The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/check_contract.py uses compile() and exec() to dynamically load and execute Python source code from a sibling directory (../jev-evidence). This dynamic loading from computed paths allows the skill to execute code that is not part of its own package, which can be misused if integrity checks are bypassed or if the companion directory is modified.
Evidence: The load_core and main functions in scripts/check_contract.py manually construct Python modules and execute compiled code into their dictionaries using exec(compile(..., 'exec'), module.__dict__).
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to search, rerank, and discover public conversation data, which are untrusted external sources. Malicious instructions could be embedded in the content being processed, potentially influencing the agent's screening, ranking, or response generation.
Ingestion points: Host search command outputs and public discovery post captures as described in SKILL.md (Workflow Step 2) and references/discovery-evaluation.md.
Boundary markers: The skill instructions recommend redacting credentials and separating interpretations from quotations, but no robust delimiters or specific "ignore embedded instructions" warnings are established for the ingested content.
Capability inventory: The skill can execute shell commands via scripts/check_contract.py and invoke host search tools.
Sanitization: There is no explicit description of sanitization, filtering, or escaping logic applied to the external search results or discovered posts before they are presented to the agent.
[COMMAND_EXECUTION]: The skill documentation in SKILL.md and references/evidence-contract.md instructs the user or agent to execute shell commands (e.g., python3 scripts/check_contract.py version) to verify dependencies and process bundles, triggering the dynamic code execution identified above.

jev-search-and-rerank