The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data including repository diffs, source code, production entry points, and fixtures. Ingestion points occur during the discovery phase and when binding contracts to specific revisions. While explicit text delimiters for this data are not defined, the instructions mitigate risk by mandating the use of 'owned isolated checkouts' and 'experiment directories' for all operations. The skill possesses capabilities for file patching and command execution, but these are gated by authorization requirements. No explicit sanitization of the ingested data is described, but the provided safety instructions for isolated execution are appropriate for the intended testing use case.
[COMMAND_EXECUTION]: The skill requires the execution of repository-specific toolchains, build commands, and test runners to generate 'red-green' evidence. These operations are explicitly restricted to 'authorized implementation/verification environments' and must be performed in isolated directories to prevent interference with other sessions or the main checkout. The instructions also emphasize preserving the integrity of protected sources and removing only owned temporary resources after execution.

test-value-evidence