diff --git a/emrg/client/app.py b/emrg/client/app.py
index c873f03..21f92cd 100644
--- a/emrg/client/app.py
+++ b/emrg/client/app.py
@@ -556,7 +556,14 @@ async def _reconnect():
                     _last_center = server_id or "emrg"
                     status.update(center=_last_center)
                     term.set_title(f"{session_title or session_id} @ {project_name}")
-                    msg_count += 1; _update_left_extra()
+                    # rant 21:52:18: daemon's done frame reports the authoritative
+                    # current-context message count (system + history + tool results);
+                    # fall back to the local +1 approximation when absent.
+                    if data.get("context_messages") is not None:
+                        msg_count = int(data["context_messages"])
+                    else:
+                        msg_count += 1
+                    _update_left_extra()
                     term.render()
                 if "error" in data:
                     err = data["error"]; logger.error("server error: %s", err)
diff --git a/emrg/server/daemon.py b/emrg/server/daemon.py
index efb3092..5220de2 100644
--- a/emrg/server/daemon.py
+++ b/emrg/server/daemon.py
@@ -2274,6 +2274,8 @@ async def _run_tool_loop(
                     "done": True,
                     "delta": False,
                     "session_id": session.session_id,
+                    # rant 21:52:18: authoritative current-context message count.
+                    "context_messages": len(messages),
                 })
 
                 # Fire-and-forget: reflect on whether to save memories
@@ -2448,6 +2450,10 @@ async def _run_tool_loop(
                 "done": True,
                 "delta": False,
                 "session_id": session.session_id,
+                # rant 21:52:18: current LLM context size (system + history +
+                # user + all tool results + assistant replies) — authoritative
+                # for the TUI status bar message count.
+                "context_messages": len(messages),
             })
 
             # Fire-and-forget: reflect on whether to save memories
diff --git a/tests/test_ws_e2e.py b/tests/test_ws_e2e.py
index 574abd1..4f357ba 100644
--- a/tests/test_ws_e2e.py
+++ b/tests/test_ws_e2e.py
@@ -222,6 +222,7 @@ async def _test():
                         }
                         await ws.send(json.dumps(task, ensure_ascii=False))
                         got_delta = got_tool = got_done = False
+                        context_messages = None
                         while True:
                             frame = await asyncio.wait_for(ws.recv(), timeout=10)
                             resp = json.loads(frame)
@@ -231,8 +232,14 @@ async def _test():
                                 got_tool = True
                             if resp.get("done"):
                                 got_done = True
+                                # rant 21:52:18: done frame must report the
+                                # current LLM context size — system(1) + user(1)
+                                # + assistant(tool_calls)(1) + tool result(1) +
+                                # final assistant(1) = 5 for this one-round flow.
+                                context_messages = resp.get("context_messages")
                                 break
                         assert got_delta and got_tool and got_done
+                        assert context_messages == 5
                     finally:
                         await ws.close()
                 finally: