hud-python 0.4.12__tar.gz → 0.4.14__tar.gz

{hud_python-0.4.12 → hud_python-0.4.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.12
+Version: 0.4.14
 Summary: SDK for the HUD platform.
 Project-URL: Homepage, https://github.com/hud-evals/hud-python
 Project-URL: Bug Tracker, https://github.com/hud-evals/hud-python/issues
@@ -35,10 +35,9 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
 Requires-Python: <3.14,>=3.11
-Requires-Dist: fastmcp>=2.11.2
 Requires-Dist: httpx<1,>=0.23.0
-Requires-Dist: hud-mcp-python-sdk>=0.1.0
-Requires-Dist: mcp>=1.13.1
+Requires-Dist: hud-fastmcp-python-sdk>=0.1.2
+Requires-Dist: hud-mcp-python-sdk>=3.13.2
 Requires-Dist: opentelemetry-api>=1.34.1
 Requires-Dist: opentelemetry-exporter-otlp-proto-http>=1.34.1
 Requires-Dist: opentelemetry-instrumentation-mcp>=0.44.1
@@ -56,6 +55,7 @@ Provides-Extra: agent
 Requires-Dist: anthropic; extra == 'agent'
 Requires-Dist: datasets>=2.14.0; extra == 'agent'
 Requires-Dist: dotenv>=0.9.9; extra == 'agent'
+Requires-Dist: hud-mcp-use-python-sdk>=2.3.13; extra == 'agent'
 Requires-Dist: ipykernel; extra == 'agent'
 Requires-Dist: ipython<9; extra == 'agent'
 Requires-Dist: jupyter-client; extra == 'agent'
@@ -63,13 +63,13 @@ Requires-Dist: jupyter-core; extra == 'agent'
 Requires-Dist: langchain; extra == 'agent'
 Requires-Dist: langchain-anthropic; extra == 'agent'
 Requires-Dist: langchain-openai; extra == 'agent'
-Requires-Dist: mcp-use; extra == 'agent'
 Requires-Dist: numpy>=1.24.0; extra == 'agent'
 Requires-Dist: openai; extra == 'agent'
 Provides-Extra: agents
 Requires-Dist: anthropic; extra == 'agents'
 Requires-Dist: datasets>=2.14.0; extra == 'agents'
 Requires-Dist: dotenv>=0.9.9; extra == 'agents'
+Requires-Dist: hud-mcp-use-python-sdk>=2.3.13; extra == 'agents'
 Requires-Dist: ipykernel; extra == 'agents'
 Requires-Dist: ipython<9; extra == 'agents'
 Requires-Dist: jupyter-client; extra == 'agents'
@@ -77,7 +77,6 @@ Requires-Dist: jupyter-core; extra == 'agents'
 Requires-Dist: langchain; extra == 'agents'
 Requires-Dist: langchain-anthropic; extra == 'agents'
 Requires-Dist: langchain-openai; extra == 'agents'
-Requires-Dist: mcp-use; extra == 'agents'
 Requires-Dist: numpy>=1.24.0; extra == 'agents'
 Requires-Dist: openai; extra == 'agents'
 Provides-Extra: dev
@@ -85,6 +84,7 @@ Requires-Dist: aiodocker>=0.24.0; extra == 'dev'
 Requires-Dist: anthropic; extra == 'dev'
 Requires-Dist: datasets>=2.14.0; extra == 'dev'
 Requires-Dist: dotenv>=0.9.9; extra == 'dev'
+Requires-Dist: hud-mcp-use-python-sdk>=2.3.13; extra == 'dev'
 Requires-Dist: inspect-ai>=0.3.80; extra == 'dev'
 Requires-Dist: ipykernel; extra == 'dev'
 Requires-Dist: ipython<9; extra == 'dev'
@@ -93,7 +93,6 @@ Requires-Dist: jupyter-core; extra == 'dev'
 Requires-Dist: langchain; extra == 'dev'
 Requires-Dist: langchain-anthropic; extra == 'dev'
 Requires-Dist: langchain-openai; extra == 'dev'
-Requires-Dist: mcp-use; extra == 'dev'
 Requires-Dist: numpy>=1.24.0; extra == 'dev'
 Requires-Dist: openai; extra == 'dev'
 Requires-Dist: pillow>=11.1.0; extra == 'dev'

