droidrun 0.3.10.dev3__py3-none-any.whl → 0.3.10.dev5__py3-none-any.whl

droidrun/agent/executor/executor_agent.py

@@ -9,6 +9,7 @@ This agent is responsible for:
 
 from __future__ import annotations
 
+import asyncio
 import json
 import logging
 from typing import TYPE_CHECKING
@@ -17,12 +18,20 @@ from llama_index.core.llms import ChatMessage, ImageBlock, TextBlock
 from llama_index.core.llms.llm import LLM
 from llama_index.core.workflow import Context, StartEvent, StopEvent, Workflow, step
 
-from droidrun.agent.executor.events import ExecutorActionEvent, ExecutorResultEvent
-from droidrun.agent.executor.prompts import build_executor_system_prompt, parse_executor_response
-from droidrun.agent.utils.tools import click, long_press, open_app, swipe, system_button, type
+from droidrun.agent.executor.events import ExecutorInternalActionEvent, ExecutorInternalResultEvent
+from droidrun.agent.executor.prompts import parse_executor_response
 from droidrun.agent.utils.inference import acall_with_retries
-from droidrun.config_manager import config
-import asyncio
+from droidrun.agent.utils.tools import (
+    ATOMIC_ACTION_SIGNATURES,
+    click,
+    long_press,
+    open_app,
+    swipe,
+    system_button,
+    type,
+)
+from droidrun.config_manager.config_manager import AgentConfig
+from droidrun.config_manager.prompt_loader import PromptLoader
 
 if TYPE_CHECKING:
     from droidrun.agent.droid.events import DroidAgentState
@@ -30,7 +39,7 @@ if TYPE_CHECKING:
 logger = logging.getLogger("droidrun")
 
 
-class ExecutorAgent(Workflow):
+class ExecutorAgent(Workflow): # TODO: Fix a bug in bad prompt
     """
     Action execution agent that performs specific actions.
 
@@ -45,22 +54,20 @@ class ExecutorAgent(Workflow):
     def __init__(
         self,
         llm: LLM,
-        vision: bool,
         tools_instance,
         shared_state: "DroidAgentState",
-        persona=None,
+        agent_config: AgentConfig,
         custom_tools: dict = None,
-        debug: bool = False,
         **kwargs
     ):
         super().__init__(**kwargs)
         self.llm = llm
-        self.vision = vision
+        self.agent_config = agent_config
+        self.config = agent_config.executor
+        self.vision = agent_config.executor.vision
         self.tools_instance = tools_instance
         self.shared_state = shared_state
-        self.persona = persona
         self.custom_tools = custom_tools or {}
-        self.debug = debug
 
         logger.info("✅ ExecutorAgent initialized successfully.")
 
@@ -70,7 +77,7 @@ class ExecutorAgent(Workflow):
         self,
         ctx: Context,
         ev: StartEvent
-    ) -> ExecutorActionEvent:
+    ) -> ExecutorInternalActionEvent:
         """
         Executor decides which action to take.
 
@@ -83,12 +90,53 @@ class ExecutorAgent(Workflow):
         subgoal = ev.get("subgoal", "")
         logger.info(f"🧠 Executor thinking about action for: {subgoal}")
 
+        # Format app card (include tags in variable value or empty string)
         app_card = ""  # TODO: Implement app card retrieval
-
-        system_prompt = build_executor_system_prompt(
-            state=self.shared_state,
-            subgoal=subgoal,
-            app_card=app_card
+        app_card_text = ""
+        if app_card.strip():
+            app_card_text = "App card gives information on how to operate the app and perform actions.\n### App Card ###\n" + app_card.strip() + "\n\n"
+
+        # Format device state (use unified state)
+        device_state_text = ""
+        if self.shared_state.formatted_device_state and self.shared_state.formatted_device_state.strip():
+            device_state_text = "### Device State ###\n" + self.shared_state.formatted_device_state.strip() + "\n\n"
+
+        # Format progress status
+        progress_status_text = self.shared_state.progress_status + "\n\n" if self.shared_state.progress_status else "No progress yet.\n\n"
+
+        # Format atomic actions
+        atomic_actions_text = chr(10).join(
+            f"- {action_name}({', '.join(action_info['arguments'])}): {action_info['description']}"
+            for action_name, action_info in ATOMIC_ACTION_SIGNATURES.items()
+        ) + "\n"
+
+        # Format action history
+        if self.shared_state.action_history:
+            action_history_text = "Recent actions you took previously and whether they were successful:\n" + "\n".join(
+                (f"Action: {act} | Description: {summ} | Outcome: Successful" if outcome
+                 else f"Action: {act} | Description: {summ} | Outcome: Failed | Feedback: {err_des}")
+                for act, summ, outcome, err_des in zip(
+                    self.shared_state.action_history[-min(5, len(self.shared_state.action_history)):],
+                    self.shared_state.summary_history[-min(5, len(self.shared_state.action_history)):],
+                    self.shared_state.action_outcomes[-min(5, len(self.shared_state.action_history)):],
+                    self.shared_state.error_descriptions[-min(5, len(self.shared_state.action_history)):], strict=True)
+            ) + "\n\n"
+        else:
+            action_history_text = "No actions have been taken yet.\n\n"
+
+        # Load and format prompt
+        system_prompt = PromptLoader.load_prompt(
+            self.agent_config.get_executor_system_prompt_path(),
+            {
+                "instruction": self.shared_state.instruction,
+                "app_card": app_card_text,
+                "device_state_text": device_state_text,
+                "plan": self.shared_state.plan,
+                "subgoal": subgoal,
+                "progress_status": progress_status_text,
+                "atomic_actions": atomic_actions_text,
+                "action_history": action_history_text
+            }
         )
 
         blocks = [TextBlock(text=system_prompt)]
@@ -112,7 +160,7 @@ class ExecutorAgent(Workflow):
             parsed = parse_executor_response(response_text)
         except Exception as e:
             logger.error(f"❌ Failed to parse executor response: {e}")
-            return ExecutorActionEvent(
+            return ExecutorInternalActionEvent(
                 action_json=json.dumps({"action": "invalid"}),
                 thought=f"Failed to parse response: {str(e)}",
                 description="Invalid response format from LLM"
@@ -122,18 +170,23 @@ class ExecutorAgent(Workflow):
         logger.info(f"🎯 Action: {parsed['action']}")
         logger.debug(f"  - Description: {parsed['description']}")
 
-        return ExecutorActionEvent(
+        event = ExecutorInternalActionEvent(
             action_json=parsed["action"],
             thought=parsed["thought"],
             description=parsed["description"]
         )
 
+        # Write event to stream for web interface
+        ctx.write_event_to_stream(event)
+
+        return event
+
     @step
     async def execute(
         self,
         ctx: Context,
-        ev: ExecutorActionEvent
-    ) -> ExecutorResultEvent:
+        ev: ExecutorInternalActionEvent
+    ) -> ExecutorInternalResultEvent:
         """
         Execute the selected action using the tools instance.
 
@@ -146,7 +199,7 @@ class ExecutorAgent(Workflow):
             action_dict = json.loads(ev.action_json)
         except json.JSONDecodeError as e:
             logger.error(f"❌ Failed to parse action JSON: {e}")
-            return ExecutorResultEvent(
+            return ExecutorInternalResultEvent(
                 action={"action": "invalid"},
                 outcome=False,
                 error=f"Invalid action JSON: {str(e)}",
@@ -155,15 +208,15 @@ class ExecutorAgent(Workflow):
                 action_json=ev.action_json
             )
 
-        # Execute the action
         outcome, error, summary = await self._execute_action(action_dict, ev.description)
 
-        if outcome:
-            await asyncio.sleep(config.agent.after_sleep_action)
+        # TODO: Add sleep after action (should be in DroidAgent.handle_executor_result)
+        # Available via: self.agent_config.after_sleep_action
+        # await asyncio.sleep(self.agent_config.after_sleep_action)
 
         logger.info(f"{'✅' if outcome else '❌'} Execution complete: {summary}")
 
