The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes code changes from the user's local branch and provides them as input to subagent reviewers. This creates a vulnerability surface where instructions hidden in the code could attempt to manipulate the subagent's analysis.\n
Ingestion points: The skill ingests the working tree, staged changes, and commits via git diff, storing the output in /tmp/branch-$SLUG.diff (SKILL.md).\n
Boundary markers: The review context is managed through a structured "Review charter" (Step 3), and subagents are directed to read the diff from a file path rather than having it interpolated into their primary instructions.\n
Capability inventory: Subagents are provided with access to .claude/tools/sandbox, allowing them to read file contents and potentially execute commands in a managed worktree.\n
Sanitization: A verification step (Step 5) is used to validate that subagent reports include evidence (proof-of-work lines) directly from the provided diff file to confirm authentic analysis.\n- [COMMAND_EXECUTION]: The skill uses the Bash tool to run git operations and manage local files for scope discovery and report generation.\n
Evidence: The skill executes git rev-parse, git diff, and git fetch, and uses local file operations like mkdir and sed to store review drafts in ~/.nx-branch-reviews (SKILL.md).

review-local-branch