Developer branch

.gitignore

@@ -89,9 +89,9 @@ _embedded_resources.py
 _build_stamp.py
 /raw_reasoning
 *.mp4
-!frontend/background.mp4
 !demo.mp4
 /Auto_Use/api_key
+/Auto_Use/agent_conversation/settings.json
 /Auto_Use/windows_use/mcp_tracking
 *.exe
 *.ico
@@ -117,3 +117,8 @@ screen_record.py
 /Auto_Use/macOS_use/scratchpad
 /cli_minion_result
 /Auto_Use/windows_use/scratchpad/cli_minion
+
+# Resumable chat memory (UI) — runtime session data lives inside the package;
+# the code (service.py / __init__.py) is tracked, the saved chats are not.
+/Auto_Use/agent_conversation/index.json
+/Auto_Use/agent_conversation/chat_*/

Auto_Use/agent_conversation/__init__.py

@@ -0,0 +1,31 @@
+# Copyright 2026 Autouse AI — https://github.com/auto-use/Auto-Use
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# If you build on this project, please keep this header and credit
+# Autouse AI (https://github.com/auto-use/Auto-Use) in forks and derivative works.
+# A small attribution goes a long way toward a healthy open-source
+# community — thank you for contributing.
+
+"""Permanent (resumable) chat memory for the UI path.
+
+All conversation-memory management lives in this package — saving, loading and
+optimizing a session's conversation, and the on-disk session folders + index.
+The agent stays a thin one-shot loop; app.py talks ONLY to ConversationService
+(load a session -> trigger the main agent -> save the run). main.py / cli.py
+never import this package, so they remain direct one-shot entry points.
+"""
+
+from .service import ConversationService, conversation
+
+__all__ = ["ConversationService", "conversation"]

Auto_Use/macOS_use/agent/coder/service.py

@@ -282,7 +282,9 @@ def _run_agent_loop(self, task: str) -> str:
 
                 return "Max iterations reached"
 
-            safe_print(f"cli agent running step {step_number}")
+            # Step heartbeat → debug log only (NOT the UI). The card shows ONLY the agent's
+            # own validated output; this "running step N" marker is ours, not the model's.
+            debug_log(f"cli agent running step {step_number}")
 
             # Get agent sitting info
             agent_sitting = self._get_agent_sitting()
@@ -353,12 +355,6 @@ def _run_agent_loop(self, task: str) -> str:
                 # Call LLM
                 raw_response = self.llm.send_request(messages)
 
-                # Stream the raw LLM response to stdout. The parent main agent's
-                # subprocess pipe reader picks it up and renders it line-by-line
-                # in the CLI streaming pill UI (no decoration, just raw text).
-                if raw_response:
-                    safe_print(raw_response)
-
                 # Check stop after LLM
                 if self.stop_event and self.stop_event.is_set():
                     break
@@ -369,10 +365,11 @@ def _run_agent_loop(self, task: str) -> str:
                 # Validate and normalize response
                 success, normalized, failed_raw = CLIAgentResponseFormatter.normalize_response(raw_response)
 
-                # If JSON parse failed, discard response and retry with fresh context
+                # If JSON parse failed, discard response and retry with fresh context.
+                # Emit NOTHING to the UI on failure — raw/partial/malformed JSON would glitch
+                # the streaming card; the raw response is already saved above for debugging.
                 if not success:
                     json_fail_count += 1
-                    safe_print(f"⚠️ JSON parse failed ({json_fail_count}/3). Discarding and retrying...")
 
                     if json_fail_count >= 3:
                         break
@@ -383,6 +380,11 @@ def _run_agent_loop(self, task: str) -> str:
                 # Reset consecutive JSON fail counter on success
                 json_fail_count = 0
 
+                # Stream the validated, complete response (action included) to the card's
+                # `>` line. The frontend parses the `action` array out of THIS JSON to drive
+                # the action-icon chain + scratchpad — no extra backend events needed.
+                safe_print(CLIAgentResponseFormatter.format_stream_json(normalized))
+
                 # Save TRUE agent memory snapshot
                 self._save_conversation_snapshot(messages, normalized, step_number)
 

Auto_Use/macOS_use/agent/coder/view.py

@@ -141,4 +141,17 @@ def get_action_block(normalized_response: str) -> list:
             return json_data.get("action", [])
         except:
             pass
-        return []
+        return []
+
+    @staticmethod
+    def format_stream_json(normalized_response: str) -> str:
+        """Per-step payload for the streaming UI terminal — ONLY the action block (a real
+        terminal shows what's being executed, not thinking/goal/memory). The frontend reads
+        `action` from this for both the `>` stream and the tool-icon chain. On parse error,
+        fall back to the input unchanged.
+        """
+        try:
+            json_data = json.loads(normalized_response)
+            return json.dumps({"action": json_data.get("action", [])}, indent=2, ensure_ascii=False)
+        except Exception:
+            return normalized_response