Fara1.5 – A family of frontier computer use agent models
clean up 1
Hussein Mozannareda8e982d853
diff --git a/src/fara/__init__.py b/src/fara/__init__.pyindex b4cbdb2..44e7a3c 100644--- a/src/fara/__init__.py+++ b/src/fara/__init__.py@@ -1,8 +1,4 @@-from .agents.fara.fara_qwen3_next import (- FaraQwen3NextAgent,- FaraQwen3NextAgentConfig,-)-from .agents.fara.fara_qwen3 import FaraQwen3Agent, FaraQwen3AgentConfig+from .agents.fara.fara15_agent import Fara15Agent, Fara15AgentConfigfrom .environments.playwright import (PlaywrightEnvironment,PlaywrightEnvironmentConfig,@@ -13,10 +9,8 @@ from .fara_7b import FARA_ACTION_DEFINITIONS, FaraAgentfrom .browser.playwright_controller import PlaywrightController__all__ = [- "FaraQwen3NextAgent",- "FaraQwen3NextAgentConfig",- "FaraQwen3Agent",- "FaraQwen3AgentConfig",+ "Fara15Agent",+ "Fara15AgentConfig","PlaywrightEnvironment","PlaywrightEnvironmentConfig","DataPoint",diff --git a/src/fara/agents/fara/fara_qwen3.py b/src/fara/agents/fara/fara15_agent.pysimilarity index 69%rename from src/fara/agents/fara/fara_qwen3.pyrename to src/fara/agents/fara/fara15_agent.pyindex 10fd5a5..0eb88fa 100644--- a/src/fara/agents/fara/fara_qwen3.py+++ b/src/fara/agents/fara/fara15_agent.py@@ -1,19 +1,25 @@-"""FaraQwen3Agent - Web automation agent using Qwen3 vision-language model."""+"""Fara-1.5 web automation agent (FaraNext browser action set).++Single-class port of agento_next's FaraQwen3NextAgent: the browser+observe-think-act loop, the expanded action space (double/right/triple+click, left_click_drag, hscroll, read_page_answer_question,+ask_user_question), and DataPoint trajectory logging via RunContext.+"""from __future__ import annotations+import astimport copyimport ioimport json-import astimport logging+import reimport shutil--import openaifrom dataclasses import dataclass, fieldfrom typing import Any, Dict, List, Tuplefrom urllib.parse import quote_plus+import openaifrom PIL import Imagefrom tenacity import (retry,@@ -23,6 +29,7 @@ from tenacity import (before_sleep_log,)+from ..._prompts import get_computer_use_system_promptfrom ...clients.wrapper import ChatCompletionClientfrom ...clients.create_utils import create_client_from_configfrom ...clients.messages import (@@ -33,7 +40,6 @@ from ...clients.messages import (ImageObj,FunctionCall,)-from ...core.agent import Agent, AgentConfigfrom ...core.run_context import RunContextfrom ...core.data_point import (@@ -47,48 +53,47 @@ from ...core.data_point import ()from ..captcha import wait_for_captchafrom ..coord_spaces import FARA_DISPLAY_SIZE+from ..computer_agent.utils import extract_from_page+from ..utils import format_text_observationfrom .fara_types import WebSurferEvent-from .prompts import get_computer_use_system_promptfrom .utils import get_trimmed_url-class FaraQwen3AgentConfig(AgentConfig):- """Configuration for FaraAgent.+def extract_allowed_actions(system_prompt_text: str) -> frozenset[str]:+ """Extract the allowed action names from the tool schema enum in the system prompt."""+ match = re.search(r'"enum":\s*\[([^\]]+)\]', system_prompt_text)+ if not match:+ raise ValueError("Could not find action enum in system prompt")+ return frozenset(json.loads(f"[{match.group(1)}]"))+++class Fara15AgentConfig(AgentConfig):+ """Configuration for Fara15Agent.Notable fields:extra_create_args: Extra kwargs merged on top of ``temperature: 0`` and- forwarded to ``chat.completions.create``. vLLM-only keys are routed- via ``extra_body``.- auto_user_reply: When True, ``ask_user_question`` does NOT halt- the run — a fixed dummy user response is injected and the agent- continues. For eval / no-user-simulator settings only. Default False- preserves ``task_workflow.py``'s ``WAITING_FOR_USER`` semantics for- data_gen.- captcha_timeout_limit: Number of consecutive captcha-wait timeouts- tolerated within a single run before the per-step captcha-wait- gate is auto-disabled for the rest of the run. Set to ``0`` to- skip the gate entirely from the start (appropriate without a- Browserbase cloud session, which is what triggers the captcha- solver). Defaults to ``2``.- raise_on_captcha_timeout: When True, restore the legacy behavior- of raising ``RuntimeError`` on the first captcha-wait timeout.- When False (default), the agent logs a warning and moves past- the timed-out captcha, falling back to ``captcha_timeout_limit``- to auto-disable the gate after repeated timeouts.+ forwarded to ``chat.completions.create``.