camel-ai 0.2.60__py3-none-any.whl → 0.2.61__py3-none-any.whl

camel/toolkits/async_browser_toolkit.py

@@ -18,9 +18,7 @@ from __future__ import annotations
 import asyncio
 import datetime
 import io
-import json
 import os
-import random
 import re
 import shutil
 import urllib.parse
@@ -28,19 +26,17 @@ from copy import deepcopy
 from typing import (
     TYPE_CHECKING,
     Any,
-    BinaryIO,
     Coroutine,
     Dict,
     List,
     Literal,
     Optional,
     Tuple,
-    TypedDict,
     Union,
     cast,
 )
 
-from PIL import Image, ImageDraw, ImageFont
+from PIL import Image
 
 if TYPE_CHECKING:
     from camel.agents import ChatAgent
@@ -57,44 +53,25 @@ from camel.utils import (
     sanitize_filename,
 )
 
-logger = get_logger(__name__)
+from .browser_toolkit_commons import (
+    ACTION_WITH_FEEDBACK_LIST,
+    AVAILABLE_ACTIONS_PROMPT,
+    GET_FINAL_ANSWER_PROMPT_TEMPLATE,
+    OBSERVE_PROMPT_TEMPLATE,
+    PLANNING_AGENT_SYSTEM_PROMPT,
+    TASK_PLANNING_PROMPT_TEMPLATE,
+    TASK_REPLANNING_PROMPT_TEMPLATE,
+    WEB_AGENT_SYSTEM_PROMPT,
+    InteractiveRegion,
+    VisualViewport,
+    _parse_json_output,
+    _reload_image,
+    add_set_of_mark,
+    interactive_region_from_dict,
+    visual_viewport_from_dict,
+)
 
-TOP_NO_LABEL_ZONE = 20
-
-AVAILABLE_ACTIONS_PROMPT = """
-1. `fill_input_id(identifier: Union[str, int], text: str)`: Fill an input
-field (e.g. search box) with the given text and press Enter.
-2. `click_id(identifier: Union[str, int])`: Click an element with the given ID.
-3. `hover_id(identifier: Union[str, int])`: Hover over an element with the
-given ID.
-4. `download_file_id(identifier: Union[str, int])`: Download a file with the
-given ID. It returns the path to the downloaded file. If the file is
-successfully downloaded, you can stop the simulation and report the path to
-the downloaded file for further processing.
-5. `scroll_to_bottom()`: Scroll to the bottom of the page.
-6. `scroll_to_top()`: Scroll to the top of the page.
-7. `scroll_up()`: Scroll up the page. It is suitable when you want to see the
-elements above the current viewport.
-8. `scroll_down()`: Scroll down the page. It is suitable when you want to see
-the elements below the current viewport. If the webpage does not change, It
-means that the webpage has scrolled to the bottom.
-9. `back()`: Navigate back to the previous page. This is useful when you want
-to go back to the previous page, as current page is not useful.
-10. `stop()`: Stop the action process, because the task is completed or failed
-(impossible to find the answer). In this situation, you should provide your
-answer in your output.
-11. `get_url()`: Get the current URL of the current page.
-12. `find_text_on_page(search_text: str)`: Find the next given text on the
-current whole page, and scroll the page to the targeted text. It is equivalent
-to pressing Ctrl + F and searching for the text, and is powerful when you want
-to fast-check whether the current page contains some specific text.
-13. `visit_page(url: str)`: Go to the specific url page.
-14. `click_blank_area()`: Click a blank area of the page to unfocus the
-current element. It is useful when you have clicked an element but it cannot
-unfocus itself (e.g. Menu bar) to automatically render the updated webpage.
-15. `ask_question_about_video(question: str)`: Ask a question about the
-current webpage which contains video, e.g. youtube websites.
-"""
+logger = get_logger(__name__)
 
 ASYNC_ACTIONS = [
     "fill_input_id",
@@ -112,46 +89,6 @@ ASYNC_ACTIONS = [
     "click_blank_area",
 ]
 
