fix(cli): preserve noted staged input on close

cli.py

@@ -12365,13 +12365,20 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin):
                 self._pending_moa_config = None
                 if _moa_cfg is None:
                     _moa_cfg = None
+                # Model/skill notes and voice instructions are API-local. Keep
+                # the original staged input as the durable transcript value so a
+                # close-path marker follows the same dict into turn setup rather
+                # than producing a second noted user row (#63766).
+                _persist_clean_user_message = (
+                    message if (_voice_prefix or agent_message != message) else None
+                )
                 try:
                     result = self.agent.run_conversation(
                         user_message=agent_message,
                         conversation_history=self.conversation_history[:-1],  # Exclude the message we just added
                         stream_callback=stream_callback,
                         task_id=self.session_id,
-                        persist_user_message=message if _voice_prefix else None,
+                        persist_user_message=_persist_clean_user_message,
                         moa_config=_moa_cfg,
                     )
                     if getattr(self, "_pending_moa_disable_after_turn", False):

tests/agent/test_turn_context.py

@@ -241,6 +241,24 @@ def test_stale_pending_cli_message_does_not_replace_new_turn_input():
     assert agent._pending_cli_user_message is None
 
 
+def test_pending_cli_message_uses_clean_override_for_api_local_note():
+    """A noted API message reuses the clean staged dict and its DB marker."""
+    agent = _FakeAgent()
+    staged = {"role": "user", "content": "clean prompt", "_db_persisted": True}
+    agent._pending_cli_user_message = staged
+
+    ctx = _build(
+        agent,
+        user_message="[MODEL NOTE]\n\nclean prompt",
+        persist_user_message="clean prompt",
+    )
+
+    assert ctx.messages[-1] is staged
+    assert ctx.messages[-1]["content"] == "[MODEL NOTE]\n\nclean prompt"
+    assert ctx.messages[-1]["_db_persisted"] is True
+    assert agent._pending_cli_user_message is None
+
+
 def test_memory_nudge_fires_at_interval():
     agent = _FakeAgent()
     agent._memory_nudge_interval = 1

tests/cli/test_cli_interrupt_ack_race.py

@@ -192,3 +192,42 @@ def test_acknowledged_interrupt_still_requeues_message():
         queued.append(cli._pending_input.get_nowait())
     assert any("redirect please" in str(q) for q in queued)
     assert cli._last_turn_interrupted is True
+
+
+def test_chat_persists_clean_input_when_a_queued_note_changes_api_message():
+    """Queued notes remain API-local and preserve close-handoff marker identity."""
+    cli = _make_cli()
+
+    class _NoteAgent(_StubAgent):
+        def __init__(self, session_id):
+            super().__init__(session_id, turn_seconds=0)
+            self.captured = None
+
+        def run_conversation(self, **kwargs):
+            self.captured = kwargs
+            return {
+                "final_response": "done",
+                "messages": [{"role": "assistant", "content": "done"}],
+                "api_calls": 1,
+                "completed": True,
+                "partial": True,
+                "response_previewed": True,
+            }
+
+    agent = _NoteAgent(cli.session_id)
+    cli.agent = agent
+    cli._interrupt_queue = queue.Queue()
+    cli._pending_input = queue.Queue()
+    cli._pending_model_switch_note = "[MODEL SWITCH NOTE]"
+
+    with patch.object(cli, "_ensure_runtime_credentials", return_value=True), \
+         patch.object(cli, "_resolve_turn_agent_config", return_value={
+             "signature": cli._active_agent_route_signature,
+             "model": None, "runtime": None, "request_overrides": None,
+         }), \
+         patch.object(cli, "_init_agent", return_value=True):
+        cli.chat("clean prompt")
+
+    assert agent.captured is not None
+    assert agent.captured["user_message"] == "[MODEL SWITCH NOTE]\n\nclean prompt"
+    assert agent.captured["persist_user_message"] == "clean prompt"

tests/cli/test_cli_shutdown_memory_messages.py

@@ -17,6 +17,7 @@ other tests keep their existing no-arg behaviour.
 from __future__ import annotations
 
 import threading
+import types
 from typing import Any
 from unittest.mock import MagicMock, patch
 