+ auto_user_reply: When True, ``ask_user_question`` does NOT halt the run —+ a fixed dummy user response is injected and the agent continues (for+ eval / no-user-simulator settings only).+ captcha_timeout_limit: Consecutive captcha-wait timeouts tolerated in a+ single run before the per-step captcha gate is auto-disabled. ``0``+ skips the gate from the start (appropriate without a Browserbase+ session, which is what triggers the captcha solver).+ raise_on_captcha_timeout: When True, raise on the first captcha-wait+ timeout instead of moving past the unsolved captcha."""- name: str = "fara"+ name: str = "fara_next"client_config: dict[str, Any] | None = None- start_page: str = "about:blank"max_rounds: int = 10max_n_images: int = 3+ max_observation_chars: int = 1000fn_call_template: str = "fara-qwen3vl"- identity: str | None = "fara_qwen3vl"+ identity: str | None = "fara_qwen35"critical_points: str | None = "fara-1.5"- computer_use_mode: str = "aurora"+ computer_use_mode: str = "fara_next_browser"save_screenshots: bool = False- animate_actions: bool = False- single_tab_mode: bool = Trueinclude_input_text_key_args: bool = Trueviewport_width: int = 1440viewport_height: int = 900@@ -104,8 +109,8 @@ class FaraQwen3AgentConfig(AgentConfig):@dataclass-class FaraQwen3AgentState:- """Mutable state for FaraQwen3Agent that persists across resume calls."""+class Fara15AgentState:+ """Mutable state that persists across resume calls."""chat_history: list[LLMMessage] = field(default_factory=list)facts: list[str] = field(default_factory=list)@@ -115,31 +120,33 @@ class FaraQwen3AgentState:mlm_height: int = 0-class FaraQwen3Agent(Agent):- """Web automation agent using vision-language models."""+class Fara15Agent(Agent):+ """Web automation agent with the FaraNext browser action set."""- DEFAULT_START_PAGE = "https://www.bing.com/"DISPLAY_SIZE = FARA_DISPLAY_SIZEPATCH_SIZE = 16MERGE_SIZE = 2USER_MESSAGE = "Here is the next screenshot. Think about what to do next."MAX_URL_LENGTH = 100+ _TEXT_OBSERVATION_ACTIONS: frozenset[str] = frozenset({"read_page_answer_question"})++ _MODE_CANONICAL_IDENTITY: dict[str, tuple[str, ...]] = {+ "fara_next_browser": ("fara_qwen3vl", "fara_qwen35"),+ "fara_next_windows": ("fara_qwen3vl_windows", "fara_qwen35_windows"),+ "fara_next_windows_core": ("fara_qwen3vl_windows", "fara_qwen35_windows"),+ }+def __init__(- self, config: FaraQwen3AgentConfig | dict[str, Any] | None = None, **kwargs: Any+ self, config: Fara15AgentConfig | dict[str, Any] | None = None, **kwargs: Any):super().__init__(config, **kwargs)- self.config: FaraQwen3AgentConfig- if type(self) is FaraQwen3Agent:- assert (- self.config.computer_use_mode == "aurora"- ), f"FaraQwen3Agent is aurora-v2 era; got computer_use_mode={self.config.computer_use_mode!r}"+ self.config: Fara15AgentConfigself.logger = logging.getLogger(__name__)self._client: ChatCompletionClient | None = None- self._state: FaraQwen3AgentState | None = None+ self._state: Fara15AgentState | None = Noneself._pending_observation: str = ""- self._os_type: str = "linux"- self._output_dir = None+ self._allowed_actions: frozenset[str] = frozenset()self._captcha_timeouts: int = 0self._captcha_disabled: bool = self.config.captcha_timeout_limit <= 0@@ -153,13 +160,13 @@ class FaraQwen3Agent(Agent):}@classmethod- def _get_config_class(cls) -> type[FaraQwen3AgentConfig]:- return FaraQwen3AgentConfig+ def _get_config_class(cls) -> type[Fara15AgentConfig]:+ return Fara15AgentConfigasync def initialize(self, run_context: RunContext) -> None:- """Initialize the agent."""+ """Initialize the agent for a run."""await super().initialize(run_context)- self._state = FaraQwen3AgentState(+ self._state = Fara15AgentState(mlm_width=self.config.viewport_width,mlm_height=self.config.viewport_height,)@@ -171,19 +178,39 @@ class FaraQwen3Agent(Agent):else:raise ValueError("Either client or client_config must be provided")+ valid_modes = set(self._MODE_CANONICAL_IDENTITY)+ if self.config.computer_use_mode not in valid_modes:+ raise ValueError(+ f"Unknown computer_use_mode '{self.config.computer_use_mode}'. "+ f"Known values: {sorted(valid_modes)}"+ )+ canonical_identities = self._MODE_CANONICAL_IDENTITY[+ self.config.computer_use_mode+ ]+ if (+ self.config.identity is not None+ and self.config.identity not in canonical_identities+ ):+ raise ValueError(+ f"identity={self.config.identity!r} does not match any canonical "+ f"identity in {list(canonical_identities)!r} for "+ f"computer_use_mode={self.config.computer_use_mode!r}. "+ f"Use one of {list(canonical_identities)!r} or set identity=None "+ f"to auto-detect."