The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to manage a background daemon that executes shell commands and file operations via the omp (Oh My Pi) agent. As described in references/architecture.md and references/safety.md, the system runs in a mode that auto-approves tool use, including the bash and file-writing tools.
[COMMAND_EXECUTION]: To mitigate risks associated with unattended execution, the skill implements a 'dangerous-command guard' (referenced in references/safety.md) that parses shell commands and blocks known destructive patterns such as rm -rf /, fork bombs, and piping remote content to a shell.
[COMMAND_EXECUTION]: A YARA alert for a fork bomb was detected in references/safety.md; however, this is a false positive. The file contains the fork bomb string (:(){ :|:& };:) as an example of a pattern that is explicitly blocked by the safety guard, not as an executable command.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it ingests untrusted data that is processed by a shell-capable agent.
Ingestion points: Untrusted data enters the agent context through the omp-rpc send "task" and omp-rpc steer "..." commands described in SKILL.md and references/commands.md.
Boundary markers: The documentation does not specify the use of strict boundary markers or 'ignore' instructions for the data sent to the sub-agent.
Capability inventory: The underlying agent has the capability to run subprocesses (bash), write files, and switch models as detailed in references/architecture.md.
Sanitization: Sanitization is provided via a pattern-based blocklist in src/danger.js, which prevents the execution of common destructive shell commands.

using-omp-rpc