The Agent Skills Directory

[DYNAMIC_EXECUTION]: The workflow engine executes scripts using a new Function constructor within an asynchronous sandbox. This is the primary mechanism for orchestration. The skill provides significant security mitigations by withholding sensitive Node.js host APIs (such as process, fs, and require) from the orchestrator context and including dedicated scripts (validate-workflow.mjs and dry-run.mjs) to verify script integrity and simulate behavior before live execution.
[PROMPT_INJECTION]: The skill defines a detailed 'Execution Standard' that establishes a 'hostile work' doctrine. This instructs AI agents to treat all artifacts as suspect and conduct adversarial reviews, effectively serving as a built-in defense against both direct and indirect prompt injection attempts within the workflow pipelines.
[DATA_EXPOSURE]: The workflow runtime implements a deterministic 'scratch' directory system (under .claude/scratch/) for managing intermediate reports and receipts. This system is carefully designed to be repo-relative and gitignored, ensuring that concurrent runs do not clobber each other's data while supporting the engine's resume and recovery capabilities.
[COMMAND_EXECUTION]: While the orchestrator script itself is sandboxed, the subagents spawned via the agent() global are designed to use standard agent tools for shell and filesystem operations. This is the intended architecture of the workflow system, where deterministic logic resides in the orchestrator and task execution resides in the agents.

workflow-creator