fix(cli): drop dead _deferred_content plumbing; cover reasoning-then-answer streaming (#47116)

hermes_cli/cli_stream_mixin.py

@@ -285,7 +285,7 @@ class CLIStreamMixin:
             _cprint(line)
 
     def _close_reasoning_box(self) -> None:
-        """Close the live reasoning box if it's open, then flush deferred content."""
+        """Close the live reasoning box if it's open (renders the buffered reasoning tail)."""
         from cli import _DIM, _RST, _cprint
         if not getattr(self, "_reasoning_box_opened", False):
             return
@@ -298,10 +298,6 @@ class CLIStreamMixin:
         self._reasoning_box_opened = False
         if not getattr(self, "_stream_box_live", False):
             self._release_held_status_lines()
-        deferred = getattr(self, "_deferred_content", "")
-        if deferred:
-            self._deferred_content = ""
-            self._emit_stream_text(deferred)
 
     def _stream_delta(self, text) -> None:
         """Line-buffered streaming callback for real-time token rendering.
@@ -531,7 +527,6 @@ class CLIStreamMixin:
         self._reasoning_box_opened = False
         self._reasoning_buf = ""
         self._reasoning_preview_buf = ""
-        self._deferred_content = ""
         # A batch cancelled/errored before any tool.started would otherwise mute the next turn's line.
         self.__dict__.pop("_tool_gen_announced", None)
         self._stream_table_buf = []

tests/hermes_cli/test_status_lines_held_during_stream_box.py

@@ -31,6 +31,31 @@ def cli_stub(monkeypatch):
     return cli, emitted
 
 
+def test_answer_streams_before_turn_end_after_reasoning(cli_stub):
+    """#47116: with show_reasoning on, the reasoning box closes on the first content token so the
+    answer streams mid-turn instead of being held until end of turn."""
+    cli, emitted = cli_stub
+    cli.show_reasoning = True
+    cli._stream_reasoning_delta("thinking about it\n")
+    cli._stream_delta("First answer line.\n")
+    # No _flush_stream(): the line must already be on screen mid-turn.
+    lines = [_plain(e) for e in emitted]
+    answer = [i for i, l in enumerate(lines) if "First answer line." in l]
+    assert answer, lines
+    reasoning = next(i for i, l in enumerate(lines) if "thinking about it" in l)
+    assert reasoning < answer[0]
+
+
+def test_late_reasoning_does_not_reopen_box_inside_answer(cli_stub):
+    cli, emitted = cli_stub
+    cli.show_reasoning = True
+    cli._stream_reasoning_delta("early thought\n")
+    cli._stream_delta("Answer.\n")
+    cli._stream_reasoning_delta("late thought\n")
+    cli._flush_stream()
+    assert not any("late thought" in _plain(e) for e in emitted), emitted
+
+
 def test_status_line_waits_for_box_footer(cli_stub):
     cli, emitted = cli_stub
     cli._stream_delta("First paragraph.\n")

tests/hermes_cli/test_stream_delta_think_tag.py

@@ -23,7 +23,6 @@ def _make_cli_stub():
     cli._reasoning_box_opened = False
     cli._reasoning_buf = ""
     cli._reasoning_preview_buf = ""
-    cli._deferred_content = ""
     cli._stream_text_ansi = ""
     cli._stream_needs_break = False
     cli._emitted = []