fix(agent): trim usage-less empty fix to guard + observability

Drop the loop-side '(empty)' rewrite (the turn-completion explainer already
owns that at delivery, and gateway/desktop match on the sentinel) and the
extra token-count persistence. Keeps: usage-absent empty streaks arm the
deterministic fast-fail after two attempts with no content or reasoning,
and every completed API call logs even when the provider omits usage
(#101898).

agent/conversation_loop.py

@@ -8761,12 +8761,7 @@ def run_conversation(
                             "answer:\n\n" + reasoning_preview
                         )
                     else:
-                        final_response = agent._format_turn_completion_explanation(
-                            _turn_exit_reason
-                        ) or (
-                            "⚠️ No reply: the model returned empty content after "
-                            "all retries and fallback attempts."
-                        )
+                        final_response = "(empty)"
                     break
                 
                 # Reset retry counter/signature on successful content

tests/run_agent/test_empty_terminal_reasoning_surface.py

@@ -11,7 +11,7 @@ Invariants pinned here:
   ``_empty_terminal_sentinel`` marker (replay semantics unchanged).
 - Raw reasoning is NEVER promoted earlier in the ladder — a reasoning-only
   response still goes through prefill continuation first.
-- A truly empty exhaustion returns an actionable no-reply explanation.
+- A truly empty exhaustion (no reasoning either) still returns "(empty)".
 """
 
 from __future__ import annotations
@@ -109,11 +109,11 @@ def test_exhausted_reasoning_only_delivers_labeled_excerpt(tmp_path, monkeypatch
     )
 
 
-def test_exhausted_truly_empty_delivers_no_reply_explanation(tmp_path, monkeypatch):
-    """No reasoning anywhere produces an honest terminal explanation even
-    when the downstream completion explainer is disabled."""
+def test_exhausted_truly_empty_keeps_existing_behavior(tmp_path, monkeypatch):
+    """No reasoning anywhere → behavior unchanged from main: the '(empty)'
+    terminal (possibly rewritten by the downstream turn-completion explainer)
+    is delivered, and no reasoning excerpt appears."""
     agent = _build_agent(tmp_path, monkeypatch)
-    monkeypatch.setattr(agent, "_turn_completion_explainer_enabled", lambda: False)
     monkeypatch.setattr(
         agent, "_interruptible_api_call",
         lambda api_kwargs: _truly_empty_response(),
@@ -122,7 +122,9 @@ def test_exhausted_truly_empty_delivers_no_reply_explanation(tmp_path, monkeypat
     result = agent.run_conversation("hello?")
 
     final = result["final_response"]
-    assert final.startswith("⚠️ No reply:")
+    # Either the raw sentinel (explainer off) or the explainer's rewrite —
+    # never the reasoning-excerpt frame, which requires reasoning to exist.
+    assert final == "(empty)" or final.startswith("⚠️ No reply:")
     assert "only internal reasoning" not in final
 
 

tests/run_agent/test_run_agent.py

@@ -3601,16 +3601,12 @@ class TestRunConversation:
             patch.object(agent, "_persist_session"),
             patch.object(agent, "_save_trajectory"),
             patch.object(agent, "_cleanup_task_resources"),
-            patch.object(
-                agent, "_turn_completion_explainer_enabled", return_value=False
-            ),
             caplog.at_level(logging.INFO, logger="agent.conversation_loop"),
         ):
             result = agent.run_conversation("answer me")
         assert result["completed"] is True
         assert result["api_calls"] == 2
         assert agent.session_api_calls == 2
-        assert result["final_response"].startswith("⚠️ No reply:")
         assert caplog.text.count("usage=unavailable") == 2
 
     def test_truly_empty_response_succeeds_on_nudge(self, agent):