The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill executes arbitrary shell commands defined in the project's development scripts (e.g., bin/dev, npm run dev) to start the server for testing. Running scripts directly from a repository being analyzed poses a risk if the repository contains malicious configuration.
[DATA_EXPOSURE]: The skill reads .env files within the repository to locate the application's PORT. Since .env files are standard locations for sensitive credentials and secrets, unselective access to these files presents a data exposure risk.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted documentation and code diffs from the repository to inform its 'autonomous fix' loop and browser automation. This allows malicious repository content to potentially manipulate the agent's actions or logic.
Ingestion points: Git diffs, STRATEGY.md, VISION.md, persona definitions, and live browser console/page content.
Boundary markers: No explicit delimiters or boundary markers are used when processing repository content.
Capability inventory: Includes arbitrary shell execution (via dev server), file system writing for reports, and Git commit operations.
Sanitization: Repository content is used directly in decision-making processes without escaping or validation.
[UNVERIFIABLE_DEPENDENCIES]: The skill relies on an external CLI tool, agent-browser, and directs the user to install it. While the user performs the installation manually, the skill's reliance on external binaries for its core browser automation functionality is a dependency risk.

ce-dogfood