{hud_python-0.4.12 → hud_python-0.4.14}/environments/README.md

@@ -351,7 +351,7 @@ from . import basic, advanced  # This registers all @setup.tool() decorated func
 
 # In setup/basic.py
 from . import setup
-from hud.tools.types import SetupResult
+from mcp.types import TextContent
 
 @setup.tool()
 async def reset(**kwargs):
@@ -361,14 +361,14 @@ async def reset(**kwargs):
         **kwargs: Additional parameters
     
     Returns:
-        SetupResult
+        TextContent
     """
     # Access environment from the hub
     env = setup.env
     await env.reset_state()
-    return SetupResult(
-        content="Environment reset to initial state",
-        info={"status": "success"}
+    return TextContent(
+        text="Environment reset to initial state",
+        type="text"
     )
 
 @setup.tool()
@@ -379,14 +379,14 @@ async def seed_data(num_items: int = 5):
         num_items: Number of items to create
     
     Returns:
-        SetupResult
+        TextContent
     """
     # Access environment from the hub
     env = setup.env
     items = await env.create_items(num_items)
-    return SetupResult(
-        content=f"Created {len(items)} items",
-        info={"items_created": len(items)}
+    return TextContent(
+        text=f"Created {len(items)} items",
+        type="text"
     )
 
 # In evaluate/__init__.py
@@ -827,13 +827,13 @@ Before making changes:
 ```python
 # In setup/my_new_setup.py
 from . import setup
-from hud.tools import BaseSetup, SetupResult
+from hud.tools import BaseSetup, TextContent
 
 @setup("my_new_setup", description="Clear description of what this does")
 class MyNewSetup(BaseSetup):
-    async def __call__(self, context, param1: str, param2: int = 10) -> SetupResult:
+    async def __call__(self, context, param1: str, param2: int = 10) -> TextContent:
         # Implementation
-        return {"status": "success", "details": "..."}
+        return TextContent(...)
 ```
 
