From 4eda1f00dfdc7e2e673f8d397a83c2e2cc3296cc Mon Sep 17 00:00:00 2001 From: Classic298 <27028174+Classic298@users.noreply.github.com> Date: Mon, 12 Oct 2026 23:24:21 +0200 Subject: [PATCH] fix: context compaction never fires during a long tool run During a long run of tool calls, context compaction between tool calls judged the conversation's size from a rough character count and ignored the token count the provider reports after every model call. That count leaves out the model's reasoning and undercounts code and command output, so on a code-heavy run it read about half the real size, compaction never started, and the reply ended with the provider's context size error. Compaction now starts from the token count the provider reported for the last model call and only estimates the new tool results on top of it. Providers that report no token counts still get the estimate. Fixes #32306 --- backend/open_webui/utils/context_compaction.py | 10 ++++------ backend/open_webui/utils/middleware.py | 6 ++++++ 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/backend/open_webui/utils/context_compaction.py b/backend/open_webui/utils/context_compaction.py index 19d44f417d..d3955148cb 100644 --- a/backend/open_webui/utils/context_compaction.py +++ b/backend/open_webui/utils/context_compaction.py @@ -321,12 +321,10 @@ def _exceeds_token_threshold(messages: list[dict], system_prompt: str, summary: if threshold <= 0: return False - # Expanded tool histories include fresh results and rebuilt prompts absent from prior usage. - if not any(message.get('role') == 'tool' for message in messages): - for idx in range(len(messages) - 1, -1, -1): - usage = messages[idx].get('usage') or (messages[idx].get('info') or {}).get('usage') - if isinstance(usage, dict) and (tokens := _usage_token_count(usage)): - return tokens + _estimate_messages_tokens(messages[idx + 1 :]) > threshold + for idx in range(len(messages) - 1, -1, -1): + usage = messages[idx].get('usage') or (messages[idx].get('info') or {}).get('usage') + if isinstance(usage, dict) and (tokens := _usage_token_count(usage)): + return tokens + _estimate_messages_tokens(messages[idx + 1 :]) > threshold estimated = _estimate_tokens(system_prompt) + _estimate_tokens(summary or '') + _estimate_messages_tokens(messages) return estimated > threshold diff --git a/backend/open_webui/utils/middleware.py b/backend/open_webui/utils/middleware.py index 57ac2e7517..1adfba67fd 100644 --- a/backend/open_webui/utils/middleware.py +++ b/backend/open_webui/utils/middleware.py @@ -5917,6 +5917,7 @@ async def streaming_chat_response_handler(response, ctx): ) tool_call_sources = [] # Track citation sources from tool results all_tool_call_sources = [] # Accumulated sources across all iterations + attached_usage = None user_message = get_last_user_message(form_data['messages']) messages = list(form_data['messages']) @@ -6280,6 +6281,11 @@ async def streaming_chat_response_handler(response, ctx): image_urls.append(part.get('image_url', '')) message['content'] = ''.join(text_parts) + # A count reported for the last request covers everything before these results. + if usage is not attached_usage and tool_messages: + tool_messages[0]['usage'] = usage + attached_usage = usage + # Refresh citation context without restoring compacted history. user_prompt = get_last_user_message_item(form_data['messages']) for message in messages: