The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/task_pipeline.py utilizes subprocess.run to orchestrate task scaffolding and verification.
Evidence: It invokes the Harbor CLI (harbor task init) and the Python interpreter for NeMo Gym scaffolding.
Controls: Commands are constructed using logic within the script, primarily targeting local executables and known CLI tools.
[DYNAMIC_EXECUTION]: Several scripts utilize dynamic importing techniques to share logic across the local repository structure.
Evidence: scripts/gym_scaffold.py uses importlib.import_module to load NeMo Gym's internal scaffolding libraries.
Evidence: scripts/candidate_admission.py, scripts/construction_plan.py, and scripts/task_evidence.py modify sys.path at runtime to import from a sibling eval-author directory.
Context: This is a functional requirement for the multi-skill architecture used by the author to maintain shared libraries.
[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data during task creation.
Ingestion points: SKILL.md and scripts/candidate_admission.py process ATIF traces (recorded interactions) and audit report JSON files.
Boundary markers: The skill instructions emphasize privacy reviews and include a "Proposal-to-construction decision boundary" to validate requirements before execution.
Capability inventory: The skill can write to the local filesystem and execute commands via the Harbor/NeMo Gym runtimes.
Sanitization: scripts/task_pipeline.py implements regex-based sanitization (re.sub) for task slugs derived from tool names.

eval-author-task-create