The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill uses finding descriptions to guide code modifications, creating an attack surface for malicious data in the input JSON files.
Ingestion points: Ingests findings from .beagle/llm-artifacts-review.json and overrides from .beagle/llm-artifacts-verification.json.
Boundary markers: Lacks explicit delimiters or instructions for the agent to ignore potentially malicious embedded instructions in the metadata fields.
Capability inventory: The agent can write to the filesystem and execute various command-line utilities including linters and test suites.
Sanitization: Validates finding IDs and file existence but does not sanitize the text content of the descriptions.
[COMMAND_EXECUTION]: The skill runs multiple development and verification tools as part of its core logic.
Evidence: Invokes git for status and stashing; jq for JSON manipulation; and ruff, mypy, eslint, tsc, go, and pytest for code verification.
[DYNAMIC_EXECUTION]: The skill executes a dynamically generated Python script for data validation.
Evidence: Section 3 passes an inline Python script via a heredoc to the python3 interpreter to process, print, and lock IDs from the review JSON file for the anti-confabulation gate.

fix-llm-artifacts