The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses npx and uvx to execute remote tools (@upstash/context7-mcp, code-review-graph, agentation-mcp) at runtime during project initialization and health checks. These packages are fetched at the @latest version, which can lead to the execution of unverified code if a package is compromised.
[COMMAND_EXECUTION]: The skill frequently uses the Bash tool to perform system-level operations such as git init, git clone, and modifying file permissions (chmod +x). It also programmatically creates and installs git pre-commit hooks (.git/hooks/pre-commit), which automates command execution during the user's git workflow.
[CREDENTIALS_UNSAFE]: Instructions direct the agent to read and write to .claude/settings.local.json. This file is used to store sensitive environment variables for MCP servers, such as GITHUB_TOKEN. The skill manages the lifecycle of these secrets, including asking the user for values and writing them to the local filesystem.
[DATA_EXFILTRATION]: The skill accesses and parses .env and .env.local files to identify required environment variables for project documentation. This results in the exposure of sensitive variable names and potentially their structures to the agent context.
[PROMPT_INJECTION]: The scanner-agent and refresh-agent perform deep scans of all source code files in the project (including .ts, .js, .py, and .model files). Because these files are untrusted and their content is interpolated into the agent's prompts without sanitization, it creates a surface for indirect prompt injection attacks.
Ingestion points: Scans all files within the project directories, including source code, configuration files, and documentation (scanner-agent.md, refresh-agent.md).
Boundary markers: No specific boundary markers or instructions to ignore embedded commands are used when the agent processes the scanned file content.
Capability inventory: The agents have access to high-privilege tools including Bash, Write, and Edit across the workspace.
Sanitization: No sanitization, validation, or filtering of external code content is performed before it is added to the prompt context.

project