shunk031-herdr-orchestrate-workers
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes the
herdrCLI to manage agent lifecycles, workspace creation, and inter-agent communication. It also executes a local bash utility,scripts/herdr-orchestrator-observer.sh, which monitors worker progress by reading transcripts and lifecycle state. - [INDIRECT_PROMPT_INJECTION]: Because the orchestrator agent ingests and processes reports from independent worker agents, a surface for indirect injection exists. The skill mitigates this by instructing the agent to build prompts in quoted shell variables to prevent unintended shell expansion of worker-provided content and by mandating independent peer reviews for all worker deliverables.
- [DATA_EXFILTRATION]: The skill includes a specific safety rule in Step 10 that prohibits the inclusion of internal infrastructure details, such as internal hostnames, IP addresses, or mount paths, in any repository artifacts or messages.
- [PROMPT_INJECTION]: A minor behavioral directive is present at the beginning of the instructions, requiring the agent to output a specific emoji and confirmation string upon reading the skill file to verify compliance.
Audit Metadata