feat(backend): 统一"发送附件"语义为上传+提及

backend/packages/harness/deerflow/agents/lead_agent/prompt.py

@@ -266,7 +266,9 @@ You: "Deploying to staging..." [proceed]
 
 **File Management:**
 - Uploaded files are automatically listed in the <uploaded_files> section before each request
-- Use `read_file` tool to read uploaded files using their paths from the list
+- Mentioned files are listed in the <mentioned_files> section when references are present
+- Treat "files the user sent" as the conversation-level union of uploaded + mentioned files (deduplicated by file path)
+- Use `read_file` tool to read listed files using their paths from the file-context sections
 - For PDF, PPT, Excel, and Word files, converted Markdown versions (*.md) are available alongside originals
 - All temporary work happens in `/mnt/user-data/workspace`
 - Final deliverables must be copied to `/mnt/user-data/outputs` and presented using `present_files` tool

backend/packages/harness/deerflow/agents/memory/prompt.py

@@ -343,11 +343,15 @@ def format_conversation_for_update(messages: list[Any]) -> str:
                         text_parts.append(text_val)
             content = " ".join(text_parts) if text_parts else str(content)
 
-        # Strip uploaded_files tags from human messages to avoid persisting
+        # Strip file-context tags from human messages to avoid persisting
-        # ephemeral file path info into long-term memory.  Skip the turn entirely
+        # ephemeral file path info into long-term memory. Skip the turn entirely
-        # when nothing remains after stripping (upload-only message).
+        # when nothing remains after stripping (file-context-only message).
         if role == "human":
-            content = re.sub(r"<uploaded_files>[\s\S]*?</uploaded_files>\n*", "", str(content)).strip()
+            content = re.sub(
+                r"<(?:uploaded_files|mentioned_files|sent_files_semantics)>[\s\S]*?</(?:uploaded_files|mentioned_files|sent_files_semantics)>\n*",
+                "",
+                str(content),
+            ).strip()
             if not content:
                 continue
 

backend/packages/harness/deerflow/agents/memory/updater.py

@@ -213,6 +213,7 @@ _UPLOAD_SENTENCE_RE = re.compile(
     r"|/mnt/user-data/uploads/"
     r"|<uploaded_files>"
     r"|<mentioned_files>"
+    r"|<sent_files_semantics>"
     r")[^.!?]*[.!?]?\s*",
     re.IGNORECASE,
 )

backend/packages/harness/deerflow/agents/middlewares/memory_middleware.py

@@ -15,7 +15,7 @@ from deerflow.config.memory_config import get_memory_config
 logger = logging.getLogger(__name__)
 
 _UPLOAD_BLOCK_RE = re.compile(
-    r"<(?:uploaded_files|mentioned_files)>[\s\S]*?</(?:uploaded_files|mentioned_files)>\n*",
+    r"<(?:uploaded_files|mentioned_files|sent_files_semantics)>[\s\S]*?</(?:uploaded_files|mentioned_files|sent_files_semantics)>\n*",
     re.IGNORECASE,
 )
 _CORRECTION_PATTERNS = (

backend/packages/harness/deerflow/agents/middlewares/uploads_middleware.py

@@ -145,6 +145,72 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
 
         return "\n".join(lines)
 
