hud-python 0.4.27__py3-none-any.whl → 0.4.29__py3-none-any.whl

hud/utils/hud_console.py

@@ -16,7 +16,9 @@ Color Palette:
 from __future__ import annotations
 
 import logging
-from typing import Any
+import time
+import traceback
+from typing import TYPE_CHECKING, Any, Literal, Self
 
 import questionary
 import typer
@@ -24,11 +26,14 @@ from rich.console import Console
 from rich.panel import Panel
 from rich.table import Table
 
+if TYPE_CHECKING:
+    from rich.status import Status
 # HUD Brand Colors - Optimized for both light and dark modes
 GOLD = "rgb(192,150,12)"  # #c0960c - Primary brand color
 RED = "rgb(220,50,47)"  # Slightly muted red that works on both backgrounds
 GREEN = "rgb(133,153,0)"  # Slightly muted green that works on both backgrounds
 DIM = "bright_black"  # Grey that's visible on both light and dark backgrounds
+YELLOW = "yellow"
 TEXT = "bright_white"  # Off-white that's readable on dark, not too bright on light
 SECONDARY = "rgb(108,113,196)"  # Muted blue-purple for secondary text
 
@@ -85,7 +90,11 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         console = self._stderr_console if stderr else self._stdout_console
-        console.print(f"[{RED}]❌ {message}[/{RED}]")
+        tb = traceback.format_exc()
+        if "NoneType: None" not in tb:
+            console.print(f"[{RED} not bold]❌ {message}\n{tb}[/{RED} not bold]")
+        else:
+            console.print(f"[{RED} not bold]❌ {message}[/{RED} not bold]")
 
     def warning(self, message: str, stderr: bool = True) -> None:
         """Print a warning message.
@@ -95,7 +104,7 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         console = self._stderr_console if stderr else self._stdout_console
-        console.print(f"[rgb(181,137,0)]⚠️  {message}[/rgb(181,137,0)]")
+        console.print(f"⚠️  [{YELLOW} not bold]{message}[/{YELLOW} not bold]")
 
     def info(self, message: str, stderr: bool = True) -> None:
         """Print an info message.
@@ -105,7 +114,7 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         console = self._stderr_console if stderr else self._stdout_console
-        console.print(f"[{TEXT}]{message}[/{TEXT}]")
+        console.print(f"[{TEXT} not bold]{message}[/{TEXT} not bold]")
 
     def print(self, message: str, stderr: bool = True) -> None:
         """Print a message.
@@ -126,7 +135,9 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         console = self._stderr_console if stderr else self._stdout_console
-        console.print(f"[{DIM}]{label}[/{DIM}] [{TEXT}]{value}[/{TEXT}]")
+        console.print(
+            f"[{DIM} not bold][default]{label}[/default][/{DIM} not bold] [default]{value}[/default]"  # noqa: E501
+        )
 
     def link(self, url: str, stderr: bool = True) -> None:
         """Print an underlined link.
@@ -150,7 +161,7 @@ class HUDConsole:
         console.print(f"[{TEXT}]{json_str}[/{TEXT}]")
 
     def key_value_table(
-        self, data: dict[str, str], show_header: bool = False, stderr: bool = True
+        self, data: dict[str, str | int | float], show_header: bool = False, stderr: bool = True
     ) -> None:
         """Print a key-value table.
 
@@ -164,7 +175,7 @@ class HUDConsole:
         table.add_column("Value")
 
         for key, value in data.items():
-            table.add_row(key, value)
+            table.add_row(key, str(value))
 
         console = self._stderr_console if stderr else self._stdout_console
         console.print(table)
@@ -210,7 +221,7 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         console = self._stderr_console if stderr else self._stdout_console
-        console.print(f"\n[rgb(181,137,0)]💡 Hint: {hint}[/rgb(181,137,0)]")
+        console.print(f"[rgb(181,137,0)]💡 Hint: {hint}[/rgb(181,137,0)]")
 
     def status_item(
         self,
@@ -305,7 +316,7 @@ class HUDConsole:
             if response_json and not details.get("Response"):
                 details["Response JSON"] = str(response_json)
             if details:
-                self.key_value_table(details, show_header=False, stderr=stderr)
+                self.key_value_table(details, show_header=False, stderr=stderr)  # type: ignore
 
         # Structured hints, if available
         hints = getattr(error, "hints", None)
@@ -315,7 +326,7 @@ class HUDConsole:
 
                 render_hints(hints, design=self)
             except Exception as render_error:
-                self.debug(f"Failed to render hints: {render_error}")
+                self.debug_log(f"Failed to render hints: {render_error}")
 
         # Standard support hint
         self.render_support_hint(stderr=stderr)
@@ -336,9 +347,35 @@ class HUDConsole:
         else:
             self._logger.setLevel(logging.WARNING)
 
+    @property
+    def prefix(self) -> str:
+        """Get the metadata of the current file."""
+        metadata = self._logger.findCaller(stacklevel=3)
+        return f"{metadata[0]}:{metadata[1]} in {metadata[2]} | "
+
     # Logging-aware methods that check logging levels before printing
+    def log(
+        self,
+        message: str,
+        level: Literal["info", "debug", "warning", "error"] = "info",
+        stderr: bool = True,
+    ) -> None:
+        """Print a message based on the logging level."""
+        prefix = self.prefix
+        if level == "info":
+            self.info_log(f"{prefix}{message}", stderr=stderr)
+        elif level == "debug":
+            self.debug_log(f"{prefix}{message}", stderr=stderr)
+        elif level == "warning":
+            self.warning_log(f"{prefix}{message}", stderr=stderr)
+        elif level == "error":
+            self.error_log(f"{prefix}{message}", stderr=stderr)
 
     def debug(self, message: str, stderr: bool = True) -> None:
+        """Print a debug message."""
+        self.debug_log(message, stderr=stderr)
+
+    def debug_log(self, message: str, stderr: bool = True) -> None:
         """Print a debug message only if DEBUG logging is enabled.
 