 **Adding New Evaluators**

hud_python-0.4.14/environments/browser/README.md

@@ -0,0 +1,213 @@
+# Browser Environment
+
+A browser automation environment for HUD that provides GUI access and web app interaction capabilities. This environment supports hot-reloading during development while maintaining persistent state.
+
+## Architecture Overview
+
+The browser environment uses a two-process architecture:
+
+1. **Context Server** (`context.py`): Long-running process that maintains persistent state
+2. **MCP Server** (`server.py`): Hot-reloadable process that handles tool requests
+
+### Key Components
+
+- **BrowserContext**: Stores persistent state (running apps, ports, playwright instance)
+- **ServiceManager**: Manages X11, VNC, and app processes
+- **BaseHub Tools**: Setup and evaluate tools organized by app (2048, todo)
+- **Multiprocessing Proxy**: Enables state sharing between processes
+
+## Context Management and Common Pitfalls
+
+### Understanding the Proxy System
+
+The browser environment uses Python's `multiprocessing.Manager` to share state between the context server and MCP server. This introduces important constraints:
+
+#### ❌ Common Pitfall: Unpicklable Objects
+
+```python
+# BAD: This will fail with "cannot pickle 'coroutine' object"
+@setup.tool("my_tool")
+async def my_tool():
+    env = setup.env
+    result = await env.call_app_api("app", "/api/endpoint")  # Returns coroutine
+    # The coroutine can't be serialized through the proxy!
+```
+
+#### ✅ Solution: Direct HTTP Calls
+
+```python
+# GOOD: Make HTTP calls directly
+@setup.tool("my_tool")
+async def my_tool():
+    import httpx
+    
+    # Get the backend port from persistent context
+    persistent_ctx = setup.env
+    backend_port = persistent_ctx.get_app_backend_port("app")
+    
+    # Make API call directly
+    url = f"http://localhost:{backend_port}/api/endpoint"
+    async with httpx.AsyncClient() as client:
+        response = await client.get(url)
+        response.raise_for_status()
+        result = response.json()
+```
+
+### State Synchronization Issues
+
+#### ❌ Common Pitfall: Direct List/Dict Manipulation
+
+```python
+# BAD: Regular Python lists don't sync through proxy
+class ServiceManager:
+    def __init__(self):
+        self._launched_apps = []  # Won't sync!
+```
+
+#### ✅ Solution: Store State in Persistent Context
+
+```python
+# GOOD: Use the persistent context for shared state
+class BrowserContext:
+    def __init__(self):
+        self._running_apps: List[str] = []
+        self._app_ports: Dict[str, Dict[str, int]] = {}
+    
+    def add_running_app(self, app_name: str) -> None:
+        """Add app to running list."""
+        if app_name not in self._running_apps:
+            self._running_apps.append(app_name)
+```
+
+### Accessing Shared Resources
+
+#### ❌ Common Pitfall: Direct Attribute Access
+
+```python
+# BAD: Direct attribute access on proxy objects
+playwright_tool = env.playwright  # May not work with proxy
+```
+
+#### ✅ Solution: Use Getter Methods
+
+```python
+# GOOD: Use proxy-friendly getter methods
+playwright_tool = persistent_ctx.get_playwright_tool()
+```
+
+## Best Practices
+
+### 1. Tool Implementation Pattern
+
+All setup and evaluate tools should follow this pattern:
+
+```python
+@setup.tool("tool_name")
+async def tool_name(param1: type, param2: type):
+    """Tool description."""
+    try:
+        # Get persistent context
+        persistent_ctx = setup.env  # or evaluate.env
+        
+        # Get app ports
+        backend_port = persistent_ctx.get_app_backend_port("app_name")
+        
+        # Make HTTP request
+        url = f"http://localhost:{backend_port}/api/endpoint"
+        async with httpx.AsyncClient() as client:
+            response = await client.method(url, json=data)
+            response.raise_for_status()
+            result = response.json()
+        
+        # Return result
+        return TextContent(
+            text=f"Success message",
+            type="text"
+        )
+    except Exception as e:
+        logger.error(f"tool_name failed: {e}")
+        return TextContent(
+            text=f"Failed: {str(e)}",
+            type="text"
+        )
+```
+
+### 2. App Launch Pattern
+
+When launching apps, ensure ports are stored in the persistent context:
+
+```python
+# In launch_app tool
+app_info = await service_manager.launch_app(app_name)
+
+# Store ports in persistent context for later access
+try:
+    backend_port = service_manager.get_app_port(app_name)
+    frontend_port = service_manager.get_app_frontend_port(app_name)
+    persistent_ctx.set_app_ports(app_name, frontend_port, backend_port)
+except Exception as e:
+    logger.error(f"Failed to store ports: {e}")
+
+# Track app in persistent context
+persistent_ctx.add_running_app(app_name)
+```
+
+### 3. Import Organization
+
+Keep imports at module level:
+
+```python
+# At top of file
+import logging
+import httpx
+from mcp.types import TextContent
+from . import setup
+
+# Not inside functions
+```
+
+## Troubleshooting
+
+### "Cannot pickle 'coroutine' object"
+
+**Cause**: Trying to return an async function result through the proxy.
+
+**Fix**: Don't use async methods on proxied objects. Make direct HTTP calls instead.
+
+### "App not launched" errors
+
+**Cause**: State synchronization issue between ServiceManager and persistent context.
+
+**Fix**: Ensure `launch_app` stores app info in the persistent context, and setup/evaluate tools check the persistent context's app list.
+
+### "Object has no attribute" on proxy objects
+
+**Cause**: Direct attribute access on multiprocessing proxy objects.
+
+**Fix**: Use getter/setter methods instead of direct attribute access.
+
+## Development Workflow
+
+1. **Start the environment**: `hud dev`
+2. **Make changes**: Edit tools in `src/hud_controller/`
+3. **Test immediately**: The MCP server hot-reloads automatically
+4. **Check logs**: Look for serialization or proxy errors
+
+## Adding New Apps
+
+1. Create app directory in `apps/`
+2. Add setup tools in `src/hud_controller/setup/app_name.py`
+3. Add evaluate tools in `src/hud_controller/evaluate/app_name.py`
+4. Follow the HTTP pattern - no `call_app_api` usage
+5. Store app ports in persistent context when launching
+
+## Key Files
+
+- `context.py`: Persistent state management
+- `server.py`: MCP server and tool definitions
+- `services.py`: Process management for X11, VNC, apps
+- `setup/`: Setup tools organized by app
+- `evaluate/`: Evaluation tools organized by app
+
+Remember: When in doubt, make direct HTTP calls and store state in the persistent context!
+

{hud_python-0.4.12 → hud_python-0.4.14}/environments/remote_browser/pyproject.toml

