The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as Git diffs and feature specifications, which could be used for indirect prompt injection attacks.
Ingestion points: The scripts/prepare-review-context.sh script extracts git diff content, which is then used in prompts for sub-agents defined in references/round-1-red-team.md and references/round-2-meta-architect.md.
Boundary markers: Untrusted content is wrapped in markdown code blocks (e.g., ```diff) within the sub-agent prompts to help distinguish it from instructions.
Capability inventory: The skill dispatches sub-agents with tool-calling capabilities and runs local shell commands via git.
Sanitization: The scripts/prepare-review-context.sh script validates commit references using git rev-parse --verify to prevent improper argument injection into git commands.
[COMMAND_EXECUTION]: The skill executes local shell commands through a helper script to manage git-related tasks.
Evidence: The scripts/prepare-review-context.sh script runs git status, git diff, git rev-parse, and git log.
Context: These commands are used as intended for gathering codebase context. The script proactively validates input arguments to ensure they are valid git objects, which is a key security measure against command injection.

dual-round-review