Messages
?Send a message
Send a message into a chat and wait for the answer. This is the same conversation the browser runs, through the same service. Nothing about how the reply is decided changes: the response policies are read from the chat, the decision model is asked what it needs, and the assistant answers under whatever those policies told it. A chat started inside a workspace is read the same way, with that workspace's policies alongside its own.
Only how the answer travels is different. In the app a reply is a stream of deltas pushed to a room; a REST caller has nowhere to be pushed to, so the request waits for the finished message.
/api/v1/chats/:chatId/messagesThe answer
Answers 201 with the three things a caller asked for: what was stored, what came back, and what shaped it.
Decision turns
The turn is what makes this API worth calling over a plain chat endpoint. It is what the decision model estimated for the message, before the assistant answered it, and it is stored per chat -- so a reply can always be read against the estimates that shaped it.
Each estimate is one decision policy's answer: the value it came down hardest on, that value's rate, and the rate for every value the policy defines. Those rates are a named probability each, not a share of a total, so they do not sum to one.
When a reply does not come
If the message was stored but the model produced nothing, the request answers 502 with the reason in error. The message is already stored at that point, so read it back before retrying: sending the same text again sends it twice.
Message errors
To watch the conversation grow as the browser sees it, rather than polling, read it back with Get a chat.