Skip to content

Components

Claw-R1 components are organized around the data flow from agent interaction to training consumption. HTTP handles agent-facing traffic, while Ray actors handle stateful data and training coordination.

  • Gateway Server

    FastAPI service that receives white-box step submissions and black-box OpenAI-compatible chat traffic, then submits normalized Step records to DataPool.

    Gateway Server

  • DataPool

    Ray actor that stores, indexes, partitions, curates, and serves step-level data by channel and prompt group.

    DataPool

  • Dashboard

    Live UI for Agentic RL data lifestyle management: collection, representation, curation, optimization preview, and training consumption.

    Dashboard

  • Reward System

    Reward workers compute or attach quality signals from rule checks, reward models, generative judges, and human feedback.

    Reward System

  • Agent Flow

    Agent execution lifecycle. White-box flows submit steps explicitly; black-box flows wrap agents that only know an OpenAI-compatible base_url.

    Agent Flow

  • Async Training

    Separate Ray actors for rollout generation and policy training, coordinated through DataPool and parameter synchronization.

    Async Training

  • Prefix Tree Merge

    Shared-prefix packing for multi-step agent training. The dashboard can preview prefix-tree structure from real DataPool steps.

    Prefix Tree Merge

Data Flow

Black-box Agent         White-box Agent
      |                       |
      v                       v
   Gateway Server  <---- explicit Step APIs
      |
      v
   DataPool  <---- Dashboard reads stats, steps, events, curation, and prefix-tree previews
      |
      v
   Async Trainer ---- Parameter Synchronizer ---- Async Rollouter / vLLM

Lifecycle Mapping

Lifecycle stage Primary component Dashboard view
Collect interactions Gateway, Agent Flow Collection
Store representation DataPool Representation
Evaluate quality Reward System Curation Signals
Curate candidates DataPool curation APIs Curation
Optimize shared context Prefix Tree Merge preview Optimization
Serve training data DataPool, Async Trainer Consumption