-        return ExecutorResultEvent(
+        result_event = ExecutorInternalResultEvent(
             action=action_dict,
             outcome=outcome,
             error=error,
@@ -172,6 +225,11 @@ class ExecutorAgent(Workflow):
             action_json=ev.action_json
         )
 
+        # Write event to stream for web interface
+        ctx.write_event_to_stream(result_event)
+
+        return result_event
+
     async def _execute_action(self, action_dict: dict, description: str) -> tuple[bool, str, str]:
         """
         Execute a single action based on the action dictionary.
@@ -312,7 +370,7 @@ class ExecutorAgent(Workflow):
     async def finalize(
         self,
         ctx: Context,
-        ev: ExecutorResultEvent
+        ev: ExecutorInternalResultEvent
     ) -> StopEvent:
         """Return executor results to parent workflow."""
         logger.debug("✅ Executor execution complete")

droidrun/agent/executor/prompts.py

@@ -76,7 +76,7 @@ The atomic action functions are listed in the format of `action(arguments): desc
 \n
 ### Latest Action History ###
 {(("Recent actions you took previously and whether they were successful:\n" + "\n".join(
-    (f"Action: {act} | Description: {summ} | Outcome: Successful" if outcome == "A"
+    (f"Action: {act} | Description: {summ} | Outcome: Successful" if outcome
      else f"Action: {act} | Description: {summ} | Outcome: Failed | Feedback: {err_des}")
     for act, summ, outcome, err_des in zip(
         state.action_history[-min(5, len(state.action_history)):],
@@ -126,7 +126,13 @@ def parse_executor_response(response: str) -> dict:
         Dictionary with 'thought', 'action', 'description' keys
     """
     thought = response.split("### Thought")[-1].split("### Action")[0].replace("\n", " ").replace("  ", " ").replace("###", "").strip()
-    action = response.split("### Action")[-1].split("### Description")[0].replace("\n", " ").replace("  ", " ").replace("###", "").strip()
+    action_raw = response.split("### Action")[-1].split("### Description")[0].replace("\n", " ").replace("  ", " ").replace("###", "").strip()
+    start_idx = action_raw.find('{')
+    end_idx = action_raw.rfind('}')
+    if start_idx != -1 and end_idx != -1:
+        action = action_raw[start_idx:end_idx + 1]
+    else:
+        action = action_raw
     description = response.split("### Description")[-1].replace("\n", " ").replace("  ", " ").replace("###", "").strip()
 
     return {

droidrun/agent/manager/__init__.py

@@ -2,17 +2,16 @@
 Manager Agent - Planning and reasoning workflow.
 """
 
-from droidrun.agent.manager.events import ManagerPlanEvent, ManagerThinkingEvent
+from droidrun.agent.droid.events import ManagerInputEvent, ManagerPlanEvent
+from droidrun.agent.manager.events import ManagerThinkingEvent, ManagerInternalPlanEvent
 from droidrun.agent.manager.manager_agent import ManagerAgent
-from droidrun.agent.manager.prompts import (
-    build_manager_system_prompt,
-    parse_manager_response,
-)
+from droidrun.agent.manager.prompts import parse_manager_response
 
 __all__ = [
     "ManagerAgent",
-    "ManagerThinkingEvent",
+    "ManagerInputEvent",
     "ManagerPlanEvent",
-    "build_manager_system_prompt",
+    "ManagerThinkingEvent",
+    "ManagerInternalPlanEvent",
     "parse_manager_response",
 ]

droidrun/agent/manager/events.py

@@ -1,5 +1,11 @@
 """
 Events for the ManagerAgent workflow.
+
+These are INTERNAL events used within ManagerAgent for:
+- Streaming to frontend/logging
+- Carrying full debug metadata
+
+For workflow coordination with DroidAgent, see droid/events.py
 """
 
 from llama_index.core.workflow.events import Event
@@ -10,11 +16,17 @@ class ManagerThinkingEvent(Event):
     pass
 