         Args:
@@ -346,7 +383,7 @@ class HUDConsole:
             stderr: If True, output to stderr (default), otherwise stdout
         """
         if self._logger.isEnabledFor(logging.DEBUG):
-            self.dim_info("[DEBUG]", message, stderr=stderr)
+            self.dim_info(message, "", stderr=stderr)
 
     def info_log(self, message: str, stderr: bool = True) -> None:
         """Print an info message only if INFO logging is enabled.
@@ -368,6 +405,25 @@ class HUDConsole:
         if self._logger.isEnabledFor(logging.INFO):
             self.progress_message(message, stderr=stderr)
 
+    def progress(self, initial: str = "", stderr: bool = True) -> _ProgressContext:
+        """Create a progress context manager for inline updates.
+
+        Args:
+            initial: Initial message to display
+            stderr: If True, output to stderr (default), otherwise stdout
+
+        Returns:
+            A context manager that provides update() method
+
+        Example:
+            with console.progress("Processing...") as progress:
+                for i in range(10):
+                    progress.update(f"Processing item {i+1}/10")
+        """
+        return _ProgressContext(
+            console=self._stderr_console if stderr else self._stdout_console, initial=initial
+        )
+
     def success_log(self, message: str, stderr: bool = True) -> None:
         """Print a success message only if INFO logging is enabled.
 
@@ -402,7 +458,7 @@ class HUDConsole:
         self,
         message: str,
         choices: list[str | dict[str, Any]] | list[str],
-        default: str | None = None,
+        default: int | None = None,
     ) -> str:
         """Interactive selection with arrow key navigation.
 
@@ -428,6 +484,7 @@ class HUDConsole:
         result = questionary.select(
             message,
             choices=q_choices,
+            default=q_choices[default] if default is not None else None,
             instruction="(Use ↑/↓ arrows, Enter to select)",
         ).ask()
 
@@ -481,6 +538,49 @@ class HUDConsole:
         else:
             return f"  [{GREEN}]✓[/{GREEN}] [{TEXT}]{content}[/{TEXT}]"
 
+    def confirm(self, message: str, default: bool = True) -> bool:
+        """Print a confirmation message.
+
+        Args:
+            message: The confirmation message
+            default: If True, the default choice is True
+        """
+        return questionary.confirm(message, default=default).ask()
+
+
+# Global design instance for convenience
+class _ProgressContext:
+    """Context manager for inline progress updates."""
+
+    def __init__(self, console: Console, initial: str = "") -> None:
+        self.console = console
+        self.initial = initial
+        self.status: Status | None = None
+        self.start_time: float | None = None
+
+    def __enter__(self) -> Self:
+        self.status = self.console.status(self.initial)
+        self.status.__enter__()
+        self.start_time = time.time()
+        return self
+
+    def __exit__(self, exc_type: object, exc_val: object, exc_tb: object) -> None:
+        if self.status:
+            self.status.__exit__(exc_type, exc_val, exc_tb)  # type: ignore
+
+    def update(self, message: str, with_elapsed: bool = True) -> None:
+        """Update the progress message.
+
+        Args:
+            message: New message to display
+            with_elapsed: If True, append elapsed time to message
+        """
+        if self.status:
+            if with_elapsed and self.start_time:
+                elapsed = time.time() - self.start_time
+                self.status.update(f"{message} [{elapsed:.1f}s]")
+            else:
+                self.status.update(message)
+
 
-# Global hud_console instance for convenience
 hud_console = HUDConsole()

hud/utils/tasks.py

@@ -0,0 +1,119 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from hud.types import Task
+from hud.utils.hud_console import HUDConsole
+
+hud_console = HUDConsole()
+
+
+def load_tasks(tasks_input: str | list[dict]) -> list[Task]:
+    """Load tasks from various sources.
+
+    Args:
+        tasks_input: Either:
+            - Path to a JSON file (array of tasks)
+            - Path to a JSONL file (one task per line)
+            - HuggingFace dataset name (format: "username/dataset" or "username/dataset:split")
+            - List of task dictionaries
+        system_prompt: Default system prompt to use if not specified in task
+
+    Returns:
+        List of validated HUD Task objects
+    """
+    tasks = []
+
+    if isinstance(tasks_input, list):
+        # Direct list of task dicts
+        hud_console.info(f"Loading {len(tasks_input)} tasks from provided list")
+        for item in tasks_input:
+            task = Task(**item)
+            tasks.append(task)
+
+    elif isinstance(tasks_input, str):
+        # Check if it's a file path
+        if Path(tasks_input).exists():
+            file_path = Path(tasks_input)
+            hud_console.info(f"Loading tasks from file: {tasks_input}")
+
+            with open(file_path) as f:
+                # Handle JSON files (array of tasks)
+                if file_path.suffix.lower() == ".json":
+                    data = json.load(f)
+                    if not isinstance(data, list):
+                        raise ValueError(
+                            f"JSON file must contain an array of tasks, got {type(data)}"
+                        )
+
+                    for item in data:
+                        task = Task(**item)
+                        tasks.append(task)
+
+                # Handle JSONL files (one task per line)
+                else:
+                    for line in f:
+                        line = line.strip()
+                        if line:  # Skip empty lines
+                            item = json.loads(line)
+
+                            # Handle case where line contains an array of tasks
+                            if isinstance(item, list):
+                                for task_item in item:
+                                    task = Task(**task_item)
+                                    tasks.append(task)
+                            # Handle normal case where line contains a single task object
+                            elif isinstance(item, dict):
+                                task = Task(**item)
+                                tasks.append(task)
+                            else:
+                                raise ValueError(
+                                    f"Invalid JSONL format: expected dict or list of dicts, got {type(item)}"  # noqa: E501
+                                )
+
+        # Check if it's a HuggingFace dataset
+        elif "/" in tasks_input:
+            hud_console.info(f"Loading tasks from HuggingFace dataset: {tasks_input}")
+            try:
+                from datasets import load_dataset
+
+                # Parse dataset name and optional split
+                if ":" in tasks_input:
+                    dataset_name, split = tasks_input.split(":", 1)
+                else:
+                    dataset_name = tasks_input
+                    split = "train"  # Default split
+
+                dataset = load_dataset(dataset_name, split=split)
+
+                # Convert dataset rows to Task objects
+                for item in dataset:
+                    if not isinstance(item, dict):
+                        raise ValueError(
+                            f"Invalid HuggingFace dataset: expected dict, got {type(item)}"
+                        )
+                    if not item["mcp_config"] or not item["prompt"]:
+                        raise ValueError(
+                            f"Invalid HuggingFace dataset: expected mcp_config and prompt, got {item}"  # noqa: E501
+                        )
+                    task = Task(**item)
+                    tasks.append(task)
+
+            except ImportError as e:
+                raise ImportError(
+                    "Please install 'datasets' to load from HuggingFace: uv pip install datasets"
+                ) from e
+            except Exception as e:
+                raise ValueError(f"Failed to load HuggingFace dataset '{tasks_input}': {e}") from e
+
+        else:
+            raise ValueError(
+                f"Invalid tasks input: '{tasks_input}' is neither a file path nor a HuggingFace dataset"  # noqa: E501
+            )
+
+    else:
+        raise TypeError(f"tasks_input must be str or list, got {type(tasks_input)}")
+
+    hud_console.info(f"Loaded {len(tasks)} tasks")
+    return tasks