+ )+async def run(self, run_context: RunContext, input: Any = None) -> Tuple[str, List, List]:"""Run the agent on a task.- Supports resuming after WAITING_FOR_USER: if ``_state.chat_history``- is already populated the fresh-start init is skipped and the user- response is appended to the conversation.+ Supports resuming after WAITING_FOR_USER: if ``_state.chat_history`` is+ already populated the fresh-start init is skipped and the user response+ is appended to the conversation."""env = run_context.environment+traj = run_context.solver_log- self._os_type = getattr(env, "os_type", "linux")- self._output_dir = run_context.output_dirpending_user_response = ""if self._state.chat_history:@@ -337,10 +364,9 @@ class FaraQwen3Agent(Agent):await super().close(run_context)def _get_final_answer(self, thoughts: str, action_description: str) -> str:- return thoughts+ return action_descriptiondef _get_observation_prefix(self) -> str:- """Return text to prepend to the next user message. Override in subclasses."""obs = self._pending_observationself._pending_observation = ""return obs@@ -401,11 +427,15 @@ class FaraQwen3Agent(Agent):)system_message = []+ full_text = ""for msg in system_prompt_info["conversation"]:tmp_content = ""for content in msg["content"]:tmp_content += content["text"]system_message.append(SystemMessage(content=tmp_content))+ full_text += tmp_content++ self._allowed_actions = extract_allowed_actions(full_text)return system_message, scaled_screenshot@@ -464,6 +494,16 @@ class FaraQwen3Agent(Agent):[tgt_x, tgt_y], im_w, im_h, og_im_w, og_im_h)+ def _proc(self, coordinate):+ """Shorthand for proc_coords with standard display/viewport args."""+ return self.proc_coords(+ coordinate,+ self.DISPLAY_SIZE,+ self.DISPLAY_SIZE,+ self.config.viewport_width,+ self.config.viewport_height,+ )+@retry(retry=retry_if_not_exception_type(openai.BadRequestError),stop=stop_after_attempt(5),@@ -479,9 +519,8 @@ class FaraQwen3Agent(Agent):"""Make a model call using the client."""if self._client is None:raise ValueError(- "No client configured for FaraAgent. Call initialize() first."+ "No client configured for Fara15Agent. Call initialize() first.")-result = await self._client.create(messages=history,extra_create_args=extra_create_args or {},@@ -491,8 +530,7 @@ class FaraQwen3Agent(Agent):def remove_screenshot_from_message(self, msg: LLMMessage) -> LLMMessage | None:"""Remove the screenshot from the message content."""if isinstance(msg.content, list):- new_content = [c for c in msg.content if not isinstance(c, ImageObj)]- msg.content = new_content+ msg.content = [c for c in msg.content if not isinstance(c, ImageObj)]return msgelif isinstance(msg.content, ImageObj):return None@@ -501,7 +539,8 @@ class FaraQwen3Agent(Agent):def maybe_remove_old_screenshots(self, history: List[LLMMessage], includes_current: bool = False) -> List[LLMMessage]:- """Remove old screenshots from the chat history."""+ """Remove old screenshots from the chat history, keeping the most recent+ ``max_n_images``. Original task and user-response messages keep their text."""if self.config.max_n_images <= 0:return history@@ -569,8 +608,7 @@ class FaraQwen3Agent(Agent):def _fit_images_to_budget(self, history: List[LLMMessage]) -> List[LLMMessage]:"""Drop oldest screenshots (keeping the most recent) until the estimated- prompt is within ``image_budget_token_cap``. Returns a new list; the- original messages in ``chat_history`` are left unchanged."""+ prompt is within ``image_budget_token_cap``. Returns a new list."""cap = self.config.image_budget_token_capif cap <= 0 or self._estimate_prompt_tokens(history) <= cap:return history@@ -642,15 +680,29 @@ class FaraQwen3Agent(Agent):async def _execute_action(self, env, function_call: list[FunctionCall]) -> tuple[bool, str]:- """Execute an action on the environment."""+ """Execute an action. Only actions in the current schema are allowed."""name = function_call[0].nameargs = function_call[0].arguments- action_description = ""+ action_type = args.get("action", "")+ if action_type not in self._allowed_actions:+ if self.config.terminate_on_parse_error:+ self.logger.warning(+ "terminate_on_parse_error=true: invalid action %r — "+ "ending trajectory with thoughts as final answer",+ action_type,+ )+ return True, args.get("thoughts", "") or args.get("answer", "")+ raise ValueError(+ f"Action '{action_type}' is not allowed in mode "+ f"'{self.config.computer_use_mode}'. Allowed: {sorted(self._allowed_actions)}"+ )++ url = (await env.get_page_context()).urlself.logger.debug(WebSurferEvent(- source="FaraAgent",- url=(await env.get_page_context()).url,+ source="Fara15Agent",+ url=url,action=name,arguments=args,message=f"{name}( {json.dumps(args)} )",@@ -658,106 +710,135 @@ class FaraQwen3Agent(Agent):)if "coordinate" in args:- args["coordinate"] = self.proc_coords(- args["coordinate"],- self.DISPLAY_SIZE,- self.DISPLAY_SIZE,- self.config.viewport_width,- self.config.viewport_height,+ args["coordinate"] = self._proc(args["coordinate"])++ is_stop_action, action_description = await self._dispatch_action(+ env, action_type, args+ )++ if action_type in self._TEXT_OBSERVATION_ACTIONS:+ self._pending_observation = format_text_observation(+ action_type,+ action_description,+ self.config.max_observation_chars,)- is_stop_action = False- action_type = args["action"]+ if not is_stop_action:+ if hasattr(env, "wait_for_load"):+ await env.wait_for_load()++ self._state.num_actions += 1+ return is_stop_action, action_description++ async def _dispatch_action(+ self, env, action_type: str, args: dict+ ) -> tuple[bool, str]:+ """Dispatch an action using the environment's canonical methods."""+ if action_type == "left_click":+ tgt_x, tgt_y = args["coordinate"]+ await env.left_click(tgt_x, tgt_y)+ return False, f"I clicked at coordinates ({tgt_x}, {tgt_y})."++ elif action_type == "key":+ keys = args.get("keys", [])+ await env.key(keys)+ return False, f"I pressed the following keys: {keys}"- if action_type == "visit_url":- url = str(args["url"])- action_description = f"I typed '{url}' into the browser address bar."+ elif action_type == "mouse_move":+ tgt_x, tgt_y = args["coordinate"]+ await env.mouse_move(tgt_x, tgt_y)+ return False, f"I moved the cursor to ({tgt_x}, {tgt_y})."++ elif action_type == "type":+ text = str(args.get("text", args.get("text_value", "")))+ await env.type(text)+ return False, f"I typed '{text}'."++ elif action_type == "scroll":+ pixels = int(args.get("pixels", 0))+ pixels = int(pixels * self.config.viewport_height / self.DISPLAY_SIZE)+ await env.scroll(pixels)+ direction = "up" if pixels > 0 else "down"+ return False, f"I scrolled {direction}."++ elif action_type == "double_click":+ tgt_x, tgt_y = args["coordinate"]+ await env.double_click(tgt_x, tgt_y)+ return False, f"I double-clicked at coordinates ({tgt_x}, {tgt_y})."++ elif action_type == "right_click":+ tgt_x, tgt_y = args["coordinate"]+ await env.right_click(tgt_x, tgt_y)+ return False, f"I right-clicked at coordinates ({tgt_x}, {tgt_y})."++ elif action_type == "triple_click":+ tgt_x, tgt_y = args["coordinate"]+ await env.triple_click(tgt_x, tgt_y)+ return False, f"I triple-clicked at coordinates ({tgt_x}, {tgt_y})."++ elif action_type == "left_click_drag":+ tgt_x, tgt_y = args["coordinate"]+ await env.left_click_drag(tgt_x, tgt_y)+ return False, f"I dragged to ({tgt_x}, {tgt_y})."++ elif action_type == "hscroll":+ pixels = int(args.get("pixels", 0))+ pixels = int(pixels * self.config.viewport_width / self.DISPLAY_SIZE)+ await env.hscroll(pixels)+ return False, f"I scrolled horizontally by {pixels} pixels."++ elif action_type == "visit_url":+ url = args.get("url", "")if url.startswith(("https://", "http://", "file://", "about:")):- await env.goto(url)+ target = urlelif " " in url:- await env.goto(- f"https://www.bing.com/search?q={quote_plus(url)}&FORM=QBLH"- )+ target = f"https://www.bing.com/search?q={quote_plus(url)}&FORM=QBLH"else:- await env.goto("https://" + url)+ target = "https://" + url+ try:+ await env.goto_url(target)+ except Exception as e:+ msg = f"visit_url to {url} failed: {type(e).__name__}: {e}"+ self._pending_observation = msg+ return False, f"I tried to navigate to {url} but it failed: {e}"+ return False, f"I navigated to {url}."elif action_type == "history_back":- action_description = "I clicked the browser back button."- await env.back()+ await env.go_back()+ return False, "I clicked the browser back button."elif action_type == "web_search":- query = args.get("query")- action_description = f"I typed '{query}' into the browser search bar."- encoded_query = quote_plus(query)- await env.goto(f"https://www.bing.com/search?q={encoded_query}&FORM=QBLH")-- elif action_type == "scroll":- pixels = int(args.get("pixels", 0))- if pixels > 0:- action_description = "I scrolled up one page in the browser."- await env.scroll_up()- elif pixels < 0:- action_description = "I scrolled down one page in the browser."- await env.scroll_down()-- elif action_type in ("keypress", "key"):- keys = args.get("keys", [])- action_description = f"I pressed the following keys: {keys}"- await env.keypress(keys)-- elif action_type in ("hover", "mouse_move"):- if "coordinate" in args:- tgt_x, tgt_y = args["coordinate"]- await env.hover(tgt_x, tgt_y)-- elif action_type in ("sleep", "wait"):- duration = args.get("duration", args.get("time", 3.0))- action_description = (- "I am waiting a short period of time before taking further action."+ query = args.get("query", "")+ await env.goto_url(+ f"https://www.bing.com/search?q={quote_plus(query)}&FORM=QBLH")- await env.wait(duration)+ return False, f"I searched for '{query}'."- elif action_type in ("click", "left_click"):- if "coordinate" in args:- tgt_x, tgt_y = args["coordinate"]- action_description = f"I clicked at coordinates ({tgt_x}, {tgt_y})."- await env.click(tgt_x, tgt_y)-- elif action_type in ("input_text", "type"):- text_value = args.get("text", args.get("text_value"))- if text_value is None:- raise ValueError(- "input_text/type action requires 'text' or 'text_value' argument"- )- text_value = str(text_value)- action_description = f"I typed '{text_value}'."- press_enter = args.get("press_enter", True)- delete_existing = args.get("delete_existing_text", False)-- if "coordinate" in args:- tgt_x, tgt_y = args["coordinate"]- await env.type_text(- tgt_x,- tgt_y,- text_value,- press_enter=press_enter,- clear_first=delete_existing,- )+ elif action_type == "read_page_answer_question":+ question = str(args.get("question", ""))+ markdown = await env.get_page_markdown()+ answer = await extract_from_page(markdown, question, self._client)+ return False, f"I read the page to answer: {question}\nAnswer: {answer}"++ elif action_type == "ask_user_question":+ question = str(args.get("question", ""))+ return True, f"I asked the user: {question}"++ elif action_type == "wait":+ duration = args.get("time", args.get("duration", 3.0))+ await env.wait(duration)+ return False, f"I waited {duration}s."elif action_type == "pause_and_memorize_fact":- fact = str(args.get("fact"))+ fact = str(args.get("fact", ""))self._state.facts.append(fact)- action_description = f"I memorized the following fact: {fact}"+ return False, f"I memorized the following fact: {fact}"- elif action_type in ("stop", "terminate"):- action_description = args.get("thoughts", "Task terminated")- is_stop_action = True+ elif action_type == "terminate":+ answer = args.get("answer")+ if answer is None:+ raise ValueError("terminate action requires 'answer' argument")+ return True, answerelse:raise ValueError(f"Unknown action: {action_type}")-- if not is_stop_action:- await env.wait_for_load()-- self._state.num_actions += 1- return is_stop_action, action_descriptiondiff --git a/src/fara/agents/fara/fara_qwen3_next.py b/src/fara/agents/fara/fara_qwen3_next.pydeleted file mode 100644index 802821a..0000000--- a/src/fara/agents/fara/fara_qwen3_next.py+++ /dev/null@@ -1,369 +0,0 @@-"""FaraQwen3NextAgent - Extends FaraQwen3Agent with expanded tool sets."""--from __future__ import annotations--import json-import re-from typing import Any-from urllib.parse import quote_plus--from ...clients.messages import FunctionCall, ImageObj, SystemMessage, UserMessage--from ...core.run_context import RunContext-from .fara_qwen3 import FaraQwen3Agent, FaraQwen3AgentConfig-from .fara_types import WebSurferEvent-from .prompts import get_computer_use_system_prompt, TOOL_SET_TO_MODE-from ..computer_agent.utils import extract_from_page-from ..utils import (- format_text_observation,- save_and_truncate_command_output,- wrap_command_with_tee,-)---def extract_allowed_actions(system_prompt_text: str) -> frozenset[str]:- """Extract the allowed action names from the tool schema enum in the system prompt."""- match = re.search(r'"enum":\s*\[([^\]]+)\]', system_prompt_text)- if not match:- raise ValueError("Could not find action enum in system prompt")- return frozenset(json.loads(f"[{match.group(1)}]"))---class FaraQwen3NextAgentConfig(FaraQwen3AgentConfig):- """Configuration for FaraQwen3NextAgent."""-- name: str = "fara_next"- max_observation_chars: int = 1000- computer_use_mode: str = "fara_next_browser"- identity: str | None = "fara_qwen3vl"- critical_points: str | None = "fara-1.5"---class FaraQwen3NextAgent(FaraQwen3Agent):- """Web/desktop automation agent with expanded tool sets (FaraNext).