loopd

loopd is an orchestration runtime for collaboration among humans, Harnesses, and Kubernetes Operators.

Its guiding idea is:

Loop = Resource(spec + status) + Reconcile

Every chat request creates a small loopd Task CRD that wakes the selected responder. A simple Operator can reconcile that Task directly; a complex Operator may create domain CRDs for its own state and completion semantics. loopd keeps visible collaboration in conversations and messages, while full execution history belongs to the Harness and AgentLedger.

Components

• loop-server is a Hertz service that owns page-visible conversations and messages. Before committing a chat request, it also creates the Task CRD; failure rolls the messages back.
• loop-runtime is a Go client embedded in an Operator. It exposes stable conversation and message capabilities through r.Loop.Chat, and lets a Reconciler watch and resolve Tasks through r.Loop.Task.
• Harness adapters connect loop-server to agentd or another intelligent execution service without leaking provider vocabulary into the public model.

AgentUE is used for the page-visible message model. AgentLedger records complete prompts, model events, tool calls, retries, and costs; it is not the chat database.

The public conversation roles are always:

user | harness | operator

Long-running execution

A question may run for minutes or days. Its lifecycle is independent from any HTTP request or browser connection:

1. loop-server creates a user message and an empty responder message with one task_id, then creates a same-ID Task CRD before committing the messages;
2. the selected Operator or Harness watches the Task and resolves its current input and conversation history through loop-runtime;
3. a complex Operator may create a domain CRD, while a simple Operator handles the shared Task directly;
4. visible progress updates the responder message while full events enter AgentLedger;
5. clients may disconnect and reload the current message snapshot;
6. the selected Harness or Operator finishes the same responder message.

Harness.Prompt returns a handle. A Reconciler can inspect it and return, or call Wait when waiting is genuinely the only remaining work. Both paths use the same (owner UID, effect key) and therefore the same external execution.