hud-python 0.4.36__tar.gz → 0.4.37__tar.gz

{hud_python-0.4.36 → hud_python-0.4.37}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.36
+Version: 0.4.37
 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
@@ -36,11 +36,13 @@ Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
 Requires-Python: <3.13,>=3.11
 Requires-Dist: anthropic
+Requires-Dist: blessed>=1.20.0
 Requires-Dist: datasets>=2.14.0
 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
 Requires-Dist: hud-mcp-use-python-sdk==2.3.19
+Requires-Dist: litellm>=1.55.0
 Requires-Dist: numpy>=1.24.0
 Requires-Dist: openai
 Requires-Dist: opentelemetry-api>=1.34.1
@@ -156,8 +158,8 @@ OSS RL environment + evals toolkit. Wrap software as environments, run benchmark
 ## Highlights
 
 - 🚀 **[MCP environment skeleton](https://docs.hud.so/core-concepts/mcp-protocol)** – any agent can call any environment.
-- ⚡️ **[Live telemetry](https://app.hud.so)** – inspect every tool call, observation, and reward in real time.
-- 🗂️ **[Public benchmarks](https://app.hud.so/leaderboards)** – OSWorld-Verified, SheetBench-50, and more.
+- ⚡️ **[Live telemetry](https://hud.so)** – inspect every tool call, observation, and reward in real time.
+- 🗂️ **[Public benchmarks](https://hud.so/leaderboards)** – OSWorld-Verified, SheetBench-50, and more.
 - 🌱 **[Reinforcement learning built-in](rl/)** – Verifiers gym pipelines for GRPO on any environment.
 - 🌐 **[Cloud browsers](environments/remote_browser/)** – AnchorBrowser, Steel, BrowserBase integrations for browser automation.
 - 🛠️ **[Hot-reload dev loop](environments/README.md#phase-5-hot-reload-development-with-cursor-agent)** – `hud dev` for iterating on environments without rebuilds.
@@ -203,14 +205,14 @@ from hud.agents import ClaudeAgent
 from hud.datasets import Task  # See docs: https://docs.hud.so/reference/tasks
 
 async def main() -> None:
-    with hud.trace("Quick Start 2048"): # All telemetry works for any MCP-based agent (see https://app.hud.so)
+    with hud.trace("Quick Start 2048"): # All telemetry works for any MCP-based agent (see https://hud.so)
         task = {
             "prompt": "Reach 64 in 2048.",
             "mcp_config": {
                 "hud": {
                     "url": "https://mcp.hud.so/v3/mcp",  # HUD's cloud MCP server (see https://docs.hud.so/core-concepts/architecture)
                     "headers": {
-                        "Authorization": f"Bearer {settings.api_key}",  # Get your key at https://app.hud.so
+                        "Authorization": f"Bearer {settings.api_key}",  # Get your key at https://hud.so
                         "Mcp-Image": "hudpython/hud-text-2048:v1.2"  # Docker image from https://hub.docker.com/u/hudpython
                     }
                 }
@@ -237,7 +239,7 @@ async def main() -> None:
 asyncio.run(main())
 ```
 
-The above example let's the agent play 2048 ([See replay](https://app.hud.so/trace/6feed7bd-5f67-4d66-b77f-eb1e3164604f))
+The above example let's the agent play 2048 ([See replay](https://hud.so/trace/6feed7bd-5f67-4d66-b77f-eb1e3164604f))
 
 ![Agent playing 2048](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/2048_1.gif)
 
@@ -268,7 +270,7 @@ Supports multi‑turn RL for both:
 - Language‑only models (e.g., `Qwen/Qwen2.5-7B-Instruct`)
 - Vision‑Language models (e.g., `Qwen/Qwen2.5-VL-3B-Instruct`)
 
-By default, `hud rl` provisions a persistant server and trainer in the cloud, streams telemetry to `app.hud.so`, and lets you monitor/manage models at `app.hud.so/models`. Use `--local` to run entirely on your machines (typically 2+ GPUs: one for vLLM, the rest for training).
+By default, `hud rl` provisions a persistant server and trainer in the cloud, streams telemetry to `hud.so`, and lets you monitor/manage models at `hud.so/models`. Use `--local` to run entirely on your machines (typically 2+ GPUs: one for vLLM, the rest for training).
 
 Any HUD MCP environment and evaluation works with our RL pipeline (including remote configurations). See the guided docs: `https://docs.hud.so/train-agents/quickstart`.
 
@@ -278,7 +280,7 @@ This is Claude Computer Use running on our proprietary financial analyst benchma
 
 ![Trace screenshot](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/trace_sheet.gif)
 
-> [See this trace on _app.hud.so_](https://app.hud.so/trace/9e212e9e-3627-4f1f-9eb5-c6d03c59070a)
+> [See this trace on _hud.so_](https://hud.so/trace/9e212e9e-3627-4f1f-9eb5-c6d03c59070a)
 
 This example runs the full dataset (only takes ~20 minutes) using [run_evaluation.py](examples/run_evaluation.py):
 
@@ -304,7 +306,7 @@ results = await run_dataset(
 print(f"Average reward: {sum(r.reward for r in results) / len(results):.2f}")
 ```
 
