diff --git a/interfaces/tui/agent.py b/interfaces/tui/agent.py
index da86ac0..87169a9 100644
--- a/interfaces/tui/agent.py
+++ b/interfaces/tui/agent.py
@@ -2,7 +2,6 @@ import json
 import threading
 from datetime import datetime
 import config
-from lib import moop
 from lib import ntro, agent_loop
 from tools.vision import ImagePayload
 
@@ -45,12 +44,39 @@ WELCOME_ART = """
 ╚══════════════════════════════════════════╝
 """
 
-def log(app, role, text):
-    app.log.append({
+def log(app, role, text, model_info=None):
+    entry = {
         "role": role,
         "text": text,
         "time": datetime.now().strftime("%H:%M"),
-    })
+    }
+    if model_info:
+        # Display pair: (provider, model)
+        entry["model_info"] = model_info
+    app.log.append(entry)
+
+
+def _pair(model_info):
+    """(provider, model) untuk display; None jika info tidak ada."""
+    if not model_info:
+        return None
+    return (model_info.get("provider"), model_info.get("model"))
+
+
+def _tag_ai_logs(app, prev_entries, pair):
+    """Tempelkan info model yang benar-benar dipakai ke entri log AI yang baru
+    dibuat pada langkah ini.
+
+    Berbasis identitas entri, bukan index: entri "Thinking..." di-pop setelah
+    llm.chat() kembali sehingga index entri AI ikut bergeser. `prev_entries`
+    menahan referensi entri lama agar id-nya tidak dipakai ulang oleh Python.
+    """
+    if not pair:
+        return
+    seen = {id(e) for e in prev_entries}
+    for entry in app.log:
+        if entry.get("role") == "ai" and id(entry) not in seen:
+            entry["model_info"] = pair
 
 
 def submit(app, stdscr):
@@ -59,12 +85,6 @@ def submit(app, stdscr):
         return
 
     log(app, "user", query)
-    model_info = (
-        moop.resolve_provider(None, app.llm.base_url, app.llm.model),
-        app.llm.model
-    )
-    if app.log:
-        app.log[-1]["model_info"] = model_info
     app.input_buffer = [""]
     app.input_line = 0
     app.input_col = 0
@@ -83,7 +103,7 @@ def submit(app, stdscr):
             app._model_info(),
         )
 
-    _add_msg(app, "user", query, model_info=app._model_info())
+    _add_msg(app, "user", query)
 
     app.agent_done.clear()
     app.agent_thread = threading.Thread(
@@ -128,8 +148,15 @@ def _agent_loop(app):
         skill = config.AGENT_SKILLS
         tool_reminder = None
 
+        log_prev = list(app.log)
         response = app.llm.chat(app.messages, tools=app.TOOLS, on_stream_chunk=on_stream_chunk, tool_reminder=tool_reminder)
 
+        # Model final: diambil dari response, bukan dari state llm saat ini,
+        # supaya auto-switch/rotasi key/fallback ikut tercatat dengan benar.
+        used = getattr(response, "model_info", None) or {}
+        pair = _pair(used)
+        akwargs = {"model_info": used} if used else {}
+
         # Hapus "Thinking..." log
         for i in range(len(app.log) - 1, -1, -1):
             if app.log[i].get('role') == 'system' and 'Thinking' in app.log[i].get('text', ''):
@@ -177,11 +204,12 @@ def _agent_loop(app):
                         if placeholder_marker in text or text == "" or text == "...":
                             app.log.pop(i)
             
-            _add_msg(app, "assistant", response.content, tool_calls=response.tool_calls)
+            _add_msg(app, "assistant", response.content, tool_calls=response.tool_calls, **akwargs)
 
             # Log content AI dulu (penjelasan sebelum tools)
             if response.content and response.content.strip():
-                log(app, "ai", response.content)
+                log(app, "ai", response.content, model_info=pair)
+                _tag_ai_logs(app, log_prev, pair)
 