-- Extends FaraQwen3Agent with:- - Dynamic tool mode selection (browser vs windows) via config.computer_use_mode- - New GUI actions: double_click, right_click, triple_click, left_click_drag, hscroll- - New non-GUI actions: read_page_answer_question, ask_user_question, run_command- - Cursor position tracking for left_click_drag- - press_enter defaults to False (model should use key(["Enter"]) explicitly)- """-- _TEXT_OBSERVATION_ACTIONS: frozenset[str] = frozenset(- {- "run_command",- "read_page_answer_question",- }- )-- _MODE_CANONICAL_IDENTITY: dict[str, tuple[str, ...]] = {- "fara_next_browser": ("fara_qwen3vl", "fara_qwen35"),- "fara_next_windows": ("fara_qwen3vl_windows", "fara_qwen35_windows"),- "fara_next_windows_core": ("fara_qwen3vl_windows", "fara_qwen35_windows"),- }-- def __init__(- self,- config: FaraQwen3NextAgentConfig | dict[str, Any] | None = None,- **kwargs: Any,- ):- super().__init__(config, **kwargs)- self.config: FaraQwen3NextAgentConfig- self._allowed_actions: frozenset[str] = frozenset()- self._cursor_x: float = 0.0- self._cursor_y: float = 0.0- self._pending_observation: str = ""-- @classmethod- def _get_config_class(cls) -> type[FaraQwen3NextAgentConfig]:- return FaraQwen3NextAgentConfig-- async def initialize(self, run_context: RunContext) -> None:- await super().initialize(run_context)- valid_modes = set(TOOL_SET_TO_MODE.values())- if self.config.computer_use_mode not in valid_modes:- raise ValueError(- f"Unknown computer_use_mode '{self.config.computer_use_mode}'. "- f"Known values: {sorted(valid_modes)}"- )- canonical_identities = self._MODE_CANONICAL_IDENTITY[- self.config.computer_use_mode- ]- if (- self.config.identity is not None- and self.config.identity not in canonical_identities- ):- raise ValueError(- f"identity={self.config.identity!r} does not match any canonical "- f"identity in {list(canonical_identities)!r} for "- f"computer_use_mode={self.config.computer_use_mode!r}. "- f"Use one of {list(canonical_identities)!r} or set identity=None "- f"to auto-detect."- )-- def _get_system_message(self, screenshot):- system_prompt_info = get_computer_use_system_prompt(- screenshot,- self.mlm_processor_im_cfg,- mode=self.config.computer_use_mode,- include_input_text_key_args=self.config.include_input_text_key_args,- fn_call_template=self.config.fn_call_template,- randomize=False,- display_size=self.DISPLAY_SIZE,- identity=self.config.identity,- critical_points=self.config.critical_points,- )- self._state.mlm_width, self._state.mlm_height = system_prompt_info["im_size"]- scaled_screenshot = screenshot.resize(- (self._state.mlm_width, self._state.mlm_height)- )-- system_message = []- full_text = ""- for msg in system_prompt_info["conversation"]:- tmp_content = ""- for content in msg["content"]:- tmp_content += content["text"]- system_message.append(SystemMessage(content=tmp_content))- full_text += tmp_content-- self._allowed_actions = extract_allowed_actions(full_text)-- return system_message, scaled_screenshot-- def _get_final_answer(self, thoughts: str, action_description: str) -> str:- return action_description-- def _get_observation_prefix(self) -> str:- obs = self._pending_observation- self._pending_observation = ""- return obs-- def maybe_remove_old_screenshots(self, history, includes_current=False):- """Override to also strip observation text from old messages in windows mode."""- result = super().maybe_remove_old_screenshots(history, includes_current)- if "windows" not in self.config.computer_use_mode:- return result-- for msg in result:- if not isinstance(msg, UserMessage):- continue- if (msg.metadata or {}).get("is_original", False):- continue- if isinstance(msg.content, list):- has_image = any(isinstance(c, ImageObj) for c in msg.content)- if not has_image:- msg.content = [- self.USER_MESSAGE if isinstance(c, str) else c- for c in msg.content- ]-- return result-- async def _execute_action(- self, env, function_call: list[FunctionCall]- ) -> tuple[bool, str]:- """Execute an action. Only actions in the current schema are allowed."""- name = function_call[0].name- args = function_call[0].arguments- action_type = args.get("action", "")-- if action_type not in self._allowed_actions:- if self.config.terminate_on_parse_error:- self.logger.warning(- "terminate_on_parse_error=true: invalid action %r — "- "ending trajectory with thoughts as final answer",- action_type,- )- return True, args.get("thoughts", "") or args.get("answer", "")- raise ValueError(- f"Action '{action_type}' is not allowed in mode "- f"'{self.config.computer_use_mode}'. Allowed: {sorted(self._allowed_actions)}"- )-- url = (await env.get_page_context()).url- self.logger.debug(- WebSurferEvent(- source="FaraQwen3NextAgent",- url=url,- action=name,- arguments=args,- message=f"{name}( {json.dumps(args)} )",- )- )-- if "coordinate" in args:- args["coordinate"] = self._proc(args["coordinate"])-- is_stop_action, action_description = await self._dispatch_action(- env, action_type, args- )-- if action_type in self._TEXT_OBSERVATION_ACTIONS:- self._pending_observation = format_text_observation(- action_type,- action_description,- self.config.max_observation_chars,- )-- if not is_stop_action:- if hasattr(env, "wait_for_load"):- await env.wait_for_load()-- self._state.num_actions += 1- return is_stop_action, action_description-- async def _dispatch_action(- self, env, action_type: str, args: dict- ) -> tuple[bool, str]:- """Dispatch action using ComputerEnvironment canonical methods."""