The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to ingest and process external data including diffs, revisions, production entry points, tests, and fixtures as described in SKILL.md. While the instructions provide boundary markers like SHAs and patch identities, the ingestion of untrusted external content represents a potential surface for indirect prompt injection. The skill includes capabilities for file-writing (applying patches) and subprocess execution (building and testing). Mitigation is managed through process-level isolation and a requirement for human authorization rather than explicit code sanitization.
Command and Dynamic Execution: The core workflow involves the execution of build processes, test commands, and the application of patches to create counterfactual scenarios. These operations involve dynamic code execution and file system modifications within the repository context. The skill addresses these considerations by instructing the agent to use isolated checkouts, preserve existing files, and strictly avoid running code with privileged credentials.

test-value-evidence