fix: context compaction not triggering with llama.cpp prompt caching

Context compaction failed to trigger for llama.cpp backends with prompt
caching enabled, letting chats grow until the server rejected the request
with a context overflow error. Two issues combined to cause this:

1. load_messages_from_db() dropped the usage field when reconstructing
   messages from the DB, so _exceeds_token_threshold() never saw any
   reported token counts and always fell back to rough char-based
   estimation.

2. The threshold check summed only input_tokens + output_tokens. llama.cpp
   reports cached prefix tokens separately (timings.cache_n, merged into
   the stored usage) and excludes them from prompt_n, so a mostly cached
   context of ~30k tokens was counted as only a few thousand.

Keep usage on DB-loaded messages so the threshold check can use it, count
cache_n towards the total, and strip usage in strip_compaction_fields()
after compaction runs so the non-standard field never reaches the LLM
payload.

Reported and verified in
https://github.com/open-webui/open-webui/discussions/24410

backend/open_webui/utils/context_compaction.py

@@ -231,8 +231,13 @@ def _exceeds_token_threshold(messages: list[dict], system_prompt: str, summary:
 
     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 usage.get('input_tokens'):
-            total = int(usage.get('input_tokens') or 0) + int(usage.get('output_tokens') or 0)
+        if isinstance(usage, dict) and (usage.get('input_tokens') or usage.get('cache_n')):
+            # llama.cpp reports cached prefix tokens separately (timings.cache_n), excluded from prompt_n
+            total = (
+                int(usage.get('input_tokens') or 0)
+                + int(usage.get('output_tokens') or 0)
+                + int(usage.get('cache_n') or 0)
+            )
             return total + _estimate_messages_tokens(messages[idx + 1 :]) > threshold
 
     estimated = _estimate_tokens(system_prompt) + _estimate_tokens(summary or '') + _estimate_messages_tokens(messages)

backend/open_webui/utils/middleware.py

@@ -1975,7 +1975,7 @@ async def load_messages_from_db(chat_id: str, message_id: str) -> Optional[list[
         return None
 
     return [
-        {k: v for k, v in msg.items() if k in ('role', 'content', 'output', 'files', 'contextSummary')}
+        {k: v for k, v in msg.items() if k in ('role', 'content', 'output', 'files', 'contextSummary', 'usage')}
         for msg in db_messages
     ]
 
@@ -2034,6 +2034,7 @@ def strip_compaction_fields(messages: list[dict]) -> list[dict]:
         clean = dict(message)
         clean.pop('contextSummary', None)
         clean.pop('context_summary', None)
+        clean.pop('usage', None)
         stripped.append(clean)
     return stripped
 
@@ -2222,7 +2223,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
                         {
                             k: v
                             for k, v in assistant_message.items()
-                            if k in ('role', 'content', 'output', 'files', 'contextSummary')
+                            if k in ('role', 'content', 'output', 'files', 'contextSummary', 'usage')
                         }
                     )