Each agent folder is named after what it drives: computer_use on the …

AutoCua/agent_launcher.py

@@ -54,11 +54,11 @@ def resolve_agent_service(mode, os=None):
     if kind == "computer":
         host = platform.system()
         if host == "Darwin":
-            from AutoCua.mac.agent.main_driver.service import AgentService
+            from AutoCua.mac.agent.computer_use.service import AgentService
         elif host == "Windows":
-            from AutoCua.windows.agent.main_driver.service import AgentService
+            from AutoCua.windows.agent.computer_use.service import AgentService
         elif host == "Linux":
-            from AutoCua.linux.agent.main_driver.service import AgentService
+            from AutoCua.linux.agent.computer_use.service import AgentService
         else:
             raise RuntimeError(f"Unsupported OS for computer use: {host}")
         return AgentService
@@ -86,7 +86,7 @@ def resolve_agent_service(mode, os=None):
     if kind == "mobile":
         device = str(os or "").strip().lower()
         if device == "ios":
-            from AutoCua.ios.agent.main_driver.service import AgentService
+            from AutoCua.ios.agent.iphone_use.service import AgentService
             return AgentService
         if device == "android":
             raise NotImplementedError("Android support is not available yet")

AutoCua/ios/agent/__init__.py

@@ -1,5 +1,5 @@
 # AutoCua/ios/agent/__init__.py
-from .main_driver.service import AgentService
-from .main_driver.view import AgentResponseFormatter
+from .iphone_use.service import AgentService
+from .iphone_use.view import AgentResponseFormatter
 
 __all__ = ['AgentService', 'AgentResponseFormatter']

AutoCua/ios/agent/__main__.py

@@ -66,7 +66,7 @@ def main():
         # pymobiledevice3) and which one.
         active_target.update({"kind": "simulation", "udid": args.sim_udid})
 
-        from AutoCua.ios.agent.main_driver.service import AgentService
+        from AutoCua.ios.agent.iphone_use.service import AgentService
 
         agent = AgentService(
             provider=args.provider,

AutoCua/ios/agent/main_driver/__init__.py → AutoCua/ios/agent/iphone_use/__init__.py

@@ -1,4 +1,4 @@
-# AutoCua/ios/agent/main_driver/__init__.py
+# AutoCua/ios/agent/iphone_use/__init__.py
 from .service import AgentService
 from .view import AgentResponseFormatter
 

AutoCua/ios/tool_registry/service.py

@@ -94,7 +94,7 @@ def _main_tools(track: dict, notes: dict = None) -> list:
 
     Descriptions are the system prompt's OWN tool text, copied verbatim from
     <tool_capability> / <os_interaction> / <task_completion> in
-    agent/main_driver/system_prompt.md - the prompt stays the source of truth
+    agent/iphone_use/system_prompt.md - the prompt stays the source of truth
     (it is user-managed and never edited from here); this just puts the same
     words where a native tool call can read them. Built twice, each with NO
     per-call params (`track` stays {}) and a dedicated notes tool in front:

AutoCua/linux/agent/__init__.py

@@ -1,5 +1,5 @@
 # AutoCua/linux/agent/__init__.py
-from .main_driver.service import AgentService
-from .main_driver.view import AgentResponseFormatter
+from .computer_use.service import AgentService
+from .computer_use.view import AgentResponseFormatter
 
 __all__ = ['AgentService', 'AgentResponseFormatter']

AutoCua/linux/agent/main_driver/__init__.py → AutoCua/linux/agent/computer_use/__init__.py

@@ -1,4 +1,4 @@
-# AutoCua/linux/agent/main_driver/__init__.py
+# AutoCua/linux/agent/computer_use/__init__.py
 from .service import AgentService
 from .view import AgentResponseFormatter
 

AutoCua/linux/remote_connection/telegram/service.py

@@ -279,7 +279,7 @@ def _build_prior_history(task, assistant_messages, tool_responses,
     surface deliberately avoids. Packages a finished run's memory lists into
     the prior_history dict AgentService expects, capped with the synthetic
     terminal bridge step + None tool slot. The bridge JSON must contain
-    main_driver's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
+    computer_use's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
     verbatim, or the <updated_user_request> run-marker logic silently stops
     firing on resume. Returns None when the run produced no
     steps — the caller keeps the previous memory then. prior_task freezes the
@@ -870,7 +870,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
     try:
         # Imported lazily — pulls in tree/element → skimage etc., which we
         # don't want to load until a task actually runs.
-        from AutoCua.linux.agent.main_driver.service import AgentService
+        from AutoCua.linux.agent.computer_use.service import AgentService
 
         # Look up the runtime API key for the chosen provider so LLMManager
         # doesn't fall back to an os.getenv() the user never set. Telegram
@@ -888,7 +888,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
         # Pipe each step's formatted response (thinking + next_goal +
         # memory, with action stripped) into the compact banner.
         # The agent already calls text_callback at
-        # main_driver/service.py:704-705 with exactly this content — same
+        # computer_use/service.py:704-705 with exactly this content — same
         # path the frontend's streamAgentText uses in app.py. update()
         # serialises onto the banner's UI thread, so the agent loop never
         # blocks on it.

AutoCua/mac/agent/__init__.py

@@ -1,5 +1,5 @@
 # AutoCua/mac/agent/__init__.py
-from .main_driver.service import AgentService
-from .main_driver.view import AgentResponseFormatter
+from .computer_use.service import AgentService
+from .computer_use.view import AgentResponseFormatter
 
 __all__ = ['AgentService', 'AgentResponseFormatter']

AutoCua/mac/agent/main_driver/__init__.py → AutoCua/mac/agent/computer_use/__init__.py

@@ -1,4 +1,4 @@
-# AutoCua/mac/agent/main_driver/__init__.py
+# AutoCua/mac/agent/computer_use/__init__.py
 from .service import AgentService
 from .view import AgentResponseFormatter
 

AutoCua/mac/controller/tool/ios_agent_child.py

@@ -106,7 +106,7 @@ def main():
         from AutoCua.ios_connector.session import active_target
         active_target.update({"kind": "hardware", "udid": os.environ.get("PYMOBILEDEVICE3_UDID")})
 
-        from AutoCua.ios.agent.main_driver.service import AgentService
+        from AutoCua.ios.agent.iphone_use.service import AgentService
 
         # Only these two: web_callback and shell_callback add sleeps, and
         # token_callback would move the main agent's memory bar

AutoCua/mac/remote_connection/telegram/service.py

@@ -277,7 +277,7 @@ def _build_prior_history(task, assistant_messages, tool_responses,
     surface deliberately avoids. Packages a finished run's memory lists into
     the prior_history dict AgentService expects, capped with the synthetic
     terminal bridge step + None tool slot. The bridge JSON must contain
-    main_driver's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
+    computer_use's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
     verbatim, or the <updated_user_request> run-marker logic silently stops
     firing on resume. Returns None when the run produced no
     steps — the caller keeps the previous memory then. prior_task freezes the
@@ -831,7 +831,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
     try:
         # Imported lazily — pulls in tree/element → skimage etc., which we
         # don't want to load until a task actually runs.
-        from AutoCua.mac.agent.main_driver.service import AgentService
+        from AutoCua.mac.agent.computer_use.service import AgentService
 
         # Look up the runtime API key for the chosen provider so LLMManager
         # doesn't fall back to an os.getenv() the user never set. Telegram
@@ -849,7 +849,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
         # Pipe each step's formatted response (thinking + next_goal +
         # memory, with action stripped) into the compact banner.
         # The agent already calls text_callback at
-        # main_driver/service.py:704-705 with exactly this content — same
+        # computer_use/service.py:704-705 with exactly this content — same
         # path the frontend's streamAgentText uses in app.py. update()
         # serialises onto the banner's UI thread, so the agent loop never
         # blocks on it.

