The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to generate and execute JavaScript or Python code within an evaluation environment to define and control workflows. This includes dynamic imports of an SDK from a path determined by the OMO_DAG_SDK_ROOT environment variable.
[INDIRECT_PROMPT_INJECTION]: The skill processes outputs from multiple sub-tasks (nodes) which may contain untrusted data from external sources or research activities. It includes the following elements related to this surface:
Ingestion points: Untrusted data enters the agent context through node outputs (result.nodes[id].output) in SKILL.md and references/planning.md.
Boundary markers: The skill requires node prompts to use explicit markers (TASK, DELIVERABLE, SCOPE, VERIFY, STOP WHEN) to delimit instructions and prevent context bleed as detailed in references/planning.md.
Capability inventory: The skill utilizes tool.workflow for task orchestration and requires the ability to execute code in JS/Python cells.
Sanitization: The skill mandates a 'verification wave' where dedicated nodes must prove claims against evidence and checks for 'drift' to ensure workers stay within their assigned scope.

mass-ulw