 
-class ManagerPlanEvent(Event):
-    """Manager has created a plan"""
+class ManagerInternalPlanEvent(Event):
+    """
+    Internal Manager planning event with full state and metadata.
+
+    This event is streamed to frontend/logging but NOT used for
+    workflow coordination between ManagerAgent and DroidAgent.
+
+    For workflow coordination, see ManagerPlanEvent in droid/events.py
+    """
     plan: str
     current_subgoal: str
-    completed_plan: str
     thought: str
     manager_answer: str = ""
-    memory_update: str = ""
+    memory_update: str = ""  # Debugging metadata: LLM's memory additions

droidrun/agent/manager/manager_agent.py

@@ -11,22 +11,26 @@ This agent is responsible for:
 from __future__ import annotations
 
 import logging
-from typing import TYPE_CHECKING, List
+from typing import TYPE_CHECKING
 
 from llama_index.core.llms.llm import LLM
 from llama_index.core.workflow import Context, StartEvent, StopEvent, Workflow, step
 
-from droidrun.agent.manager.events import ManagerPlanEvent, ManagerThinkingEvent
-from droidrun.agent.manager.prompts import build_manager_system_prompt, parse_manager_response
+from droidrun.agent.manager.events import ManagerInternalPlanEvent, ManagerThinkingEvent
+from droidrun.agent.manager.prompts import parse_manager_response
 from droidrun.agent.utils import convert_messages_to_chatmessages
 from droidrun.agent.utils.chat_utils import remove_empty_messages
-from droidrun.agent.utils.device_state_formatter import get_device_state_exact_format
+from droidrun.agent.utils.device_state_formatter import format_device_state
 from droidrun.agent.utils.inference import acall_with_retries
 from droidrun.agent.utils.tools import build_custom_tool_descriptions
+from droidrun.config_manager.prompt_loader import PromptLoader
+from droidrun.config_manager.app_card_loader import AppCardLoader
 
 if TYPE_CHECKING:
     from droidrun.agent.droid.events import DroidAgentState
     from droidrun.tools import Tools
+    from droidrun.config_manager.config_manager import AgentConfig
+
 
 logger = logging.getLogger("droidrun")
 
@@ -45,22 +49,21 @@ class ManagerAgent(Workflow):
     def __init__(
         self,
         llm: LLM,
-        vision: bool,
-        personas: List,
         tools_instance: "Tools",
         shared_state: "DroidAgentState",
+        agent_config: "AgentConfig",
         custom_tools: dict = None,
-        debug: bool = False,
         **kwargs
     ):
         super().__init__(**kwargs)
         self.llm = llm
-        self.vision = vision
-        self.personas = personas
+        self.config = agent_config.manager
+        self.vision = self.config.vision
         self.tools_instance = tools_instance
         self.shared_state = shared_state
         self.custom_tools = custom_tools or {}
-        self.debug = debug
+        self.agent_config = agent_config
+        self.app_card_loader = self.agent_config.app_cards
 
         logger.info("✅ ManagerAgent initialized successfully.")
 
@@ -70,23 +73,23 @@ class ManagerAgent(Workflow):
 
     def _build_system_prompt(
         self,
-        has_text_to_modify: bool
+        has_text_to_modify: bool,
+        app_card: str = ""
     ) -> str:
         """
         Build system prompt with all context.
 
         Args:
             has_text_to_modify: Whether text manipulation mode is enabled
-
+            app_card: App card content
         Returns:
             Complete system prompt
         """
-
-        # Get error history if error_flag_plan is set
-        error_history = []
+        # Format error history
+        error_history_text = ""
         if self.shared_state.error_flag_plan:
             k = self.shared_state.err_to_manager_thresh
-            error_history = [
+            errors = [
                 {
                     "action": act,
                     "summary": summ,
@@ -98,19 +101,76 @@ class ManagerAgent(Workflow):
                     self.shared_state.error_descriptions[-k:], strict=True
                 )
             ]
-
-        # Build custom tools descriptions
-        custom_tools_descriptions = build_custom_tool_descriptions(self.custom_tools)
-
-        return build_manager_system_prompt(
-            instruction=self.shared_state.instruction,
-            has_text_to_modify=has_text_to_modify,
-            app_card="",  # TODO: implement app card retrieval system
-            device_date=self.tools_instance.get_date(),
-            important_notes="",  # TODO: expose important_notes in DroidAgentState if needed
-            error_flag=self.shared_state.error_flag_plan,
-            error_history=error_history,
-            custom_tools_descriptions=custom_tools_descriptions
+            error_history_text = (
+                "<potentially_stuck>\n"
+                "You have encountered several failed attempts. Here are some logs:\n"
+            )
+            for error in errors:
+                error_history_text += (
+                    f"- Attempt: Action: {error['action']} | "
+                    f"Description: {error['summary']} | "
+                    f"Outcome: Failed | "
+                    f"Feedback: {error['error']}\n"
+                )
+            error_history_text += "</potentially_stuck>\n\n"
+
+        # Text manipulation section
+        text_manipulation_section = ""
+        if has_text_to_modify:
+            text_manipulation_section = """
+
+<text_manipulation>
+1. Use **TEXT_TASK:** prefix in your plan when you need to modify text in the currently focused text input field
+2. TEXT_TASK is for editing, formatting, or transforming existing text content in text boxes using Python code
+3. Do not use TEXT_TASK for extracting text from messages, typing new text, or composing messages
+4. The focused text field contains editable text that you can modify
+5. Example plan item: 'TEXT_TASK: Add "Hello World" at the beginning of the text'
+6. Always use TEXT_TASK for modifying text, do not try to select the text to copy/cut/paste or adjust the text
+</text_manipulation>"""
+
+        # Device date (include tags in variable value or empty string)
+        device_date = self.tools_instance.get_date()
+        device_date_text = ""
+        if device_date.strip():
+            device_date_text = f"<device_date>\n{device_date}\n</device_date>\n\n"
+
+        # App card (include tags in variable value or empty string)
+        app_card = app_card
+        app_card_text = ""
+        if app_card.strip():
+            app_card_text = "App card gives information on how to operate the app and perform actions.\n<app_card>\n" + app_card.strip() + "\n</app_card>\n\n"
+
+        # Important notes (include tags in variable value or empty string)
+        important_notes = ""  # TODO: implement
+        important_notes_text = ""
+        if important_notes.strip():
+            important_notes_text = "<important_notes>\n" + important_notes + "\n</important_notes>\n\n"
+
+        # Custom tools
+        custom_tools_desc = build_custom_tool_descriptions(self.custom_tools)
+        custom_tools_text = ""
+        if custom_tools_desc.strip():
+            custom_tools_text = """
+
+<custom_actions>
+The executor has access to these additional custom actions beyond the standard actions (click, type, swipe, etc.):
+""" + custom_tools_desc + """
+
+You can reference these custom actions or tell the Executer agent to use them in your plan when they help achieve the user's goal.
+</custom_actions>"""
+
+        # Load and format prompt
+        return PromptLoader.load_prompt(
+            self.agent_config.get_manager_system_prompt_path(),
+            {
+                "instruction": self.shared_state.instruction,
+                "device_date": device_date_text,
+                "app_card": app_card_text,
+                "important_notes": important_notes_text,
+                "error_history": error_history_text,
+                "text_manipulation_section": text_manipulation_section,
+                "custom_tools_descriptions": custom_tools_text
+            }
         )
 
     def _build_messages_with_context(
@@ -155,28 +215,28 @@ class ManagerAgent(Workflow):
                 else:
                     messages[last_user_idx]['content'].insert(0, {"text": f"<memory>\n{current_memory}\n</memory>\n"})
 
