The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions to execute local shell commands for starting a FastAPI server using python3 execution/local_server.py and establishing an external tunnel via cloudflared.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by exposing local endpoints to the internet, allowing untrusted external data to reach the agent's execution environment.
Ingestion points: External webhook access is provided via the /directive and /general-agent endpoints through a Cloudflare tunnel.
Boundary markers: There are no instructions for using delimiters or boundary markers to distinguish untrusted external input from system instructions.
Capability inventory: The documentation explicitly grants the server "full access to local files and credentials" and identifies that it reads from a local .env file.
Sanitization: The instructions do not include any measures for sanitizing, validating, or filtering data received from the external tunnel before processing.
[DATA_EXPOSURE]: The skill's stated purpose involves granting a network-accessible server direct access to sensitive local environment variables and filesystem data, which increases the risk of credential leakage if the exposed endpoints are exploited.

local-server