Agent

Every conversation the app's agent-chat interface has held is recorded here, and these two reads are how you work from what actually happened instead of a guess. threads list pages the conversations newest-activity-first, transcripts excluded. Each row carries who spoke, a message count, and two flags that mark a thread worth opening: a tool-error count, and whether a turn broke on a model error, a rate limit, or credits instead of replying. devSession separates your own tunnel test threads from real traffic.

threads get opens one conversation in full: every user message, every agent reply, and each tool the agent called, with the arguments it chose and what came back. A method tool call carries a request id, so its input, output, console output, and error are one requests get away; client tools run in the browser and carry none.

This is the primary loop for iterating on an agent: read the transcript, then fix the system prompt and the tool descriptions from the evidence.