@@ -3,7 +3,7 @@ name = "hud-remote-browser"
 version = "0.1.0"
 description = "HUD Remote Browser Controller with MCP tools for cloud browser providers"
 requires-python = ">=3.11,<3.13"
-dependencies = [ "hud-python==0.4.12", "pyautogui", "playwright", "httpx", "typer", "google-api-python-client", "google-auth",]
+dependencies = [ "hud-python>=0.4.12", "pyautogui", "playwright", "httpx", "typer", "google-api-python-client", "google-auth",]
 
 [build-system]
 requires = [ "hatchling",]

{hud_python-0.4.12 → hud_python-0.4.14}/hud/cli/__init__.py

@@ -348,6 +348,11 @@ def dev(
     ),
     port: int = typer.Option(8765, "--port", "-p", help="HTTP server port (ignored for stdio)"),
     no_reload: bool = typer.Option(False, "--no-reload", help="Disable hot-reload"),
+    full_reload: bool = typer.Option(
+        False,
+        "--full-reload",
+        help="Restart entire container on file changes (instead of just server process)",
+    ),
     verbose: bool = typer.Option(False, "--verbose", "-v", help="Show server logs"),
     inspector: bool = typer.Option(
         False, "--inspector", help="Launch MCP Inspector (HTTP mode only)"
@@ -375,12 +380,13 @@ def dev(
         hud dev . --inspector        # Launch MCP Inspector (HTTP mode only)
         hud dev . --interactive      # Launch interactive testing mode (HTTP mode only)
         hud dev . --no-logs          # Disable Docker log streaming
+        hud dev . --full-reload      # Restart entire container on file changes (instead of just server)
 
         # With Docker arguments (after all options):
         hud dev . -e BROWSER_PROVIDER=anchorbrowser -e ANCHOR_API_KEY=xxx
         hud dev . -e API_KEY=secret -v /tmp/data:/data --network host
         hud dev . --build -e DEBUG=true --memory 2g
-    """
+    """  # noqa: E501
     # Parse directory and Docker arguments
     if params:
         directory = params[0]
@@ -397,6 +403,7 @@ def dev(
         transport,
         port,
         no_reload,
+        full_reload,
         verbose,
         inspector,
         no_logs,

{hud_python-0.4.12 → hud_python-0.4.14}/hud/cli/dev.py

@@ -35,6 +35,7 @@ def create_proxy_server(
     directory: str | Path,
     image_name: str,
     no_reload: bool = False,
+    full_reload: bool = False,
     verbose: bool = False,
     docker_args: list[str] | None = None,
     interactive: bool = False,
@@ -48,8 +49,12 @@ def create_proxy_server(
         design.warning(f"Could not extract CMD from {image_name}, using default")
         original_cmd = ["python", "-m", "hud_controller.server"]
 
-    # Generate container name from image
-    container_name = f"{image_name.replace(':', '-').replace('/', '-')}"
+    # Generate unique container name from image to avoid conflicts between multiple instances
+    import os
+
+    pid = str(os.getpid())[-6:]  # Last 6 digits of process ID for uniqueness
+    base_name = image_name.replace(":", "-").replace("/", "-")
+    container_name = f"{base_name}-{pid}"
 
     # Build the docker run command
     docker_cmd = [
@@ -73,14 +78,20 @@ def create_proxy_server(
     if interactive:
         no_reload = True
 
-    if not no_reload:
-        # Inject our supervisor into the CMD
+    # Validate reload options
+    if no_reload and full_reload:
+        design.warning("Cannot use --full-reload with --no-reload, ignoring --full-reload")
+        full_reload = False
+
+    if not no_reload and not full_reload:
+        # Standard hot-reload: inject supervisor for server restart within container
         modified_cmd = inject_supervisor(original_cmd)
         docker_cmd.extend(["--entrypoint", modified_cmd[0]])
         docker_cmd.append(image_name)
         docker_cmd.extend(modified_cmd[1:])
     else:
-        # No reload - use original CMD
+        # No reload or full reload: use original CMD without supervisor
+        # Note: Full reload logic (container restart) would be implemented here in the future
         docker_cmd.append(image_name)
 
