fix: chat stays stuck on the context size error after a long tool run

When context compaction summarized part of a long run of tool calls, that summary was thrown away once the reply finished. The next message then sent the whole run again, every tool result included, and failed with the provider's context size error on every retry, so the chat stayed stuck. The summary is now kept with the reply, and the next message picks up from it instead of resending the whole run.

Fixes #32306

backend/open_webui/utils/context_compaction.py

@@ -305,8 +305,13 @@ def _build_context_usage(tokens: int, threshold: int) -> dict:
 def _apply_latest_summary_checkpoint(messages: list[dict]) -> tuple[list[dict], str | None]:
     summary = None
     summary_idx = None
+    cut_idx = None
+    kept_output = None
 
     for idx, message in enumerate(messages):
+        if checkpoint := _find_tool_loop_checkpoint(messages, idx):
+            summary, cut_idx, kept_output = checkpoint
+            summary_idx = cut_idx - 1
         value = message.get('contextSummary') or message.get('context_summary')
         if isinstance(value, str) and value.strip():
             summary = value
@@ -314,7 +319,36 @@ def _apply_latest_summary_checkpoint(messages: list[dict]) -> tuple[list[dict],
 
     if summary_idx is None:
         return messages, None
-    return messages[summary_idx:], summary
+    return [
+        {**message, 'output': kept_output} if idx == cut_idx else message
+        for idx, message in enumerate(messages)
+        if idx >= summary_idx
+    ], summary
+
+
+def _find_tool_loop_checkpoint(messages: list[dict], idx: int) -> tuple[str, int, list[dict]] | None:
+    """Tool loop summary saved in this reply, with the reply it cut and that reply's remaining output."""
+    output = messages[idx].get('output') or []
+    marker = next((item for item in reversed(output) if item.get('type') == 'open_webui:context_compaction'), None)
+    if not marker:
+        return None
+    # The loop can cut into an earlier reply when this one started near the threshold.
+    for cut_idx in range(idx, -1, -1):
+        cut_output = messages[cut_idx].get('output') or []
+        start = next(
+            (
+                position
+                for position, item in enumerate(cut_output)
+                if item.get('type') == 'function_call' and item.get('call_id') == marker['call_id']
+            ),
+            None,
+        )
+        if start is not None:
+            while start and cut_output[start - 1].get('type') != 'function_call_output':
+                start -= 1
+            kept_output = [item for item in cut_output[start:] if item.get('type') != 'open_webui:context_compaction']
+            return marker['summary'], cut_idx, kept_output
+    return None
 
 
 def _exceeds_token_threshold(messages: list[dict], system_prompt: str, summary: str | None, threshold: int) -> bool:

backend/open_webui/utils/middleware.py

@@ -6326,6 +6326,20 @@ async def streaming_chat_response_handler(response, ctx):
                                         idx for idx, message in enumerate(messages) if message.get('role') != 'system'
                                     )
                                     messages[checkpoint] = {**messages[checkpoint], 'contextSummary': context_summary}
+                                    kept_call_id = next(
+                                        message['tool_calls'][0]['id']
+                                        for message in messages
+                                        if message.get('tool_calls')
+                                    )
+                                    output.append(
+                                        {
+                                            'type': 'open_webui:context_compaction',
+                                            'id': output_id('cc'),
+                                            'call_id': kept_call_id,
+                                            'summary': context_summary,
+                                        }
+                                    )
+                                    await emit_output()
                             except Exception:
                                 log.exception('Tool loop context compaction failed; keeping the current context')