fix: small wiki fixes

application/agents/tools/wiki.py

@@ -201,7 +201,7 @@ class WikiTool(Tool):
     def _reembed(self, path: str, content_hash: str) -> None:
         """Enqueue an async re-embed for ``path``, authored as the source owner.
 
-        The Unit-2 worker loads the source via ``get_any(source_id, user)``
+        The re-embed worker loads the source via ``get_any(source_id, user)``
         (owner-scoped), so the owner — not the caller — must be passed as
         ``user`` or a team editor's edit would fail to re-embed. A per-page
         idempotency key guards each edit independently and dedups broker

application/api/answer/services/stream_processor.py

@@ -912,13 +912,14 @@ class StreamProcessor:
         return value or "prefetch"
 
     def _build_wiki_config(self) -> Optional[Dict[str, Any]]:
-        """Resolve the WikiTool config for the first writable wiki source (D18).
+        """Resolve the WikiTool config for the first writable wiki source.
 
         A source qualifies when ``SourceConfig.parse(config).kind == "wiki"`` and
         the principal can write it (``effective_write_owner`` returns an owner —
         owner or team editor; viewers get None and no tool). v1 supports one
-        writable wiki source; extras are logged and skipped. Returns None when no
-        writable wiki source is present.
+        writable wiki source; the first match wins and the scan stops there so
+        this runs at most one owner+source lookup per chat on the hot path.
+        Returns None when no writable wiki source is present.
         """
         from application.api.user.team_sharing import effective_write_owner
 
@@ -943,19 +944,13 @@ class StreamProcessor:
                         continue
                     if SourceConfig.parse(source_doc.get("config")).kind != "wiki":
                         continue
-                    if wiki_config is not None:
-                        logger.info(
-                            "Multiple writable wiki sources for this agent; "
-                            "using the first and skipping source %s",
-                            str(source_doc["id"]),
-                        )
-                        continue
                     wiki_config = {
                         "source_id": str(source_doc["id"]),
                         "source_owner_id": owner,
                         "decoded_token": self.decoded_token,
                         "user": caller,
                     }
+                    break
         except Exception:
             logger.exception("Failed to resolve wiki tool config")
             return None
@@ -1602,7 +1597,7 @@ class StreamProcessor:
             "tool_executor": tool_executor,
         }
 
-        # Wiki tool injection + authz (D18): only for agent types that build a
+        # Wiki tool injection + authz: only for agent types that build a
         # tools_dict (classic/agentic/research), and only when a writable wiki
         # source is present for the principal (viewers get nothing).
         if agent_type in ("classic", "agentic", "research"):

application/storage/db/repositories/wiki_pages.py

@@ -26,6 +26,17 @@ def _content_hash(content: str) -> str:
     return hashlib.sha256(content.encode("utf-8")).hexdigest()
 
 
+def _escape_like(value: str) -> str:
+    """Escape ``LIKE`` wildcards so a path prefix matches literally.
+
+    Without this a path containing ``_`` or ``%`` (e.g. ``/api_v1/``) would be
+    a wildcard pattern, so a prefix list/delete could over-match sibling paths.
+    Backslash is escaped first so the wildcard escapes are not double-escaped;
+    callers must pair the result with ``LIKE ... ESCAPE '\\'``.
+    """
+    return value.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
+
+
 class WikiPagesRepository:
     def __init__(self, conn: Connection) -> None:
         self._conn = conn
@@ -149,10 +160,11 @@ class WikiPagesRepository:
         result = self._conn.execute(
             text(
                 "SELECT * FROM wiki_pages "
-                "WHERE source_id = CAST(:source_id AS uuid) AND path LIKE :prefix "
+                "WHERE source_id = CAST(:source_id AS uuid) "
+                "AND path LIKE :prefix ESCAPE '\\' "
                 "ORDER BY path"
             ),
-            {"source_id": source_id, "prefix": prefix + "%"},
+            {"source_id": source_id, "prefix": _escape_like(prefix) + "%"},
         )
         return [row_to_dict(r) for r in result.fetchall()]
 