-ACTION_WITH_FEEDBACK_LIST = [
-    'ask_question_about_video',
-    'download_file_id',
-    'find_text_on_page',
-]
-
-
-# Code from magentic-one
-class DOMRectangle(TypedDict):
-    x: Union[int, float]
-    y: Union[int, float]
-    width: Union[int, float]
-    height: Union[int, float]
-    top: Union[int, float]
-    right: Union[int, float]
-    bottom: Union[int, float]
-    left: Union[int, float]
-
-
-class VisualViewport(TypedDict):
-    height: Union[int, float]
-    width: Union[int, float]
-    offsetLeft: Union[int, float]
-    offsetTop: Union[int, float]
-    pageLeft: Union[int, float]
-    pageTop: Union[int, float]
-    scale: Union[int, float]
-    clientWidth: Union[int, float]
-    clientHeight: Union[int, float]
-    scrollWidth: Union[int, float]
-    scrollHeight: Union[int, float]
-
-
-class InteractiveRegion(TypedDict):
-    tag_name: str
-    role: str
-    aria_name: str
-    v_scrollable: bool
-    rects: List[DOMRectangle]
-
 
 def extract_function_name(s: str) -> str:
     r"""Extract the pure function name from a string (without parameters or
@@ -159,7 +96,7 @@ def extract_function_name(s: str) -> str:
 
     Args:
         s (str): Input string, e.g., `1.`**`click_id(14)`**, `scroll_up()`,
-        `'visit_page(url)'`, etc.
+        `\'visit_page(url)\'`, etc.
 
     Returns:
         str: Pure function name (e.g., `click_id`, `scroll_up`, `visit_page`)
@@ -179,303 +116,6 @@ def extract_function_name(s: str) -> str:
     return re.split(r'[ (\n]', s, maxsplit=1)[0]
 
