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.

A /v2/chat response ends with finish_reason='MAX_TOKENS' and an obviously truncated final sentence. The agent presents it as a complete answer.

Inspect finish_reason on message-end (e.g., COMPLETE, MAX_TOKENS, STOP_SEQUENCE, TOOL_CALL). On MAX_TOKENS, mark the output partial and either raise to the caller or continue by appending the truncated assistant turn — never surface a truncated answer as complete.

A /v2/chat response carries a usage object (billed_units / tokens with input and output counts). The agent estimates cost from response string length instead of reading usage.

Read input and output token counts from the response usage object for cost telemetry. Do not estimate tokens from character length. Aggregate usage from message-end when streaming.