The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from Git diffs and repository files, creating a surface for indirect prompt injection attacks.\n
Ingestion points: SKILL.md (Section 2) and references/git_operations.md describe reading staged changes, worktree files, and historical Git blobs.\n
Boundary markers: The instructions do not explicitly require the agent to wrap untrusted data in XML tags or similar delimiters with "ignore instructions" warnings when processing file content.\n
Capability inventory: The skill utilizes Git commands and file system navigation tools (rg, codemap-search).\n
Sanitization: No specific sanitization or filtering of input code is mentioned before analysis.\n- [EXTERNAL_DOWNLOADS]: The evaluation viewer evals/results/review.html references the SheetJS library from a Content Delivery Network (CDN).\n
Evidence: <script src="https://cdn.sheetjs.com/xlsx-0.20.3/package/dist/xlsx.full.min.js" ...>\n
Context: This is used for rendering spreadsheet data within the local evaluation report viewer and originates from a well-known service provider.\n- [COMMAND_EXECUTION]: The skill package includes Python scripts for building evaluation fixtures that execute Git commands via the subprocess module.\n
Evidence: evals/build_fixtures.py uses subprocess.check_output and evals/verify_git_examples.py uses subprocess.run to automate Git repository initialization and commits.\n
Context: These are development and testing utilities used to create synthetic repositories for regression testing.

code-review