The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from code diffs and issue/PR descriptions, creating an attack surface for instructions embedded in external content to influence the automated review process. Ingestion points: Code diffs (SKILL.md Step 5), issue content (SKILL.md Step 3), and PR metadata (queue-mode.md). Boundary markers: Diffs are handled via file isolation and processed by distinct reviewer LLM instances. Prompt instructions in reviewer-briefs.md mandate independent judgment and provide specific rubrics to reduce the impact of pre-judged findings in the input text. Capability inventory: The skill has the capability to merge PRs (rota ship pr-merge), record verdicts to local storage (rota verdict add), and post feedback to remote trackers (rota tracker call). Sanitization: No explicit content sanitization or escaping of processed diffs or issue text is mentioned.\n- [COMMAND_EXECUTION]: The skill invokes local CLI tools to manage the review lifecycle. Evidence includes frequent use of the rota toolset and git commands (e.g., rota review package, git checkout, rota tracker call) documented across SKILL.md and queue-mode.md.\n- [DATA_EXFILTRATION]: Automated review findings are transmitted to external hosting providers. The queue-mode.md file details the use of rota tracker call to post comments and notes to GitHub and GitLab. These operations target established, well-known services and are necessary for the skill's intended purpose.

rota-review