Does Tessary's frustration classifier see the whole conversation, or just the last few turns?

Just the last two turns before the one it’s scoring, not the whole thread. The classifier sends its decision model the scored message plus the four messages from the two turns right before it, using only each turn’s actual reply and skipping the tool calls inside those turns and any sub-agent’s own output. A fix this month closed a real gap in that: the old version scanned a fixed batch of recent spans instead of counting turns, so a turn packed with tool calls could push the real replies out of the window entirely and leave a message scored with too little context, or none at all.

The tradeoff is reach, not accuracy. Frustration that surfaced several turns back and hasn’t recurred since won’t feed into the current turn’s score, the same kind of context boundary that trips up an agent losing track of instructions given much earlier in a conversation. Each new turn gets scored again, though, so frustration that keeps recurring stays visible as the conversation continues.