AutoCua/ui/service.py

@@ -2246,10 +2246,11 @@ def monitor_todo():
                     agent_glow.start()
                 # The browser agent is the Rust extension (AutoCua/web/agent
                 # holds .rs, not .py): its AgentService is re-exported from the
-                # package facade. Every other agent is a Python main_driver.
+                # package facade. iOS is agent/iphone_use; mac, windows and linux are agent/computer_use.
                 AgentService = importlib.import_module(
                     "AutoCua.web.agent" if run_pkg == 'web'
-                    else f"AutoCua.{run_pkg}.agent.main_driver.service"
+                    else "AutoCua.ios.agent.iphone_use.service" if run_pkg == 'ios'
+                    else f"AutoCua.{run_pkg}.agent.computer_use.service"
                 ).AgentService
 
                 agent = AgentService(
@@ -3684,10 +3685,10 @@ def delete_chat(chat_id):
 
 
 def _read_main_system_prompt():
-    """Best-effort read of the active platform's main_driver system prompt, so a
+    """Best-effort read of the active platform's computer_use system prompt, so a
     downloaded conversation log shows the same SYSTEM PROMPT block as a main.py run."""
     try:
-        p = get_platform_path() / "agent" / "main_driver" / "system_prompt.md"
+        p = get_platform_path() / "agent" / "computer_use" / "system_prompt.md"
         with open(p, "r", encoding="utf-8") as f:
             return f.read()
     except Exception:

AutoCua/windows/agent/__init__.py

@@ -1,4 +1,4 @@
-from .main_driver.service import AgentService
-from .main_driver.view import AgentResponseFormatter
+from .computer_use.service import AgentService
+from .computer_use.view import AgentResponseFormatter
 
 __all__ = ['AgentService', 'AgentResponseFormatter']

AutoCua/windows/agent/main_driver/__init__.py → AutoCua/windows/agent/computer_use/__init__.py

@@ -1,4 +1,4 @@
-# AutoCua/windows/agent/main_driver/__init__.py
+# AutoCua/windows/agent/computer_use/__init__.py
 from .service import AgentService
 from .view import AgentResponseFormatter
 

AutoCua/windows/agent/minions/claude_prompt.md

@@ -276,7 +276,7 @@ Rules for the report:
 **OS: Windows PowerShell. You are READ-ONLY.**
 1. You MUST NEVER modify the filesystem. No editing, creating, deleting, moving, or renaming files. No `Set-Content`, `Add-Content`, `Out-File`, `New-Item`, `Remove-Item`, `Move-Item`, `Rename-Item`, redirection (`>`, `>>`), or any side-effecting shell command. You have NO `write` tool and NO `replace` tool. If you find yourself wanting to edit, instead record the exact location in your final report so the parent agent can apply the change.
 2. Drill-down workflow: start broad (`glob`/`grep` to find candidates), then narrow (`view` exact ranges) - never dump whole large files into context. Standard pair: `grep` (locate the line), then `view` (read a 20-50 line range around the hit).
-3. Always anchor findings to `path:line_no` (e.g. `AutoCua\windows\agent\main_driver\service.py:418`). Vague references like "somewhere in service.py" are never acceptable.
+3. Always anchor findings to `path:line_no` (e.g. `AutoCua\windows\agent\computer_use\service.py:418`). Vague references like "somewhere in service.py" are never acceptable.
 4. For change requests: trace every connection - definition site, every caller, every place that reads/writes the affected state, related tests, related prompts. Report ALL of them, not just the obvious one. Missing one place = parent agent ships a broken change.
 5. Keep running notes in `<scratchpad>` after every confirmed finding so they survive across iterations and assemble into the final report.
 6. When `view` shows `[line_number] text`, those numbers are the file's real line numbers - quote them exactly in your report. For files over 2000 lines, whole-file `view` returns only the first 2000 plus a footer showing the total — re-call `view` with `start`/`end` for other sections.

AutoCua/windows/agent/minions/system_prompt.md

@@ -302,7 +302,7 @@ Rules for the report:
 **OS: Windows PowerShell. You are READ-ONLY.**
 1. You MUST NEVER modify the filesystem. No editing, creating, deleting, moving, or renaming files. No `Set-Content`, `Add-Content`, `Out-File`, `New-Item`, `Remove-Item`, `Move-Item`, `Rename-Item`, redirection (`>`, `>>`), or any side-effecting shell command. You have NO `write` tool and NO `replace` tool. If you find yourself wanting to edit, instead record the exact location in your final report so the parent agent can apply the change.
 2. Drill-down workflow: start broad (`glob`/`grep` to find candidates), then narrow (`view` exact ranges) - never dump whole large files into context. Standard pair: `grep` (locate the line), then `view` (read a 20-50 line range around the hit).
-3. Always anchor findings to `path:line_no` (e.g. `AutoCua\windows\agent\main_driver\service.py:418`). Vague references like "somewhere in service.py" are never acceptable.
+3. Always anchor findings to `path:line_no` (e.g. `AutoCua\windows\agent\computer_use\service.py:418`). Vague references like "somewhere in service.py" are never acceptable.
 4. For change requests: trace every connection - definition site, every caller, every place that reads/writes the affected state, related tests, related prompts. Report ALL of them, not just the obvious one. Missing one place = parent agent ships a broken change.
 5. Keep running notes in `<scratchpad>` after every confirmed finding so they survive across iterations and assemble into the final report.
 6. When `view` shows `[line_number] text`, those numbers are the file's real line numbers - quote them exactly in your report. For files over 2000 lines, whole-file `view` returns only the first 2000 plus a footer showing the total — re-call `view` with `start`/`end` for other sections.

