The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documentation describes a harness that ingests untrusted data from the application under test, which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The harness reads the application's stdout logs (e.g., using app.stdout.waitFor or grep on log files) and captures semantic UI snapshots via ToolkitDriver.snapshot(), which includes labels and text values from the app.
Boundary markers: No specific delimiters or instructions are provided to the agent to treat application logs or UI content as untrusted data.
Capability inventory: The skill possesses significant capabilities including process management (launching binaries, pkill for reaping processes), file system interaction (reading/writing pid files and logs in .showcase/), and UI automation (clicking, typing, and executing expressions via the Dart VM service).
Sanitization: There is no mention of sanitizing or escaping the text read from the application before it is processed by the agent or used in logic.
[COMMAND_EXECUTION]: The skill instructs the agent to use various shell commands for environment management, process control, and log parsing.
Evidence: Examples include the use of pgrep to find process IDs, pkill and SIGKILL for process termination, and complex grep patterns to extract VM service URIs from log files. While these are standard for a testing harness, they represent powerful capabilities that could be abused if the agent is misled by malicious input.

flutter-mcp-e2e-harness