The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes bundled Python scripts (scripts/scan_agents.py and scripts/bump_version.py) to perform its primary function. These scripts are local to the skill folder and perform deterministic analysis of agent definition files without spawning shell commands or making network requests.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests external agent definition files from standard directories (e.g., ~/.claude/agents/, ~/.codex/agents/).
Ingestion points: The scripts scan_agents.py and bump_version.py read the content of Markdown and TOML agent definitions.
Boundary markers: The scripts utilize strict regular expressions to parse specific metadata fields (name, description, version) and treat the rest of the file content as inert text for token counting.
Capability inventory: scan_agents.py is read-only. bump_version.py performs file writes but only to update the version field. No scripts utilize network access, eval, exec, or subprocess calls.
Sanitization: The skill does not execute the content of the audited files, mitigating the risk of injection during the audit process.
[SAFE]: The skill demonstrates safe practices, including explicit refusal to follow symlinks to prevent directory traversal and the use of atomic writes when updating version numbers in agent files.

subagent-optimizer