hud/utils/tests/test_version.py

@@ -5,4 +5,4 @@ def test_import():
     """Test that the package can be imported."""
     import hud
 
-    assert hud.__version__ == "0.4.27"
+    assert hud.__version__ == "0.4.29"

hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.27"
+__version__ = "0.4.29"

{hud_python-0.4.27.dist-info → hud_python-0.4.29.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.27
+Version: 0.4.29
 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
@@ -34,7 +34,7 @@ Classifier: Programming Language :: Python :: 3
 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-Python: <3.13,>=3.11
 Requires-Dist: httpx<1,>=0.23.0
 Requires-Dist: hud-fastmcp-python-sdk>=0.1.2
 Requires-Dist: hud-mcp-python-sdk>=3.13.2
@@ -107,6 +107,24 @@ Requires-Dist: pytest<9,>=8.1.1; extra == 'dev'
 Requires-Dist: ruff>=0.11.8; extra == 'dev'
 Requires-Dist: setuptools; extra == 'dev'
 Requires-Dist: textdistance<5,>=4.5.0; extra == 'dev'
+Provides-Extra: rl
+Requires-Dist: anthropic; extra == 'rl'
+Requires-Dist: bitsandbytes>=0.41.0; (sys_platform == 'linux') and extra == 'rl'
+Requires-Dist: datasets>=2.14.0; extra == 'rl'
+Requires-Dist: dotenv>=0.9.9; extra == 'rl'
+Requires-Dist: ipykernel; extra == 'rl'
+Requires-Dist: ipython<9; extra == 'rl'
+Requires-Dist: jupyter-client; extra == 'rl'
+Requires-Dist: jupyter-core; extra == 'rl'
+Requires-Dist: langchain; extra == 'rl'
+Requires-Dist: langchain-anthropic; extra == 'rl'
+Requires-Dist: langchain-openai; extra == 'rl'
+Requires-Dist: liger-kernel>=0.5.0; (sys_platform == 'linux') and extra == 'rl'
+Requires-Dist: numpy>=1.24.0; extra == 'rl'
+Requires-Dist: openai; extra == 'rl'
+Requires-Dist: peft>=0.17.1; extra == 'rl'
+Requires-Dist: pillow>=11.1.0; extra == 'rl'
+Requires-Dist: vllm==0.10.1.1; extra == 'rl'
 Description-Content-Type: text/markdown
 
