Merge pull request #46 from browser-use/feature/output-variables

Type safe outputs from a workflow

Author: gregpr07

workflows/backend/service.py

@@ -180,14 +180,14 @@ async def run_workflow_in_background(
 					'extracted_content': s.extracted_content,
 					'status': 'completed',
 				}
-				for i, s in enumerate(result)
+				for i, s in enumerate(result.step_results)
 			]
 			for step in formatted_result:
 				await self._write_log(log_file, f'[{ts}] Completed step {step["step_id"]}: {step["extracted_content"]}\n')
 
 			self.active_tasks[task_id].status = 'completed'
 			self.active_tasks[task_id].result = formatted_result
-			await self._write_log(log_file, f'[{ts}] Workflow completed successfully with {len(result)} steps\n')
+			await self._write_log(log_file, f'[{ts}] Workflow completed successfully with {len(result.step_results)} steps\n')
 
 		except asyncio.CancelledError:
 			await self._write_log(log_file, f'[{time.strftime("%Y-%m-%d %H:%M:%S")}] Workflow force‑cancelled\n')

workflows/cli.py

@@ -32,12 +32,14 @@
 llm_instance = None
 try:
 	llm_instance = ChatOpenAI(model='gpt-4o')
+	page_extraction_llm = ChatOpenAI(model='gpt-4o-mini')
 except Exception as e:
 	typer.secho(f'Error initializing LLM: {e}. Would you like to set your OPENAI_API_KEY?', fg=typer.colors.RED)
 	set_openai_api_key = input('Set OPENAI_API_KEY? (y/n): ')
 	if set_openai_api_key.lower() == 'y':
 		os.environ['OPENAI_API_KEY'] = input('Enter your OPENAI_API_KEY: ')
 		llm_instance = ChatOpenAI(model='gpt-4o')
+		page_extraction_llm = ChatOpenAI(model='gpt-4o-mini')
 
 builder_service = BuilderService(llm=llm_instance) if llm_instance else None
 # recorder_service = RecorderService() # Placeholder
@@ -151,7 +153,6 @@ def create_workflow():
 	"""
 	if not recording_service:
 		# Adjusted RecordingService initialization check assuming it doesn't need LLM
-		# If it does, this check should be more robust (e.g. based on llm_instance)
 		typer.secho(
 			'RecordingService not available. Cannot create workflow.',
 			fg=typer.colors.RED,
@@ -279,7 +280,7 @@ def run_as_tool_command(
 
 	try:
 		# Pass llm_instance to ensure the workflow can use it if needed for as_tool() or run_with_prompt()
-		workflow_obj = Workflow.load_from_file(str(workflow_path), llm=llm_instance)
+		workflow_obj = Workflow.load_from_file(str(workflow_path), llm=llm_instance, page_extraction_llm=page_extraction_llm)
 	except Exception as e:
 		typer.secho(f'Error loading workflow: {e}', fg=typer.colors.RED)
 		raise typer.Exit(code=1)
@@ -328,7 +329,11 @@ def run_workflow_command(
 		browser_instance = Browser()  # Add any necessary config if required
 		controller_instance = WorkflowController()  # Add any necessary config if required
 		workflow_obj = Workflow.load_from_file(
-			str(workflow_path), llm=llm_instance, browser=browser_instance, controller=controller_instance
+			str(workflow_path),
+			llm=llm_instance,
+			browser=browser_instance,
+			controller=controller_instance,
+			page_extraction_llm=page_extraction_llm,
 		)
 	except Exception as e:
 		typer.secho(f'Error loading workflow: {e}', fg=typer.colors.RED)
@@ -389,7 +394,7 @@ def run_workflow_command(
 		typer.secho('\nWorkflow execution completed!', fg=typer.colors.GREEN, bold=True)
 		typer.echo(typer.style('Result:', bold=True))
 		# Output the number of steps executed, similar to previous behavior
-		typer.echo(f'{typer.style(str(len(result)), bold=True)} steps executed.')
+		typer.echo(f'{typer.style(str(len(result.step_results)), bold=True)} steps executed.')
 		# For more detailed results, one might want to iterate through the 'result' list
 		# and print each item, or serialize the whole list to JSON.
 		# For now, sticking to the step count as per original output.
@@ -415,7 +420,9 @@ def mcp_server_command(
 	typer.echo()  # Add space
 
 	llm_instance = ChatOpenAI(model='gpt-4o')
-	mcp = get_mcp_server(llm_instance, workflow_dir='./tmp')
+	page_extraction_llm = ChatOpenAI(model='gpt-4o-mini')
+
+	mcp = get_mcp_server(llm_instance, page_extraction_llm=page_extraction_llm, workflow_dir='./tmp')
 
 	mcp.run(
 		transport='sse',

workflows/pyproject.toml

@@ -45,6 +45,9 @@ quote-style = "single"
 indent-style = "tab"
 docstring-code-format = true
 
+[tool.pyright]
+typeCheckingMode = "basic"
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"

workflows/workflow_use/builder/prompts.py

@@ -34,6 +34,7 @@
        - Evaluating content to match user input (e.g., finding a specific item based on its name or attributes).
      - Break complex tasks into multiple specific agentic steps rather than one broad task.
      - **Use the user’s goal (if provided) or inferred intent from the recording** to identify where agentic steps are needed for dynamic content, even if the recording uses deterministic steps.
+   - **extract_page_content** - Use this type when you want to extract data from the page. If the task is simply extracting data from the page, use this instead of agentic steps (never create agentic step for simple data extraction).
    - **Deterministic events** → keep the original recorder event structure. The
      value of `"type"` MUST match **exactly** one of the available action
      names listed below; all additional keys are interpreted as parameters for