The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/fetch_ipsae.sh uses curl to download the ipsae.py utility from a third-party GitHub repository (dunbracklab/IPSAE). Additionally, scripts/setup_af2_params.sh uses wget to retrieve AlphaFold2 model parameters from a Google Cloud Storage bucket.
[REMOTE_CODE_EXECUTION]: Code downloaded via scripts/fetch_ipsae.sh is stored in the vendor/ipsae/ directory and subsequently executed using subprocess.run by the validation script scripts/validate_binders.py.
[COMMAND_EXECUTION]: Multiple scripts, including scripts/pipeline.py, scripts/boltz2_refold.py, and scripts/validate_binders.py, utilize subprocess.run to execute the complexa CLI and other internal Python utilities. This creates an orchestration layer that relies on shell-level execution for core functionality.
[DYNAMIC_EXECUTION]: The orchestration scripts dynamically construct command-line arguments using variables such as task_name and run_dir which are influenced by user input. While the scripts generally use list-based command construction (mitigating shell injection), the execution paths are determined at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill represents a significant attack surface for indirect injection as it ingests and processes complex biological data from external databases (UniProt, PDB). This data directly influences the generation parameters, including hotspot residues and target cropping, which steer the agent's downstream computational actions.

complexa-binder-design