 <div align="left">

{hud_python-0.4.27.dist-info → hud_python-0.4.29.dist-info}/RECORD

@@ -1,47 +1,51 @@
-hud/__init__.py,sha256=KU7G-_Mj6Mjf7trXA6X0ufN6QUmqhVi19NKbnNIvD74,532
+hud/__init__.py,sha256=JMDFUE1pP0J1Xl_miBdt7ERvoffZmTzSFe8yxz512A8,552
 hud/__main__.py,sha256=YR8Dq8OhINOsVfQ55PmRXXg4fEK84Rt_-rMtJ5rvhWo,145
-hud/settings.py,sha256=bgq_zstNGlan7UDZOiElY8Aw5ZU4xL7Ds5HP_Xzph1A,2535
-hud/types.py,sha256=jta4Hgj_rBdIMcf7mz0tsdgfA1pssp327bxcT6Mfp00,6107
-hud/version.py,sha256=-lY50vN_2vhScYD-TJkkkeLQohONEsT3WAAxfldqpVg,105
+hud/settings.py,sha256=sMS31iW1m-5VpWk-Blhi5-obLcUA0fwxWE1GgJz-vqU,2708
+hud/types.py,sha256=Cn9suZ_ZitLnxmnknfbCYVvmLsXRWI56kJ3LXtdfI6M,10157
+hud/version.py,sha256=I65QTO8NMhRiDAN0vYH7G47SRwt713WHRXjHgx1VsL0,105
 hud/agents/__init__.py,sha256=UoIkljWdbq4bM0LD-mSaw6w826EqdEjOk7r6glNYwYQ,286
-hud/agents/base.py,sha256=hUCFJZV3L69YFqMd2Room4WZmdmr6rxjiUvhf9cVXys,33811
-hud/agents/claude.py,sha256=yTVbSNMwNjF9qZAnwDSsgrscaenJpMGnGdW_4rFW_pY,15603
+hud/agents/base.py,sha256=y3Aq3DtsDT5X2Q-_C-jqxFtDcHUlnBDOczmZIhqC0-w,35118
+hud/agents/claude.py,sha256=wHiw8iAnjnRmZyKRKcOhagCDQMhz9Z6rlSBWqH1X--M,15781
 hud/agents/grounded_openai.py,sha256=U-FHjB2Nh1_o0gmlxY5F17lWJ3oHsNRIB2a7z-IKB64,11231
 hud/agents/langchain.py,sha256=1EgCy8jfjunsWxlPC5XfvfLS6_XZVrIF1ZjtHcrvhYw,9584
 hud/agents/openai.py,sha256=ovARRWNuHqKkZ2Q_OCYSVCIZckrh8XY2jUB2p2x1m88,14259
-hud/agents/openai_chat_generic.py,sha256=BDtMQnf9ddYN4fKy-CN9IotLyYvHGdZqjju5PODEnU4,10683
+hud/agents/openai_chat_generic.py,sha256=T2HTZaSUmvqBhYIfurtreb-gxnlVBqDO64GySu_uFIA,11514
 hud/agents/misc/__init__.py,sha256=BYi4Ytp9b_vycpZFXnr5Oyw6ncKLNNGml8Jrb7bWUb4,136
 hud/agents/misc/response_agent.py,sha256=OJdQJ76jP9xxQxVYJ-qPcdBxvFr8ABcwbP1f1I5zU5A,3227
 hud/agents/tests/__init__.py,sha256=W-O-_4i34d9TTyEHV-O_q1Ai1gLhzwDaaPo02_TWQIY,34
-hud/agents/tests/test_base.py,sha256=F39ajSqASGUbPyPoWSY9KARFav62qNTK74W11Tr1Tg4,28970
+hud/agents/tests/test_base.py,sha256=bDznxQDv2ickRkw98joH9zfuZT6ItHbmWvQ67iboa4g,28733
 hud/agents/tests/test_claude.py,sha256=wqEKlzEvx8obz1sSm4NY0j-Zyt1qWNfDOmRqYIuAEd0,13069
 hud/agents/tests/test_client.py,sha256=uikgh6yhjPPX2RBU4XJQMz1mNox9uXjuwsP8t93id18,13337
 hud/agents/tests/test_grounded_openai_agent.py,sha256=VK8lUvHIjWicMX00VKPE-FZyjiJqTEhb80MuRRa9fVc,5437
-hud/agents/tests/test_openai.py,sha256=lhHowQyLp09oDVwBUjUECoPeH01F16i4O9YIHeugK6E,7028
-hud/cli/__init__.py,sha256=-32_tNDsFc38AMAvJFcHg38EVS5PN3n6pjW7eN2Nxj4,36259
+hud/agents/tests/test_openai.py,sha256=1S5IZuc3O3moSp70gqVGjc6m-_b49dCfz2fgX5IGvl4,7036
+hud/cli/__init__.py,sha256=aAqoGtdoSP8omcmKqiLp8OhxHr4nCXGupGUwhNAOEQM,40119
 hud/cli/__main__.py,sha256=fDH7XITyuDITwSDIVwRso06aouADO0CzTHKqp5TOwJE,143
 hud/cli/analyze.py,sha256=4u5oYfJMquOjT9PzzRTYVcTZDxDi0ilNP_g532_hpOU,14716
 hud/cli/build.py,sha256=SsZLQKYyMvIikLaSPYmg7-S1iJn-KkWXURhniQ0Ifi8,18015
 hud/cli/clone.py,sha256=AwVDIuhr8mHb1oT2Af2HrD25SiTdwATpE6zd93vzLgA,6099
 hud/cli/debug.py,sha256=jtFW8J5F_3rhq1Hf1_SkJ7aLS3wjnyIs_LsC8k5cnzc,14200
-hud/cli/dev.py,sha256=-T7s3aiEq0X_lA6xQHOP4evvgP2wO76OYdsL7eeVmJ8,31740
-hud/cli/eval.py,sha256=1Kifhr3tm2Md3r_lC1P9ET93P_eEB3is3MFu_Hg9GH4,15947
-hud/cli/hf.py,sha256=wbHNXEI5K17cA9micfeVG4KRbUiI3MymLKI7TO-tW5E,15446
+hud/cli/dev.py,sha256=56vQdH9oe_XGnOcRcFbNIsLEoBnpCl1eANlRFUeddHQ,31734
+hud/cli/eval.py,sha256=W_eY4uoIQwHcSCvxNaQeRfWC10uQA1UhBWiNQzQPuXM,22694
+hud/cli/get.py,sha256=sksKrdzBGZa7ZuSoQkc0haj-CvOGVSSikoVXeaUd3N4,6274
 hud/cli/init.py,sha256=XswZB2ZnzdE0pxP1kRmO3bHfWGCrKAyrME34ZyPzs98,19715