 
-def _get_str(d: Any, k: str) -> str:
-    r"""Safely retrieve a string value from a dictionary."""
-    if k not in d:
-        raise KeyError(f"Missing required key: '{k}'")
-    val = d[k]
-    if isinstance(val, str):
-        return val
-    raise TypeError(
-        f"Expected a string for key '{k}', " f"but got {type(val).__name__}"
-    )
-
-
-def _get_number(d: Any, k: str) -> Union[int, float]:
-    r"""Safely retrieve a number (int or float) from a dictionary"""
-    val = d[k]
-    if isinstance(val, (int, float)):
-        return val
-    raise TypeError(
-        f"Expected a number (int/float) for key "
-        f"'{k}', but got {type(val).__name__}"
-    )
-
-
-def _get_bool(d: Any, k: str) -> bool:
-    r"""Safely retrieve a boolean value from a dictionary."""
-    val = d[k]
-    if isinstance(val, bool):
-        return val
-    raise TypeError(
-        f"Expected a boolean for key '{k}', " f"but got {type(val).__name__}"
-    )
-
-
-def _parse_json_output(text: str) -> Dict[str, Any]:
-    r"""Extract JSON output from a string."""
-
-    markdown_pattern = r'```(?:json)?\s*(.*?)\s*```'
-    markdown_match = re.search(markdown_pattern, text, re.DOTALL)
-    if markdown_match:
-        text = markdown_match.group(1).strip()
-
-    triple_quotes_pattern = r'"""(?:json)?\s*(.*?)\s*"""'
-    triple_quotes_match = re.search(triple_quotes_pattern, text, re.DOTALL)
-    if triple_quotes_match:
-        text = triple_quotes_match.group(1).strip()
-
-    try:
-        return json.loads(text)
-    except json.JSONDecodeError:
-        try:
-            fixed_text = re.sub(
-                r'`([^`]*?)`(?=\s*[:,\[\]{}]|$)', r'"\1"', text
-            )
-            return json.loads(fixed_text)
-        except json.JSONDecodeError:
-            result = {}
-            try:
-                bool_pattern = r'"(\w+)"\s*:\s*(true|false)'
-                for match in re.finditer(bool_pattern, text, re.IGNORECASE):
-                    key, value = match.groups()
-                    result[key] = value.lower() == "true"
-
-                str_pattern = r'"(\w+)"\s*:\s*"([^"]*)"'
-                for match in re.finditer(str_pattern, text):
-                    key, value = match.groups()
-                    result[key] = value
-
-                num_pattern = r'"(\w+)"\s*:\s*(-?\d+(?:\.\d+)?)'
-                for match in re.finditer(num_pattern, text):
-                    key, value = match.groups()
-                    try:
-                        result[key] = int(value)
-                    except ValueError:
-                        result[key] = float(value)
-
-                empty_str_pattern = r'"(\w+)"\s*:\s*""'
-                for match in re.finditer(empty_str_pattern, text):
-                    key = match.group(1)
-                    result[key] = ""
-
-                if result:
-                    return result
-
-                logger.warning(f"Failed to parse JSON output: {text}")
-                return {}
-            except Exception as e:
-                logger.warning(f"Error while extracting fields from JSON: {e}")
-                return {}
-
-
-def _reload_image(image: Image.Image):
-    buffer = io.BytesIO()
-    image.save(buffer, format="PNG")
-    buffer.seek(0)
-    return Image.open(buffer)
-
-
-def dom_rectangle_from_dict(rect: Dict[str, Any]) -> DOMRectangle:
-    r"""Create a DOMRectangle object from a dictionary."""
-    return DOMRectangle(
-        x=_get_number(rect, "x"),
-        y=_get_number(rect, "y"),
-        width=_get_number(rect, "width"),
-        height=_get_number(rect, "height"),
-        top=_get_number(rect, "top"),
-        right=_get_number(rect, "right"),
-        bottom=_get_number(rect, "bottom"),
-        left=_get_number(rect, "left"),
-    )
-
-
-def interactive_region_from_dict(region: Dict[str, Any]) -> InteractiveRegion:
-    r"""Create an :class:`InteractiveRegion` object from a dictionary."""
-    typed_rects: List[DOMRectangle] = []
-    for rect in region["rects"]:
-        typed_rects.append(dom_rectangle_from_dict(rect))
-
-    return InteractiveRegion(
-        tag_name=_get_str(region, "tag_name"),
-        role=_get_str(region, "role"),
-        aria_name=_get_str(region, "aria-name"),
-        v_scrollable=_get_bool(region, "v-scrollable"),
-        rects=typed_rects,
-    )
-
-
-def visual_viewport_from_dict(viewport: Dict[str, Any]) -> VisualViewport:
-    r"""Create a :class:`VisualViewport` object from a dictionary."""
-    return VisualViewport(
-        height=_get_number(viewport, "height"),
-        width=_get_number(viewport, "width"),
-        offsetLeft=_get_number(viewport, "offsetLeft"),
-        offsetTop=_get_number(viewport, "offsetTop"),
-        pageLeft=_get_number(viewport, "pageLeft"),
-        pageTop=_get_number(viewport, "pageTop"),
-        scale=_get_number(viewport, "scale"),
-        clientWidth=_get_number(viewport, "clientWidth"),
-        clientHeight=_get_number(viewport, "clientHeight"),
-        scrollWidth=_get_number(viewport, "scrollWidth"),
-        scrollHeight=_get_number(viewport, "scrollHeight"),
-    )
-
-
-def add_set_of_mark(
-    screenshot: Union[bytes, Image.Image, io.BufferedIOBase],
-    ROIs: Dict[str, InteractiveRegion],
-) -> Tuple[Image.Image, List[str], List[str], List[str]]:
-    if isinstance(screenshot, Image.Image):
-        return _add_set_of_mark(screenshot, ROIs)
-
-    if isinstance(screenshot, bytes):
-        screenshot = io.BytesIO(screenshot)
-
-    image = Image.open(cast(BinaryIO, screenshot))
-    comp, visible_rects, rects_above, rects_below = _add_set_of_mark(
-        image, ROIs
-    )
-    image.close()
-    return comp, visible_rects, rects_above, rects_below
-
-
-def _add_set_of_mark(
-    screenshot: Image.Image, ROIs: Dict[str, InteractiveRegion]
-) -> Tuple[Image.Image, List[str], List[str], List[str]]:
-    r"""Add a set of marks to the screenshot.
-
-    Args:
-        screenshot (Image.Image): The screenshot to add marks to.
-        ROIs (Dict[str, InteractiveRegion]): The regions to add marks to.
-
-    Returns:
-        Tuple[Image.Image, List[str], List[str], List[str]]: A tuple
-            containing the screenshot with marked ROIs, ROIs fully within the
-            images, ROIs located above the visible area, and ROIs located below
-            the visible area.
-    """
-    visible_rects: List[str] = list()
-    rects_above: List[str] = list()  # Scroll up to see
-    rects_below: List[str] = list()  # Scroll down to see
-
-    fnt = ImageFont.load_default(14)
-    base = screenshot.convert("L").convert("RGBA")
-    overlay = Image.new("RGBA", base.size)
-
-    draw = ImageDraw.Draw(overlay)
-    for r in ROIs:
-        for rect in ROIs[r]["rects"]:
-            # Empty rectangles
-            if not rect or rect["width"] == 0 or rect["height"] == 0:
-                continue
-
-            # TODO: add scroll left and right?
-            horizontal_center = (rect["right"] + rect["left"]) / 2.0
-            vertical_center = (rect["top"] + rect["bottom"]) / 2.0
-            is_within_horizon = 0 <= horizontal_center < base.size[0]
-            is_above_viewport = vertical_center < 0
-            is_below_viewport = vertical_center >= base.size[1]
-
-            if is_within_horizon:
-                if is_above_viewport:
-                    rects_above.append(r)
-                elif is_below_viewport:
-                    rects_below.append(r)
-                else:  # Fully visible
-                    visible_rects.append(r)
-                    _draw_roi(draw, int(r), fnt, rect)
-
-    comp = Image.alpha_composite(base, overlay)
-    overlay.close()
-    return comp, visible_rects, rects_above, rects_below
-
-
-def _draw_roi(
-    draw: ImageDraw.ImageDraw,
-    idx: int,
-    font: ImageFont.FreeTypeFont | ImageFont.ImageFont,
-    rect: DOMRectangle,
-) -> None:
-    r"""Draw a ROI on the image.
-
-    Args:
-        draw (ImageDraw.ImageDraw): The draw object.
-        idx (int): The index of the ROI.
-        font (ImageFont.FreeTypeFont | ImageFont.ImageFont): The font.
-        rect (DOMRectangle): The DOM rectangle.
-    """
-    color = _get_random_color(idx)
-    text_color = _get_text_color(color)
-
-    roi = ((rect["left"], rect["top"]), (rect["right"], rect["bottom"]))
-
-    label_location = (rect["right"], rect["top"])
-    label_anchor = "rb"
-
-    if label_location[1] <= TOP_NO_LABEL_ZONE:
-        label_location = (rect["right"], rect["bottom"])
-        label_anchor = "rt"
-
-    draw.rectangle(
-        roi, outline=color, fill=(color[0], color[1], color[2], 48), width=2
-    )
-
-    bbox = draw.textbbox(
-        label_location,
-        str(idx),
-        font=font,
-        anchor=label_anchor,
-        align="center",
-    )
-    bbox = (bbox[0] - 3, bbox[1] - 3, bbox[2] + 3, bbox[3] + 3)
-    draw.rectangle(bbox, fill=color)
-
-    draw.text(
-        label_location,
-        str(idx),
-        fill=text_color,
-        font=font,
-        anchor=label_anchor,
-        align="center",
-    )
-
-
-def _get_text_color(
-    bg_color: Tuple[int, int, int, int],
-) -> Tuple[int, int, int, int]:
-    r"""Determine the ideal text color (black or white) for contrast.
-
-    Args:
-        bg_color: The background color (R, G, B, A).
-
-    Returns:
-        A tuple representing black or white color for text.
-    """
-    luminance = bg_color[0] * 0.3 + bg_color[1] * 0.59 + bg_color[2] * 0.11
-    return (0, 0, 0, 255) if luminance > 120 else (255, 255, 255, 255)
-
-
-def _get_random_color(identifier: int) -> Tuple[int, int, int, int]:
-    r"""Generate a consistent random RGBA color based on the identifier.
-
-    Args:
-        identifier: The ID used as a seed to ensure color consistency.
-
-    Returns:
-        A tuple representing (R, G, B, A) values.
-    """
-    rnd = random.Random(int(identifier))
-    r = rnd.randint(0, 255)
-    g = rnd.randint(125, 255)
-    b = rnd.randint(0, 50)
-    color = [r, g, b]
-    # TODO: check why shuffle is needed?
-    rnd.shuffle(color)
-    color.append(255)
-    return cast(Tuple[int, int, int, int], tuple(color))
-
-
 class AsyncBaseBrowser:
     def __init__(
         self,
@@ -505,12 +145,19 @@ class AsyncBaseBrowser:
             async_playwright,
         )
 