     # Create configuration following MCPConfig schema
@@ -96,9 +107,14 @@ def create_proxy_server(
 
     # Debug output - only if verbose
     if verbose:
-        if not no_reload:
+        if not no_reload and not full_reload:
+            design.info("Mode: Hot-reload (server restart within container)")
             design.info("Watching: /app/src for changes")
+        elif full_reload:
+            design.info("Mode: Full reload (container restart on file changes)")
+            design.info("Note: Full container restart not yet implemented, using no-reload mode")
         else:
+            design.info("Mode: No reload")
             design.info("Container will run without hot-reload")
         design.command_example(f"docker logs -f {container_name}", "View container logs")
 
@@ -127,6 +143,7 @@ async def start_mcp_proxy(
     transport: str,
     port: int,
     no_reload: bool = False,
+    full_reload: bool = False,
     verbose: bool = False,
     inspector: bool = False,
     no_logs: bool = False,
@@ -212,8 +229,12 @@ async def start_mcp_proxy(
         design.error(f"Source directory not found: {src_path}")
         raise click.Abort
 
-    # Extract container name from the proxy configuration
-    container_name = f"{image_name.replace(':', '-').replace('/', '-')}"
+    # Extract container name from the proxy configuration (must match create_proxy_server naming)
+    import os
+
+    pid = str(os.getpid())[-6:]  # Last 6 digits of process ID for uniqueness
+    base_name = image_name.replace(":", "-").replace("/", "-")
+    container_name = f"{base_name}-{pid}"
 
     # Remove any existing container with the same name (silently)
     # Note: The proxy creates containers on-demand when clients connect
@@ -347,6 +368,7 @@ async def start_mcp_proxy(
         # Always show waiting message
         log_design.info("")  # Empty line for spacing
         log_design.progress_message("⏳ Waiting for first client connection to start container...")
+        log_design.info(f"📋 Looking for container: {container_name}")  # noqa: G004
 
         # Keep trying to stream logs - container is created on demand
         has_shown_started = False
@@ -397,7 +419,8 @@ async def start_mcp_proxy(
 
                         # Show all logs with gold formatting like hud debug
                         # Format all logs in gold/dim style like hud debug's stderr
-                        log_design.console.print(
+                        # Use stdout console to avoid stderr redirection when not verbose
+                        log_design._stdout_console.print(
                             f"[rgb(192,150,12)]■[/rgb(192,150,12)] {decoded_line}", highlight=False
                         )
 
@@ -408,16 +431,19 @@ async def start_mcp_proxy(
                 await asyncio.sleep(1)
                 continue  # Loop back to check if container exists
 
-            except Exception:
-                # Some unexpected error
+            except Exception as e:
+                # Some unexpected error - show it so we can debug
+                log_design.warning(f"Failed to stream Docker logs: {e}")  # noqa: G004
                 if verbose:
-                    log_design.warning("Failed to stream logs")
+                    import traceback
+
+                    log_design.warning(f"Traceback: {traceback.format_exc()}")  # noqa: G004
                 await asyncio.sleep(1)
 
     # CRITICAL: Create proxy AFTER all logging setup to prevent it from resetting logging config
     # This is important because FastMCP might initialize loggers during creation
     proxy = create_proxy_server(
-        directory, image_name, no_reload, verbose, docker_args or [], interactive
+        directory, image_name, no_reload, full_reload, verbose, docker_args or [], interactive
     )
 
     # One more attempt to suppress the FastMCP server log
@@ -548,6 +574,7 @@ def run_mcp_dev_server(
     transport: str = "http",
     port: int = 8765,
     no_reload: bool = False,
+    full_reload: bool = False,
     verbose: bool = False,
     inspector: bool = False,
     no_logs: bool = False,
@@ -706,6 +733,7 @@ def run_mcp_dev_server(
                 transport,
                 port,
                 no_reload,
+                full_reload,
                 verbose,
                 inspector,
                 no_logs,

{hud_python-0.4.12 → hud_python-0.4.14}/hud/cli/eval.py

@@ -26,15 +26,6 @@ def build_agent(
     """Create and return the requested agent type."""
 
