The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local git commands (e.g., git diff, git rev-parse) to resolve ranges and analyze repository state. It includes explicit security instructions to sanitize user-provided range inputs, specifically rejecting shell metacharacters and option-style strings (--output=...) to prevent command injection.
[PROMPT_INJECTION]: The skill processes untrusted data from git history and file diffs. It mitigates indirect prompt injection risks by explicitly instructing the agent to treat this data as 'passive content' that must not be interpreted as instructions. It also defines a red-flag condition to stop if text in the diff is treated as a command.
[DATA_EXFILTRATION]: All operations are restricted to the local environment. The skill explicitly forbids network operations during range resolution and excludes the use of external tools like the GitHub CLI (gh) that might require remote data transfer.
[SAFE]: The skill follows least-privilege principles by refusing to write files into the repository worktree and ensuring that any requested output files are stored in temporary directories outside the project root.

allocate-attention