The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function is to execute arbitrary shell lines defined in a JSON descriptor. The documentation notes that these commands have process authority in the current working directory and are not filesystem or process sandboxed.
[DYNAMIC_CONTEXT_INJECTION]: The skill instructions direct the agent to use the ! command syntax (e.g., ! omc graph run <file>) for pipelines requiring human approval. This syntax executes shell commands at the moment of invocation within the agent session.
[DYNAMIC_EXECUTION]: The skill interprets external JSON files to dynamically determine and execute command sequences at runtime, effectively acting as a script interpreter for the defined pipeline.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted JSON data which can define the behavior of both shell commands and agent nodes. 1. Ingestion points: The descriptor.json file provided by the user or found in the project environment (e.g., .omc/graphs/*.json). 2. Boundary markers: The skill suggests the agent show the descriptor to the user before running, which acts as a manual review step. 3. Capability inventory: The skill can execute arbitrary shell commands via the omc CLI and launch agent nodes with custom instructions. 4. Sanitization: Sub-processes receive an allowlisted environment (PATH, HOME, USER, etc.), though the command strings themselves are not sanitized.

graph