The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted external sources, specifically 'referee reports,' 'audit findings,' and 'AI reviewer' comments. Although the instructions emphasize verification, an attacker providing a report could attempt to inject malicious instructions that influence the agent's behavior during the adjudication process.
Ingestion points: Incoming review findings, referee reports, code reviews, and audit findings (SKILL.md).
Boundary markers: Present. The skill includes explicit warnings to treat findings as 'candidates' and 'hypotheses' rather than verdicts, and warns against 'poisoned fixes.'
Capability inventory: The agent has access to Bash, Write, Read, Grep, Agent, and Task tools.
Sanitization: The skill mandates 'Mechanical checks' and verification against the actual source code, which serves as a procedural sanitization layer.
[DYNAMIC_EXECUTION]: The skill directs the agent to verify findings by executing the code being reviewed.
Evidence: Section 2 ('Mechanical checks beat opinion') instructs the agent to 'execute the consuming code' and 'run the identity on a small adversarial case' to confirm the validity of a finding. This is a standard procedure for code review but involves the dynamic execution of reviewed artifacts.
[COMMAND_EXECUTION]: The skill uses the Bash tool to perform verification tasks and mechanical checks on code and artifacts. This capability is necessary for the skill's primary purpose but represents a tool that can be used for execution.

adjudicate-review