The Agent Skills Directory

[SAFE]: The skill implements a robust 'verify-then-report' pattern for all findings generated by external LLM providers. Findings are explicitly labeled as CONFIRMED or REJECTED only after the host session validates them against the actual source code, ensuring that model hallucinations or malicious suggestions are caught during the review process.
[COMMAND_EXECUTION]: The skill makes extensive use of CLI tools including git, gh, and grok for repository analysis and PR management. These operations are scoped to read-only actions for information gathering (diffing, viewing PRs) and reporting (updating PR descriptions). The inclusion of a watchdog timer in parallel executions in the council action demonstrates defensive programming to prevent resource exhaustion from hung processes.
[INDIRECT_PROMPT_INJECTION]: The skill explicitly recognizes the risk of malicious instructions embedded in the code diffs or PR descriptions it processes. It instructs the agent to treat delegate output as 'untrusted input' and provides a mandatory verification step for every finding. This architecture directly addresses the risk of an attacker attempting to compromise the agent through crafted code review targets by placing the local session as a verification gate.

check