@@ -491,6 +492,98 @@ def test_cli_close_persists_pending_user_when_agent_snapshot_is_empty(tmp_path,
     assert staged["_db_persisted"] is True
 
 
+def test_cli_close_preserves_clean_staged_user_across_noted_worker_turn(tmp_path, monkeypatch):
+    """A noted API-only turn reuses the close-marked clean staged user row."""
+    monkeypatch.setenv("HERMES_HOME", str(tmp_path / ".hermes"))
+
+    import cli as cli_mod
+    from hermes_state import SessionDB
+
+    db = SessionDB(db_path=tmp_path / "state.db")
+    session_id = "cli-close-noted-staged-user"
+    db.create_session(session_id=session_id, source="cli")
+    prefix = [
+        {"role": "user", "content": "old prompt"},
+        {"role": "assistant", "content": "old answer"},
+    ]
+    agent = _real_agent(db, session_id, prefix)
+    agent._flush_messages_to_session_db(prefix, [])
+    staged = {"role": "user", "content": "new prompt"}
+    agent._pending_cli_user_message = staged
+
+    cli = object.__new__(cli_mod.HermesCLI)
+    cli.conversation_history = list(prefix) + [staged]
+    cli.session_id = session_id
+    cli.agent = agent
+
+    cli._persist_active_session_before_close()
+    assert staged["_db_persisted"] is True
+
+    # A queued model/skills note changes only the API message. The worker
+    # reuses the marked clean dict, so the normal persistence seam cannot append
+    # a second noted user row.
+    from agent.turn_context import build_turn_context
+
+    agent.quiet_mode = True
+    agent.max_iterations = 1
+    agent.provider = "test"
+    agent.base_url = ""
+    agent.api_key = ""
+    agent.api_mode = "chat_completions"
+    agent.tools = []
+    agent.valid_tool_names = set()
+    agent.enabled_toolsets = None
+    agent.disabled_toolsets = None
+    agent._skip_mcp_refresh = True
+    agent.compression_enabled = False
+    agent.context_compressor = types.SimpleNamespace(protect_first_n=2, protect_last_n=2)
+    agent._memory_store = None
+    agent._memory_manager = None
+    agent._memory_nudge_interval = 0
+    agent._turns_since_memory = 0
+    agent._user_turn_count = 0
+    agent._todo_store = types.SimpleNamespace(has_items=lambda: True)
+    agent._tool_guardrails = types.SimpleNamespace(reset_for_turn=lambda: None)
+    agent._compression_warning = None
+    agent._interrupt_requested = False
+    agent._memory_write_origin = "assistant_tool"
+    agent._stream_context_scrubber = None
+    agent._stream_think_scrubber = None
+    agent._restore_primary_runtime = lambda: None
+    agent._cleanup_dead_connections = lambda: False
+    agent._emit_status = lambda _message: None
+    agent._replay_compression_warning = lambda: None
+    agent._hydrate_todo_store = lambda *_args: None
+    agent._safe_print = lambda *_args: None
+
+    worker = build_turn_context(
+        agent,
+        "[MODEL SWITCH NOTE]\n\nnew prompt",
+        None,
+        prefix,
+        "task",
+        None,
+        "new prompt",
+        None,
+        restore_or_build_system_prompt=lambda *_args: None,
+        install_safe_stdio=lambda: None,
+        sanitize_surrogates=lambda value: value,
+        summarize_user_message_for_log=lambda value: value,
+        set_session_context=lambda _session_id: None,
+        set_current_write_origin=lambda _origin: None,
+        ra=lambda: types.SimpleNamespace(_set_interrupt=lambda *_args: None),
+    )
+    assert worker.messages[-1] is staged
+    assert worker.messages[-1]["content"] == "[MODEL SWITCH NOTE]\n\nnew prompt"
+
+    stored = db.get_messages_as_conversation(session_id)
+    assert [m["content"] for m in stored] == [
+        "old prompt",
+        "old answer",
+        "new prompt",
+    ]
+
+
 def test_cli_close_builds_prompt_before_creating_first_session_row(tmp_path, monkeypatch):
     """First-turn close persistence must not leave a NULL prompt snapshot."""
     monkeypatch.setenv("HERMES_HOME", str(tmp_path / ".hermes"))