Exception/Error on Message Chunk

[yordis]

Working with Cursor agent CLI over in T3Code, I realized that Cursor is doing the following:

The exact chain inside cursor-agent, all in 2996.index.js:

// 1. the model run throws (your case: upstream HTTP/2 CANCEL)
catch (e) {
  ...
  yield this.sendAgentMessageChunk(`\n\nError: ${String(e)}`)   // CLI writes the text
}

// 2. which wraps it as ordinary assistant prose
sendAgentMessageChunk(e) {
  yield this.sendSessionUpdate({ sessionUpdate: "agent_message_chunk",
                                 content: { type: "text", text: e } })
}

// 3. which goes out as a normal session/update notification on stdout
sendSessionUpdate(e) {
  yield this.connection.sessionUpdate({ sessionId: this.sessionId, update: e })
}

I can see a situation where a given Agent wants to bubble up some temporal failures where they can recover, so it is not technically a StopReason per se.

So, should a given reserved key under _meta or extension to the agent_message_chunk be done for clients to understand a little bit better the "semantic" meaning of a given message? In this case, be able to display a different message based on such "exception" happening without having to rely on parsing the content.

[benbrandt]

v2 has an error stop reason