fix(tui): persist model switch marker as user message

tests/test_tui_gateway_server.py

@@ -3497,11 +3497,11 @@ def fake_switch_model(**kwargs):
             "Model: anthropic/claude-sonnet-4.6\nProvider: anthropic"
         )
         assert agent._cached_system_prompt == db.system_prompt
-        assert session["history"][-1]["role"] == "system"
+        assert session["history"][-1]["role"] == "user"
         assert "changed to anthropic/claude-sonnet-4.6" in session["history"][-1]["content"]
         assert db.messages[-1] == {
             "session_id": "session-key",
-            "role": "system",
+            "role": "user",
             "content": session["history"][-1]["content"],
         }
         # ...and the shared process env was NOT touched.

tests/tui_gateway/test_model_switch_marker.py

@@ -0,0 +1,40 @@
+"""Regression tests for gateway model-switch history markers."""
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+from tui_gateway.server import _append_model_switch_marker
+
+
+def test_model_switch_marker_is_user_message_not_mid_history_system():
+    """Model switches must not inject a system role after conversation turns.
+
+    Strict OpenAI-compatible providers such as vLLM reject message arrays that
+    contain system messages after the beginning of the conversation.
+    """
+    db = MagicMock()
+    session = {
+        "session_key": "sess-1",
+        "history": [
+            {"role": "user", "content": "hello"},
+            {"role": "assistant", "content": "hi"},
+        ],
+        "history_version": 7,
+        "agent": SimpleNamespace(_session_db=db),
+    }
+
+    _append_model_switch_marker(
+        session,
+        model="qwen3.6-35b",
+        provider="api.lucasnicolas.dev",
+    )
+
+    marker = session["history"][-1]
+    assert marker["role"] == "user"
+    assert "active model for this chat has changed to qwen3.6-35b" in marker["content"]
+    assert session["history_version"] == 8
+    db.append_message.assert_called_once_with(
+        session_id="sess-1",
+        role="user",
+        content=marker["content"],
+    )

tui_gateway/server.py

@@ -1721,7 +1721,11 @@ def _append_model_switch_marker(session: dict | None, *, model: str, provider: s
         f"{model}{provider_part}. From this point forward, use this runtime "
         "metadata when answering questions about what model/provider is active.]"
     )
-    entry = {"role": "system", "content": marker}
+    # Persist as a user message, not a system message.  The gateway appends
+    # this marker after prior conversation turns, and strict OpenAI-compatible
+    # providers such as vLLM reject system messages that are not at the
+    # beginning of the API message list.
+    entry = {"role": "user", "content": marker}
 
     lock = session.get("history_lock")
     if lock is not None:
@@ -1736,14 +1740,14 @@ def _append_model_switch_marker(session: dict | None, *, model: str, provider: s
         agent = session.get("agent")
         db = getattr(agent, "_session_db", None) if agent is not None else None
         if db is not None:
-            db.append_message(session_id=session_key, role="system", content=marker)
+            db.append_message(session_id=session_key, role="user", content=marker)
             return
 
         _ensure_session_db_row(session)
         with _session_db(session) as scoped_db:
             if scoped_db is not None:
                 scoped_db.append_message(
-                    session_id=session_key, role="system", content=marker
+                    session_id=session_key, role="user", content=marker
                 )
     except Exception:
         logger.debug("failed to persist model switch marker", exc_info=True)