The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation framework in scripts/evaluation.py enables the execution of arbitrary local commands to launch and interact with MCP servers.
Evidence: The script utilizes CLI arguments (-c/--command and -a/--args) to parameterize shell commands that are subsequently executed via the mcp library's stdio_client transport.
[INDIRECT_PROMPT_INJECTION]: The skill's evaluation workflow ingests content from XML files, exposing the agent to indirect prompt injection from untrusted test suites.
Ingestion points: scripts/evaluation.py parses question and answer fields from <qa_pair> elements in local XML files.
Boundary markers: While the evaluation loop uses a system prompt to define the assistant's role, it lacks robust delimiters or instructions to prevent the agent from obeying malicious directives embedded within the test questions.
Capability inventory: The internal agent has the authority to invoke any tool exposed by the MCP server connection established by the harness.
Sanitization: Input from XML files is processed directly without filtering or adversarial sanitization.
[DYNAMIC_EXECUTION]: The provided scripts dynamically initialize process environments and execute code based on runtime configuration.
Evidence: scripts/connections.py manages a MCPConnection abstraction that dynamically selects transport layers and spawns subprocesses with user-provided environment variables and arguments.

mcp-builder