The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to invoke the complexa CLI and Python modules (protein_mpnn, colabdesign, etc.). These are the intended primary functions of the skill and target a first-party toolset.
[EXTERNAL_DOWNLOADS]: The preflight.sh script and instructions reference downloading configurations and checkpoints from official vendor sources (e.g., NVIDIA-managed paths) or trusted repositories. These align with the [TRUST-SCOPE-RULE] as they involve well-known technology organizations and official tool installation paths.
[DYNAMIC_EXECUTION]: The scripts/write_manifest.py and scripts/preflight.sh use subprocess.run to gather system information (like nvidia-smi and git rev-parse). These are limited to local diagnostics and metadata recording, posing no privilege escalation or arbitrary execution risk.
[DATA_EXPOSURE]: The skill accesses .env files and checkpoint paths. The preflight.sh script includes specific logic to parse the .env file as data (using a Python regex parser) rather than sourcing it, and it redacts sensitive values (keys, tokens) before reporting, which is a defensive best practice.
[INDIRECT_PROMPT_INJECTION]: The skill ingests PDB data from local directories. It includes 'dry-run' capabilities and boundary markers (thresholds) that limit the influence of external data on the agent's logic. Severity is rated as SAFE as it only processes structural biology data formats.

complexa-evaluate-pdbs