Runtime

One core, multiple surfaces

The MiniClaw Agent Core does not know which platform produced a message. TUI, Feishu, Telegram, and Discord convert external messages into an internal form before handing them to the same AgentRuntime. The model and Tool Loop therefore keep one mental model.

TUI / Feishu / Telegram / Discord
                  ↓
            AgentRuntime
                  ↓
      Provider → Policy → Tool
                  ↓
        result → original surface

Shared and isolated

The surfaces share AgentRuntime, while each platform keeps an isolated Transport, Delivery, Manager, queue, and runtime failure state. Reconnection or delivery trouble in one Channel must not mutate the state of another.

Provider boundary

The current scope uses one OpenAI-compatible Provider. A Provider owns model protocol behavior; it does not own Channel identity, Workspace authorization, or Tool safety policy.

Request order

1. A Channel or TUI accepts the message and resolves conversation identity.
2. The Agent decides whether a Tool is needed.
3. Policy checks workspace, network, or command boundaries.
4. When required, Approval binds the exact parameters.
5. The Tool acts and returns the result to the original conversation.

See the system architecture for module responsibilities.