-> Running a dataset creates a job and streams results to the [app.hud.so](https://app.hud.so) platform for analysis and [leaderboard submission](https://docs.hud.so/evaluate-agents/leaderboards).
+> Running a dataset creates a job and streams results to the [hud.so](https://hud.so) platform for analysis and [leaderboard submission](https://docs.hud.so/evaluate-agents/leaderboards).
 
 ## Building Environments (MCP)
 
@@ -395,7 +397,7 @@ Tools
 hud push # needs docker login, hud api key
 ```
 
-5. Now you can use `mcp.hud.so` to launch 100s of instances of this environment in parallel with any agent, and see everything live on [app.hud.so](https://app.hud.so):
+5. Now you can use `mcp.hud.so` to launch 100s of instances of this environment in parallel with any agent, and see everything live on [hud.so](https://hud.so):
 
 ```python
 from hud.agents import ClaudeAgent
@@ -426,7 +428,7 @@ result = await ClaudeAgent().run({  # See all agents: https://docs.hud.so/refere
 
 ## Leaderboards & benchmarks
 
-All leaderboards are publicly available on [app.hud.so/leaderboards](https://app.hud.so/leaderboards) (see [docs](https://docs.hud.so/evaluate-agents/leaderboards))
+All leaderboards are publicly available on [hud.so/leaderboards](https://hud.so/leaderboards) (see [docs](https://docs.hud.so/evaluate-agents/leaderboards))
 
 ![Leaderboard](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/leaderboards_3.png)
 
@@ -440,7 +442,7 @@ Using the [`run_dataset`](https://docs.hud.so/reference/tasks#run_dataset) funct
 %%{init: {"theme": "neutral", "themeVariables": {"fontSize": "14px"}} }%%
 graph LR
     subgraph "Platform"
-        Dashboard["📊 app.hud.so"]
+        Dashboard["📊 hud.so"]
         API["🔌 mcp.hud.so"]
     end
   

{hud_python-0.4.36 → hud_python-0.4.37}/README.md

@@ -23,8 +23,8 @@ OSS RL environment + evals toolkit. Wrap software as environments, run benchmark
 ## Highlights
 
 - 🚀 **[MCP environment skeleton](https://docs.hud.so/core-concepts/mcp-protocol)** – any agent can call any environment.
-- ⚡️ **[Live telemetry](https://app.hud.so)** – inspect every tool call, observation, and reward in real time.
-- 🗂️ **[Public benchmarks](https://app.hud.so/leaderboards)** – OSWorld-Verified, SheetBench-50, and more.
+- ⚡️ **[Live telemetry](https://hud.so)** – inspect every tool call, observation, and reward in real time.
+- 🗂️ **[Public benchmarks](https://hud.so/leaderboards)** – OSWorld-Verified, SheetBench-50, and more.
 - 🌱 **[Reinforcement learning built-in](rl/)** – Verifiers gym pipelines for GRPO on any environment.
 - 🌐 **[Cloud browsers](environments/remote_browser/)** – AnchorBrowser, Steel, BrowserBase integrations for browser automation.
 - 🛠️ **[Hot-reload dev loop](environments/README.md#phase-5-hot-reload-development-with-cursor-agent)** – `hud dev` for iterating on environments without rebuilds.
@@ -70,14 +70,14 @@ from hud.agents import ClaudeAgent
 from hud.datasets import Task  # See docs: https://docs.hud.so/reference/tasks
 