 hud/cli/list_func.py,sha256=EVi2Vc3Lb3glBNJxFx4MPnZknZ4xmuJz1OFg_dc8a_E,7177
 hud/cli/pull.py,sha256=Vd1l1-IwskyACzhtC8Df1SYINUZEYmFxrLl0s9cNN6c,12151
 hud/cli/push.py,sha256=JXUxu1QGU7BPWb0erSJq42CIq0sLbaDAO42yYDcvA1g,18347
 hud/cli/remove.py,sha256=8vGQyXDqgtjz85_vtusoIG8zurH4RHz6z8UMevQRYM4,6861
-hud/cli/rl/README.md,sha256=3pqRZMrnwD-lJwWGCCNZNhGdZG6zyydLBOer0e8BkLw,5983
-hud/cli/rl/__init__.py,sha256=PISQwpkMX9JQYMBK5j8ziP8vbLit6Sh1b7a94MgogBU,3424
-hud/cli/rl/init.py,sha256=xm2hQisL1VVOnKu7cjOuEadwMnqHQjlr5ekbzoAd6dA,13996
-hud/cli/rl/pod.py,sha256=601aiVhUI7ihempzBW-KFqbuU4RFJT0PZAzE5OVFEX4,19411
-hud/cli/rl/ssh.py,sha256=cUsw_tUuuSJlYaQJSzdDQeixGoVwx3GvKYZIPkQOp1U,11755
-hud/cli/rl/train.py,sha256=YyUSVjRtq1Sqg7zvKMdUupbE9eQ5unZOOldw4JLAv8M,20284
-hud/cli/rl/utils.py,sha256=V28ihGb02bwYEYnvLJOYVMQSIiEGM1KjDJBr8Tkbn5I,4888
+hud/cli/flows/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+hud/cli/flows/tasks.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+hud/cli/rl/__init__.py,sha256=HRIWMx5Dmmu2IorrmcyVu-Bs_3B45pwEsCC67M1z3Oo,22670
+hud/cli/rl/config.py,sha256=iNhCxotM33OEiP9gqPvn8A_AxrBVe6fcFCQTvc13xzA,2884
+hud/cli/rl/display.py,sha256=OTFcoIn5k3ehfi5-PmuClIkY04f4tpSMuNf7yd-bBr4,5212
+hud/cli/rl/gpu.py,sha256=peXS-NdUF5RyuSs0aZoCzGLboneBUpCy8f9f99WMrG0,2009
+hud/cli/rl/gpu_utils.py,sha256=H5ckPwgj5EVP3yJ5eVihR5R7Y6Gp6pt8ZUfWCCwcLG4,11072
+hud/cli/rl/presets.py,sha256=DzOO82xL5QyzdVtlX-Do1CODMvDz9ILMPapjU92jcZg,3051
+hud/cli/rl/remote_runner.py,sha256=ILjWpjbdZyH7db-C-x1T0GwsrLu3PBKx3IoTZl-qamg,12733
+hud/cli/rl/rl_api.py,sha256=INJobvSa50ccR037u_GPsDa_9WboWyNwqEaoh9hcXj0,4306
+hud/cli/rl/vllm.py,sha256=Gq_M6KsQArGz7FNIdemuM5mk16mu3xe8abpO2GCCuOE,6093
 hud/cli/tests/__init__.py,sha256=ZrGVkmH7DHXGqOvjOSNGZeMYaFIRB2K8c6hwr8FPJ-8,68
 hud/cli/tests/test_analyze.py,sha256=inbRvi7KJKoMYrcqXU6RSayoh7mAOGVrRknm6BLQFes,11055
-hud/cli/tests/test_analyze_metadata.py,sha256=PqeovqpPIitu3YkQ7LvzD9potNxoU_KvDXH4_-lvYFs,9992
+hud/cli/tests/test_analyze_metadata.py,sha256=RtJ5PiOWu-AyOijLyZZwNfYazqwSMvtDS0krMMw0mak,9943
 hud/cli/tests/test_build.py,sha256=9B_-jkCDMsNNn087SWbPEJVsbD_D1u5Mv1gBIySPuhU,14409
 hud/cli/tests/test_cli_init.py,sha256=_H0bAn5_skJ91Zj8P5P_wtZoPWvrN7jMhPZvmnnf0n8,11289
 hud/cli/tests/test_cli_main.py,sha256=0wMho9p9NcGjp0jLiUtCQh_FYdbMaCJtSY3sBbSgPwA,697
@@ -66,42 +70,56 @@ hud/cli/utils/registry.py,sha256=p6IaWmhUbf0Yh6aGa3jIPoSFT2uJPTOVBLS0jpKDUqc,437
 hud/cli/utils/remote_runner.py,sha256=50Fqp7LG9dwiDyNM-LspNiKrSHmEo00C_Iuu5jg9mXU,9407
 hud/cli/utils/runner.py,sha256=7HxVGa6OTflQnO0FSuRs273wnAmbm7cFRU9RTGL3-Wo,4390
 hud/cli/utils/server.py,sha256=EE5DJ0RAmXCEjMcZycpAsAxxCj6sOdIsXqPh38kK2ew,7416
+hud/cli/utils/tasks.py,sha256=bpH2mQkfgKUpgh1J0NtVxMxcM1jDZk2GAAPQMcqAUz4,826
 hud/clients/README.md,sha256=XNE3mch95ozDgVqfwCGcrhlHY9CwT1GKfNANNboowto,3826
 hud/clients/__init__.py,sha256=N5M_gZv4nP7dLRwpAiaqqaxyaLieGW6397FszeG7JGw,364
-hud/clients/base.py,sha256=vt9mRMZwZg2DtVxQmccR_VZZGamXhx3dvlFJPulbOd8,14131
+hud/clients/base.py,sha256=Q4iJ78wxDYj_RTnj000ILetvhxZpGX2mubcfGzp3JYw,14207
 hud/clients/fastmcp.py,sha256=KJGi8bmds0Q6rHnkTXb_Hw9ZqWmSo0OfjW05SSuyEJU,9182
-hud/clients/mcp_use.py,sha256=TsIIItqmt93hcdW1hj2HlMPJppVo4JDoUuCqQ-ZBoBI,13521
+hud/clients/mcp_use.py,sha256=DgQ5q868jutH5aiDu3NVNePPqlOI_0iTWKmIuqCEJQ0,13559
 hud/clients/tests/__init__.py,sha256=sKOtJFFa4mDIXh1U6O8ZUHjigE8CiRMQ2PzJTIBZuVE,33
 hud/clients/tests/test_client_integration.py,sha256=kohU6jfCNfwSnAushHeB1_CmDlRfQc7VBL0GEdJYSeI,4198
