The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several utility scripts to perform system health checks and usage auditing, including scripts/skill-sprawl-audit.py, scripts/check-routing-drift.py, scripts/usage-report.py, and scripts/governance-report.py.
[COMMAND_EXECUTION]: It interacts with the local environment using the ~/.claude/scripts/learning-db.py script to manage interaction history and system learnings.
[COMMAND_EXECUTION]: The skill utilizes the gh (GitHub CLI) tool to list PRs, view review comments, and create or merge Pull Requests on remote repositories.
[COMMAND_EXECUTION]: It manages project branches and commits changes using the git command-line tool as part of its automated implementation pipeline.
[COMMAND_EXECUTION]: Persistence is established by the skill using a ~/.claude/scripts/crontab-manager.py script to configure a nightly execution schedule.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted content from external sources to inform its evolution process. This includes processing GitHub Pull Request metadata and review comments, as well as incorporating signals from web searches performed by the 'Community' perspective agent.
Ingestion points: GitHub PR metadata and comments (via gh pr list/view), and web search results.
Boundary markers: None identified. The instructions do not specify delimiters or 'ignore' instructions for the processed external text.
Capability inventory: The skill has broad capabilities including file modification (Write/Edit), bash command execution, and managing remote repository states via PRs.
Sanitization: No evidence of sanitization or verification of the external input strings is performed before they are used to generate system improvement proposals.
[DATA_EXFILTRATION]: The skill's ability to create and merge Pull Requests involves pushing code and analysis to remote repositories, which presents a potential vector for exfiltrating internal data if the 'winner' proposal is influenced by malicious input.
[REMOTE_CODE_EXECUTION]: The skill performs autonomous code generation and validation. It implements new skills or modifications based on its diagnosis and then executes this generated code (e.g., using python3 -m scripts.skill_eval.run_eval) to verify success before promotion. This behavior, when combined with untrusted input ingestion, creates a risk of executing malicious instructions.

toolkit-evolution