The Agent Skills Directory

[COMMAND_EXECUTION]: Discovery collectors such as collectors/linux.sh, collectors/windows.ps1, and collectors/network-device.sh execute system commands and show-commands (e.g., systemctl, Get-CimInstance, show version) to gather asset information from remote targets.
[DATA_EXFILTRATION]: The skill's primary function is to collect and consolidate infrastructure metadata (OS versions, hardware specs, installed packages, and services) from remote hosts into a central inventory ledger (CMDB).
[PROMPT_INJECTION]: An indirect prompt injection surface exists because the agent processes inventory data collected from external systems (such as hostnames, package versions, or service names) during reconciliation and querying tasks.
Ingestion points: Host data enters the agent context via scripts in the collectors/ directory.
Boundary markers: Discovery data is serialized to JSON, but the skill instructions lack explicit boundary markers or instructions for the agent to ignore potentially malicious content within discovered strings.
Capability inventory: The agent has the capability to execute RPC methods (inventory:upsert, inventory:query) and start new tasks via agent:startTask based on discovery results.
Sanitization: The skill uses Python's json.dumps in linux.sh and a custom escaping function jstr in shell scripts to normalize collected data into valid JSON format.

rterm-discovery