Skip to content

A compaction never cuts between a tool call and its results, a summary that fails leaves the conversation as it was, and compacted keeps a large tool result as a blob

Decision 774 · accepted · 2026-10-04 · issue #399

split_for_compaction moved its cut back to the nearest user message, and a tool-results message is a user message, so in a turn of many calls the part it kept could start with results whose call had gone into the summary, while the summariser's own request ended on that call with no results after it. Anthropic's and OpenAI's APIs refuse both with a 400 that is not a context overflow: the summary failed, and every later request carried the same results without their call and failed too, so a session on those endpoints stalled until a person started another. A gateway that does not check passed it, which is how it went unnoticed. - The cut still moves back to the person's message a reply answers, but never onto tool results; results at the head of what would be kept go into the summary with their call. That is the old rule without its fault: an input stays with the replies to it, a compaction with nothing older to summarise does not happen, and a turn of nothing but calls can still be compacted. Cutting back past the results to their call keeps the pair too, but can leave nothing to summarise but the input or the last summary, a call that saves nothing. Cutting only at a person's input means a turn of nothing but calls, the turn #389 is about, could never be compacted, and an overflow in it would end the turn. - The conversation is replaced when the summary arrives, not when it is asked for. Nothing is added to the conversation while an agent compacts (input waits), so apply_compaction takes the same split the summariser was sent. A summary that fails, or a cancel while it is asked for, leaves the conversation whole, which is what a replay rebuilds, since nothing was logged; before, memory held only the kept part and a restart had everything. The next request may not fit, and the provider says so (659). - compacted logs what it keeps as tool_results logs results: a tool result over 16 KiB is a blob reference, resolved on replay, as PROTOCOL.md's rule for large payloads says. A log written before has its results inline, and they resolve to themselves.

Governs: apps/troupe_core/lib/troupe/agent/server.ex