fix(knowledge): delete the document directory even before it is indexed

src/everos/service/knowledge.py

@@ -508,7 +508,9 @@ async def delete_document(
 ) -> DeleteResult:
     """Remove a document directory; cascade handles SQLite/LanceDB cleanup.
 
-    Idempotent: returns ``deleted_topics=0`` when the document does not exist.
+    Idempotent: returns ``deleted_topics=0`` when neither the index nor the
+    disk has the document, and also when the directory was removed before the
+    cascade had indexed its topics.
 
     Args:
         doc_id: Document primary key.
@@ -519,12 +521,28 @@ async def delete_document(
         DeleteResult with the topic count that was present before deletion.
     """
     row = await knowledge_document_repo.get_by_doc_id(doc_id)
+    memory_root = MemoryRoot.resolve()
     if row is None:
-        return DeleteResult(doc_id=doc_id, deleted_topics=0)
+        # The index trails the markdown by seconds, so a document created a
+        # moment ago has a directory but no row yet. Deleting by the index
+        # alone leaves that directory behind for the cascade to index right
+        # back in, and the "deleted" document reappears. Markdown is the
+        # truth: find the directory by its name and count the topic files
+        # it holds, so the response says what was actually removed.
+        topic_count = await anyio.to_thread.run_sync(
+            _remove_unindexed_doc_dirs,
+            memory_root.knowledge_dir(app_id, project_id),
+            doc_id,
+        )
+        logger.info(
+            "document deleted",
+            doc_id=doc_id,
+            topic_count=topic_count,
+            indexed=False,
+        )
+        return DeleteResult(doc_id=doc_id, deleted_topics=topic_count)
 
     topic_count = await knowledge_topic_sqlite_repo.count_by_doc_id(doc_id)
-
-    memory_root = MemoryRoot.resolve()
     doc_dir = memory_root.root / Path(row.md_path).parent
     if await anyio.Path(doc_dir).is_dir():
         await anyio.to_thread.run_sync(shutil.rmtree, doc_dir)
@@ -537,6 +555,26 @@ async def delete_document(
     return DeleteResult(doc_id=doc_id, deleted_topics=topic_count)
 
 
+def _remove_unindexed_doc_dirs(knowledge_dir: Path, doc_id: str) -> int:
+    """Remove every ``knowledge/<category>/<title>_<doc_id>/`` directory.
+
+    Returns the number of topic files (``N_*.md``) that were on disk. The
+    directory name is compared literally — ``delete_document`` is a public
+    service function, so a caller-supplied ``doc_id`` must not act as a glob
+    pattern or a path; only the HTTP route validates the id's shape.
+    """
+    if not knowledge_dir.is_dir():
+        return 0
+    suffix = f"_{doc_id}"
+    topics = 0
+    for doc_dir in knowledge_dir.glob("*/*"):
+        if not doc_dir.is_dir() or not doc_dir.name.endswith(suffix):
+            continue
+        topics += sum(1 for _ in doc_dir.glob("[0-9]*.md"))
+        shutil.rmtree(doc_dir)
+    return topics
+
+
 async def replace_document(
     *,
     extractor: KnowledgeExtractor,

tests/unit/test_service/test_knowledge_crud.py

@@ -220,8 +220,11 @@ async def test_delete_document_success(tmp_path: Path) -> None:
     mock_anyio.to_thread.run_sync.assert_awaited_once()
 
 
-async def test_delete_document_idempotent() -> None:
+async def test_delete_document_idempotent(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
     """Returns deleted_topics=0 without error when document does not exist."""
+    monkeypatch.setenv("EVEROS_ROOT", str(tmp_path))
     with patch(f"{_MOD}.knowledge_document_repo") as mock_doc_repo:
         mock_doc_repo.get_by_doc_id = AsyncMock(return_value=None)
 
@@ -232,6 +235,58 @@ async def test_delete_document_idempotent() -> None:
     assert result.deleted_topics == 0
 
 
+async def test_delete_document_removes_dir_before_the_index_has_the_row(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """A document created a moment ago has a directory but no SQLite row yet
+    (the cascade trails the markdown by seconds). Delete must still remove the
+    directory — and only that directory — and report the topic files it held.
+    """
+    monkeypatch.setenv("EVEROS_ROOT", str(tmp_path))
+    kdir = tmp_path / "app1" / "proj1" / "knowledge"
+    doc_dir = kdir / "Technology" / "Release_checklist_d_lagging00001"
+    doc_dir.mkdir(parents=True)
+    (doc_dir / "index.md").write_text("---\ndoc_id: d_lagging00001\n---\n")
+    (doc_dir / "1_before.md").write_text("# before\n")
+    (doc_dir / "2_after.md").write_text("# after\n")
+    sibling = kdir / "Technology" / "Other_notes_d_sibling000001"
+    sibling.mkdir()
+    (sibling / "index.md").write_text("---\ndoc_id: d_sibling000001\n---\n")
+    (kdir / ".taxonomy.md").write_text("# taxonomy\n")
+
+    with patch(f"{_MOD}.knowledge_document_repo") as mock_doc_repo:
+        mock_doc_repo.get_by_doc_id = AsyncMock(return_value=None)
+        result = await delete_document("d_lagging00001", "app1", "proj1")
+
+    assert result.deleted_topics == 2
+    assert not doc_dir.exists()
+    assert sibling.is_dir()
+    assert (kdir / ".taxonomy.md").exists()
+
+
+@pytest.mark.parametrize("doc_id", ["d_absent0000001", "*", "*/_original"])
+async def test_delete_document_unindexed_id_not_on_disk_removes_nothing(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, doc_id: str
+) -> None:
+    """No row and no directory for the id → nothing is touched, including
+    when the id looks like a glob pattern or a path: the name match is literal.
+    """
+    monkeypatch.setenv("EVEROS_ROOT", str(tmp_path))
+    kdir = tmp_path / "app1" / "proj1" / "knowledge"
+    sibling = kdir / "Technology" / "Other_notes_d_sibling000001"
+    sibling.mkdir(parents=True)
+    (sibling / "index.md").write_text("---\ndoc_id: d_sibling000001\n---\n")
+    (sibling / "_original").mkdir()
+
+    with patch(f"{_MOD}.knowledge_document_repo") as mock_doc_repo:
+        mock_doc_repo.get_by_doc_id = AsyncMock(return_value=None)
+        result = await delete_document(doc_id, "app1", "proj1")
+
+    assert result.deleted_topics == 0
+    assert (sibling / "index.md").exists()
+    assert (sibling / "_original").is_dir()
+
+
 # ── list_documents ────────────────────────────────────────────────────────────