sentienceapi 0.90.12__py3-none-any.whl → 0.92.2__py3-none-any.whl

sentience/__init__.py

@@ -3,7 +3,7 @@ Sentience Python SDK - AI Agent Browser Automation
 """
 
 from .actions import click, click_rect, press, type_text
-from .agent import SentienceAgent
+from .agent import SentienceAgent, SentienceAgentAsync
 from .agent_config import AgentConfig
 
 # Agent Layer (Phase 1 & 2)
@@ -14,9 +14,6 @@ from .browser import SentienceBrowser
 from .cloud_tracing import CloudTraceSink, SentienceLogger
 from .conversational_agent import ConversationalAgent
 from .expect import expect
-
-# Formatting (v0.12.0+)
-from .formatting import format_snapshot_for_llm
 from .generator import ScriptGenerator, generate
 from .inspector import Inspector, inspect
 from .llm_provider import (
@@ -55,12 +52,14 @@ from .query import find, query
 from .read import read
 from .recorder import Recorder, Trace, TraceStep, record
 from .screenshot import screenshot
+from .sentience_methods import AgentAction, SentienceMethod
 from .snapshot import snapshot
 from .text_search import find_text_rect
 from .tracer_factory import SENTIENCE_API_URL, create_tracer
 from .tracing import JsonlTraceSink, TraceEvent, Tracer, TraceSink
 
 # Utilities (v0.12.0+)
+# Import from utils package (re-exports from submodules for backward compatibility)
 from .utils import (
     canonical_snapshot_loose,
     canonical_snapshot_strict,
@@ -68,9 +67,13 @@ from .utils import (
     save_storage_state,
     sha256_digest,
 )
+
+# Formatting (v0.12.0+)
+from .utils.formatting import format_snapshot_for_llm
+from .visual_agent import SentienceVisualAgent, SentienceVisualAgentAsync
 from .wait import wait_for
 
-__version__ = "0.90.12"
+__version__ = "0.92.2"
 
 __all__ = [
     # Core SDK
@@ -117,6 +120,9 @@ __all__ = [
     "AnthropicProvider",
     "LocalLLMProvider",
     "SentienceAgent",
+    "SentienceAgentAsync",
+    "SentienceVisualAgent",
+    "SentienceVisualAgentAsync",
     "ConversationalAgent",
     # Agent Layer Models
     "AgentActionResult",
@@ -150,4 +156,7 @@ __all__ = [
     "format_snapshot_for_llm",
     # Agent Config (v0.12.0+)
     "AgentConfig",
+    # Enums
+    "SentienceMethod",
+    "AgentAction",
 ]

sentience/_extension_loader.py

@@ -0,0 +1,40 @@
+"""
+Shared extension loading logic for sync and async implementations
+"""
+
+from pathlib import Path
+
+
+def find_extension_path() -> Path:
+    """
+    Find Sentience extension directory (shared logic for sync and async).
+
+    Checks multiple locations:
+    1. sentience/extension/ (installed package)
+    2. ../sentience-chrome (development/monorepo)
+
+    Returns:
+        Path to extension directory
+
+    Raises:
+        FileNotFoundError: If extension not found in any location
+    """
+    # 1. Try relative to this file (installed package structure)
+    # sentience/_extension_loader.py -> sentience/extension/
+    package_ext_path = Path(__file__).parent / "extension"
+
+    # 2. Try development root (if running from source repo)
+    # sentience/_extension_loader.py -> ../sentience-chrome
+    dev_ext_path = Path(__file__).parent.parent.parent / "sentience-chrome"
+
+    if package_ext_path.exists() and (package_ext_path / "manifest.json").exists():
+        return package_ext_path
+    elif dev_ext_path.exists() and (dev_ext_path / "manifest.json").exists():
+        return dev_ext_path
+    else:
+        raise FileNotFoundError(
+            f"Extension not found. Checked:\n"
+            f"1. {package_ext_path}\n"
+            f"2. {dev_ext_path}\n"
+            "Make sure the extension is built and 'sentience/extension' directory exists."
+        )