 async def main() -> None:
-    with hud.trace("Quick Start 2048"): # All telemetry works for any MCP-based agent (see https://app.hud.so)
+    with hud.trace("Quick Start 2048"): # All telemetry works for any MCP-based agent (see https://hud.so)
         task = {
             "prompt": "Reach 64 in 2048.",
             "mcp_config": {
                 "hud": {
                     "url": "https://mcp.hud.so/v3/mcp",  # HUD's cloud MCP server (see https://docs.hud.so/core-concepts/architecture)
                     "headers": {
-                        "Authorization": f"Bearer {settings.api_key}",  # Get your key at https://app.hud.so
+                        "Authorization": f"Bearer {settings.api_key}",  # Get your key at https://hud.so
                         "Mcp-Image": "hudpython/hud-text-2048:v1.2"  # Docker image from https://hub.docker.com/u/hudpython
                     }
                 }
@@ -104,7 +104,7 @@ async def main() -> None:
 asyncio.run(main())
 ```
 
-The above example let's the agent play 2048 ([See replay](https://app.hud.so/trace/6feed7bd-5f67-4d66-b77f-eb1e3164604f))
+The above example let's the agent play 2048 ([See replay](https://hud.so/trace/6feed7bd-5f67-4d66-b77f-eb1e3164604f))
 
 ![Agent playing 2048](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/2048_1.gif)
 
@@ -135,7 +135,7 @@ Supports multi‑turn RL for both:
 - Language‑only models (e.g., `Qwen/Qwen2.5-7B-Instruct`)
 - Vision‑Language models (e.g., `Qwen/Qwen2.5-VL-3B-Instruct`)
 
-By default, `hud rl` provisions a persistant server and trainer in the cloud, streams telemetry to `app.hud.so`, and lets you monitor/manage models at `app.hud.so/models`. Use `--local` to run entirely on your machines (typically 2+ GPUs: one for vLLM, the rest for training).
+By default, `hud rl` provisions a persistant server and trainer in the cloud, streams telemetry to `hud.so`, and lets you monitor/manage models at `hud.so/models`. Use `--local` to run entirely on your machines (typically 2+ GPUs: one for vLLM, the rest for training).
 
 Any HUD MCP environment and evaluation works with our RL pipeline (including remote configurations). See the guided docs: `https://docs.hud.so/train-agents/quickstart`.
 
@@ -145,7 +145,7 @@ This is Claude Computer Use running on our proprietary financial analyst benchma
 
 ![Trace screenshot](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/trace_sheet.gif)
 
-> [See this trace on _app.hud.so_](https://app.hud.so/trace/9e212e9e-3627-4f1f-9eb5-c6d03c59070a)
+> [See this trace on _hud.so_](https://hud.so/trace/9e212e9e-3627-4f1f-9eb5-c6d03c59070a)
 
 This example runs the full dataset (only takes ~20 minutes) using [run_evaluation.py](examples/run_evaluation.py):
 
@@ -171,7 +171,7 @@ results = await run_dataset(
 print(f"Average reward: {sum(r.reward for r in results) / len(results):.2f}")
 ```
 
-> Running a dataset creates a job and streams results to the [app.hud.so](https://app.hud.so) platform for analysis and [leaderboard submission](https://docs.hud.so/evaluate-agents/leaderboards).
+> Running a dataset creates a job and streams results to the [hud.so](https://hud.so) platform for analysis and [leaderboard submission](https://docs.hud.so/evaluate-agents/leaderboards).
 
