The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data artifacts such as CLAIMS.md and MISSING_EVIDENCE.md to generate a final review report. This ingestion of untrusted data represents an attack surface where malicious content in the inputs could potentially influence the output report.\n
Ingestion points: The script scripts/run.py reads multiple markdown and TSV files from the output/ directory of the provided workspace.\n
Boundary markers: The script uses structured parsing for TSV and JSONL data, but lacks explicit delimiters or "ignore instructions" markers when handling text blocks from markdown files.\n
Capability inventory: The skill's capabilities are limited to reading and writing local text files. There are no network requests, shell command executions, or dynamic code evaluation routines present in the code.\n
Sanitization: The tool uses specific parsing logic (e.g., csv.DictReader and parse_item_blocks) which acts as a basic structural validation of the input data before rendering the final output.\n- [DYNAMIC_EXECUTION]: The execution script dynamically modifies the Python module search path (sys.path) by searching upwards from its own location for a marker file (AGENTS.md).\n
Evidence: The script scripts/run.py traverses up to 10 parent directories to identify the repository root and inserts this path into sys.path.\n
Purpose: This mechanism is used to enable the import of local helper modules in the tooling package relative to the repository structure.\n
Risk Assessment: While this pattern involves dynamic path resolution, it is used here for internal library discovery within a project structure and does not incorporate untrusted external input into the search path logic.

rubric-writer