 hud/clients/tests/test_fastmcp.py,sha256=4q3TzDjuieTZa89taiNJIrzbUncNkYOG4MaubypA21k,13030
-hud/clients/tests/test_mcp_use_retry.py,sha256=v9vwKa9nrW0EYiWzkq_vpyju40NaOm3SBiJh8VJYX2E,13103
+hud/clients/tests/test_mcp_use_retry.py,sha256=9FxLAz4L5Vv3OTtj4wdhRY23wDYALUpE12TYWl7fbJA,13299
 hud/clients/tests/test_protocol.py,sha256=aK4CS4g3j1D5jPo83ykzZuHUvcZFAulYtIq9T9Hb_fQ,6640
 hud/clients/utils/__init__.py,sha256=-zZjcKIWGj2tXbVDOW45UgoGghhLJzFQVZ6miKenuA4,595
 hud/clients/utils/mcp_use_retry.py,sha256=sBCjtgnAXiXASjzFF_AtBEtmizay0Fi0nPL6sVoooeI,6675
 hud/clients/utils/retry.py,sha256=mMs2T_mAlb8AYhSqMR4AmCw7838gqCC4mdG3zjMAYM4,5744
 hud/clients/utils/retry_transport.py,sha256=Rsq25eiKKt_pM1bas78QEZvO0illK97X_3opmaS3A3w,6809
-hud/datasets/__init__.py,sha256=74T4mrjELKtE04XkZKwU8QAJcg2wjqXLqRO9s4GlPr4,678
-hud/datasets/task.py,sha256=bDVLy4EOBfjiU4i8hrFqcQ3dc077vzVNRXIbyNXFnp8,3916
-hud/datasets/utils.py,sha256=3hKvZTkZuCRkTeITB86nNdA1dtHZAqFfAdSPMtcTUhs,4275
-hud/datasets/execution/__init__.py,sha256=4m1AEpMQaUSJFVN_iAXvY6zFttVgZKwE6oQtC0Rrk7U,330
-hud/datasets/execution/parallel.py,sha256=ZEMMmH018QXOL1QoD_AyspQkGs_41F-GB_mTgdXh6J4,25780
-hud/datasets/execution/runner.py,sha256=EEvb90vvAqFXXx8NyVKLfK5p-gtsfJqiFJAoqSjyfXg,4695
+hud/datasets/__init__.py,sha256=-g05iDy76CU4JiRHjKBBhgh3STtiIjmWhUfPqgf5hJE,697
+hud/datasets/parallel.py,sha256=m7_z2QwjaRuM9gJFYyiPIJUwrlTxZSvFMAd9L2IDZEo,25772
+hud/datasets/runner.py,sha256=2KhGEDzYW_qrSCaNJmsKqiAYZE_-h5VaQ7kv8rSe7Fw,4687
+hud/datasets/utils.py,sha256=hdZfjWH5l3FVJaWBSHEEpjujAG7DqEam_vHgslL8MLs,4279
 hud/misc/__init__.py,sha256=m_pprQQ-G-Y0Sd0NEiR8MtAMbElnuFZ2OWT8TXrw7c4,43
 hud/misc/claude_plays_pokemon.py,sha256=IthAkjDVr2Q-GNvX-QLJyMzN7-0pHqqJbagGNv2m7yo,10453
 hud/native/__init__.py,sha256=TqM0KaiQnDb2Nv1zOgpEMiLVq8JPd4j_aaK4rUZ0IiA,232
 hud/native/comparator.py,sha256=A16wFGINLHPyuD23e8sEYxhRwWCUBYzYLb6TpvzJG9c,18297
 hud/native/tests/__init__.py,sha256=gBTLMm6w5f6D-02Se2WleYsEEYyFt95JDcFzp3C2L_k,40
 hud/native/tests/test_comparator.py,sha256=x1gFLXEDRIiJhH8tg5Rd3ptY-modYaHgSm6-hCJ1EdY,18568
-hud/native/tests/test_native_init.py,sha256=Is8fcDZimp1Oi2Bv4zavqM3KrpS86_DUXFnqc0AsCH0,2736
+hud/native/tests/test_native_init.py,sha256=00zSF6jBLIYYbDaslI7fk-ANTEofIC9SPrRtX942WM4,2738
 hud/otel/__init__.py,sha256=ii17ayoWiS5vAhA7UAmZ8TkmP52gs2pWyHsD46-uYbE,1003
 hud/otel/collector.py,sha256=jLZymZ8r7xt2VDuWexfbnT7PY1-0aiyLMgjBy8KDY1M,4497
-hud/otel/config.py,sha256=BPEyINdlxjsTj-IOO7F_JXAqrbMEWNMuhi2GDGqB5Kk,6785
+hud/otel/config.py,sha256=mricuAmtFd1yIfOYKw2aHI-u4piku0GXHWv6hjsWQLM,6806
 hud/otel/context.py,sha256=C9MvO99cRSNNDEDC7ehO3eoTPnb6J7AemUYvEp57yEU,17774
 hud/otel/exporters.py,sha256=RLAjWa8b2DJEU21740Idq4fmeIuabLEqGGUspcFDcH4,14331
-hud/otel/instrumentation.py,sha256=bBWxQ5vkiP-2WZ_2ztR5LAw0Wu02SZNKUgs5f-rU-ro,3734
+hud/otel/instrumentation.py,sha256=fsFG9W89RdewFDxWKN9Ft4GUb7WbIKpfucTc16WxaZU,5093
 hud/otel/processors.py,sha256=-gGRbwifplcExDQBLfx_9tqWreDImULJNcENgO9q7VU,4700
 hud/otel/tests/__init__.py,sha256=VNJKBMaxTtbn7trW-1Ph50zCvCok_wTSGcI1HD6GOLA,43
 hud/otel/tests/test_processors.py,sha256=np0R4ssd9j6LJSJykJ5bNjl0POwNYNhgb7BqOZHwcMY,6778
+hud/rl/README.md,sha256=OLOmRGOWuQGE-xjLqQJbazzX8ygUG17ECP6kjbM2C0g,1163
+hud/rl/__init__.py,sha256=yYL7U1WV6L3mr3Hig48-4lhnryTaWj4nCXm4lG5vrYI,25
+hud/rl/actor.py,sha256=0YChXyxCz1wVBQ9lKb7vSl64_HQ24-DmYqCCxuORzJc,6747
+hud/rl/buffer.py,sha256=xz4FlvO9l945VsSS4lzRFMwH3rA9HafgbUfADSauXok,15210
