fix: skip oneshot memory prefetch

agent/agent_init.py

@@ -235,6 +235,7 @@ def init_agent(
     checkpoint_max_total_size_mb: int = 500,
     checkpoint_max_file_size_mb: int = 10,
     pass_session_id: bool = False,
+    prefetch_after_turn: bool = True,
 ):
     """
     Initialize the AI Agent.
@@ -317,6 +318,7 @@ def init_agent(
     agent.skip_context_files = skip_context_files
     agent.load_soul_identity = load_soul_identity
     agent.pass_session_id = pass_session_id
+    agent.prefetch_after_turn = prefetch_after_turn
     agent._credential_pool = credential_pool
     agent.log_prefix_chars = log_prefix_chars
     agent.log_prefix = f"{log_prefix} " if log_prefix else ""

hermes_cli/oneshot.py

@@ -361,6 +361,8 @@ def _run_agent(
         #   - dangerous-command approval → bypassed via HERMES_YOLO_MODE=1
         #   - skill secret capture → returns gracefully when no callback set
         clarify_callback=_oneshot_clarify_callback,
+        # Oneshot has no next turn, so don't warm one.
+        prefetch_after_turn=False,
     )
 
     # Belt-and-braces: make sure AIAgent doesn't invoke any streaming
@@ -369,8 +371,14 @@ def _run_agent(
     agent.stream_delta_callback = None
     agent.tool_gen_callback = None
 
-    result = agent.run_conversation(prompt)
-    return (result.get("final_response") or "", result)
+    try:
+        result = agent.run_conversation(prompt)
+        return (result.get("final_response") or "", result)
+    finally:
+        try:
+            agent.shutdown_memory_provider(getattr(agent, "messages", []))
+        except Exception:
+            pass
 
 
 def _oneshot_clarify_callback(question: str, choices=None) -> str:

run_agent.py

@@ -495,6 +495,7 @@ def __init__(
         checkpoint_max_total_size_mb: int = 500,
         checkpoint_max_file_size_mb: int = 10,
         pass_session_id: bool = False,
+        prefetch_after_turn: bool = True,
     ):
         """Forwarder — see ``agent.agent_init.init_agent``."""
         from agent.agent_init import init_agent
@@ -570,6 +571,7 @@ def __init__(
             checkpoint_max_total_size_mb=checkpoint_max_total_size_mb,
             checkpoint_max_file_size_mb=checkpoint_max_file_size_mb,
             pass_session_id=pass_session_id,
+            prefetch_after_turn=prefetch_after_turn,
         )
 
     def _get_session_db_for_recall(self):
@@ -3375,10 +3377,11 @@ def _sync_external_memory_for_turn(
                 response_text,
                 **sync_kwargs,
             )
-            self._memory_manager.queue_prefetch_all(
-                user_text,
-                session_id=self.session_id or "",
-            )
+            if getattr(self, "prefetch_after_turn", True):
+                self._memory_manager.queue_prefetch_all(
+                    user_text,
+                    session_id=self.session_id or "",
+                )
         except Exception:
             pass
 

tests/hermes_cli/test_tui_resume_flow.py

@@ -702,6 +702,41 @@ def test_oneshot_exit_code_zero_when_failed_with_error_text(monkeypatch, capsys)
     assert "HTTP 404" in capsys.readouterr().out
 
 
+def test_oneshot_agent_disables_next_turn_prefetch(monkeypatch):
+    _stub_plugin_discovery(monkeypatch)
+    import hermes_cli.oneshot as oneshot_mod
+
+    captured = {}
+    result = {"final_response": "ok"}
+
+    class FakeAgent:
+        def __init__(self, **kwargs):
+            captured.update(kwargs)
+            self.suppress_status_output = False
+            self.stream_delta_callback = object()
+            self.tool_gen_callback = object()
+            self.messages = [{"role": "user", "content": "hello"}]
+            self.shutdown_calls = []
+
+        def run_conversation(self, _prompt):
+            return result
+
+        def shutdown_memory_provider(self, messages):
+            self.shutdown_calls.append(messages)
+            captured["shutdown_calls"] = list(self.shutdown_calls)
+
+    import hermes_cli.runtime_provider as runtime_provider_mod
+    import run_agent as run_agent_mod
+
+    monkeypatch.setattr(run_agent_mod, "AIAgent", FakeAgent)
+    monkeypatch.setattr(runtime_provider_mod, "resolve_runtime_provider", lambda **_kw: {})
+    monkeypatch.setattr(oneshot_mod, "_create_session_db_for_oneshot", lambda: object())
+
+    assert oneshot_mod._run_agent("hello") == ("ok", result)
+    assert captured["prefetch_after_turn"] is False
+    assert captured["shutdown_calls"] == [[{"role": "user", "content": "hello"}]]
+
+
 def test_oneshot_reraises_keyboard_interrupt(monkeypatch):
     _stub_plugin_discovery(monkeypatch)
     import hermes_cli.oneshot as oneshot_mod

tests/run_agent/test_memory_sync_interrupted.py

@@ -35,6 +35,7 @@ def _bare_agent():
     # providers that cache per-session state can update it mid-process
     # (see #6672).
     agent.session_id = "test_session_001"
+    setattr(agent, "prefetch_after_turn", True)
     return agent
 
 
@@ -91,6 +92,23 @@ def test_completed_turn_syncs_and_queues_prefetch(self):
             session_id="test_session_001",
         )
 
+    def test_completed_oneshot_turn_syncs_but_skips_next_turn_prefetch(self):
+        """Oneshot should save memory, but not warm a turn that cannot exist."""
+        agent = _bare_agent()
+        setattr(agent, "prefetch_after_turn", False)
+
+        agent._sync_external_memory_for_turn(
+            original_user_message="remember this",
+            final_response="remembered",
+            interrupted=False,
+        )
+
+        agent._memory_manager.sync_all.assert_called_once_with(
+            "remember this", "remembered",
+            session_id="test_session_001",
+        )
+        agent._memory_manager.queue_prefetch_all.assert_not_called()
+
     def test_completed_turn_syncs_messages_when_present(self):
         agent = _bare_agent()
         messages = [