-        self.history: list = []
+        self.history: list[Any] = []
         self.headless = headless
         self.channel = channel
         self.playwright = async_playwright()
-        self.page_history: list = []
+        self.page_history: list[Any] = []
         self.cookie_json_path = cookie_json_path
+        self.playwright_server: Any = None
+        self.playwright_started: bool = False
+        self.browser: Any = None
+        self.context: Any = None
+        self.page: Any = None
+        self.page_url: str = ""
+        self.web_agent_model: Optional[BaseModelBackend] = None
 
         # Set the cache directory
         self.cache_dir = "tmp/" if cache_dir is None else cache_dir
@@ -631,7 +278,10 @@ class AsyncBaseBrowser:
             return "User cancelled the video analysis."
 
         model = None
-        if hasattr(self, 'web_agent_model'):
+        if (
+            hasattr(self, 'web_agent_model')
+            and self.web_agent_model is not None
+        ):
             model = self.web_agent_model
 
         video_analyzer = VideoAnalysisToolkit(model=model)
@@ -723,7 +373,8 @@ class AsyncBaseBrowser:
             )
 
             _, file_path = await self.get_screenshot(save_image=True)
-            screenshots.append(file_path)
+            if file_path is not None:
+                screenshots.append(file_path)
 
             await self.page.evaluate(f"window.scrollBy(0, {scroll_step})")
             # Allow time for content to load
@@ -800,7 +451,7 @@ class AsyncBaseBrowser:
         for k in result:
             typed_results[k] = interactive_region_from_dict(result[k])
 
-        return typed_results  # type: ignore[return-value]
+        return typed_results
 
     def get_interactive_elements(
         self,
@@ -830,13 +481,13 @@ class AsyncBaseBrowser:
         """
 
         await self.wait_for_load()
-        screenshot, _ = await self.get_screenshot(save_image=False)
-        rects = await self.get_interactive_elements()
+        screenshot, _ = await self.async_get_screenshot(save_image=False)
+        rects = await self.async_get_interactive_elements()
 
         file_path: str | None = None
         comp, _, _, _ = add_set_of_mark(
             screenshot,
-            rects,  # type: ignore[arg-type]
+            rects,
         )
         if save_image:
             parsed_url = urllib.parse.urlparse(self.page_url)
@@ -902,7 +553,7 @@ class AsyncBaseBrowser:
 
         try:
             await target.wait_for(timeout=5000)
-        except (TimeoutError, Exception) as e:
+        except (TimeoutError, Exception) as e:  # type: ignore[misc]
             logger.debug(f"Error during click operation: {e}")
             raise ValueError("No such element.") from None
 
@@ -926,7 +577,7 @@ class AsyncBaseBrowser:
                 self.page_history.append(deepcopy(self.page.url))
                 self.page = new_page
 
-        except (TimeoutError, Exception) as e:
+        except (TimeoutError, Exception) as e:  # type: ignore[misc]
             logger.debug(f"Error during click operation: {e}")
             pass
 