+hud/rl/chat_template.jinja,sha256=XTdzI8oFGEcSA-exKxyHaprwRDmX5Am1KEb0VxvUc6U,4965
+hud/rl/config.py,sha256=PAKYPCsKl8yg_j3gJSE5SJUgLM7j0lFy0K_Vt4-otDM,5384
+hud/rl/distributed.py,sha256=8avhrb0lHYkhW22Z7MfkqSnlczWj5jMrUMEtkcoCf74,2473
+hud/rl/learner.py,sha256=JBlzPFgX16uk6f6xYINLXWdDD1i1tiyzm6GLryrzgYg,24545
+hud/rl/train.py,sha256=oZQGo0Wvb2LSrhh-7FLOsGCvI4G4AjgAAvF9P0k9l1Q,13436
+hud/rl/types.py,sha256=lrLKo7iaqodYth2EyeuOQfLiuzXfYM2eJjPmpObrD7c,3965
+hud/rl/utils.py,sha256=IsgVUUibxnUzb32a4mu1sYrgJC1CwoG9E-Dd5y5VDOA,19115
+hud/rl/vllm_adapter.py,sha256=TBNo5lyNzszg6ATk9JoEZAm-xk_tcUJmq9YXwF1NB5w,3961
+hud/rl/tests/__init__.py,sha256=PXmD3Gs6xOAwaYKb4HnwZERDjX05N1QF-aU6ya0dBtE,27
+hud/rl/tests/test_learner.py,sha256=qfSHFFROteRb98TjBuAKjFmZjCGfuWXPysVvTAWJ7wQ,6025
+hud/rl/utils/start_vllm_server.sh,sha256=ThPokrLK_Qm_uh916fHXXBfMlw1TC97P57-AEI5MuOc,910
 hud/server/__init__.py,sha256=8LUwgsXO8xiViWP7uImDwcOsWLu01r5F4r8U8qH3rSY,91
 hud/server/context.py,sha256=6bCdSzv1FGyItu9472HbbYef279H7QuMGJDR8EtYg5Y,3210
 hud/server/low_level.py,sha256=XYs2pOJ9kN4OcJ6ahDmXM5mWkzq5wJLpKFInUYrWEok,4701
@@ -115,14 +133,14 @@ hud/shared/requests.py,sha256=HWrPp7nBSK4jhv9wqZdFiNrVaaxV0vWS8fcgGtoztBc,9479
 hud/shared/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hud/shared/tests/test_exceptions.py,sha256=bCKGqw2RwBmlGIMNpddzyfnWAOQwSIYB76r6iaBE2is,15761
 hud/shared/tests/test_requests.py,sha256=nKFcSN1sjrOouVU2xik9lE5Wxapy3EWsO8iIXrM_Sts,9114
-hud/telemetry/__init__.py,sha256=pPqkq4XekqJEkzwoCwGHGAv1NcytZ55yIB1aLxBpo6U,663
+hud/telemetry/__init__.py,sha256=uWiloBMXgEzPRsRIOpiSBhcTxJDyHfBqTg7qi8kxSTc,683
 hud/telemetry/instrument.py,sha256=m3u6YK02PTk39Jr4L3se7l-cYyKx0maCaqf5Z5JqWNA,14096
-hud/telemetry/job.py,sha256=DSZU_yVpag4VOGkCC5egcjVYuZ02nuzMqFZCgJqcQvM,9998
+hud/telemetry/job.py,sha256=0c4Z69uNDwMeY_c_JpyuBvwvYTYk1YIkwotbXo3TAS4,11593
 hud/telemetry/replay.py,sha256=YW17s314s5Wy6Rl8MXHqg1FU8EF9_XcHBMJI0rrkyS4,2306
-hud/telemetry/trace.py,sha256=gem9pcioNI71hLs18vNidv-7KAZBJAn35stjVvOu_Ic,2208
+hud/telemetry/trace.py,sha256=N2b_kc1JQKqxGb0mQjJ2HQrAJR94_Ai-1UCIs3LdANI,4671
 hud/telemetry/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hud/telemetry/tests/test_replay.py,sha256=eREc6qgSJDRT1pOPdyhiEoEJ9H2yT1ospaU1RvTKlvg,1328
-hud/telemetry/tests/test_trace.py,sha256=uKPjVqJ7kSWhiLXxMVQjQ3PcuKM-ryNs38LNaJanByg,2477
+hud/telemetry/tests/test_trace.py,sha256=0rxR77CjcStat3ILA9QAswieOJ3J_386QmjmNDp34oA,2486
 hud/tools/__init__.py,sha256=i6lE0GxYcPnlLLd-55ryCCHo7o9anC4RfqkuYFXvzMQ,1009
 hud/tools/base.py,sha256=4qm5LS3SAkrq_lyfToWYCN9tNvTHohKJNH2siHkE364,15824
 hud/tools/bash.py,sha256=LJViMGb3lTGBm_gequVVTM7ySh1Xh9bOOIZXU29Lmrw,5209
@@ -134,7 +152,7 @@ hud/tools/types.py,sha256=g-CWnUUDSxxIfUy54S1bpY1nfTzdYO1R_nPKYReABjQ,2734
 hud/tools/utils.py,sha256=bfVyYMcBOJvr1QdptCjVb6jaHVGIL5WUxmY59kzMekQ,1447
 hud/tools/computer/__init__.py,sha256=3tQBXPtGX0WPCwFXzEs3-duwg0rPEgj_0-K7aHskeQE,367
 hud/tools/computer/anthropic.py,sha256=oJfNMnjNFAn9mW1xY1nbWnTY2IqwFqdDR0mWSf8lu-s,17352
-hud/tools/computer/hud.py,sha256=8oA-ofitdUxOfP_3QVdYShMyYLEVS8tdz7KSYcmAmbw,16455
+hud/tools/computer/hud.py,sha256=v2VdLMsc8-3J4_k2vpcZRr_RJzXsxdVNs6IfbCL-WTc,16466
 hud/tools/computer/openai.py,sha256=QEsF45LWOHftDrAoIOnCFZZT1cL--s-ArSov5aluWb8,11189
 hud/tools/computer/settings.py,sha256=b1XJsEQjB9qhN1xHfVENATkzinEebe0ZPyLzMgCGkKY,2763
 hud/tools/executors/__init__.py,sha256=jHxfus9SLhkL6YGtebR5RyKYyVAix3yu5EkUp2Q27Kg,732
@@ -154,8 +172,8 @@ hud/tools/tests/__init__.py,sha256=eEYYkxX5Hz9woXVOBJ2H2_CQoEih0vH6nRt3sH2Z8v8,4
