The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection attack surface as it is designed to ingest and act upon data from external, potentially untrusted sources.
Ingestion points: Analyzes repository files, diffs, tests, comments, PR metadata (titles, bodies, and comments), project rules, and supplied web content within SKILL.md.
Boundary markers: Present. The skill features a robust 'Untrusted data boundary' section that explicitly instructs the agent to treat external content as data rather than operational instructions, and explicitly mandates ignoring embedded prompts that attempt to widen scope or override system requirements.
Capability inventory: Executes local environment Git commands (git rev-parse, git status, git diff, git ls-files) and orchestrates parallel sub-tasks via specialized sub-agents (code-explorer, code-architect, code-reviewer) within SKILL.md.
Sanitization: Instructions require filtering for facts, enforcing strict completion criteria, and redacting sensitive data/credentials using a [REDACTED] placeholder.
[COMMAND_EXECUTION]: The skill instructs the agent to run local Git commands (git rev-parse HEAD, git status --short, git diff, git diff --cached, git ls-files --others --exclude-standard) in SKILL.md during the implementation phase. These commands are benign, non-privileged, and strictly necessary for establishing a baseline version control status before code modification.

feature-dev