The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/check_contract.py uses exec() and compile() to load and run Python code from a computed file system path pointing to a sibling directory (../jev-evidence).
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for collecting and labeling external evidence, which creates a vulnerability surface for indirect prompt injection as untrusted data is processed into the agent's context.
Ingestion points: Untrusted data includes source content, navigable source URLs, and fetched candidates as described in the Workflow section of SKILL.md.
Boundary markers: There are no explicit instructions for using delimiters or instruction-ignore markers when processing the evidence collected.
Capability inventory: The skill has access to shell tools (Bash, Grep) and can execute custom Python code, providing a target for malicious instructions.
Sanitization: Instructions mention type and probability validation but lack requirements for sanitizing natural language content from external sources.

jev-label-and-act