The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/kev_modal.py uses subprocess.Popen to execute shell commands for training, benchmarking, and deployment tasks.
Evidence: The train_subprocess function in scripts/kev_modal.py executes commands built as lists of strings, including user-controlled parameters like learning rates and batch sizes.
Context: While the use of list-based arguments with shell=False (the default) is a safer practice, the execution of commands derived from external configuration remains a capability to monitor.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from user codebases and data files, creating a vulnerability surface where malicious instructions could influence agent behavior or the training process.
Ingestion points: scripts/extract_workload.py scans a user-provided root directory for code patterns; scripts/convert_data.py reads local CSV/JSONL files.
Boundary markers: The skill does not implement explicit delimiters or "ignore embedded instructions" warnings when processing these files.
Capability inventory: The skill has the ability to execute subprocesses (scripts/kev_modal.py) and perform network operations via Modal and OpenAI-compatible APIs.
Sanitization: The skill uses ast.literal_eval for safe parsing of literals, but the natural language content from processed files is interpolated into prompts for LLMs in scripts/generate_data.py without additional escaping or sanitization.
[EXTERNAL_DOWNLOADS]: The skill downloads its core code and dependencies from external repositories during the training and deployment phases.
Evidence: scripts/kev_modal.py clones the https://github.com/jaredpalmer/kev.git repository to provide the underlying model logic.
Context: The repository is owned by the skill's author (jaredpalmer), representing a legitimate vendor resource.

kev-finetune