The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process data from external web applications, including DOM snapshots, console logs, and component states via the agent-browser tool. This creates a surface for indirect prompt injection, where a target application could include malicious instructions to influence the agent's behavior.
Ingestion points: Target web pages accessed via agent-browser open, console output retrieved via agent-browser console, and React component inspection.
Capability inventory: The agent has extensive capabilities, including shell command execution (git, npm, portless, curl), local file system writes (screenshots, reports), and direct modification of source code to perform bug fixes.
Boundary markers: No explicit delimiters or boundary markers are defined to isolate untrusted browser content from the agent's control logic.
Sanitization: The instructions recommend redacting passwords from generated reports, but no automated sanitization is applied to browser-provided data before it is interpreted.
[COMMAND_EXECUTION]: The skill performs various shell operations to manage the testing lifecycle. This includes starting local development servers via npm run dev or portless, managing source code state with git (add, commit, diff, status), and interacting with the Supabase Admin API via curl to manage test users.
[DYNAMIC_EXECUTION]: The skill utilizes agent-browser eval to execute arbitrary JavaScript code within the target application's browser context. This is used for advanced diagnostics, such as checking component properties, monitoring network requests, or simulating complex user interactions like Drag-and-Drop.
[CREDENTIALS_UNSAFE]: The skill provides templates and instructions for the agent to access and utilize administrative credentials (such as the Supabase SERVICE_ROLE key) to perform automated testing tasks like creating or querying users in the local environment.

wachi-qa