Merge pull request #3786 from ningyixi/fix/docx-block-content-controls

fix: retain block content controls when reading Word documents

common/office_text.py

@@ -8,7 +8,7 @@ from contextlib import contextmanager
 
 
 def iter_docx_body_text(document):
-    """Yield paragraphs and table rows in the order they occur in the body.
+    """Yield paragraphs and table rows in body order, including content controls.
 
     Builds python-docx's block objects directly, which also works on versions
     that predate ``Document.iter_inner_content()``.
@@ -25,6 +25,10 @@ def _iter_docx_blocks(container, parent):
     for element in container.iterchildren():
         if element.tag == qn("w:p"):
             yield Paragraph(element, parent).text
+        elif element.tag == qn("w:sdt"):
+            content = element.find(qn("w:sdtContent"))
+            if content is not None:
+                yield from _iter_docx_blocks(content, parent)
         elif element.tag == qn("w:tbl"):
             seen_cells = set()
             for row in Table(element, parent).rows:

tests/test_word_content_controls.py

@@ -0,0 +1,82 @@
+"""Document tools retain block-level Word content controls in reading order."""
+
+from io import BytesIO
+
+import pytest
+import requests
+
+from agent.tools.read.read import Read
+from agent.tools.web_fetch.web_fetch import WebFetch
+from common.office_text import iter_docx_body_text
+
+
+@pytest.mark.parametrize("tool_name", ["read", "web_fetch"])
+@pytest.mark.parametrize("placement", ["body", "cell"])
+def test_block_content_controls_are_read(tmp_path, monkeypatch, tool_name, placement):
+    docx = pytest.importorskip("docx")
+    from docx.oxml import OxmlElement
+
+    def wrap(blocks):
+        control = OxmlElement("w:sdt")
+        properties = OxmlElement("w:sdtPr")
+        alias = OxmlElement("w:alias")
+        alias.set(docx.oxml.ns.qn("w:val"), "Control metadata")
+        properties.append(alias)
+        control.append(properties)
+        content = OxmlElement("w:sdtContent")
+        control.append(content)
+        for block in blocks:
+            content.append(block)
+        return control
+
+    document = docx.Document()
+    if placement == "body":
+        parent = document
+    else:
+        parent = document.add_table(rows=1, cols=1).cell(0, 0)
+
+    parent.add_paragraph("Before control")
+    first = parent.add_paragraph("Controlled paragraph")
+    table = parent.add_table(rows=1, cols=1)
+    table.cell(0, 0).text = "Controlled table"
+    nested = parent.add_paragraph("Nested control")
+    control = wrap([first._p, table._tbl, wrap([nested._p])])
+    parent.add_paragraph("After control")._p.addprevious(control)
+    path = tmp_path / "report.docx"
+    document.save(path)
+
+    if tool_name == "read":
+        result = Read({"cwd": str(tmp_path)}).execute({"path": str(path)})
+        text = result.result.get("content", "")
+    else:
+        tool = WebFetch({"cwd": str(tmp_path)})
+        response = requests.Response()
+        response.status_code = 200
+        response.raw = BytesIO(path.read_bytes())
+        response.headers["Content-Length"] = str(path.stat().st_size)
+        monkeypatch.setenv("WEB_SECURITY_SSRF_PROTECTION", "false")
+        monkeypatch.setattr(tool, "_safe_get", lambda *_args, **_kwargs: response)
+        result = tool.execute({"url": "https://example.test/report.docx"})
+        text = result.result
+
+    assert result.status == "success", result.result
+    expected = ["Before control", "Controlled paragraph", "Controlled table", "Nested control", "After control"]
+    positions = [text.index(piece) for piece in expected]
+    assert positions == sorted(positions)
+    assert all(text.count(piece) == 1 for piece in expected)
+    assert "Control metadata" not in text
+
+
+@pytest.mark.parametrize("has_content", [False, True])
+def test_empty_content_controls_preserve_surrounding_text(has_content):
+    docx = pytest.importorskip("docx")
+    from docx.oxml import OxmlElement
+
+    document = docx.Document()
+    document.add_paragraph("Before control")
+    control = OxmlElement("w:sdt")
+    if has_content:
+        control.append(OxmlElement("w:sdtContent"))
+    document.add_paragraph("After control")._p.addprevious(control)
+
+    assert list(iter_docx_body_text(document)) == ["Before control", "After control"]