The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from project files and script outputs to determine necessary code fixes. This creates a vulnerability where malicious project content could influence the agent to perform unintended edits.
Ingestion points: The skill parses stdout and stderr from package.json check scripts (e.g., linting, type-checking, building) and audit findings (SKILL.md Step 3 and Step 4).
Boundary markers: None explicitly defined in the instructions for isolating tool output from agent instructions.
Capability inventory: The agent has Read and Edit capabilities for the local filesystem and can execute shell commands via Bash (including pnpm, npm, and yarn).
Sanitization: No specific sanitization or escaping of the parsed error messages is mentioned before they are interpreted for auto-fixing.
[COMMAND_EXECUTION]: The skill executes various package manager commands and local helper scripts to perform discovery, auditing, and fixing operations.
Evidence: The skill uses pnpm/npm/yarn run check, pnpm/npm/yarn audit, and pnpm/npm/yarn add/update commands to manage project dependencies and verify build status.
Allowed Tools: It utilizes predefined bash scripts located in ${CLAUDE_PLUGIN_ROOT}/scripts/ for tasks like discovering scripts and managing peer synchronization.

running-check-script