fix(agent): re-sync model identity when LM Studio swaps model mid-session

agent/chat_completion_helpers.py

@@ -1140,6 +1140,44 @@ def rewrite_prompt_model_identity(agent, model: str, provider: str) -> None:
     agent._cached_system_prompt = sp
 
 
+def sync_lmstudio_active_model(agent, response) -> None:
+    """Re-point the cached identity when LM Studio serves a different model.
+
+    LM Studio (and other local OpenAI-compatible servers) let the user swap the
+    loaded model from the app while a Hermes session stays open.  ``agent.model``
+    and the ``Model:`` line in the cached system prompt are resolved once at
+    session start, so after a swap the agent keeps reporting the original model
+    when asked about its inference engine — only a brand-new session picks up the
+    change (#54454).
+
+    Every chat-completion response echoes the model that actually served it, so
+    when that live name diverges from ``agent.model`` we adopt it and rewrite the
+    cached identity via :func:`rewrite_prompt_model_identity` (the same in-place,
+    non-persisted rewrite already used for provider failover).  Scoped to the
+    ``lmstudio`` provider — the only path where the server-side model changes
+    underneath a live session — so no other provider's identity is touched.
+    """
+    if (getattr(agent, "provider", "") or "").strip().lower() != "lmstudio":
+        return
+    live_model = getattr(response, "model", None)
+    if not isinstance(live_model, str) or not live_model.strip():
+        return
+    live_model = live_model.strip()
+    current = (getattr(agent, "model", "") or "").strip()
+    if not current:
+        return
+    try:
+        from agent.model_metadata import _model_id_matches
+
+        if _model_id_matches(live_model, current) or _model_id_matches(current, live_model):
+            return
+    except Exception:
+        if live_model == current:
+            return
+    agent.model = live_model
+    rewrite_prompt_model_identity(agent, live_model, agent.provider)
+
+
 def _fallback_entry_key(fb: dict) -> tuple[str, str, str]:
     return (
         str(fb.get("provider") or "").strip().lower(),

agent/conversation_loop.py

@@ -2263,6 +2263,15 @@ def _perform_api_call(next_api_kwargs):
                     except Exception:
                         pass
                 agent._touch_activity(f"API call #{api_call_count} completed")
+                # LM Studio users can swap the loaded model from the app while a
+                # session is open; the response echoes the live model, so re-sync
+                # the cached identity when it drifts (#54454).
+                if agent.provider == "lmstudio":
+                    try:
+                        from agent.chat_completion_helpers import sync_lmstudio_active_model
+                        sync_lmstudio_active_model(agent, response)
+                    except Exception:
+                        pass
                 break  # Success, exit retry loop
 
             except InterruptedError:

tests/agent/test_failover_identity.py

@@ -10,7 +10,10 @@
 
 from types import SimpleNamespace
 
-from agent.chat_completion_helpers import rewrite_prompt_model_identity
+from agent.chat_completion_helpers import (
+    rewrite_prompt_model_identity,
+    sync_lmstudio_active_model,
+)
 from agent.conversation_loop import _sync_failover_system_message
 
 
@@ -70,6 +73,62 @@ def test_empty_values_leave_lines_unchanged(self):
         assert agent._cached_system_prompt == _PROMPT
 
 
+_LMSTUDIO_PROMPT = (
+    "You are a helpful assistant.\n"
+    "\n"
+    "Conversation started: Wednesday, June 10, 2026\n"
+    "Model: gemma-3-4b\n"
+    "Provider: lmstudio"
+)
+
+
+def _lmstudio_agent(model="gemma-3-4b", prompt=_LMSTUDIO_PROMPT):
+    return SimpleNamespace(
+        provider="lmstudio",
+        model=model,
+        _cached_system_prompt=prompt,
+        ephemeral_system_prompt=None,
+    )
+
+
+class TestSyncLmstudioActiveModel:
+    def test_adopts_swapped_model_and_rewrites_identity(self):
+        # User swapped Gemma -> Qwen in the LM Studio app mid-session (#54454).
+        agent = _lmstudio_agent()
+        sync_lmstudio_active_model(agent, SimpleNamespace(model="qwen3-6b"))
+        assert agent.model == "qwen3-6b"
+        assert "Model: qwen3-6b" in agent._cached_system_prompt
+        assert "Model: gemma-3-4b" not in agent._cached_system_prompt
+
+    def test_noop_when_model_unchanged(self):
+        agent = _lmstudio_agent()
+        sync_lmstudio_active_model(agent, SimpleNamespace(model="gemma-3-4b"))
+        assert agent.model == "gemma-3-4b"
+        assert agent._cached_system_prompt == _LMSTUDIO_PROMPT
+
+    def test_noop_on_slug_vs_basename_match(self):
+        # LM Studio native API returns "publisher/slug"; the configured value is
+        # the bare slug — these are the same model, not a drift.
+        agent = _lmstudio_agent(model="gemma-3-4b")
+        sync_lmstudio_active_model(agent, SimpleNamespace(model="google/gemma-3-4b"))
+        assert agent.model == "gemma-3-4b"
+        assert agent._cached_system_prompt == _LMSTUDIO_PROMPT
+
+    def test_noop_for_non_lmstudio_provider(self):
+        agent = _lmstudio_agent()
+        agent.provider = "openai"
+        sync_lmstudio_active_model(agent, SimpleNamespace(model="qwen3-6b"))
+        assert agent.model == "gemma-3-4b"
+        assert agent._cached_system_prompt == _LMSTUDIO_PROMPT
+
+    def test_noop_when_response_has_no_model(self):
+        agent = _lmstudio_agent()
+        sync_lmstudio_active_model(agent, SimpleNamespace(model=None))
+        assert agent.model == "gemma-3-4b"
+        sync_lmstudio_active_model(agent, SimpleNamespace(model="   "))
+        assert agent.model == "gemma-3-4b"
+
+
 class TestSyncFailoverSystemMessage:
     def test_patches_in_flight_system_message(self):
         agent = _agent()