The Agent Skills Directory

[DYNAMIC_EXECUTION]: The hooks/response-lint.py script implements dynamic code loading. It reads an installation path from a local configuration file (installed_plugins.json), adds that path to the Python system path, and imports a module named metrics_v2 at runtime. The hooks/comment-lint.py script also uses importlib.util to dynamically execute and import commit-pr-lint.py from a relative path.
[INDIRECT_PROMPT_INJECTION]: The skill defines a procedure for the agent to update its own global instructions (rules.md) and a regex-based dictionary (dict.txt) based on user-provided feedback. Because these rules are persistent and applied to future sessions, a malicious user could potentially inject instructions or word replacements that alter the agent's future behavior. 1. Ingestion points: User messages triggered by phrases like 'tone fix' or 'don't use this word'. 2. Boundary markers: None specifically defined for the user input added to the rules file. 3. Capability inventory: Writing to instructions and dictionary files; modifying agent configuration files via install.py. 4. Sanitization: Relies on the agent to interpret and correctly place the user-suggested rules without sanitizing potential injection patterns.
[COMMAND_EXECUTION]: The install.py script performs significant file system modifications, including creating symbolic links and editing JSON configuration files within the user's home directory (e.g., in .claude, .codex, and .config/opencode). The adapters/opencode.js component also executes Python scripts via sub-processes during the tool execution lifecycle.

tone-fix