@@ -180,9 +192,10 @@ class WikiPagesRepository:
         result = self._conn.execute(
             text(
                 "DELETE FROM wiki_pages "
-                "WHERE source_id = CAST(:source_id AS uuid) AND path LIKE :prefix"
+                "WHERE source_id = CAST(:source_id AS uuid) "
+                "AND path LIKE :prefix ESCAPE '\\'"
             ),
-            {"source_id": source_id, "prefix": prefix + "%"},
+            {"source_id": source_id, "prefix": _escape_like(prefix) + "%"},
         )
         return result.rowcount
 

application/worker.py

@@ -2205,6 +2205,12 @@ def convert_source_to_wiki_worker(self, source_id, user):
     grouped: dict[str, list[tuple[object, str]]] = {}
     original_doc_ids: list[object] = []
     for chunk in store.get_chunks() or []:
+        # Track every original chunk so a skipped one (empty/no-path/invalid)
+        # is purged on convert too, not left orphaned in the vector store after
+        # the source flips to wiki. Only deleted once pages are created below.
+        doc_id = _chunk_doc_id(chunk)
+        if doc_id is not None:
+            original_doc_ids.append(doc_id)
         metadata = _chunk_metadata(chunk)
         rel_path = _chunk_page_path(metadata)
         text = _chunk_text(chunk)
@@ -2216,9 +2222,6 @@ def convert_source_to_wiki_worker(self, source_id, user):
         if _wiki_page_path_from_rel(rel_path) is None:
             skipped.append({"file": rel_path, "reason": "invalid path"})
             continue
-        doc_id = _chunk_doc_id(chunk)
-        if doc_id is not None:
-            original_doc_ids.append(doc_id)
         grouped.setdefault(rel_path, []).append((_chunk_order_hint(metadata), text))
 
     with db_session() as conn:

tests/storage/db/repositories/test_wiki_pages.py

@@ -117,6 +117,16 @@ class TestListByPrefix:
         results = repo.list_by_prefix(sid, "/docs/")
         assert {r["path"] for r in results} == {"/docs/a.md", "/docs/b.md"}
 
+    def test_prefix_underscore_is_literal_not_wildcard(self, pg_conn):
+        # ``_`` is a LIKE single-char wildcard; without escaping, listing
+        # ``/api_v1/`` would also match the sibling ``/apiXv1/``.
+        repo = _repo(pg_conn)
+        sid = _make_source(pg_conn)
+        repo.upsert(sid, "/api_v1/a.md", "a")
+        repo.upsert(sid, "/apiXv1/b.md", "b")
+        results = repo.list_by_prefix(sid, "/api_v1/")
+        assert {r["path"] for r in results} == {"/api_v1/a.md"}
+
     def test_list_for_source_returns_all(self, pg_conn):
         repo = _repo(pg_conn)
         sid = _make_source(pg_conn)
@@ -184,6 +194,16 @@ class TestDelete:
         assert repo.delete_by_prefix(sid, "/dir/") == 2
         assert repo.get_by_path(sid, "/other/c.md") is not None
 
+    def test_delete_by_prefix_underscore_is_literal_not_wildcard(self, pg_conn):
+        # Deleting ``/api_v1/`` must not also drop the sibling ``/apiXv1/``
+        # via the unescaped ``_`` LIKE wildcard.
+        repo = _repo(pg_conn)
+        sid = _make_source(pg_conn)
+        repo.upsert(sid, "/api_v1/a.md", "a")
+        repo.upsert(sid, "/apiXv1/b.md", "b")
+        assert repo.delete_by_prefix(sid, "/api_v1/") == 1
+        assert repo.get_by_path(sid, "/apiXv1/b.md") is not None
+
 
 class TestExpectedVersion:
     def test_matching_version_updates(self, pg_conn):

tests/worker/test_convert_source_to_wiki.py

@@ -132,6 +132,29 @@ class TestConvertSourceToWikiWorker:
         assert result["status"] == "converted"
         store.delete_chunk.assert_not_called()
 
+    def test_skipped_chunk_with_doc_id_still_deleted(
+        self, pg_conn, patch_worker_db, task_self, monkeypatch
+    ):
+        # A chunk skipped for an invalid path must still have its original
+        # vector chunk purged, not left orphaned after the source flips to wiki.
+        from application import worker
+
+        source_id = _seed_source(pg_conn)
+        store = _patch_store(
+            monkeypatch,
+            [
+                _chunk("good", source="ok.md", doc_id="ok"),
+                _chunk("bad", source="../evil.md", doc_id="evil"),
+            ],
+        )
+        _patch_reembed(monkeypatch)
+
+        result = worker.convert_source_to_wiki_worker(task_self, source_id, "alice")
+
+        assert result["pages_created"] == 1
+        deleted = {c.args[0] for c in store.delete_chunk.call_args_list}
+        assert deleted == {"ok", "evil"}
+
     def test_no_pages_path_deletes_nothing(
         self, pg_conn, patch_worker_db, task_self, monkeypatch
     ):