 hud/tools/tests/test_base.py,sha256=m6EelJ47F_hMqvRjrr6vEdiW1AtLgz3ZH1V1IUzTxzI,8983
 hud/tools/tests/test_bash.py,sha256=-g9a6sYgKKXRXmqYGYQBgpKEF_OlKE_uDDQXYMx6rT0,5113
 hud/tools/tests/test_bash_extended.py,sha256=G1pgl2e7_7ILYYS2FE6pTwI3_IPzmkPjBjGnxMGwGq8,7000
-hud/tools/tests/test_computer.py,sha256=BmrX2PV3wzHC4-xPJnWvenDV_kWt8LF8ia0kELBgHho,16394
-hud/tools/tests/test_computer_actions.py,sha256=YtUNFL7anhpXrcvg8EoUY1CqIV-TAAyaNFLZO9CiJ40,1194
+hud/tools/tests/test_computer.py,sha256=6BglMqodHkUT7365Ub0HpFe9dWUBiD3rL7_eDUsP07k,16397
+hud/tools/tests/test_computer_actions.py,sha256=gufOrd4pOgOv_dOKonRAYb1g_Ph0FBTisa61YgnFitw,1195
 hud/tools/tests/test_edit.py,sha256=pHw1MSs-P8mDKrwKUDW77vQfoMNwmbrEBt_MkKr2rE0,9734
 hud/tools/tests/test_init.py,sha256=fl4Tf4IUUFOKhdSRHu9GE4mkaTDiXw-2auxj4s84HuE,698
 hud/tools/tests/test_playwright_tool.py,sha256=TG0uieerc5wXq_JX66BLfXxphbSYGlDPhSbuoeizMu4,6737
@@ -166,10 +184,12 @@ hud/tools/tests/test_utils.py,sha256=qaujM1uyTMaKqWIeEgxty5GOFyfSUtrYCEHhmIazoy4
 hud/utils/__init__.py,sha256=nk9Re6ls2RudAWnAHDWYbLG28AwNF4qMFYf5xQIJhQA,181
 hud/utils/agent_factories.py,sha256=cvfXByqG6gOYHtm1VGeJjCpxoLxM4aJez8rH-AerP_A,3186
 hud/utils/async_utils.py,sha256=5cKrJcnaHV2eJNxeyx0r7fPcdPTDBK7kM9-nLaF51X4,2409
-hud/utils/hud_console.py,sha256=1JlIphJNtqS2cm4DsuQnvD0dJgyYD1Kf7575vPO0fz4,18202
+hud/utils/group_eval.py,sha256=oaoBqlQN6g5gRQmuY_JmqM5bpuf2sFIgu4uDZ7X-3a0,8360
+hud/utils/hud_console.py,sha256=ywTrzyNhWFoQN2PpzpDDKp_32b-ACDvfKQuWxDoF8iE,21898
 hud/utils/mcp.py,sha256=jvCWb5MXlMMObhrbYoiTlI-L9HNkEjLVx8GJ-HbdQ7U,2626
 hud/utils/pretty_errors.py,sha256=WGeL4CTHtlA6KgPuV_JSX5l6H4-xbuTp6Y6tw1bkiFg,2430
 hud/utils/progress.py,sha256=suikwFM8sdSfkV10nAOEaInDhG4XKgOSvFePg4jSj1A,5927
+hud/utils/tasks.py,sha256=JwFIq0cpPMpMYnICUmx_G4CF6uy9MtiCmmmN7eA6FsA,4682
 hud/utils/telemetry.py,sha256=hrVIx2rUjSGyy9IVxTZ_3Jii83PiHjyFRd5ls2whimM,1863
 hud/utils/tests/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 hud/utils/tests/test_async_utils.py,sha256=RkdSnYErRV3Jn7dfg6CPlcE1RSUL__2B627oIqAyy1s,5945
@@ -177,10 +197,10 @@ hud/utils/tests/test_init.py,sha256=2QLQSGgyP9wJhOvPCusm_zjJad0qApOZi1BXpxcdHXQ,
 hud/utils/tests/test_mcp.py,sha256=0pUa16mL-bqbZDXp5NHBnt1gO5o10BOg7zTMHZ1DNPM,4023
 hud/utils/tests/test_progress.py,sha256=QSF7Kpi03Ff_l3mAeqW9qs1nhK50j9vBiSobZq7T4f4,7394
 hud/utils/tests/test_telemetry.py,sha256=5jl7bEx8C8b-FfFUko5pf4UY-mPOR-9HaeL98dGtVHM,2781
-hud/utils/tests/test_version.py,sha256=nedCusQa_U-8ulETFb0keW6EYCWFc9kcQN3lIEKgFgw,160
+hud/utils/tests/test_version.py,sha256=cXnWb-tAwTPK40YNpgJ0EwXe7op_7xDXBGf7sf5-ECU,160
 hud/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-hud_python-0.4.27.dist-info/METADATA,sha256=Nxr0fHdGSkv0PQAyNvKPT1pW26sV6Tz_gwCOcgWk1Iw,20239
-hud_python-0.4.27.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-hud_python-0.4.27.dist-info/entry_points.txt,sha256=jJbodNFg1m0-CDofe5AHvB4zKBq7sSdP97-ohaQ3ae4,63
-hud_python-0.4.27.dist-info/licenses/LICENSE,sha256=yIzBheVUf86FC1bztAcr7RYWWNxyd3B-UJQ3uddg1HA,1078
-hud_python-0.4.27.dist-info/RECORD,,
+hud_python-0.4.29.dist-info/METADATA,sha256=xT3yJCdo-1-xJ1q54lgIyFALitPOiU2-YSParihkYvM,21069
+hud_python-0.4.29.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+hud_python-0.4.29.dist-info/entry_points.txt,sha256=jJbodNFg1m0-CDofe5AHvB4zKBq7sSdP97-ohaQ3ae4,63
+hud_python-0.4.29.dist-info/licenses/LICENSE,sha256=yIzBheVUf86FC1bztAcr7RYWWNxyd3B-UJQ3uddg1HA,1078
+hud_python-0.4.29.dist-info/RECORD,,