AutoSwarm-Shadow-Adapter
Zero-modification Human-in-the-Loop adapter for OpenOPC's agentic DAG runtime.
What is AutoSwarm-Shadow-Adapter?
Zero-modification Human-in-the-Loop adapter for OpenOPC's agentic DAG runtime.
Overview of AutoSwarm-Shadow-Adapter
What it is
AutoSwarm-Shadow-Adapter is a non-blocking Human-in-the-Loop (HITL) and Bring Your Own Compute (BYOC) adapter for the AutoSwarm agentic DAG runtime. It allows AutoSwarm DAGs to pause execution when a task requires human approval or remote GPU execution, without causing engine timeouts or blocking parallel branches. The adapter intercepts task execution, parks the task in an isolated SQLite database, and returns control to the engine immediately. It also provides a React-based human portal for review and submission of deliverables, and supports remote worker nodes that can execute tasks using their own compute resources. The project is built for AutoSwarm and requires zero modifications to the core AutoSwarm engine.
Key features
- Non-blocking intercept: Intercepts AutoSwarm task execution in under 50ms and returns control to the engine immediately, preventing timeouts.
- Isolated persistence: Task state is stored in a separate SQLite WAL database, ensuring resilience across system restarts.
- Multi-user access control: Supports multiple users with JWT authentication and a queue-based system for human review.
- File attachments: Allows up to 5 files per submission with a total payload limit of 50MB.
- Audit log compliance: Provides an immutable timeline with user attribution for all actions.
- Iterative rework loop: Supports a built-in
rework_requestedstate transition for iterative human-AI collaboration. - Remote worker support: Enables remote nodes to execute tasks using their own compute (e.g., local GPU, private database) via the ShadowWorker CLI or Python SDK.
Use cases
- Human-in-the-loop approval workflows where a human must review and approve AI-generated outputs before they proceed downstream.
- Remote compute execution where tasks are offloaded to remote machines with specialized hardware (e.g., GPU) or private data sources.
- Multi-agent DAGs that require parallel human and AI contributions, such as document review, code review, or legal counsel.
- Corporate knowledge management where deliverables (PDFs, CSVs, code files) are stored in a centralized artifact registry and injected into downstream tasks.
Getting started
See the repository README for detailed installation and setup instructions.
License and pricing
License: MIT. Pricing: open source (no cost).
Based on the project's public README and repository metadata.