Chat and Responses

Run threaded conversational execution through SOIT agents and response projections.

Chat is the user-facing conversation surface. Responses are the runtime projection that turns agent execution, streaming events, tool calls, citations, and trace state into a consumable interaction model.

When to use chat and responses

Use chat for human-in-the-loop interaction with agents. Use responses when an application needs a structured API surface for agent output, streaming, citations, and execution status.

The conversation model

A thread is the workspace-scoped session container. It carries the conversation's agent binding, model defaults (temperature, max tokens, context window), knowledge and tool configuration, and message counters.
Thread messages form an append-only ledger with a strict sequence_no per thread. Each message records its role, content, model usage (prompt/completion tokens, finish reason), citations, attachments, and tool calls.
A response links user input to agent runtime execution and projects the execution state — streaming events, tool call progress, citations — into a consumable shape.
Behind every assistant message sits a run: messages link to run_id, task_id, and response_id, so a chat answer is always traceable to its full execution evidence.

Streaming and evidence

Responses stream over SSE or WebSocket: response events arrive as the run progresses, tool calls surface as they execute, and citations attach as retrieval completes. When the turn finishes, the run detail page shows the same events as a permanent replayable record — the stream and the ledger are two views of one execution.

How they connect

Threads execute through the bound Agent's published version.
Citations connect response output to Knowledge sources.
Tool calls flow through the same governed routing as every capability — see Plugins.
Runs and run steps preserve the trace behind each response in Tasks and Runs.

API and UI surfaces

API prefixes: /api/v1/threads (sessions and messages), /api/v1/responses (execution projections)
Related prefixes: /api/v1/agents, /api/v1/runs
Web route: /chat