             # Log dan execute tool_calls
             for tc in response.tool_calls:
@@ -190,7 +218,7 @@ def _agent_loop(app):
                 log(app, "tool_call", json.dumps({
                     "name": tname,
                     "arguments": targs,
-                }))
+                }), model_info=pair)
                 if app.scroll_follow:
                     app.scroll = 999999
                 result = agent_loop.execute_tool(tc, app.TOOL_HANDLERS)
@@ -205,8 +233,9 @@ def _agent_loop(app):
             has_content = bool(response.content) or (isinstance(response.content, list) and len(response.content) > 0)
             if has_content:
                 if isinstance(response.content, list):
-                    log(app, "ai", "[Image generated]")
-                _add_msg(app, "assistant", response.content)
+                    log(app, "ai", "[Image generated]", model_info=pair)
+                _add_msg(app, "assistant", response.content, **akwargs)
+                _tag_ai_logs(app, log_prev, pair)
             log(app, "sep", "")
             ntro.end(stamp)
             app.agent_done.set()
diff --git a/interfaces/tui/app.py b/interfaces/tui/app.py
index bf8c2e3..f29f702 100644
--- a/interfaces/tui/app.py
+++ b/interfaces/tui/app.py
@@ -209,22 +209,23 @@ class HendrikTUI:
                 if i > 0:
                     log(self, "sep", "")
                 log(self, "user", msg["content"])
-                mi = msg.get("model_info")
-                if mi:
-                    self.log[-1]["model_info"] = (mi.get("provider"), mi.get("model"))
             elif msg["role"] == "assistant":
+                mi = msg.get("model_info")
+                pair = (mi.get("provider"), mi.get("model")) if mi else None
                 # Log tool calls jika ada
                 if msg.get("tool_calls"):
                     for tc in msg["tool_calls"]:
                         log(self, "tool_call", json.dumps({
                             "name": tc["function"]["name"],
                             "arguments": tc["function"]["arguments"],
-                        }))
+                        }), model_info=pair)
                 # Log AI content jika ada dan ini adalah response final
                 next_role = session.messages[i + 1]["role"] if i + 1 < len(session.messages) else None
                 if next_role in ("user", None):
                     if msg.get("content"):
                         log(self, "ai", msg["content"])
+                        if pair:
+                            self.log[-1]["model_info"] = pair
             elif msg["role"] == "tool":
                 # Tool result - tidak perlu di-log secara eksplisit
                 pass
diff --git a/interfaces/tui/render.py b/interfaces/tui/render.py
index aff6627..c663e79 100644
--- a/interfaces/tui/render.py
+++ b/interfaces/tui/render.py
@@ -55,6 +55,19 @@ def draw_header(app, stdscr):
     draw_menubar(app, stdscr)
 
 
+def _model_info_line(item):
+    """Baris " Provider - Model " dari entri log; "" jika info tidak ada.
+
+    Dipakai bersama oleh blok `ai` dan `tool_call` agar tampilan konsisten."""
+    model_info = item.get("model_info")
+    if not model_info:
+        return ""
+    p_name, m_name = model_info
+    if p_name:
+        return f" {p_name} - {m_name} "
+    return f" {m_name} " if m_name else ""
+
+
 def draw_chat(app, stdscr):
     # Area chat — dari baris 2 sampai baris (h - 11).
     # Bisa di-scroll dengan Page Up / Page Down.
@@ -136,20 +149,13 @@ def draw_chat(app, stdscr):
             _add_blank()
 
         if role == "user":
-            model_info = item.get("model_info", None)
-            if model_info:
-                p_name, m_name = model_info
-                if p_name:
-                    info_line = f" {p_name} - {m_name} "
-                else:
-                    info_line = f" {m_name} "
-            else:
-                info_line = " Unknown - Model info not found "
-            _add_row([(C_USER, info_line, False)])
-            label = f" You ({item['time']}) "
-            _add_row([(C_USER, label)])
+            _add_row([(C_USER, f" You ({item['time']}) ")])
             _wrap_render(text, indent=1, color=C_INPUT, bold=False)
         elif role == "ai":
+            info_line = _model_info_line(item)
+            if info_line:
+                _add_row([(C_AI, info_line, False)])
+
             label = f" {app.character_name} ({item['time']}) "
             _add_row([(C_AI, label)])
             _wrap_render(text, indent=1, color=C_INPUT, bold=False)
@@ -165,6 +171,7 @@ def draw_chat(app, stdscr):
         elif role == "tool_call":
             # Format:
             #   (blank line)
+            #    Ollama Cloud - gemma4:31b-cloud
             #   Hendrik run_bash (HH:MM)   ← "Hendrik" hijau, "run_bash (HH:MM)" kuning
             #   {                         ← arguments indent 1 spasi, kuning
             #     "command": "ls -la"
@@ -173,6 +180,9 @@ def draw_chat(app, stdscr):
             # Blank line di atas (kecuali sebelumnya sudah blank dari role lain)
             if idx > 0 and app.log[idx - 1]["role"] not in ("sep", "welcome"):
                 _add_blank()
+            info_line = _model_info_line(item)
+            if info_line:
+                _add_row([(C_AI, info_line, False)])
             try:
                 tc = json.loads(text)
                 tname = tc["name"]
diff --git a/services/llm_client.py b/services/llm_client.py
index c32c60e..87f25b9 100644
--- a/services/llm_client.py
+++ b/services/llm_client.py
@@ -1,6 +1,6 @@
 import json
 import socket
-from lib import gadget
+from lib import gadget, moop
 import urllib.request
 import urllib.error
 
@@ -30,11 +30,14 @@ class LLMClient:
         self._chain_index = 0
 
     class Message:
-        def __init__(self, msg):
+        def __init__(self, msg, model_info=None):
             raw_content     = msg.get('content', '')      # Ambil konten mentah
             self.content    = gadget.strip_thinking(raw_content) if isinstance(raw_content, str) else raw_content # Auto-strip <thinking> dari content
             self.tool_calls = msg.get('tool_calls', None) # Ambil tool calls
             self.warning    = None
+            # Model yang benar-benar dipakai menghasilkan pesan ini. Bisa berbeda
+            # dengan model awal permintaan karena auto-switch/fallback.
+            self.model_info = model_info or {}
             
     # ------------------------------------------------------------ config ---
 
@@ -86,6 +89,23 @@ class LLMClient:
         self.model = cand.get("model", "")
         self.api_key = cand.get("api_key", "")
 
+    def _model_info(self, cand=None):
+        """Info model yang dipakai untuk satu permintaan.
+
+        Dari kandidat chain jika ada (provider sudah dibawa moop), jika tidak
+        (klien non-chain) pakai state aktif dan cari nama provider-nya."""
+        if cand:
+            return {
+                "provider": cand.get("provider"),
+                "base_url": cand.get("base_url", ""),
+                "model": cand.get("model", ""),
+            }
+        return {
+            "provider": moop.resolve_provider(None, self.base_url, self.model),
+            "base_url": self.base_url,
+            "model": self.model,
+        }
+
     # ----------------------------------------------------- chain selection ---
 
     def _pick_chain_type(self, messages):
@@ -108,7 +128,7 @@ class LLMClient:
     def chat(self, messages, tools=None, on_stream_chunk=None, disable_reasoning=False, tool_reminder=None):
 
         if not self._chains: # kalau bukan chain
-            return self._chat_once(
+            result = self._chat_once(
                 self.base_url, self.model, self.api_key,
                 messages,
                 tools=tools,
@@ -116,6 +136,8 @@ class LLMClient:
                 disable_reasoning=disable_reasoning,
                 tool_reminder=tool_reminder,
             )
+            result.model_info = self._model_info()
+            return result
 
         chain_type = self._pick_chain_type(messages)
         if chain_type != self._chain_key:
@@ -176,6 +198,9 @@ class LLMClient:
                     disable_reasoning=disable_reasoning, tool_reminder=tool_reminder,
                 )
                 self._chain_index = idx
+                # Catat model yang benar-benar menjawab (bukan model awal
+                # permintaan) supaya pemanggil bisa menampilkan data yang jujur.
+                result.model_info = self._model_info(cand)
                 return result
             except _LLMFail as e:
                 last = e