The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by scanning and refactoring entire codebases, which are external, untrusted sources of data. A malicious file within a repository could attempt to manipulate the agent's risk assessment or refactor logic via embedded instructions.
Ingestion points: The skill maps repository boundaries, including root paths, subtrees, submodules, and file contents as described in SKILL.md.
Boundary markers: The instructions mandate a "propose-and-approve protocol" and "explicit approval gates," requiring a human to review all changes before execution.
Capability inventory: The skill possesses the capability to suggest and execute repository-wide edits as defined in the instructions.
Sanitization: The skill relies on "before/after diff previews" and "risk annotations" to surface changes for human verification, though it does not specify automated sanitization of ingested content.

codebase-orchestrator