The Agent Skills Directory

[PROMPT_INJECTION]: The skill manages a workflow where an orchestrator agent ingests data from subagents via logs, artifacts, and return values (SKILL.md). This creates an indirect prompt injection surface. The orchestrator uses a specific dispatch template to bound subagent roles and prevent unauthorized actions. Capabilities include shell execution for verification and Git operations. The skill mandates 'Iron Laws' requiring the orchestrator to verify all claims against ground-truth files on disk, such as checking 'git log' and re-running verification commands, to mitigate risks from subagent hallucinations or malicious instructions.\n- [COMMAND_EXECUTION]: The skill utilizes shell commands for node verification and repository state management (e.g., 'git status', 'git log'). These operations are restricted to a dedicated '.harness' directory and the project's worktree, following a least-privilege model for subagent dispatch.

workgraph