perf(dashboard): keyset pagination for streaming session export

OFFSET paging made the streaming export O(n^2) on huge transcripts;
after_id keyset paging keeps each page seek O(1). Adds after_id to
SessionDB.get_messages (ascending-only, guarded against latest/offset
combos).

hermes_cli/web_routers/sessions.py

@@ -746,13 +746,15 @@ async def export_session_endpoint(session_id: str, profile: Optional[str] = None
             )
             yield metadata[:-1] + ',"messages":['
 
-            offset = 0
+            # Keyset pagination (id > last_seen): O(n) total over the
+            # transcript, vs OFFSET's O(n²) on huge sessions.
+            last_id = None
             first = True
             while True:
                 messages = db.get_messages(
                     sid,
                     limit=500,
-                    offset=offset,
+                    after_id=last_id if last_id is not None else 0,
                 )
                 for message in messages:
                     if not first:
@@ -765,7 +767,9 @@ async def export_session_endpoint(session_id: str, profile: Optional[str] = None
                     first = False
                 if len(messages) < 500:
                     break
-                offset += len(messages)
+                last_id = messages[-1].get("id")
+                if last_id is None:
+                    break  # defensive: cannot keyset without row ids
 
             yield "]}"
         finally:

hermes_state.py

@@ -7421,6 +7421,7 @@ class SessionDB(SessionSearchMixin, SessionSchemaMixin, SessionPortabilityMixin)
         limit: Optional[int] = None,
         offset: int = 0,
         latest: bool = False,
+        after_id: Optional[int] = None,
     ) -> List[Dict[str, Any]]:
         """Load messages for a session in insertion order.
 
@@ -7440,13 +7441,22 @@ class SessionDB(SessionSearchMixin, SessionSchemaMixin, SessionPortabilityMixin)
         order. ``offset`` alone (without ``limit``) also pages — SQLite
         requires a LIMIT clause for OFFSET, so it's emitted as ``LIMIT -1``
         (unbounded).
+
+        ``after_id`` enables keyset pagination (``id > after_id``): O(1)
+        page seeks on huge transcripts where OFFSET degrades to O(n) per
+        page. Ascending order only (incompatible with ``latest``/``offset``).
         """
+        if after_id is not None and (latest or offset):
+            raise ValueError("after_id is incompatible with latest/offset paging")
         active_clause = "" if include_inactive else " AND active = 1"
+        keyset_clause = " AND id > ?" if after_id is not None else ""
         sql = (
             "SELECT * FROM messages WHERE session_id = ?"
-            f"{active_clause} ORDER BY id {'DESC' if latest else 'ASC'}"
+            f"{active_clause}{keyset_clause} ORDER BY id {'DESC' if latest else 'ASC'}"
         )
         params: list = [session_id]
+        if after_id is not None:
+            params.append(after_id)
         if limit is not None or offset:
             # SQLite's OFFSET requires LIMIT; -1 means "no limit".
             sql += " LIMIT ? OFFSET ?"

tests/hermes_cli/test_web_server.py

@@ -1895,7 +1895,7 @@ class TestWebServerEndpoints:
         original_get_messages = SessionDB.get_messages
 
         def tracked_get_messages(self, session_id, *args, **kwargs):
-            calls.append((kwargs.get("limit"), kwargs.get("offset")))
+            calls.append((kwargs.get("limit"), kwargs.get("after_id")))
             return original_get_messages(self, session_id, *args, **kwargs)
 
         monkeypatch.setattr(SessionDB, "get_messages", tracked_get_messages)

tests/test_hermes_state.py

@@ -3775,6 +3775,19 @@ class TestGetMessagesPagination:
         assert [m["content"] for m in page2] == ["msg-2", "msg-3", "msg-4", "msg-5"]
         assert [m["content"] for m in page3] == ["msg-0", "msg-1"]
 
+    def test_after_id_keyset_pages_forward_in_insertion_order(self, db):
+        self._seed(db)
+        page1 = db.get_messages("s1", limit=4, after_id=0)
+        assert [m["content"] for m in page1] == ["msg-0", "msg-1", "msg-2", "msg-3"]
+        page2 = db.get_messages("s1", limit=4, after_id=page1[-1]["id"])
+        assert [m["content"] for m in page2] == ["msg-4", "msg-5", "msg-6", "msg-7"]
+        page3 = db.get_messages("s1", limit=4, after_id=page2[-1]["id"])
+        assert [m["content"] for m in page3] == ["msg-8", "msg-9"]
+        with pytest.raises(ValueError):
+            db.get_messages("s1", limit=4, after_id=0, latest=True)
+        with pytest.raises(ValueError):
+            db.get_messages("s1", limit=4, after_id=0, offset=2)
+
     def test_resume_safety_counts_active_rows_across_lineage(self, db):
         db.create_session(session_id="root", source="cli")
         db.append_messages_batch(