How do I replay a failed agent session turn by turn?
When an agent’s tab goes idle and misses a run of steps, you recover them the same way you replay any failed turn: feed the model the exact message history and tool outputs it had at that point, in order, instead of letting live tools answer again. Capture three things per turn: the messages up to that point, the tool call and its recorded output, and the configuration, model, prompt version, settings, that was active. Replay the turns in order, holding tool outputs fixed, so each step sees exactly what the original run saw. If a step calls a tool again instead of reading the recorded result, you’re not replaying, you’re running a new session that happens to start the same way and can diverge from turn one. Doing it turn by turn, rather than replaying the whole session as one unit, lets you stop at the turn where things went wrong and inspect the model’s state right there.