The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from 'handoff files' generated by sub-agents to determine task completion and the overall run verdict.
Ingestion points: Handoff files are read by the omh_dag action=settle operation after a child task completes (documented in references/run-loop.md).
Boundary markers: The skill uses [dag:<run>/<node>] identifiers and requires a dedicated role: verify node that must depend on producers to judge the goal independently.
Capability inventory: The skill can orchestrate multiple sub-processes via delegate_task and manage file system changes within a defined write_scope.
Sanitization: Implements a rigorous 'Goal loop' (verify, repair, extend) where completion is only claimed when the omh_dag tool provides a verified verdict based on evidence-gated checks.
[DYNAMIC_EXECUTION]: The skill generates prompts and task definitions dynamically at runtime based on the planned graph nodes.
Evidence: Instructions in SKILL.md guide the creation of self-contained prompts using structured fields: TASK, DELIVERABLE, SCOPE, VERIFY, and STOP WHEN.
Control: The workflow involves a 'lint' phase where the omh_dag tool validates the plan and roles before any execution occurs.
[COMMAND_EXECUTION]: The skill facilitates the execution of external coding CLI tools like Codex or Claude Code under specific conditions.
Evidence: The 'External coding owner' logic in references/run-loop.md describes exporting nodes to these tools using omh coding fanout commands.
Control: These tools are only invoked if the user explicitly names the external owner in the specific request that starts the run, preventing unauthorized or silent tool usage.

dag-ulw