_INJECTION]: The skill processes untrusted tool execution data, representing a surface for indirect prompt injection (Category 8).
Ingestion points: The scripts/block-dangerous-git.sh script ingests JSON data via stdin to evaluate commands before they are executed.
Boundary markers: None present in the script; it assumes the input format is consistent with the agent's internal tool-use schema.
Capability inventory: The script performs pattern matching and exits with code 0 (allow) or 2 (block). It does not perform network operations or unauthorized file writes.
Sanitization: Input is parsed using jq to extract the command string, which is then compared against a hardcoded list of dangerous patterns using regex.
[SAFE]: The skill modifies configuration files (.claude/settings.json or ~/.claude/settings.json) to install the hook. While this creates persistence for the guardrail, it is the stated and transparent purpose of the skill and is restricted to the application's own configuration settings.

git-guardrails-claude-code