Name

JIMS stands for "Just an Integrated Multiagent System"

Concepts

Thread

Thread is a conversation between user (or users) and agentic system Each interaction between agents and users happens in thread

Event

Something that happens in a thread Each event has id, type and data

Example:

{
    "event_id": "...",
    "event_type": "comm.user_message",
    "event_data": {
        "content": "Hello!"
    }
}

event_type and event_domain

event_type is structured as follows:

• <event_domain>.<event_name>[.<subname>...]

event_domain is the first segment and is used for routing and filtering. The rest of the string identifies the concrete event.

event_domain values include:

comm.* - user–assistant communication

Visible to the end user.

• comm.user_message: a user text message (input);
• comm.assistant_message: a text message produced by an agent/pipeline.

These events are:

• Exposed as CommunicationEvent (role, content).
• Collected into ThreadContext.history.

context.* - comm.* extended with pipeline context

Conversational AI Pipelines can benefit from preserving some extra data (context) about the conversation in between messages (previous reasoning, extra data, etc.). For this, we use a superset of comm. domain called context.. It includes everything in comm. (everything in ThreadContext.history), and adds context. events passed from Pipeline. context. events are internal and not sent/shown to the end user.

state.* - internal state of the pipeline

Used to persist some data between pipeline calls.

• ThreadContext.set_state(state_name, state) emits state.set.<state_name> with a dict payload.
• ThreadContext.get_state(state_name, state_type) scans events in reverse for state.set.<state_name> and validates the payload into a Pydantic model.

Define your own state schema inside the Pipeline

jims.* - lifecycle and system domain:

Lifecycle and system‑level events

Example:

• jims.lifecycle.thread_created: emitted when a thread is first created by ThreadController.new_thread(...).

Pipeline‑specific domains

Defined in a Pipeline, these events are produced via ThreadContext.send_event(...) and consumed elsewhere.

example: rag.*

• rag.query_processed
• rag.error