--- if action_type == "left_click":- tgt_x, tgt_y = args["coordinate"]- await env.left_click(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I clicked at coordinates ({tgt_x}, {tgt_y})."-- elif action_type == "key":- keys = args.get("keys", [])- await env.key(keys)- return False, f"I pressed the following keys: {keys}"-- elif action_type == "mouse_move":- tgt_x, tgt_y = args["coordinate"]- await env.mouse_move(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I moved the cursor to ({tgt_x}, {tgt_y})."-- elif action_type == "type":- text = str(args.get("text", args.get("text_value", "")))- await env.type(text)- return False, f"I typed '{text}'."-- elif action_type == "scroll":- pixels = int(args.get("pixels", 0))- pixels = int(pixels * self.config.viewport_height / self.DISPLAY_SIZE)- await env.scroll(pixels)- direction = "up" if pixels > 0 else "down"- return False, f"I scrolled {direction}."-- elif action_type == "double_click":- tgt_x, tgt_y = args["coordinate"]- await env.double_click(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I double-clicked at coordinates ({tgt_x}, {tgt_y})."-- elif action_type == "right_click":- tgt_x, tgt_y = args["coordinate"]- await env.right_click(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I right-clicked at coordinates ({tgt_x}, {tgt_y})."-- elif action_type == "triple_click":- tgt_x, tgt_y = args["coordinate"]- await env.triple_click(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I triple-clicked at coordinates ({tgt_x}, {tgt_y})."-- elif action_type == "left_click_drag":- tgt_x, tgt_y = args["coordinate"]- await env.left_click_drag(tgt_x, tgt_y)- self._cursor_x, self._cursor_y = tgt_x, tgt_y- return False, f"I dragged to ({tgt_x}, {tgt_y})."-- elif action_type == "hscroll":- pixels = int(args.get("pixels", 0))- pixels = int(pixels * self.config.viewport_width / self.DISPLAY_SIZE)- await env.hscroll(pixels)- return False, f"I scrolled horizontally by {pixels} pixels."--- elif action_type == "visit_url":- url = args.get("url", "")- if url.startswith(("https://", "http://", "file://", "about:")):- target = url- elif " " in url:- target = f"https://www.bing.com/search?q={quote_plus(url)}&FORM=QBLH"- else:- target = "https://" + url- try:- await env.goto_url(target)- except Exception as e:- msg = f"visit_url to {url} failed: {type(e).__name__}: {e}"- self._pending_observation = msg- return False, f"I tried to navigate to {url} but it failed: {e}"- return False, f"I navigated to {url}."-- elif action_type == "history_back":- await env.go_back()- return False, "I clicked the browser back button."-- elif action_type == "web_search":- query = args.get("query", "")- await env.goto_url(- f"https://www.bing.com/search?q={quote_plus(query)}&FORM=QBLH"- )- return False, f"I searched for '{query}'."-- elif action_type == "read_page_answer_question":- question = str(args.get("question", ""))- markdown = await env.get_page_markdown()- answer = await extract_from_page(markdown, question, self._client)- return False, f"I read the page to answer: {question}\nAnswer: {answer}"--- elif action_type == "run_command":- command = str(args.get("command", ""))- step = self._state.current_step- cmd = wrap_command_with_tee(command, step, self._os_type)- result = await env.execute(cmd)- if isinstance(result, dict):- stdout = result.get("stdout", "")- stderr = result.get("stderr", "")- result = f"{stdout}\n{stderr}".strip() if stderr else stdout- else:- result = str(result)- result = save_and_truncate_command_output(- result,- self._output_dir,- step,- max_chars=self.config.max_observation_chars,- os_type=self._os_type,- )- return False, f"I executed command: {command}\nOutput: {result}"--- elif action_type == "ask_user_question":- question = str(args.get("question", ""))- return True, f"I asked the user: {question}"-- elif action_type == "wait":- duration = args.get("time", args.get("duration", 3.0))- await env.wait(duration)- return False, f"I waited {duration}s."