@@ -962,7 +613,7 @@ class AsyncBaseBrowser:
             identifier = str(identifier)
         try:
             target = self.page.locator(f"[__elementId='{identifier}']")
-        except (TimeoutError, Exception) as e:
+        except (TimeoutError, Exception) as e:  # type: ignore[misc]
             logger.debug(f"Error during download operation: {e}")
             logger.warning(
                 f"Element with identifier '{identifier}' not found."
@@ -1015,7 +666,7 @@ class AsyncBaseBrowser:
 
         try:
             target = self.page.locator(f"[__elementId='{identifier}']")
-        except (TimeoutError, Exception) as e:
+        except (TimeoutError, Exception) as e:  # type: ignore[misc]
             logger.debug(f"Error during fill operation: {e}")
             logger.warning(
                 f"Element with identifier '{identifier}' not found."
@@ -1079,7 +730,7 @@ class AsyncBaseBrowser:
             identifier = str(identifier)
         try:
             target = self.page.locator(f"[__elementId='{identifier}']")
-        except (TimeoutError, Exception) as e:
+        except (TimeoutError, Exception) as e:  # type: ignore[misc]
             logger.debug(f"Error during hover operation: {e}")
             logger.warning(
                 f"Element with identifier '{identifier}' not found."
@@ -1316,7 +967,7 @@ class AsyncBrowserToolkit(BaseToolkit):
                 login.
                 (default: :obj:`None`)
         """
-
+        super().__init__()
         self.browser = AsyncBaseBrowser(
             headless=headless,
             cache_dir=cache_dir,
@@ -1328,8 +979,9 @@ class AsyncBrowserToolkit(BaseToolkit):
         self.web_agent_model = web_agent_model
         self.planning_agent_model = planning_agent_model
         self.output_language = output_language
+        self.browser.web_agent_model = web_agent_model
 
-        self.history: list = []
+        self.history: list[Any] = []
         self.web_agent, self.planning_agent = self._initialize_agent()
 
     def _reset(self):
@@ -1359,11 +1011,7 @@ class AsyncBrowserToolkit(BaseToolkit):
         else:
             planning_model = self.planning_agent_model
 
-        system_prompt = """
-You are a helpful web agent that can assist users in browsing the web.
-Given a high-level task, you can leverage predefined browser tools to help
-users achieve their goals.
-        """
+        system_prompt = WEB_AGENT_SYSTEM_PROMPT
 
         web_agent = ChatAgent(
             system_message=system_prompt,
@@ -1371,10 +1019,7 @@ users achieve their goals.
             output_language=self.output_language,
         )
 
-        planning_system_prompt = """
-You are a helpful planning agent that can assist users in planning complex
-tasks which need multi-step browser interaction.
-        """
+        planning_system_prompt = PLANNING_AGENT_SYSTEM_PROMPT
 
         planning_agent = ChatAgent(
             system_message=planning_system_prompt,
@@ -1390,97 +1035,23 @@ tasks which need multi-step browser interaction.
         r"""Let agent observe the current environment, and get the next
         action."""
 
-        detailed_plan_prompt = ""
+        detailed_plan_prompt_str = ""
 
         if detailed_plan is not None:
-            detailed_plan_prompt = f"""
+            detailed_plan_prompt_str = f"""
 Here is a plan about how to solve the task step-by-step which you must follow:
-<detailed_plan>{detailed_plan}<detailed_plan>
+<detailed_plan>{detailed_plan}</detailed_plan>
         """
 