sentience/action_executor.py

@@ -0,0 +1,215 @@
+"""
+Action Executor for Sentience Agent.
+
+Handles parsing and execution of action commands (CLICK, TYPE, PRESS, FINISH).
+This separates action execution concerns from LLM interaction.
+"""
+
+import re
+from typing import Any, Union
+
+from .actions import click, click_async, press, press_async, type_text, type_text_async
+from .browser import AsyncSentienceBrowser, SentienceBrowser
+from .models import Snapshot
+from .protocols import AsyncBrowserProtocol, BrowserProtocol
+
+
+class ActionExecutor:
+    """
+    Executes actions and handles parsing of action command strings.
+
+    This class encapsulates all action execution logic, making it easier to:
+    - Test action execution independently
+    - Add new action types in one place
+    - Handle action parsing errors consistently
+    """
+
+    def __init__(
+        self,
+        browser: SentienceBrowser | AsyncSentienceBrowser | BrowserProtocol | AsyncBrowserProtocol,
+    ):
+        """
+        Initialize action executor.
+
+        Args:
+            browser: SentienceBrowser, AsyncSentienceBrowser, or protocol-compatible instance
+                    (for testing, can use mock objects that implement BrowserProtocol)
+        """
+        self.browser = browser
+        # Check if browser is async - support both concrete types and protocols
+        # Check concrete types first (most reliable)
+        if isinstance(browser, AsyncSentienceBrowser):
+            self._is_async = True
+        elif isinstance(browser, SentienceBrowser):
+            self._is_async = False
+        else:
+            # For protocol-based browsers, check if methods are actually async
+            # This is more reliable than isinstance checks which can match both protocols
+            import inspect
+
+            start_method = getattr(browser, "start", None)
+            if start_method and inspect.iscoroutinefunction(start_method):
+                self._is_async = True
+            elif isinstance(browser, BrowserProtocol):
+                # If it implements BrowserProtocol and start is not async, it's sync
+                self._is_async = False
+            else:
+                # Default to sync for unknown types
+                self._is_async = False
+
+    def execute(self, action_str: str, snap: Snapshot) -> dict[str, Any]:
+        """
+        Parse action string and execute SDK call (synchronous).
+
+        Args:
+            action_str: Action string from LLM (e.g., "CLICK(42)", "TYPE(15, \"text\")")
+            snap: Current snapshot (for context, currently unused but kept for API consistency)
+
+        Returns:
+            Execution result dictionary with keys:
+            - success: bool
+            - action: str (e.g., "click", "type", "press", "finish")
+            - element_id: Optional[int] (for click/type actions)
+            - text: Optional[str] (for type actions)
+            - key: Optional[str] (for press actions)
+            - outcome: Optional[str] (action outcome)
+            - url_changed: Optional[bool] (for click actions)
+            - error: Optional[str] (if action failed)
+            - message: Optional[str] (for finish action)
+
+        Raises:
+            ValueError: If action format is unknown
+            RuntimeError: If called on async browser (use execute_async instead)
+        """
+        if self._is_async:
+            raise RuntimeError(
+                "ActionExecutor.execute() called on async browser. Use execute_async() instead."
+            )
+
+        # Parse CLICK(42)
+        if match := re.match(r"CLICK\s*\(\s*(\d+)\s*\)", action_str, re.IGNORECASE):
+            element_id = int(match.group(1))
+            result = click(self.browser, element_id)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "click",
+                "element_id": element_id,
+                "outcome": result.outcome,
+                "url_changed": result.url_changed,
+            }
+
+        # Parse TYPE(42, "hello world")
+        elif match := re.match(
+            r'TYPE\s*\(\s*(\d+)\s*,\s*["\']([^"\']*)["\']\s*\)',
+            action_str,
+            re.IGNORECASE,
+        ):
+            element_id = int(match.group(1))
+            text = match.group(2)
+            result = type_text(self.browser, element_id, text)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "type",
+                "element_id": element_id,
+                "text": text,
+                "outcome": result.outcome,
+            }
+
+        # Parse PRESS("Enter")
+        elif match := re.match(r'PRESS\s*\(\s*["\']([^"\']+)["\']\s*\)', action_str, re.IGNORECASE):
+            key = match.group(1)
+            result = press(self.browser, key)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "press",
+                "key": key,
+                "outcome": result.outcome,
+            }
+
+        # Parse FINISH()
+        elif re.match(r"FINISH\s*\(\s*\)", action_str, re.IGNORECASE):
+            return {
+                "success": True,
+                "action": "finish",
+                "message": "Task marked as complete",
+            }
+
+        else:
+            raise ValueError(
+                f"Unknown action format: {action_str}\n"
+                f'Expected: CLICK(id), TYPE(id, "text"), PRESS("key"), or FINISH()'
+            )
+
+    async def execute_async(self, action_str: str, snap: Snapshot) -> dict[str, Any]:
+        """
+        Parse action string and execute SDK call (asynchronous).
+
+        Args:
+            action_str: Action string from LLM (e.g., "CLICK(42)", "TYPE(15, \"text\")")
+            snap: Current snapshot (for context, currently unused but kept for API consistency)
+
+        Returns:
+            Execution result dictionary (same format as execute())
+
+        Raises:
+            ValueError: If action format is unknown
+            RuntimeError: If called on sync browser (use execute() instead)
+        """
+        if not self._is_async:
+            raise RuntimeError(
+                "ActionExecutor.execute_async() called on sync browser. Use execute() instead."
+            )
+
+        # Parse CLICK(42)
+        if match := re.match(r"CLICK\s*\(\s*(\d+)\s*\)", action_str, re.IGNORECASE):
+            element_id = int(match.group(1))
+            result = await click_async(self.browser, element_id)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "click",
+                "element_id": element_id,
+                "outcome": result.outcome,
+                "url_changed": result.url_changed,
+            }
+
+        # Parse TYPE(42, "hello world")
+        elif match := re.match(
+            r'TYPE\s*\(\s*(\d+)\s*,\s*["\']([^"\']*)["\']\s*\)',
+            action_str,
+            re.IGNORECASE,
+        ):
+            element_id = int(match.group(1))
+            text = match.group(2)
+            result = await type_text_async(self.browser, element_id, text)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "type",
+                "element_id": element_id,
+                "text": text,
+                "outcome": result.outcome,
+            }
+
+        # Parse PRESS("Enter")
+        elif match := re.match(r'PRESS\s*\(\s*["\']([^"\']+)["\']\s*\)', action_str, re.IGNORECASE):
+            key = match.group(1)
+            result = await press_async(self.browser, key)  # type: ignore
+            return {
+                "success": result.success,
+                "action": "press",
+                "key": key,
+                "outcome": result.outcome,
+            }
+
+        # Parse FINISH()
+        elif re.match(r"FINISH\s*\(\s*\)", action_str, re.IGNORECASE):
+            return {
+                "success": True,
+                "action": "finish",
+                "message": "Task marked as complete",
+            }
+
+        else:
+            raise ValueError(
+                f"Unknown action format: {action_str}\n"
+                f'Expected: CLICK(id), TYPE(id, "text"), PRESS("key"), or FINISH()'
+            )