blacksmith-orchestrate
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill performs several shell-based operations using git to manage parallel worktrees, including 'git rev-parse' to locate configuration directories, and 'git fetch', 'git merge-base', and 'git rebase' to verify and synchronize code between branches during the relay phase.
- [PRIVILEGE_ESCALATION]: The 'afk' flag enables a high-autonomy mode where the agent can autonomously perform git commits and pull request merges. While this is documented as an opt-in feature with a 5-minute inactivity timeout and a 7-point verification checklist (verifying CI status, goal completion, and review convergence), it grants the AI agent significant write access to remote repositories and the ability to bypass per-action human review of code changes.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from external tracker containers (milestones, epics, labels), implementation plans, and free-form user goals which could contain malicious instructions designed to influence the orchestrator's behavior.
- Ingestion points: Data is pulled from issue trackers (GitHub, GitLab, Jira, Linear) in Step 2, and from plan files or free-form text in Step 3.
- Boundary markers: The primary boundary is the Step 5 'Battle Plan' gate, which requires a user to review and approve the task decomposition and analysis before any implementation dispatch occurs.
- Capability inventory: The skill possesses the ability to spawn sub-agents, write to the local filesystem (ledger and worktrees), and perform remote git operations (push, merge).
- Sanitization: The skill uses structured JSON schemas to validate sub-agent responses and implements triage logic to detect mismatches between task descriptions and file counts, though it does not explicitly sanitize prose descriptions before passing them to sub-agents.
- [DYNAMIC_EXECUTION]: The skill executes a JavaScript Workflow script ('scripts/scout-fanout.mjs') to perform parallel analysis of tasks. This script utilizes the 'agent()' and 'parallel()' primitives to dispatch sub-agents and process their structured output according to a defined schema.
Audit Metadata