     # Import agents lazily to avoid dependency issues
-    try:
-        from hud.agents.misc.response_agent import ResponseAgent
-    except ImportError as e:
-        design.error(
-            "Agent dependencies are not installed. "
-            "Please install with: pip install 'hud-python[agent]'"
-        )
-        raise typer.Exit(1) from e
-
     if agent_type == "openai":
         try:
             from hud.agents import OperatorAgent
@@ -45,12 +36,12 @@ def build_agent(
             )
             raise typer.Exit(1) from e
 
-        allowed_tools = allowed_tools or ["openai_computer"]
-
-        return OperatorAgent(
-            allowed_tools=allowed_tools,
-            response_agent=ResponseAgent(),
-        )
+        if allowed_tools:
+            return OperatorAgent(
+                allowed_tools=allowed_tools,
+            )
+        else:
+            return OperatorAgent()
 
     # Fallback Claude agent (Anthropic)
     try:
@@ -63,13 +54,16 @@ def build_agent(
         raise typer.Exit(1) from e
 
     model = model or "claude-sonnet-4-20250514"
-    allowed_tools = allowed_tools or ["anthropic_computer"]
 
-    return ClaudeAgent(
-        model=model,
-        allowed_tools=allowed_tools,
-        response_agent=ResponseAgent(),
-    )
+    if allowed_tools:
+        return ClaudeAgent(
+            model=model,
+            allowed_tools=allowed_tools,
+        )
+    else:
+        return ClaudeAgent(
+            model=model,
+        )
 
 
 async def run_single_task(
@@ -100,8 +94,8 @@ async def run_single_task(
         with open(path) as f:  # noqa: ASYNC230
             json_data = json.load(f)
 
-        # Check if JSON contains a list of tasks
-        if isinstance(json_data, list):
+        # Check if JSON contains multiple tasks (list with more than 1 task)
+        if isinstance(json_data, list) and len(json_data) > 1:
             design.info(f"Found {len(json_data)} tasks in JSON file, running as dataset…")
 
             # Build agent class and config for run_dataset
@@ -118,8 +112,10 @@ async def run_single_task(
                     raise typer.Exit(1) from e
 
                 agent_config: dict[str, Any] = {
-                    "allowed_tools": allowed_tools or ["openai_computer"],
                 }
+                if allowed_tools:
+                    agent_config["allowed_tools"] = allowed_tools
+
             else:
                 try:
                     from hud.agents import ClaudeAgent
@@ -134,8 +130,9 @@ async def run_single_task(
 
                 agent_config = {
                     "model": model or "claude-sonnet-4-20250514",
-                    "allowed_tools": allowed_tools or ["anthropic_computer"],
                 }
+                if allowed_tools:
+                    agent_config["allowed_tools"] = allowed_tools
 
             # Run as dataset with single-task concurrency to maintain debug behavior
             results = await run_dataset(
@@ -146,7 +143,6 @@ async def run_single_task(
                 max_concurrent=1,  # Run sequentially for debug mode
                 metadata={"source": str(path)},
                 max_steps=max_steps,
-                auto_respond=True,
             )
 
             # Display summary
@@ -154,8 +150,15 @@ async def run_single_task(
             design.success(f"Completed {len(results)} tasks: {successful} successful")
             return
 
-        # Single task JSON
-        task = Task(**json_data)
+        # Single task JSON (either direct object or list with 1 task)
+        if isinstance(json_data, list) and len(json_data) == 1:
+            design.info("Found 1 task in JSON file, running as single task…")
+            task = Task(**json_data[0])
+        elif isinstance(json_data, dict):
+            task = Task(**json_data)
+        else:
+            design.error("JSON file must contain a list of tasks when using --full flag")
+            raise typer.Exit(1)
     else:
         # Load from HuggingFace dataset
         try:
@@ -238,8 +241,10 @@ async def run_full_dataset(
             raise typer.Exit(1) from e
 
         agent_config: dict[str, Any] = {
-            "allowed_tools": allowed_tools or ["openai_computer"],
         }
+        if allowed_tools:
+            agent_config["allowed_tools"] = allowed_tools
+
     else:
         try:
             from hud.agents import ClaudeAgent
@@ -254,8 +259,9 @@ async def run_full_dataset(
 
         agent_config = {
             "model": model or "claude-sonnet-4-20250514",
-            "allowed_tools": allowed_tools or ["anthropic_computer"],
         }
+        if allowed_tools:
+            agent_config["allowed_tools"] = allowed_tools
 
     design.info("🚀 Running evaluation…")
     return await run_dataset(
@@ -266,7 +272,6 @@ async def run_full_dataset(
         max_concurrent=max_concurrent,
         metadata={"dataset": source},
         max_steps=max_steps,
-        auto_respond=True,
     )