fix(attachments): parse channel uploads for LLM and stop echoing user files

WhatsApp/channel binary_ref uploads now expand into text_ref/tabular summaries for the model, and outbound replies only auto-attach generated image/video media instead of lookup tool text_ref results.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_binary_ref_attachments.py

@@ -0,0 +1,98 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+
+from runtime.application.gateway.inbound_service import _channel_attachments_for_gateway
+from runtime.chat.agent_messages import build_llm_messages
+from svc.files.attachment_assets import AttachmentAssetStore
+from svc.files.file_attachments import expand_attachment_ref
+from svc.llm.chat_models import RuleBasedChatModel
+
+
+@pytest.fixture
+def att_store(tmp_path, monkeypatch):
+    store = AttachmentAssetStore(root_dir=tmp_path / "att")
+
+    def _factory(root_dir=None):
+        if root_dir is None:
+            return store
+        return AttachmentAssetStore(root_dir=root_dir)
+
+    monkeypatch.setattr("svc.files.file_attachments.AttachmentAssetStore", _factory)
+    monkeypatch.setattr("runtime.chat.agent_messages.AttachmentAssetStore", _factory)
+    return store
+
+
+def _msg_text(message: dict) -> str:
+    content = message.get("content")
+    if isinstance(content, list):
+        return "\n".join(str(x.get("text") or "") for x in content if isinstance(x, dict))
+    return str(content or "")
+
+
+def test_expand_attachment_ref_large_txt_becomes_text_ref(att_store) -> None:
+    text = ("line\n" * 5000).strip()
+    data = text.encode("utf-8")
+    meta = att_store.save_bytes(data, filename="sites.txt", mime="text/plain")
+    got = expand_attachment_ref({"type": "binary_ref", "attachment_id": meta.attachment_id})
+    types = {str(x.get("type") or "") for x in got}
+    assert "text_ref" in types
+    assert "text" in types
+
+
+def test_channel_attachments_for_gateway_parses_local_txt(tmp_path) -> None:
+    p = tmp_path / "Site_List.txt"
+    p.write_text("a\nb\nc\n", encoding="utf-8")
+    got = _channel_attachments_for_gateway([{"local_path": str(p), "kind": "document", "mime": "text/plain"}])
+    assert got
+    assert any(str(x.get("type") or "") == "text" for x in got)
+
+
+def test_build_llm_messages_binary_ref_meta_for_historical_turn(att_store) -> None:
+    data = b"hello attachment"
+    meta = att_store.save_bytes(data, filename="note.txt", mime="text/plain")
+    rows = [
+        SimpleNamespace(
+            role="user",
+            event_type="user_text",
+            content="first",
+            attachments=[{"type": "binary_ref", "attachment_id": meta.attachment_id}],
+            tool_calls=None,
+            turn_uuid="t1",
+        ),
+        SimpleNamespace(
+            role="user",
+            event_type="user_text",
+            content="second",
+            attachments=None,
+            tool_calls=None,
+            turn_uuid="t2",
+        ),
+    ]
+    msgs = build_llm_messages(store_messages=rows, system_prompt="s", model=RuleBasedChatModel(), lang="zh")
+    first_user = next(m for m in msgs if m.get("role") == "user" and "first" in _msg_text(m))
+    content = _msg_text(first_user)
+    assert "BinaryAttachment" in content
+    assert meta.attachment_id in content
+
+
+def test_build_llm_messages_binary_ref_expands_on_last_turn(att_store) -> None:
+    text = ("row\n" * 4000).strip()
+    data = text.encode("utf-8")
+    meta = att_store.save_bytes(data, filename="Site_List.txt", mime="text/plain")
+    rows = [
+        SimpleNamespace(
+            role="user",
+            event_type="user_text",
+            content="帮看一下有多少数据",
+            attachments=[{"type": "binary_ref", "attachment_id": meta.attachment_id}],
+            tool_calls=None,
+            turn_uuid="t1",
+        ),
+    ]
+    msgs = build_llm_messages(store_messages=rows, system_prompt="s", model=RuleBasedChatModel(), lang="zh")
+    user = msgs[-1]
+    content = _msg_text(user)
+    assert "LongTextAttachment" in content or "text_id" in content