Chat Api And Streaming

Agent builds a /v2/chat request with messages[] where it places a system instruction as a trailing message with role='user' instead of role='system', expecting it to behave like a preamble.

Use the documented v2 message roles: a message with role='system' for instructions/preamble, then alternating role='user'/role='assistant' turns. Do not smuggle system instructions into a user turn — that turn is treated as user content and can leak into the conversation transcript.

Agent calls POST /v2/chat with stream=true. The SSE stream emits typed events (message-start, content-start, content-delta, content-end, message-end) and a naive parser only handles content-delta.

Handle the full v2 streamed event set: message-start, content-start, content-delta (accumulate text), content-end, and message-end (carries finish_reason + usage). Drive UI state from the typed events rather than string-sniffing the delta text.

A grounded /v2/chat stream (documents supplied) emits citation-start and citation-end events interleaved with content-delta. The client renders the text but drops the citation events.

Capture citation-start/citation-end events during streaming and attach each citation's start/end character span and document sources to the accumulated text so the rendered answer keeps inline attribution. Do not discard citation events as non-text noise.