AutoCua/windows/remote_connection/telegram/service.py

@@ -278,7 +278,7 @@ def _build_prior_history(task, assistant_messages, tool_responses,
     surface deliberately avoids. Packages a finished run's memory lists into
     the prior_history dict AgentService expects, capped with the synthetic
     terminal bridge step + None tool slot. The bridge JSON must contain
-    main_driver's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
+    computer_use's _BRIDGE_SIGNATURE ('"next_goal": "Previous run concluded.')
     verbatim, or the <updated_user_request> run-marker logic silently stops
     firing on resume. Returns None when the run produced no
     steps — the caller keeps the previous memory then. prior_task freezes the
@@ -832,7 +832,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
     try:
         # Imported lazily — pulls in tree/element → skimage etc., which we
         # don't want to load until a task actually runs.
-        from AutoCua.windows.agent.main_driver.service import AgentService
+        from AutoCua.windows.agent.computer_use.service import AgentService
 
         # Look up the runtime API key for the chosen provider so
         # LLMManager doesn't fall back to an os.getenv() that the user
@@ -852,7 +852,7 @@ def _run_agent(task, provider, model, chat_id, bot, loop, prior_history=None,
         # Pipe each step's formatted response (thinking + next_goal +
         # memory, with action stripped) into the compact banner.
         # The agent already calls text_callback at
-        # main_driver/service.py with exactly this content — same path the
+        # computer_use/service.py with exactly this content — same path the
         # frontend's streamAgentText uses in app.py. update() forwards via
         # the MSG stdin command to the banner subprocess; the call returns
         # quickly so the agent loop never blocks on it.

README.md

@@ -1159,7 +1159,7 @@ AutoCua/
   default_skills/          the shipped skills, read-only; yours go in AutoCua_data/skills/
   llm_provider/            every LLM endpoint and the model tables, one copy for all platforms
   mac/  windows/           computer use: agent, controller, tree, sandbox, tool_registry
-    agent/main_driver/       the desktop loop
+    agent/computer_use/      the desktop loop
     agent/coder/             the coding agent
     agent/minions/           read-only scouts
     controller/tool/         shell, open_app, screenshot, applescript, kernel_input

agent_capabilities.md

@@ -366,7 +366,7 @@ run_agent(
 ## Under the hood
 
 Each platform's `AgentService.__init__` (e.g.
-[AutoCua/windows/agent/main_driver/service.py](AutoCua/windows/agent/main_driver/service.py))
+[AutoCua/windows/agent/computer_use/service.py](AutoCua/windows/agent/computer_use/service.py))
 normalises `speed`, loads `fast_system_prompt.md` or `system_prompt.md` from
 the same folder, and passes the mode to `LLMManager`, which sends the
 `MAIN_TOOLS_FAST` registry (notes tool `memory`) or `MAIN_TOOLS` (notes tool
@@ -446,7 +446,7 @@ Remove the argument (or pass `False`) to stop writing logs.
 ## Under the hood
 
 Each `AgentService.__init__` (e.g.
-[AutoCua/windows/agent/main_driver/service.py](AutoCua/windows/agent/main_driver/service.py))
+[AutoCua/windows/agent/computer_use/service.py](AutoCua/windows/agent/computer_use/service.py))
 resets `conversation/` and `raw_reasoning/` and, when the flag is on, calls
 `_save_conversation_snapshot()` after every LLM step to write
 `conversation_N.txt`. The web agent does the same in Rust