refactor-board
Pass
Audited by Gen Agent Trust Hub on Jul 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute local Node.js scripts (
orchestrate.mjs,board.mjs) located in a specific directory ($HOME/.claude/skills/refactor-chain/scripts). These scripts manage the planning, recording, and aggregation of audit findings. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it audits external codebases. This risk is managed through a structured approach:
- Ingestion points: The
board-plancommand takes a--target <dir>argument, where untrusted codebase content is ingested into the agent context via subagent finders. - Boundary markers: The system uses explicitly defined 'finder prompts' which include a mandatory rubric and output contract to constrain the behavior of subagents.
- Capability inventory: The skill primarily performs file reads and script execution via Node.js for orchestration, but does not perform direct file writes or network operations.
- Sanitization: Every finding must undergo 'adversarial verification' where a separate Coordinator persona attempts to refute the Lead finder's evidence before it is recorded.
Audit Metadata