-            # Add device state to last user message
-            current_a11y = (self.shared_state.ui_elements_list_after or self.shared_state.device_state_text or "").strip()
-            if current_a11y:
+            # Add CURRENT device state to last user message (use unified state)
+            current_state = self.shared_state.formatted_device_state.strip()
+            if current_state:
                 if messages[last_user_idx]['content'] and 'text' in messages[last_user_idx]['content'][0]:
-                    messages[last_user_idx]['content'][0]['text'] += f"\n<device_state>\n{current_a11y}\n</device_state>\n"
+                    messages[last_user_idx]['content'][0]['text'] += f"\n<device_state>\n{current_state}\n</device_state>\n"
                 else:
-                    messages[last_user_idx]['content'].insert(0, {"text": f"<device_state>\n{current_a11y}\n</device_state>\n"})
+                    messages[last_user_idx]['content'].insert(0, {"text": f"<device_state>\n{current_state}\n</device_state>\n"})
 
             # Add screenshot to last user message
             if screenshot and self.vision:
                 messages[last_user_idx]['content'].append({"image": screenshot})
 
-            # Add previous device state to SECOND-TO-LAST user message (if exists)
+            # Add PREVIOUS device state to SECOND-TO-LAST user message (if exists)
             if len(user_indices) >= 2:
                 second_last_user_idx = user_indices[-2]
-                prev_a11y = (self.shared_state.ui_elements_list_before or "").strip()
+                prev_state = self.shared_state.previous_formatted_device_state.strip()
 
-                if prev_a11y:
+                if prev_state:
                     if messages[second_last_user_idx]['content'] and 'text' in messages[second_last_user_idx]['content'][0]:
-                        messages[second_last_user_idx]['content'][0]['text'] += f"\n<device_state>\n{prev_a11y}\n</device_state>\n"
+                        messages[second_last_user_idx]['content'][0]['text'] += f"\n<device_state>\n{prev_state}\n</device_state>\n"
                     else:
-                        messages[second_last_user_idx]['content'].insert(0, {"text": f"<device_state>\n{prev_a11y}\n</device_state>\n"})
+                        messages[second_last_user_idx]['content'].insert(0, {"text": f"<device_state>\n{prev_state}\n</device_state>\n"})
         messages = remove_empty_messages(messages)
         return messages
 
@@ -263,9 +323,23 @@ class ManagerAgent(Workflow):
         logger.info("💬 Preparing manager input...")
 
         # ====================================================================
-        # Step 1: Get device state (UI elements accessibility tree)
+        # Step 1: Get and format device state using unified formatter
         # ====================================================================
-        device_state_text, focused_text = get_device_state_exact_format(self.tools_instance.get_state())
+        raw_state = self.tools_instance.get_state()
+        formatted_text, focused_text, a11y_tree, phone_state = format_device_state(raw_state)
+
+        # Update shared state (previous ← current, current ← new)
+        self.shared_state.previous_formatted_device_state = self.shared_state.formatted_device_state
+        self.shared_state.formatted_device_state = formatted_text
+        self.shared_state.focused_text = focused_text
+        self.shared_state.a11y_tree = a11y_tree
+        self.shared_state.phone_state = phone_state
+
+        # Extract and store package/app name
+        self.shared_state.current_package_name = phone_state.get('packageName', 'Unknown')
+        self.shared_state.current_app_name = phone_state.get('currentApp', 'Unknown')
+
+        # App cards
 
         # ====================================================================
         # Step 2: Capture screenshot if vision enabled
@@ -278,6 +352,7 @@ class ManagerAgent(Workflow):
                     success, screenshot = result
                     if not success:
                         screenshot = None
+
                 else:
                     screenshot = result
                 logger.debug("📸 Screenshot captured for Manager")
@@ -288,29 +363,9 @@ class ManagerAgent(Workflow):
         # ====================================================================
         # Step 3: Detect text manipulation mode
         # ====================================================================
-        focused_text = focused_text or ""
         focused_text_clean = focused_text.replace("'", "").strip()
-
-        # Check if focused text differs from last typed text
-        # last_typed_text = ""
-        # if self.shared_state.action_history:
-        #     recent_actions = self.shared_state.action_history[-1:] if len(self.shared_state.action_history) >= 1 else []
-        #     for action in reversed(recent_actions):
-        #         if isinstance(action, dict) and action.get('action') == 'type':
-        #             last_typed_text = action.get('text', '')
-        #             break
-
         has_text_to_modify = (focused_text_clean != "")
 
