fix(async-context-compression): strengthen summary path robustness

- Add comprehensive error logging for LLM response validation failures
- Thread __request__ context through entire summary generation pipeline
- Load and merge previous_summary from DB when not in outlet payload
- Use real request object instead of minimal synthetic context

plugins/filters/async-context-compression/README.md

@@ -1,13 +1,14 @@
 # Async Context Compression Filter
 
-**Author:** [Fu-Jie](https://github.com/Fu-Jie/openwebui-extensions) | **Version:** 1.4.1 | **Project:** [OpenWebUI Extensions](https://github.com/Fu-Jie/openwebui-extensions) | **License:** MIT
+**Author:** [Fu-Jie](https://github.com/Fu-Jie/openwebui-extensions) | **Version:** 1.4.2 | **Project:** [OpenWebUI Extensions](https://github.com/Fu-Jie/openwebui-extensions) | **License:** MIT
 
 This filter reduces token consumption in long conversations through intelligent summarization and message compression while keeping conversations coherent.
 
-## What's new in 1.4.1
+## What's new in 1.4.2
 
-- **Reverse-Unfolding Mechanism**: Accurately reconstructs the expanded native tool-calling sequence during the outlet phase to permanently fix coordinate drift and missing summaries for long tool-based conversations.
+- **Enhanced Summary Path Robustness**: Thread `__request__` context through entire summary generation pipeline for reliable authentication and provider handling.
-- **Safer Tool Trimming**: Refactored `enable_tool_output_trimming` to strictly use atomic block groups for safe trimming, completely preventing JSON payload corruption.
+- **Improved Error Diagnostics**: LLM response validation failures now include complete response body in error logs for transparent troubleshooting.
+- **Smart Previous Summary Loading**: Automatically load and merge previous summaries from DB when not present in outlet payload, enabling incremental state merging across summary generations.
 
 ---
 

plugins/filters/async-context-compression/README_CN.md

@@ -1,15 +1,16 @@
 # 异步上下文压缩过滤器
 
-**作者:** [Fu-Jie](https://github.com/Fu-Jie/openwebui-extensions) | **版本:** 1.4.1 | **项目:** [OpenWebUI Extensions](https://github.com/Fu-Jie/openwebui-extensions) | **许可证:** MIT
+**作者:** [Fu-Jie](https://github.com/Fu-Jie/openwebui-extensions) | **版本:** 1.4.2 | **项目:** [OpenWebUI Extensions](https://github.com/Fu-Jie/openwebui-extensions) | **许可证:** MIT
 
 > **重要提示**：为了确保所有过滤器的可维护性和易用性，每个过滤器都应附带清晰、完整的文档，以确保其功能、配置和使用方法得到充分说明。
 
 本过滤器通过智能摘要和消息压缩技术，在保持对话连贯性的同时，显著降低长对话的 Token 消耗。
 
-## 1.4.1 版本更新
+## 1.4.2 版本更新
 
-- **逆向展开机制**: 引入 `_unfold_messages` 机制以在 `outlet` 阶段精确对齐坐标系，彻底解决了由于前端视图折叠导致长轮次工具调用对话出现进度漂移或跳过生成摘要的问题。
+- **强化摘要路径健壮性**: 在整个摘要生成管道中透传 `__request__` 上下文，确保认证和提供商处理的可靠性。
-- **更安全的工具内容裁剪**: 重构了 `enable_tool_output_trimming`，现在严格使用原子级分组进行安全的原生工具内容裁剪，替代了激进的正则表达式匹配，防止 JSON 载荷损坏。
+- **改进错误诊断**: LLM 响应校验失败时，错误日志现在包含完整的响应体，便于透明的故障排除。
+- **智能旧摘要加载**: 当 outlet payload 中缺失摘要消息时，自动从 DB 加载并合并前一代摘要，实现增量式状态合并。
 
 ---
 

plugins/filters/async-context-compression/async_context_compression.py

@@ -5,16 +5,17 @@ author: Fu-Jie
 author_url: https://github.com/Fu-Jie/openwebui-extensions
 funding_url: https://github.com/open-webui
 description: Reduces token consumption in long conversations while maintaining coherence through intelligent summarization and message compression.
-version: 1.4.1
+version: 1.4.2
 openwebui_id: b1655bc8-6de9-4cad-8cb5-a6f7829a02ce
 license: MIT
 
 ═══════════════════════════════════════════════════════════════════════════════
-📌 What's new in 1.4.1
+📌 What's new in 1.4.2
 ═══════════════════════════════════════════════════════════════════════════════
 
-  ✅ Reverse-Unfolding Mechanism: Accurately reconstructs the expanded native tool-calling sequence during the outlet phase to permanently fix coordinate drift and missing summaries for long tool-based conversations.
+  ✅ Enhanced Summary Path Robustness: Thread __request__ context through entire summary generation pipeline for reliable authentication and provider handling.
-  ✅ Safer Tool Trimming: Refactored `enable_tool_output_trimming` to strictly use atomic block groups for safe trimming, completely preventing JSON payload corruption.
+  ✅ Improved Error Diagnostics: LLM response validation failures now include complete response body in error logs for better troubleshooting.
+  ✅ Smart Previous Summary Loading: Automatically load and merge previous summaries from DB when not present in outlet payload, enabling incremental state merging.
 
 ═══════════════════════════════════════════════════════════════════════════════
 📌 Overview
@@ -2502,6 +2503,7 @@ class Filter:
         __model__: dict = None,
         __event_emitter__: Callable[[Any], Awaitable[None]] = None,
         __event_call__: Callable[[Any], Awaitable[None]] = None,
+        __request__: Request = None,
     ) -> dict:
         """
         Executed after the LLM response is complete.
@@ -2614,6 +2616,7 @@ class Filter:
                 lang,
                 __event_emitter__,
                 __event_call__,
+                __request__,
             )
         )
 
@@ -2630,6 +2633,7 @@ class Filter:
         lang: str,
         __event_emitter__: Callable,
         __event_call__: Callable,
+        __request__: Request = None,
     ):
         """Wrapper to run summary generation with an async lock."""
         async with lock:
@@ -2642,6 +2646,7 @@ class Filter:
                 lang,
                 __event_emitter__,
                 __event_call__,
+                __request__,
             )
 
     async def _check_and_generate_summary_async(
@@ -2654,6 +2659,7 @@ class Filter:
         lang: str = "en-US",
         __event_emitter__: Callable[[Any], Awaitable[None]] = None,
         __event_call__: Callable[[Any], Awaitable[None]] = None,
+        __request__: Request = None,
     ):
         """
         Background processing: Calculates Token count and generates summary (does not block response).
@@ -2757,6 +2763,7 @@ class Filter:
                     lang,
                     __event_emitter__,
                     __event_call__,
+                    __request__,
                 )
             else:
                 await self._log(
@@ -2788,6 +2795,7 @@ class Filter:
         lang: str = "en-US",
         __event_emitter__: Callable[[Any], Awaitable[None]] = None,
         __event_call__: Callable[[Any], Awaitable[None]] = None,
+        __request__: Request = None,
     ):
         """
         Generates summary asynchronously (runs in background, does not block response).
@@ -2941,8 +2949,25 @@ class Filter:
             # 4. Build conversation text
             conversation_text = self._format_messages_for_summary(middle_messages)
 
-            # 5. Call LLM to generate new summary
+            # 5. Determine previous_summary to pass to LLM.
-            # Note: previous_summary is not passed here because old summary (if any) is already included in middle_messages
+            # When summary_index is not None, the old summary message is already the first
+            # entry of middle_messages (protected_prefix=1), so it appears verbatim in
+            # conversation_text — no need to inject separately.
+            # When summary_index is None the outlet messages come from raw DB history that
+            # has never had the summary injected, so we must load it from DB explicitly.
+            if summary_index is None:
+                previous_summary = await asyncio.to_thread(
+                    self._load_summary, chat_id, body
+                )
+                if previous_summary:
+                    await self._log(
+                        "[🤖 Async Summary Task] Loaded previous summary from DB to pass as context (summary not in messages)",
+                        event_call=__event_call__,
+                    )
+            else:
+                previous_summary = None  # already embedded in middle_messages[0]
+
+            # 6. Call LLM to generate new summary
 
             # Send status notification for starting summary generation
             if __event_emitter__:
@@ -2959,11 +2984,12 @@ class Filter:
                 )
 
             new_summary = await self._call_summary_llm(
-                None,
                 conversation_text,
                 {**body, "model": summary_model_id},
                 user_data,
                 __event_call__,
+                __request__,
+                previous_summary=previous_summary,
             )
 
             if not new_summary:
@@ -3186,11 +3212,12 @@ class Filter:
 
     async def _call_summary_llm(
         self,
-        previous_summary: Optional[str],
         new_conversation_text: str,
         body: dict,
         user_data: dict,
         __event_call__: Callable[[Any], Awaitable[None]] = None,
+        __request__: Request = None,
+        previous_summary: Optional[str] = None,
     ) -> str:
         """
         Calls the LLM to generate a summary using Open WebUI's built-in method.
@@ -3201,33 +3228,52 @@ class Filter:
         )
 
         # Build summary prompt (Optimized for State/Working Memory and Tool Calling)
-        summary_prompt = f"""
+        previous_summary_block = (
-You are an expert Context Compression Engine. Your goal is to create a high-fidelity, highly dense "Working Memory" from the provided conversation.
+            f"<previous_working_memory>\n{previous_summary}\n</previous_working_memory>\n\n"
-This conversation may contain previous Working Memories and raw native tool-calling sequences (JSON arguments and results).
+            if previous_summary
+            else ""
+        )
+        summary_prompt = f"""You are an expert Context Compression Engine. Produce a high-fidelity, maximally dense "Working Memory" snapshot from the inputs below.
 
-### Rules of Engagement
+### Processing Rules
-1.  **Incremental Integration**: If the conversation begins with an existing Working Memory/Summary, you must PRESERVE its core facts and MERGE the new conversation events into it. Do not discard older facts.
+1.  **State-Aware Merging**: If `<previous_working_memory>` is provided, you MUST merge it with the new conversation. Preserve facts that are still true; UPDATE or SUPERSEDE facts whose state has changed (e.g., "bug X exists" → "bug X fixed in commit abc"); REMOVE facts fully resolved with no future relevance.
-2.  **Tool-Call Decompression**: Raw JSON/Text outputs from tools are noisy. Extract ONLY the definitive facts, actionable data, or root causes of errors. Ignore the structural payload.
+2.  **Goal Tracking**: Reflect the LATEST user intent as "Current Goal". If the goal has shifted, move the old goal to Working Memory as "Prior Goal (completed/abandoned)".
-3.  **Ruthless Denoising**: Completely eliminate greetings, apologies ("I'm sorry for the error"), acknowledgments ("Sure, I can do that"), and redundant confirmations.
+3.  **Tool-Call Decompression**: From raw JSON tool arguments/results, extract ONLY: definitive facts, concrete return values, error codes, root causes. Discard structural boilerplate.
-4.  **Verbatim Retention**: ANY code snippets, shell commands, file paths, specific parameters, and Message IDs (e.g., [ID: ...]) MUST be kept exactly as they appear to maintain traceability.
+4.  **Error & Exception Verbatim**: Stack traces, error messages, exception types, and exit codes MUST be quoted exactly — they are primary debugging artifacts.
-5.  **Logic Preservation**: Clearly link "what the user asked" -> "what the tool found" -> "how the system reacted".
+5.  **Ruthless Denoising**: Delete greetings, apologies, acknowledgments, and any phrase that carries zero information.
+6.  **Verbatim Retention**: Code snippets, shell commands, file paths, config values, and Message IDs (e.g., [ID: ...]) MUST appear character-for-character.
+7.  **Causal Chain**: For each tool call or action, record: trigger → operation → outcome (one line per event).
 
 ### Output Constraints
-*   **Format**: Strictly follow the Markdown structure below.
+*   **Format**: Follow the Required Structure below — omit a section only if it has zero content.
-*   **Length**: Maximum {self.valves.max_summary_tokens} Tokens.
+*   **Token Budget**: Stay under {self.valves.max_summary_tokens} tokens. Prioritize recency and actionability when trimming.
-*   **Tone**: Robotic, objective, dense.
+*   **Tone**: Terse, robotic, third-person where applicable.
-*   **Language**: Consistent with the conversation language.
+*   **Language**: Match the dominant language of the conversation.
-*   **Forbidden**: NO conversational openings/closings (e.g., "Here is the summary", "Hope this helps"). Output the data directly.
+*   **Forbidden**: No preamble, no closing remarks, no meta-commentary. Start directly with the first section header.
 
-### Suggested Summary Structure
+### Required Output Structure
-*   **Current Goal**: What is the user ultimately trying to achieve?
+## Current Goal
-*   **Working Memory & Facts**: (Bullet points of established facts, parsed tool results, and constraints. Cite Message IDs if critical).
+(Single sentence: what the user is trying to achieve RIGHT NOW)
-*   **Code & Artifacts**: (Only if applicable. Include exact code blocks).
+
-*   **Recent Actions**: (e.g., "Attempted to run script, failed with SyntaxError, applied fix").
+## Working Memory & Facts
-*   **Pending/Next Steps**: What is waiting to be done.
+(Bullet list — each item: one established fact, constraint, or parsed tool result. Mark superseded items as ~~old~~ → new. Cite [ID: ...] when critical.)
+
+## Code & Artifacts
+(Only if present. Exact code blocks with language tags. File paths as inline code.)
+
+## Causal Log
+(Chronological. Format: `[MSG_ID?] action → result`. One line per event. Keep only the last N events that remain causally relevant.)
+
+## Errors & Exceptions
+(Only if unresolved. Exact quoted text. Include error type, message, and last known stack frame.)
+
+## Pending / Next Steps
+(Ordered list. First item = most immediate action.)
 
 ---
+{previous_summary_block}<new_conversation>
 {new_conversation_text}
+</new_conversation>
 ---
 
 Generate the Working Memory:
@@ -3277,8 +3323,8 @@ Generate the Working Memory:
                 event_call=__event_call__,
             )
 
-            # Create Request object
+            # Use the injected request if available, otherwise fall back to a minimal synthetic one
-            request = Request(scope={"type": "http", "app": webui_app})
+            request = __request__ or Request(scope={"type": "http", "app": webui_app})
 
             # Call generate_chat_completion
             response = await generate_chat_completion(request, payload, user)
@@ -3299,8 +3345,13 @@ Generate the Working Memory:
                 or "choices" not in response
                 or not response["choices"]
             ):
+                try:
+                    response_repr = json_module.dumps(response, ensure_ascii=False, indent=2)
+                except Exception:
+                    response_repr = repr(response)
                 raise ValueError(
-                    f"LLM response format incorrect or empty: {type(response).__name__}"
+                    f"LLM response format incorrect or empty: {type(response).__name__}\n"
+                    f"Full response:\n{response_repr}"
                 )
 
             summary = response["choices"][0]["message"]["content"].strip()