Bugfix: avoid stale tool-turn fallback on empty responses

.github/workflows/contributor-check.yml

@@ -35,7 +35,7 @@ jobs:
           while IFS= read -r email; do
             # Skip teknium and bot emails
             case "$email" in
-              *teknium*|*noreply@github.com*|*dependabot*|*github-actions*|*anthropic.com*|*cursor.com*)
+              *teknium*|*noreply@github.com*|*@users.noreply.github.com|*dependabot*|*github-actions*|*anthropic.com*|*cursor.com*)
                 continue ;;
             esac
 

run_agent.py

@@ -9985,25 +9985,29 @@ def _stop_spinner():
 
                     assistant_msg = self._build_assistant_message(assistant_message, finish_reason)
 
-                    # If this turn has both content AND tool_calls, capture the content
-                    # as a fallback final response. Common pattern: model delivers its
-                    # answer and calls memory/skill tools as a side-effect in the same
-                    # turn. If the follow-up turn after tools is empty, we use this.
+                    # Only carry forward earlier turn content as a fallback
+                    # when EVERY tool call is post-response housekeeping
+                    # (memory, todo, skill_manage, etc.). Using substantive
+                    # tool turns as fallback content silently abandons
+                    # unfinished multi-step tasks on the next empty response.
+                    _HOUSEKEEPING_TOOLS = frozenset({
+                        "memory", "todo", "skill_manage", "session_search",
+                    })
+                    _all_housekeeping = all(
+                        tc.function.name in _HOUSEKEEPING_TOOLS
+                        for tc in assistant_message.tool_calls
+                    )
                     turn_content = assistant_message.content or ""
                     if turn_content and self._has_content_after_think_block(turn_content):
-                        self._last_content_with_tools = turn_content
+                        if _all_housekeeping:
+                            self._last_content_with_tools = turn_content
+                        else:
+                            self._last_content_with_tools = None
                         # Only mute subsequent output when EVERY tool call in
-                        # this turn is post-response housekeeping (memory, todo,
-                        # skill_manage, etc.).  If any substantive tool is present
-                        # (search_files, read_file, write_file, terminal, ...),
-                        # keep output visible so the user sees progress.
-                        _HOUSEKEEPING_TOOLS = frozenset({
-                            "memory", "todo", "skill_manage", "session_search",
-                        })
-                        _all_housekeeping = all(
-                            tc.function.name in _HOUSEKEEPING_TOOLS
-                            for tc in assistant_message.tool_calls
-                        )
+                        # this turn is post-response housekeeping. If any
+                        # substantive tool is present (search_files, read_file,
+                        # write_file, terminal, ...), keep output visible so
+                        # the user sees progress.
                         if _all_housekeeping and self._has_stream_consumers():
                             self._mute_post_response = True
                         elif self.quiet_mode:

tests/run_agent/test_run_agent.py

@@ -2079,6 +2079,72 @@ def _fake_api_call(api_kwargs):
         assert result["final_response"] == "Fresh partial content from this turn"
         assert result["api_calls"] == 1
 
+    def test_empty_after_substantive_tool_turn_retries_instead_of_using_stale_content(self, agent):
+        """Substantive tool turns must not become the final answer fallback."""
+        self._setup_agent(agent)
+        tc = _mock_tool_call(name="web_search", arguments='{"q":"test"}', call_id="c1")
+        resp1 = _mock_response(
+            content="I found a lead and I'm checking one more source.",
+            finish_reason="tool_calls",
+            tool_calls=[tc],
+        )
+        resp2 = _mock_response(content=None, finish_reason="stop")
+        resp3 = _mock_response(
+            content="Here is the completed answer after the follow-up retry.",
+            finish_reason="stop",
+        )
+        agent.client.chat.completions.create.side_effect = [resp1, resp2, resp3]
+
+        status_messages = []
+
+        with (
+            patch("run_agent.handle_function_call", return_value="search result"),
+            patch.object(agent, "_persist_session"),
+            patch.object(agent, "_save_trajectory"),
+            patch.object(agent, "_cleanup_task_resources"),
+            patch.object(agent, "_emit_status", side_effect=status_messages.append),
+        ):
+            result = agent.run_conversation("search something")
+
+        assert result["completed"] is True
+        assert result["final_response"] == "Here is the completed answer after the follow-up retry."
+        assert result["api_calls"] == 3
+        assert not any(
+            "using earlier content as final answer" in msg.lower()
+            for msg in status_messages
+        )
+
+    def test_empty_after_housekeeping_tool_turn_uses_prior_content_fallback(self, agent_with_memory_tool):
+        """Housekeeping-only tool turns may still reuse prior visible content."""
+        self._setup_agent(agent_with_memory_tool)
+        tc = _mock_tool_call(name="memory", arguments='{"content":"saved note"}', call_id="c1")
+        resp1 = _mock_response(
+            content="Saved that to memory.",
+            finish_reason="tool_calls",
+            tool_calls=[tc],
+        )
+        resp2 = _mock_response(content=None, finish_reason="stop")
+        agent_with_memory_tool.client.chat.completions.create.side_effect = [resp1, resp2]
+
+        status_messages = []
+
+        with (
+            patch("run_agent.handle_function_call", return_value="saved"),
+            patch.object(agent_with_memory_tool, "_persist_session"),
+            patch.object(agent_with_memory_tool, "_save_trajectory"),
+            patch.object(agent_with_memory_tool, "_cleanup_task_resources"),
+            patch.object(agent_with_memory_tool, "_emit_status", side_effect=status_messages.append),
+        ):
+            result = agent_with_memory_tool.run_conversation("remember this")
+
+        assert result["completed"] is True
+        assert result["final_response"] == "Saved that to memory."
+        assert result["api_calls"] == 2
+        assert any(
+            "using earlier content as final answer" in msg.lower()
+            for msg in status_messages
+        )
+
     def test_nous_401_refreshes_after_remint_and_retries(self, agent):
         self._setup_agent(agent)
         agent.provider = "nous"