Open WebUI Clarifies Context Compaction Limits: A Huge Single-Turn Reply Can Still Push a Conversation Beyond the Window

In a technical discussion on October 20, Open WebUI clarified the limits of automatic context compaction: exceeding the configured token threshold does not guarantee that a summary will be generated. Community reporters said tool-heavy agent tasks can store extensive output in a single assistant message, leaving a conversation with few messages whose actual content has already exceeded the model’s context window. Project members reclassified the report as a feature request, explaining that the current behavior matches the documentation. Discussion and project response

The report concerns two checks: compaction is skipped when there are three or fewer messages, and the cutoff for retaining recent content is chosen by message count rather than by each message’s share of the tokens. As a result, even when a longer conversation is summarized, retaining one enormous tool reply may still push the next request over the limit. The reporter said they observed this in a deployment running 0.12.0 and provided reproduction steps that do not require calling a model; these observations should not be taken as evidence of universal impact across all versions. Report details

Version differences are key. Project members said that starting with 0.12.1, the native tool loop also checks and compacts context between tool turns, reducing the chance that the entire tool history accumulates into a giant reply. But oversized existing messages face another limitation: the summarization request itself must fit the full text to be summarized, so manually running /compact may also fail. Members suggested that affected users edit the first message and resubmit it to create a new branch without the huge reply, or start a new conversation. Workarounds

The official documentation states that compaction is disabled by default and, at the start of a turn, requires a suitable user message to be found as the cutoff; if none is found, the full history is sent. Engineering teams should therefore treat the threshold as a condition for starting a summary and separately validate the size of the final request, especially with workloads that produce extensive tool output in a single turn. For input-size limits that must be enforced, the documentation recommends implementing a trimming policy with a filter Function. Context management documentation

The community proposed a fix that adjusts the cutoff based on token proportions, but that PR was closed without being merged and should not be considered an official fix. The next points to watch are whether the project adopts this kind of splitting strategy and whether existing oversized messages get a more direct recovery path. Proposed fix