The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a multi-agent orchestration loop where the driver subagent ingests data from external plan files, executor reports, and verifier verdicts to determine the next iteration's actions. This architecture presents an attack surface where malicious content within a project's implementation or test output could influence the orchestrator's behavior or subsequent subagent contexts.
Ingestion points: The orchestrator reads task state via omh_state, project evidence via omh_gather_evidence, and subagent results via the delegate_task function.
Boundary markers: The skill utilizes structured templates and platform-specific role markers (e.g., [omh-role:executor]) to separate contexts, though it relies on verbatim interpolation of task descriptions into subagent goals.
Capability inventory: The system is capable of executing arbitrary project shell commands during evidence gathering and spawning subagents with broad file and repository access.
Sanitization: Instructions emphasize high-fidelity data transfer from plans to subagents to ensure specification compliance, with minimal explicit sanitization of subagent-supplied feedback before re-processing.
[COMMAND_EXECUTION]: The skill guides the agent to perform repository management tasks using git and project-level test execution via the omh_gather_evidence tool. These operations are standard for a development orchestrator but involve direct interaction with the filesystem and shell.
[DYNAMIC_EXECUTION]: The skill dynamically assembles task goals for subagents using plan data and iteration-specific learnings, then executes these goals through the delegate_task capability.

omh-ralph-driver