-- elif action_type == "pause_and_memorize_fact":- fact = str(args.get("fact", ""))- self._state.facts.append(fact)- return False, f"I memorized the following fact: {fact}"-- elif action_type == "terminate":- answer = args.get("answer")- if answer is None:- raise ValueError("terminate action requires 'answer' argument")- return True, answer-- else:- raise ValueError(f"Unknown action: {action_type}")-- def _proc(self, coordinate):- """Shorthand for proc_coords with standard display/viewport args."""- return self.proc_coords(- coordinate,- self.DISPLAY_SIZE,- self.DISPLAY_SIZE,- self.config.viewport_width,- self.config.viewport_height,- )diff --git a/src/fara/agents/fara/prompts.py b/src/fara/agents/fara/prompts.pydeleted file mode 100644index e64f2e3..0000000--- a/src/fara/agents/fara/prompts.py+++ /dev/null@@ -1,22 +0,0 @@-"""Fara prompt entry points.--``get_computer_use_system_prompt`` is re-exported from this repo's-``fara._prompts`` (which ships the FaraNext browser tool schema, the model-identities, the critical-points block, and the fn-call template). Only the-browser tool set is supported here.-"""--from ..._prompts import get_computer_use_system_prompt--__all__ = ["get_computer_use_system_prompt", "TOOL_SET_TO_MODE"]--TOOL_SET_TO_MODE = {- "BROWSER_TOOLS_CORE": "fara_next_browser",- "GPT54_BROWSER_TOOLS_CORE": "fara_next_browser",- "BROWSER_TOOLS_WITH_READ_PAGE": "fara_next_browser",- "GPT54_BROWSER_TOOLS_WITH_READ_PAGE": "fara_next_browser",- "WINDOWS_TOOLS_CORE": "fara_next_windows_core",- "WINDOWS_TOOLS_WITH_RUN_COMMAND": "fara_next_windows",- "GPT54_WINDOWS_TOOLS_CORE": "fara_next_windows_core",- "GPT54_WINDOWS_TOOLS_WITH_RUN_COMMAND": "fara_next_windows",-}diff --git a/src/fara/run_fara.py b/src/fara/run_fara.pyindex a919d3d..4b9745c 100755--- a/src/fara/run_fara.py+++ b/src/fara/run_fara.py@@ -7,7 +7,7 @@ from typing import Dictfrom pathlib import Pathimport json-from .agents.fara.fara_qwen3_next import FaraQwen3NextAgent, FaraQwen3NextAgentConfig+from .agents.fara.fara15_agent import Fara15Agent, Fara15AgentConfigfrom .environments.playwright import PlaywrightEnvironmentfrom .core.run_context import RunContextfrom .core.data_point import Task, SolverStatus, UserMessage, UserMessageType@@ -20,7 +20,7 @@ logging.basicConfig(format="%(message)s",)-for _logger_name in ("fara.agents.fara.fara_qwen3", "fara.fara_7b.fara_agent"):+for _logger_name in ("fara.agents.fara.fara15_agent", "fara.fara_7b.fara_agent"):_agent_logger = logging.getLogger(_logger_name)_agent_logger.setLevel(logging.INFO)_handler = logging.StreamHandler()@@ -55,7 +55,7 @@ async def run_fara15_agent(max_rounds: int = 100,use_browser_base: bool = False,) -> None:- """Interactive loop for the Fara-1.5 agent (FaraQwen3NextAgent)."""+ """Interactive loop for the Fara-1.5 agent."""output_root = Path(output_folder or "./fara_runs")print("Initializing Browser...")@@ -71,8 +71,8 @@ async def run_fara15_agent(await env.initialize()print("Browser Running... Starting Fara-1.5 Agent...")- agent = FaraQwen3NextAgent(- FaraQwen3NextAgentConfig(+ agent = Fara15Agent(+ Fara15AgentConfig(client_config=endpoint_config,max_rounds=max_rounds,identity="fara_qwen35",diff --git a/tests/test_fara15.py b/tests/test_fara15.pyindex eabdf6c..661551f 100644--- a/tests/test_fara15.py+++ b/tests/test_fara15.py@@ -1,4 +1,4 @@-"""Tests for the Fara-1.5 agent (vendored FaraQwen3NextAgent).+"""Tests for the Fara-1.5 agent (Fara15Agent).Covers the two things the port must get exactly right: the system promptthe model sees (identity + critical points + tool schema), and the@@ -10,9 +10,9 @@ import loggingfrom PIL import Image-from fara import DataPoint, FaraQwen3NextAgent, Task-from fara.agents.fara.fara_qwen3_next import (- FaraQwen3NextAgentConfig,+from fara import DataPoint, Fara15Agent, Task+from fara.agents.fara.fara15_agent import (+ Fara15AgentConfig,extract_allowed_actions,)from fara._prompts import (@@ -49,8 +49,8 @@ BROWSER_ACTIONS = {def _make_agent(identity="fara_qwen35"):- agent = FaraQwen3NextAgent(- FaraQwen3NextAgentConfig(+ agent = Fara15Agent(+ Fara15AgentConfig(client_config={"model": "m", "base_url": "u", "api_key": "k"},identity=identity,)@@ -122,7 +122,7 @@ def test_proc_coords_scales_display_space_to_viewport():def test_parse_thoughts_and_action():- agent = FaraQwen3NextAgent.__new__(FaraQwen3NextAgent)+ agent = Fara15Agent.__new__(Fara15Agent)agent.logger = logging.getLogger("test")action_json = json.dumps({