-        observe_prompt = f"""
-Please act as a web agent to help me complete the following high-level task:
-<task>{task_prompt}</task>
-Now, I have made screenshot (only the current viewport, not the full webpage)
-based on the current browser state, and marked interactive elements in the
-webpage.
-Please carefully examine the requirements of the task, and current state of
-the browser, and provide the next appropriate action to take.
-
-{detailed_plan_prompt}
-
-Here are the current available browser functions you can use:
-{AVAILABLE_ACTIONS_PROMPT}
-
-Here are the latest {self.history_window} trajectory (at most) you have taken:
-<history>
-{self.history[-self.history_window :]}
-</history>
-
-Your output should be in json format, including the following fields:
-- `observation`: The detailed image description about the current viewport. Do
-not over-confident about the correctness of the history actions. You should
-always check the current viewport to make sure the correctness of the next
-action.
-- `reasoning`: The reasoning about the next action you want to take, and the
-possible obstacles you may encounter, and how to solve them. Do not forget to
-check the history actions to avoid the same mistakes.
-- `action_code`: The action code you want to take. It is only one step action
-code, without any other texts (such as annotation)
-
-Here is two example of the output:
-```json
-{{
-    "observation": [IMAGE_DESCRIPTION],
-    "reasoning": [YOUR_REASONING],
-    "action_code": "fill_input_id([ID], [TEXT])"
-}}
-
-{{
-    "observation":  "The current page is a CAPTCHA verification page on Amazon. It asks the user to ..",
-    "reasoning": "To proceed with the task of searching for products, I need to complete..",
-    "action_code": "fill_input_id(3, 'AUXPMR')"
-}}
-
-Here are some tips for you:
-- Never forget the overall question: **{task_prompt}**
-- Maybe after a certain operation (e.g. click_id), the page content has not
-changed. You can check whether the action step is successful by looking at the
-`success` of the action step in the history. If successful, it means that the
-page content is indeed the same after the click. You need to try other methods.
-- If using one way to solve the problem is not successful, try other ways.
-Make sure your provided ID is correct!
-- Some cases are very complex and need to be achieve by an iterative process.
-You can use the `back()` function to go back to the previous page to try other
-methods.
-- There are many links on the page, which may be useful for solving the
-problem. You can use the `click_id()` function to click on the link to see if
-it is useful.
-- Always keep in mind that your action must be based on the ID shown in the
-current image or viewport, not the ID shown in the history.
-- Do not use `stop()` lightly. Always remind yourself that the image only
-shows a part of the full page. If you cannot find the answer, try to use
-functions like `scroll_up()` and `scroll_down()` to check the full content of
-the webpage before doing anything else, because the answer or next key step
-may be hidden in the content below.
-- If the webpage needs human verification, you must avoid processing it.
-Please use `back()` to go back to the previous page, and try other ways.
-- If you have tried everything and still cannot resolve the issue, please stop
-the simulation, and report issues you have encountered.
-- Check the history actions carefully, detect whether you have repeatedly made
-the same actions or not.
-- When dealing with wikipedia revision history related tasks, you need to
-think about the solution flexibly. First, adjust the browsing history
-displayed on a single page to the maximum, and then make use of the
-find_text_on_page function. This is extremely useful which can quickly locate
-the text you want to find and skip massive amount of useless information.
-- Flexibly use interactive elements like slide down selection bar to filter
-out the information you need. Sometimes they are extremely useful.
-```
-        """  # noqa: E501
-
+        observe_prompt = OBSERVE_PROMPT_TEMPLATE.format(
+            task_prompt=task_prompt,
+            detailed_plan_prompt=detailed_plan_prompt_str,
+            AVAILABLE_ACTIONS_PROMPT=AVAILABLE_ACTIONS_PROMPT,
+            history_window=self.history_window,
+            history=self.history[-self.history_window :],
+        )
         # get current state
-        som_screenshot, _ = await self.browser.get_som_screenshot(
+        som_screenshot, _ = await self.browser.async_get_som_screenshot(
             save_image=True
         )
         img = _reload_image(som_screenshot)
@@ -1493,7 +1064,7 @@ out the information you need. Sometimes they are extremely useful.
 
         resp_content = resp.msgs[0].content
 
-        resp_dict = _parse_json_output(resp_content)
+        resp_dict = _parse_json_output(resp_content, logger)
         observation_result: str = resp_dict.get("observation", "")
         reasoning_result: str = resp_dict.get("reasoning", "")
         action_code: str = resp_dict.get("action_code", "")
@@ -1632,12 +1203,9 @@ out the information you need. Sometimes they are extremely useful.
         current viewport.
         """
 