+    def _merge_sent_files(self, uploaded_files: list[dict], mention_files: list[dict]) -> list[dict]:
+        """Build conversation-level sent-files view (uploads ∪ mentions, deduped by path)."""
+
+        merged: dict[str, dict] = {}
+
+        def _upsert(file: dict, source: str) -> None:
+            path = file.get("path") or ""
+            if not path:
+                return
+            entry = merged.get(path)
+            if entry is None:
+                entry = {
+                    "filename": file.get("filename") or Path(path).name,
+                    "path": path,
+                    "size": int(file.get("size") or 0),
+                    "sent_sources": set(),
+                }
+                merged[path] = entry
+            entry["sent_sources"].add(source)
+            entry["size"] = max(entry["size"], int(file.get("size") or 0))
+            if source == "mention" and file.get("ref_source"):
+                entry["ref_source"] = file["ref_source"]
+
+        for file in uploaded_files:
+            _upsert(file, "upload")
+        for file in mention_files:
+            _upsert(file, "mention")
+
+        ordered = sorted(
+            merged.values(),
+            key=lambda f: (str(f.get("filename", "")).lower(), str(f.get("path", "")).lower()),
+        )
+        for file in ordered:
+            sources = file.get("sent_sources") or set()
+            if "upload" in sources and "mention" in sources:
+                file["sent_source_label"] = "upload+mention"
+            elif "upload" in sources:
+                file["sent_source_label"] = "upload"
+            else:
+                file["sent_source_label"] = "mention"
+        return ordered
+
+    def _create_sent_files_summary(self, sent_files: list[dict]) -> str:
+        """Create policy block describing unified 'sent files' semantics."""
+        lines = [
+            "<sent_files_semantics>",
+            "Conversation attachment semantics:",
+            "- Treat uploaded files and mentioned files as one unified concept of files the user has sent.",
+            "- For questions like 'what files did I send' or 'how many files did I send', use the conversation-level union of uploaded + mentioned files.",
+            "- Count unique files by path (deduplicated).",
+            "",
+            "Conversation-level sent files (deduplicated):",
+        ]
+        if sent_files:
+            for file in sent_files:
+                size_kb = file["size"] / 1024
+                size_str = f"{size_kb:.1f} KB" if size_kb < 1024 else f"{size_kb / 1024:.1f} MB"
+                lines.append(
+                    f"- {file['filename']} ({size_str}, source: {file['sent_source_label']})"
+                )
+                lines.append(f"  Path: {file['path']}")
+        else:
+            lines.append("- (none)")
+        lines.append("</sent_files_semantics>")
+        return "\n".join(lines)
+
     def _mentioned_files_from_kwargs(self, message: HumanMessage) -> list[dict]:
         """Extract mention references from additional_kwargs.files.
 
@@ -186,7 +252,7 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
         return references
 
     def _create_mentions_message(self, mention_files: list[dict]) -> str:
-        lines = ["<mentioned_files>", "The following files were referenced in this message:", ""]
+        lines = ["<mentioned_files>", "The following files were referenced by the user in this conversation:", ""]
         for file in mention_files:
             size_kb = file["size"] / 1024
             size_str = f"{size_kb:.1f} KB" if size_kb < 1024 else f"{size_kb / 1024:.1f} MB"
@@ -199,6 +265,21 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
         lines.append("</mentioned_files>")
         return "\n".join(lines)
 
+    def _mentioned_files_from_messages(self, messages: list) -> list[dict]:
+        """Extract mention references across conversation messages."""
+        references: list[dict] = []
+        seen: set[tuple[str, str]] = set()
+        for message in messages:
+            if not isinstance(message, HumanMessage):
+                continue
+            for file in self._mentioned_files_from_kwargs(message):
+                key = (file["filename"], file["path"])
+                if key in seen:
+                    continue
+                seen.add(key)
+                references.append(file)
+        return references
+
     def _files_from_kwargs(self, message: HumanMessage, uploads_dir: Path | None = None) -> list[dict] | None:
         """Extract file info from message additional_kwargs.files.
 
@@ -282,7 +363,7 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
 
         # Get newly uploaded files from the current message's additional_kwargs.files
         new_files = self._files_from_kwargs(last_message, uploads_dir) or []
-        mention_files = self._mentioned_files_from_kwargs(last_message)
+        mention_files = self._mentioned_files_from_messages(messages)
 
         # Collect historical files from the uploads directory (all except the new ones)
         new_filenames = {f["filename"] for f in new_files}
@@ -311,13 +392,18 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
                 file["outline"] = outline
                 file["outline_preview"] = preview
 
-        if not new_files and not historical_files and not mention_files:
+        sent_files = self._merge_sent_files(new_files + historical_files, mention_files)
+
+        if not new_files and not historical_files and not mention_files and not sent_files:
             return None
 
         logger.debug(f"New files: {[f['filename'] for f in new_files]}, historical: {[f['filename'] for f in historical_files]}")
 
         # Create context message(s) and prepend to the last human message content.
-        message_parts = [self._create_files_message(new_files, historical_files)]
+        message_parts = [
+            self._create_files_message(new_files, historical_files),
+            self._create_sent_files_summary(sent_files),
+        ]
         if mention_files:
             message_parts.append(self._create_mentions_message(mention_files))
         files_message = "\n\n".join(message_parts)

backend/tests/test_memory_updater.py

@@ -510,6 +510,22 @@ class TestFormatConversationForUpdate:
         assert "raw user text" in result
         assert "structured text" in result
 
+    def test_strips_uploaded_mentioned_and_sent_semantics_tags(self):
+        msg = MagicMock()
+        msg.type = "human"
+        msg.content = (
+            "<uploaded_files>\nfile list\n</uploaded_files>\n"
+            "<sent_files_semantics>\nsummary\n</sent_files_semantics>\n"
+            "<mentioned_files>\nmentions\n</mentioned_files>\n"
+            "actual question"
+        )
+
+        result = format_conversation_for_update([msg])
+        assert "actual question" in result
+        assert "uploaded_files" not in result
+        assert "mentioned_files" not in result
+        assert "sent_files_semantics" not in result
+
 
 # ---------------------------------------------------------------------------
 # update_memory - structured LLM response handling

backend/tests/test_uploads_middleware_core_logic.py

@@ -278,10 +278,91 @@ class TestBeforeAgent:
         assert result is not None
         content = result["messages"][-1].content
         assert "<mentioned_files>" in content
-        assert "referenced in this message" in content
+        assert "referenced by the user in this conversation" in content
         assert "/mnt/user-data/uploads/saten-ruiko.jpg" in content
         assert "Do not re-upload them." in content
 
+    def test_injects_sent_files_semantics_for_upload_and_mention_union(self, tmp_path):
+        mw = _middleware(tmp_path)
+        uploads_dir = _uploads_dir(tmp_path)
+        (uploads_dir / "uploaded.txt").write_bytes(b"u")
+        msg = _human(
+            "how many files did I send?",
+            files=[
+                {
+                    "filename": "uploaded.txt",
+                    "size": 1,
+                    "path": "/mnt/user-data/uploads/uploaded.txt",
+                },
+                {
+                    "filename": "mentioned.jpg",
+                    "size": 0,
+                    "path": "/mnt/user-data/outputs/mentioned.jpg",
+                    "ref_kind": "mention",
+                    "ref_source": "artifact",
+                },
+            ],
+        )
+        result = mw.before_agent(self._state(msg), _runtime())
+
+        assert result is not None
+        content = result["messages"][-1].content
+        assert "<sent_files_semantics>" in content
+        assert "union of uploaded + mentioned files" in content
+        assert "uploaded.txt (0.0 KB, source: upload)" in content
+        assert "mentioned.jpg (0.0 KB, source: mention)" in content
+
+    def test_sent_files_union_dedupes_same_file_path_and_marks_both(self, tmp_path):
+        mw = _middleware(tmp_path)
+        uploads_dir = _uploads_dir(tmp_path)
+        (uploads_dir / "same.txt").write_bytes(b"x")
+        msg = _human(
+            "count files",
+            files=[
+                {
+                    "filename": "same.txt",
+                    "size": 1,
+                    "path": "/mnt/user-data/uploads/same.txt",
+                },
+                {
+                    "filename": "same.txt",
+                    "size": 1,
+                    "path": "/mnt/user-data/uploads/same.txt",
+                    "ref_kind": "mention",
+                    "ref_source": "upload",
+                },
+            ],
+        )
+        result = mw.before_agent(self._state(msg), _runtime())
+
+        assert result is not None
+        content = result["messages"][-1].content
+        assert "same.txt (0.0 KB, source: upload+mention)" in content
+        assert content.count("same.txt (0.0 KB, source: upload+mention)") == 1
+
+    def test_historical_mentions_are_included_for_follow_up_count_question(self, tmp_path):
+        mw = _middleware(tmp_path)
+        prev = _human(
+            "analyse this",
+            files=[
+                {
+                    "filename": "history.png",
+                    "size": 0,
+                    "path": "/mnt/user-data/outputs/history.png",
+                    "ref_kind": "mention",
+                    "ref_source": "artifact",
+                }
+            ],
+        )
+        current = _human("我总共发送了多少个附件？")
+        result = mw.before_agent(self._state(prev, current), _runtime())
+
+        assert result is not None
+        content = result["messages"][-1].content
+        assert "<mentioned_files>" in content
+        assert "history.png" in content
+        assert "source: mention" in content
+
     def test_mentioned_files_do_not_enter_uploaded_files_state(self, tmp_path):
         mw = _middleware(tmp_path)
         msg = _human(

frontend/src/core/messages/utils.ts

@@ -337,12 +337,17 @@ export interface FileInMessage {
 }
 
 /**
- * Strip <uploaded_files> tag from message content.
+ * Strip internal file-context tags from message content.
- * Returns the content with the tag removed.
+ * Returns the content with these tags removed:
+ * - <uploaded_files>...</uploaded_files>
+ * - <mentioned_files>...</mentioned_files>
+ * - <sent_files_semantics>...</sent_files_semantics>
  */
 export function stripUploadedFilesTag(content: string): string {
   return content
     .replace(/<uploaded_files>[\s\S]*?<\/uploaded_files>/g, "")
+    .replace(/<mentioned_files>[\s\S]*?<\/mentioned_files>/g, "")
+    .replace(/<sent_files_semantics>[\s\S]*?<\/sent_files_semantics>/g, "")
     .trim();
 }