-        # ====================================================================
-        # Step 4: Update state with device info
-        # ====================================================================
-        self.shared_state.device_state_text = device_state_text
-        self.shared_state.focused_text = focused_text
-        # Shift UI elements: before ← after, after ← current
-        self.shared_state.ui_elements_list_before = self.shared_state.ui_elements_list_after
-        self.shared_state.ui_elements_list_after = device_state_text
-
         # ====================================================================
         # Step 5: Build user message entry
         # ====================================================================
@@ -328,7 +383,7 @@ class ManagerAgent(Workflow):
         if self.shared_state.last_summary:
             parts.append(f"<last_action_description>\n{self.shared_state.last_summary}\n</last_action_description>\n")
 
-        
+
         self.shared_state.message_history.append({
             "role": "user",
             "content": [{"text": "".join(parts)}]
@@ -346,7 +401,7 @@ class ManagerAgent(Workflow):
         self,
         ctx: Context,
         ev: ManagerThinkingEvent
-    ) -> ManagerPlanEvent:
+    ) -> ManagerInternalPlanEvent:
         """
         Manager reasons and creates plan.
 
@@ -362,11 +417,15 @@ class ManagerAgent(Workflow):
 
         has_text_to_modify = self.shared_state.has_text_to_modify
         screenshot = self.shared_state.screenshot
+        if self.app_card_loader.enabled:
+            app_card = AppCardLoader.load_app_card(self.shared_state.current_package_name, self.app_card_loader.app_cards_dir)
+        else:
+            app_card = ""
 
         # ====================================================================
         # Step 1: Build system prompt
         # ====================================================================
-        system_prompt = self._build_system_prompt(has_text_to_modify)
+        system_prompt = self._build_system_prompt(has_text_to_modify, app_card)
 
         # ====================================================================
         # Step 2: Build messages with context
@@ -423,7 +482,6 @@ class ManagerAgent(Workflow):
         # Update planning fields
         self.shared_state.plan = parsed["plan"]
         self.shared_state.current_subgoal = parsed["current_subgoal"]
-        self.shared_state.completed_plan = parsed.get("completed_subgoal", "No completed subgoal.")
         self.shared_state.finish_thought = parsed["thought"]
         self.shared_state.manager_answer = parsed["answer"]
 
@@ -431,20 +489,24 @@ class ManagerAgent(Workflow):
         logger.debug(f"  - Current subgoal: {parsed['current_subgoal']}")
         logger.debug(f"  - Manager answer: {parsed['answer'][:50] if parsed['answer'] else 'None'}")
 
-        return ManagerPlanEvent(
+        event = ManagerInternalPlanEvent(
             plan=parsed["plan"],
             current_subgoal=parsed["current_subgoal"],
-            completed_plan=parsed.get("completed_subgoal", "No completed subgoal."),
             thought=parsed["thought"],
             manager_answer=parsed["answer"],
             memory_update=memory_update
         )
 
+        # Write event to stream for web interface
+        ctx.write_event_to_stream(event)
+
+        return event
+
     @step
     async def finalize(
         self,
         ctx: Context,
-        ev: ManagerPlanEvent
+        ev: ManagerInternalPlanEvent
     ) -> StopEvent:
         """Return manager results to parent workflow."""
         logger.debug("✅ Manager planning complete")
@@ -452,7 +514,6 @@ class ManagerAgent(Workflow):
         return StopEvent(result={
             "plan": ev.plan,
             "current_subgoal": ev.current_subgoal,
-            "completed_plan": ev.completed_plan,
             "thought": ev.thought,
             "manager_answer": ev.manager_answer,
             "memory_update": ev.memory_update