-        prompt = f"""
-We are solving a complex web task which needs multi-step browser interaction. After the multi-step observation, reasoning and acting with web browser, we think that the task is currently solved.
-Here are all trajectory we have taken:
-<history>{self.history}</history>
-Please find the final answer, or give valuable insights and founds (e.g. if previous actions contain downloading files, your output should include the path of the downloaded file) about the overall task: <task>{task_prompt}</task>
-        """  # noqa: E501
+        prompt = GET_FINAL_ANSWER_PROMPT_TEMPLATE.format(
+            history=self.history, task_prompt=task_prompt
+        )
 
         message = BaseMessage.make_user_message(
             role_name='user',
@@ -1653,13 +1221,9 @@ Please find the final answer, or give valuable insights and founds (e.g. if prev
         # Here are the available browser functions we can
         # use: {AVAILABLE_ACTIONS_PROMPT}
 
-        planning_prompt = f"""
-<task>{task_prompt}</task>
-According to the problem above, if we use browser interaction, what is the general process of the interaction after visiting the webpage `{start_url}`? 
-
-Please note that it can be viewed as Partially Observable MDP. Do not over-confident about your plan.
-Please first restate the task in detail, and then provide a detailed plan to solve the task.
-"""  # noqa: E501
+        planning_prompt = TASK_PLANNING_PROMPT_TEMPLATE.format(
+            task_prompt=task_prompt, start_url=start_url
+        )
 
         message = BaseMessage.make_user_message(
             role_name='user', content=planning_prompt
@@ -1685,28 +1249,16 @@ Please first restate the task in detail, and then provide a detailed plan to sol
 
         # Here are the available browser functions we can
         # use: {AVAILABLE_ACTIONS_PROMPT}
-        replanning_prompt = f"""
-We are using browser interaction to solve a complex task which needs multi-step actions.
-Here are the overall task:
-<overall_task>{task_prompt}</overall_task>
-
-In order to solve the task, we made a detailed plan previously. Here is the detailed plan:
-<detailed plan>{detailed_plan}</detailed plan>
-
-According to the task above, we have made a series of observations, reasonings, and actions. Here are the latest {self.history_window} trajectory (at most) we have taken:
-<history>{self.history[-self.history_window :]}</history>
-
-However, the task is not completed yet. As the task is partially observable, we may need to replan the task based on the current state of the browser if necessary.
-Now please carefully examine the current task planning schema, and our history actions, and then judge whether the task needs to be fundamentally replanned. If so, please provide a detailed replanned schema (including the restated overall task).
-
-Your output should be in json format, including the following fields:
-- `if_need_replan`: bool, A boolean value indicating whether the task needs to be fundamentally replanned.
-- `replanned_schema`: str, The replanned schema for the task, which should not be changed too much compared with the original one. If the task does not need to be replanned, the value should be an empty string. 
-"""  # noqa: E501
+        replanning_prompt = TASK_REPLANNING_PROMPT_TEMPLATE.format(
+            task_prompt=task_prompt,
+            detailed_plan=detailed_plan,
+            history_window=self.history_window,
+            history=self.history[-self.history_window :],
+        )
         # Reset the history message of planning_agent.
         self.planning_agent.reset()
         resp = self.planning_agent.step(replanning_prompt)
-        resp_dict = _parse_json_output(resp.msgs[0].content)
+        resp_dict = _parse_json_output(resp.msgs[0].content, logger)
 
         if_need_replan = resp_dict.get("if_need_replan", False)
         replanned_schema = resp_dict.get("replanned_schema", "")
@@ -1791,7 +1343,7 @@ Your output should be in json format, including the following fields:
                 The task is not completed within the round limit. Please check 
                 the last round {self.history_window} information to see if 
                 there is any useful information:
-                <history>{self.history[-self.history_window:]}</history>
+                <history>{self.history[-self.history_window :]}</history>
             """
 
         else: