The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill processes untrusted trace data which acts as a vector for indirect injection.
Ingestion points: Untrusted data is ingested from traces.json via scripts/extract_issue_context.py and referenced workflow steps.
Boundary markers: None. The skill lacks delimiters or safety instructions to separate untrusted trace content from the agent's internal logic.
Capability inventory: The skill is explicitly authorized to modify and overwrite executable Python scripts in the scripts/ directory.
Sanitization: None. The skill does not perform validation or sanitization of trace content before using it to generate code modifications.
Dynamic Code Modification (LOW): The skill performs automated script modification (Category 10). While this is the primary intended purpose of the skill, it creates a powerful capability that can be misdirected by adversarial inputs in the execution traces.
Missing Component (SAFE): The script scripts/analyze_traces.py is referenced in the workflow but was not provided in the source files, preventing a full audit of the primary analysis logic.

skill-evolver