The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest, analyze, and modify the entire contents of a user-provided repository. This creates a risk where malicious instructions hidden in the source code could influence the agent's analysis or refactoring implementation.
Ingestion points: Full repository scope (SKILL.md).
Boundary markers: The skill does not define explicit boundary markers or instructions for subagents to ignore embedded prompts in the analyzed data.
Capability inventory: The agent can write to files (Phase 2), execute git commands, and run arbitrary shell commands for verification (verification.md).
Sanitization: No sanitization of ingested code content is performed before it is passed to subagents.
[COMMAND_EXECUTION]: The skill identifies and executes the project's verification tools to ensure refactors don't break functionality. This involves executing shell commands (e.g., npm test, pytest, cargo test) that run scripts defined in the repository's configuration files (e.g., package.json, Makefile, pyproject.toml).
Evidence: The references/verification.md file contains a catalog of common test and build commands for various ecosystems that the lead agent is instructed to run.

improve-code-structure