The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the @playwright/test package and official Playwright Docker images from Microsoft's Container Registry (mcr.microsoft.com/playwright). These are well-known, official, and trusted sources.
[INDIRECT_PROMPT_INJECTION]: The skill takes user-provided strings (such as URLs and commit reasons) and incorporates them into shell commands, creating a potential surface for command injection.
Ingestion points: The skill prompts the user for a baseURL, app-host, port, and a USER_REASON for git commits in SKILL.md (Workflows A, B, and C).
Boundary markers: The skill provides explicit instructions to the agent to validate the input: "If the user's reason contains backticks, $(, or newlines, ask for a simpler version — those characters cause shell injection in the commit command."
Capability inventory: The skill possesses capabilities to execute shell commands (git commit, curl, npm install, docker run) across multiple workflows.
Sanitization: Sanitization is handled via natural language instructions to the agent, which is a common but imperfect mitigation against adversarial input.
[COMMAND_EXECUTION]: The skill performs various shell operations for project setup, dependency installation, and version control. While these are necessary for its stated purpose, they involve direct command execution on the host machine.

pixel-perfect