The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process findings from potentially untrusted external sources, such as other AI models or referee reports, which could contain malicious or incorrect instructions.
Ingestion points: Incoming findings from AI reviewers, referees, or linters described in the metadata and core instructions.
Boundary markers: The skill emphasizes a "do not ingest" policy, requiring findings to be treated as hypotheses until verified against the source.
Capability inventory: The skill uses high-privilege tools including Bash, Write, and Agent to verify and apply fixes.
Sanitization: Prescribes mechanical checks, grepping, and source-code verification to ensure findings are valid before any changes are committed.
[DYNAMIC_EXECUTION]: Section 2 ("Mechanical checks beat opinion") encourages the agent to execute code to verify review findings.
Evidence: Instructions to "execute the consuming code" and "run the identity on a small adversarial case" involve runtime execution of project code to validate logic.
[COMMAND_EXECUTION]: The skill utilizes shell-based tools to perform audits and verification steps.
Evidence: Inclusion of Bash and Grep in allowed-tools to facilitate mechanical checks such as "grep for the symbol" and running tests.

adjudicate-review