The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a conduit for untrusted data from a database into the agent's context, creating a surface for injection attacks.
Ingestion points: The scripts/inspect_best_programs.py script reads the code and text_feedback columns from a SQLite database specified by the user.
Boundary markers: The script uses standard Markdown code blocks to wrap the ingested content, but it lacks specific instructions or delimiters that warn the reading agent to ignore potential instructions embedded within those blocks.
Capability inventory: The script writes files to the local system. The output is intended for 'coding agents' which typically have capabilities for file system access and shell command execution, making them vulnerable to instructions found in the context bundle.
Sanitization: There is no evidence of sanitization, filtering, or escaping performed on the content loaded from the database before it is included in the output Markdown.
[DYNAMIC_EXECUTION]: The script employs dynamic path manipulation to handle its dependency imports.
Evidence: The script scripts/inspect_best_programs.py uses sys.path.insert(0, str(repo_root)) to dynamically add a computed repository root to the search path for the shinka.utils module.

shinka-inspect