Merge pull request #240 from vvincent1234/feat/deep-research

Feat/deep research

Author: warmshao

src/agent/custom_agent.py

@@ -111,6 +111,8 @@ def __init__(
 
         # record last actions
         self._last_actions = None
+        # record extract content
+        self.extracted_content = ""
         # custom new info
         self.add_infos = add_infos
         # agent_state for Stop
@@ -261,9 +263,16 @@ async def step(self, step_info: Optional[CustomAgentStepInfo] = None) -> None:
             if len(actions) == 0:
                 # TODO: fix no action case
                 result = [ActionResult(is_done=True, extracted_content=step_info.memory, include_in_memory=True)]
+            for ret_ in result:
+                if "Extracted page" in ret_.extracted_content:
+                    # record every extracted page
+                    self.extracted_content += ret_.extracted_content
             self._last_result = result
             self._last_actions = actions
             if len(result) > 0 and result[-1].is_done:
+                if not self.extracted_content:
+                    self.extracted_content = step_info.memory
+                result[-1].extracted_content = self.extracted_content
                 logger.info(f"📄 Result: {result[-1].extracted_content}")
 
             self.consecutive_failures = 0
@@ -338,6 +347,10 @@ async def run(self, max_steps: int = 100) -> AgentHistoryList:
                     break
             else:
                 logger.info("❌ Failed to complete task in maximum steps")
+                if not self.extracted_content:
+                    self.history.history[-1].result[-1].extracted_content = step_info.memory
+                else:
+                    self.history.history[-1].result[-1].extracted_content = self.extracted_content
 
             return self.history
 

src/agent/custom_prompts.py

@@ -5,6 +5,7 @@
 from browser_use.agent.views import ActionResult, ActionModel
 from browser_use.browser.views import BrowserState
 from langchain_core.messages import HumanMessage, SystemMessage
+from datetime import datetime
 
 from .custom_views import CustomAgentStepInfo
 
@@ -116,15 +117,11 @@ def get_system_message(self) -> SystemMessage:
         Returns:
             str: Formatted system prompt
         """
-        time_str = self.current_date.strftime("%Y-%m-%d %H:%M")
-
         AGENT_PROMPT = f"""You are a precise browser automation agent that interacts with websites through structured commands. Your role is to:
     1. Analyze the provided webpage elements and structure
     2. Plan a sequence of actions to accomplish the given task
     3. Your final result MUST be a valid JSON as the **RESPONSE FORMAT** described, containing your action sequence and state assessment, No need extra content to expalin. 
 
-    Current date and time: {time_str}
-
     {self.input_format()}
 
     {self.important_rules()}
@@ -159,6 +156,9 @@ def get_user_message(self) -> HumanMessage:
             step_info_description = f'Current step: {self.step_info.step_number}/{self.step_info.max_steps}\n'
         else:
             step_info_description = ''
+            
+        time_str = datetime.now().strftime("%Y-%m-%d %H:%M")
+        step_info_description += "Current date and time: {time_str}"
 
         elements_text = self.state.element_tree.clickable_elements_to_string(include_attributes=self.include_attributes)
 

src/controller/custom_controller.py

@@ -1,15 +1,33 @@
+import pdb
+
 import pyperclip
 from typing import Optional, Type
 from pydantic import BaseModel
 from browser_use.agent.views import ActionResult
 from browser_use.browser.context import BrowserContext
 from browser_use.controller.service import Controller, DoneAction
+from main_content_extractor import MainContentExtractor
+from browser_use.controller.views import (
+    ClickElementAction,
+    DoneAction,
+    ExtractPageContentAction,
+    GoToUrlAction,
+    InputTextAction,
+    OpenTabAction,
+    ScrollAction,
+    SearchGoogleAction,
+    SendKeysAction,
+    SwitchTabAction,
+)
+import logging
+
+logger = logging.getLogger(__name__)
 
 
 class CustomController(Controller):
     def __init__(self, exclude_actions: list[str] = [],
-                output_model: Optional[Type[BaseModel]] = None
-                ):
+                 output_model: Optional[Type[BaseModel]] = None
+                 ):
         super().__init__(exclude_actions=exclude_actions, output_model=output_model)
         self._register_custom_actions()
 
@@ -29,3 +47,25 @@ async def paste_from_clipboard(browser: BrowserContext):
             await page.keyboard.type(text)
 
             return ActionResult(extracted_content=text)
+
+        @self.registry.action(
+            'Extract page content to get the pure text or markdown with links if include_links is set to true',
+            param_model=ExtractPageContentAction,
+            requires_browser=True,
+        )
+        async def extract_content(params: ExtractPageContentAction, browser: BrowserContext):
+            page = await browser.get_current_page()
+            # use jina reader
+            url = page.url
+            jina_url = f"https://r.jina.ai/{url}"
+            await page.goto(jina_url)
+            output_format = 'markdown' if params.include_links else 'text'
+            content = MainContentExtractor.extract(  # type: ignore
+                html=await page.content(),
+                output_format=output_format,
+            )
+            # go back to org url
+            await page.go_back()
+            msg = f'📄  Extracted page content as {output_format}\n: {content}\n'
+            logger.info(msg)
+            return ActionResult(extracted_content=msg)