From cb5bb1e7b0fdf897dc8b5ca423b69820f33b977d Mon Sep 17 00:00:00 2001
From: Classic298 <27028174+Classic298@users.noreply.github.com>
Date: Thu, 2 Jul 2026 21:22:48 +0200
Subject: [PATCH] fix: preserve system prompt across context compaction

Context compaction dropped the user's system prompt and replaced it with the summary. compact_messages_for_request() treated messages[0] (the chat system prompt, re-prepended in middleware) as ordinary history: _find_compaction_boundary always folded it into the compacted slice, and _apply_latest_summary_checkpoint's messages[summary_idx:] slice cut it off on every subsequent turn. The function then returned only recent_messages with no system role, so add_or_update_system_message(..., append=True) had nothing to append to and inserted a summary-only system message instead. The system_prompt argument was used purely for token estimation, never to exempt it from compaction.

Pop a leading system message on entry and re-prepend it on every return path, so the summarizer never sees it and append=True appends the summary to the real system prompt as intended.

Fixes #26710
---
 backend/open_webui/utils/context_compaction.py | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/backend/open_webui/utils/context_compaction.py b/backend/open_webui/utils/context_compaction.py
index e20f6b9cdc..7affb4b3e3 100644
--- a/backend/open_webui/utils/context_compaction.py
+++ b/backend/open_webui/utils/context_compaction.py
@@ -52,16 +52,24 @@ async def compact_messages_for_request(
     if not config['enable']:
         return messages, None, False
 
+    # Exempt a leading system message from compaction; re-prepend on every return.
+    system_message = None
+    if messages and messages[0].get('role') == 'system':
+        system_message, messages = messages[0], messages[1:]
+
+    def _restore(msgs: list[dict]) -> list[dict]:
+        return [system_message, *msgs] if system_message else msgs
+
     messages, previous_summary = _apply_latest_summary_checkpoint(messages)
     token_threshold = _resolve_token_threshold(config['token_threshold'], metadata)
     if not _exceeds_token_threshold(messages, system_prompt, previous_summary, token_threshold) or len(messages) <= 3:
-        return messages, previous_summary, False
+        return _restore(messages), previous_summary, False
 
     boundary = _find_compaction_boundary(messages)
     compacted_messages = messages[:boundary]
     recent_messages = messages[boundary:]
     if not compacted_messages or not recent_messages:
-        return messages, previous_summary, False
+        return _restore(messages), previous_summary, False
 
     event_emitter = None
     if metadata.get('chat_id') and metadata.get('message_id'):
@@ -138,7 +146,7 @@ async def compact_messages_for_request(
             }
         )
 
-    return recent_messages, summary, True
+    return _restore(recent_messages), summary, True
 
 
 async def compact_chat_branch(request, user, chat: Any, model_id: str, models: dict) -> dict: