The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests source code from a repository, which is considered untrusted data. It implements robust mitigations by including a mandatory notice field in its JSON output and a preamble in Markdown output that explicitly warns the agent to ignore any instructions found within the source code slices. The coordinator agent is further instructed to validate all worker subagent claims against the original packet.\n- [COMMAND_EXECUTION]: The skill uses uv run to execute a bundled Python script (scripts/build_slice_packet.py) for deterministic context selection. The script includes a safe_source_path function that resolves and validates file paths, preventing directory traversal attacks by ensuring the tool only reads files within the intended target root directory.

slicing-code-context