 ## Building Environments (MCP)
 
@@ -262,7 +262,7 @@ Tools
 hud push # needs docker login, hud api key
 ```
 
-5. Now you can use `mcp.hud.so` to launch 100s of instances of this environment in parallel with any agent, and see everything live on [app.hud.so](https://app.hud.so):
+5. Now you can use `mcp.hud.so` to launch 100s of instances of this environment in parallel with any agent, and see everything live on [hud.so](https://hud.so):
 
 ```python
 from hud.agents import ClaudeAgent
@@ -293,7 +293,7 @@ result = await ClaudeAgent().run({  # See all agents: https://docs.hud.so/refere
 
 ## Leaderboards & benchmarks
 
-All leaderboards are publicly available on [app.hud.so/leaderboards](https://app.hud.so/leaderboards) (see [docs](https://docs.hud.so/evaluate-agents/leaderboards))
+All leaderboards are publicly available on [hud.so/leaderboards](https://hud.so/leaderboards) (see [docs](https://docs.hud.so/evaluate-agents/leaderboards))
 
 ![Leaderboard](https://raw.githubusercontent.com/hud-evals/hud-python/main/docs/src/images/leaderboards_3.png)
 
@@ -307,7 +307,7 @@ Using the [`run_dataset`](https://docs.hud.so/reference/tasks#run_dataset) funct
 %%{init: {"theme": "neutral", "themeVariables": {"fontSize": "14px"}} }%%
 graph LR
     subgraph "Platform"
-        Dashboard["📊 app.hud.so"]
+        Dashboard["📊 hud.so"]
         API["🔌 mcp.hud.so"]
     end
   

{hud_python-0.4.36 → hud_python-0.4.37}/environments/README.md

@@ -495,7 +495,7 @@ from hud.agents import ClaudeAgent
 from hud.clients import MCPClient
 
 async def main():
-    # `trace` captures *everything* that happens and sends it to app.hud.so
+    # `trace` captures *everything* that happens and sends it to hud.so
     with hud.trace("local_test"):
         task = Task(
             prompt="Complete the task",
@@ -524,7 +524,7 @@ async def main():
 asyncio.run(main())
 ```
 
-The `trace` context manager sends a full timeline of agent actions, tool calls, and rewards to app.hud.so – perfect for debugging.
+The `trace` context manager sends a full timeline of agent actions, tool calls, and rewards to hud.so – perfect for debugging.
 
 See `examples/01_hello_2048.py` and `examples/task_with_setup_eval.py` for larger end-to-end demos.
 
@@ -532,7 +532,7 @@ See `examples/01_hello_2048.py` and `examples/task_with_setup_eval.py` for large
 
 ## Phase 4 – Remote Deployment & HUD Runner
 
-**Goal →** the exact same image runs in parallel on hundreds of instances, and exposes more telemetry so the app.hud.so can visualise the whole lifecycle.
+**Goal →** the exact same image runs in parallel on hundreds of instances, and exposes more telemetry so the hud.so can visualise the whole lifecycle.
 
 ### 1. Publish your image
 
@@ -595,11 +595,11 @@ async def initialize_environment(session=None, progress_token=None):
     await send(100, "ready")
 ```
 
-Those messages are displayed live on app.hud.so alongside resource graphs – perfect feedback while you wait.
+Those messages are displayed live on hud.so alongside resource graphs – perfect feedback while you wait.
 
 ### 4. Live telemetry (`telemetry://live`) (Optional)
 
-Expose a resource named `telemetry://live` exactly like in `environments/browser/src/hud_controller/server.py` to return live url to be displayed on app.hud.so.
+Expose a resource named `telemetry://live` exactly like in `environments/browser/src/hud_controller/server.py` to return live url to be displayed on hud.so.
 
 Once all of the above works you can unleash *hundreds* of concurrent agents on your new environment.
 

{hud_python-0.4.36 → hud_python-0.4.37}/environments/blank/README.md

@@ -10,7 +10,7 @@
 
 IMPORTANT: Make sure all logs are going to stderr instead of stdio, which is reserved for MCP communication
 
-### Interactive Development
+### Testing your environment
 ```bash
 # 1. Configure your API keys (optional - only needed for evaluation)
 # Edit .env file to add your HUD_API_KEY and ANTHROPIC_API_KEY
@@ -24,13 +24,29 @@ hud dev --build --interactive
 hud eval tasks.json --agent claude
 
 # Option B: Interactive notebook test_env.ipynb (great for learning!)
-# Requires installation:
-pip install hud-python[agents]
 
 # Option C: Simple Python script (runs all tasks from tasks.json)
 python test_task.py
 ```
 
+## Iterating on your environment
+This is usually the process for making any environment better:
+```bash
+# 1. Start the environment and interact with it directly (or give MCP server to an agent):
+hud dev --build --interactive
+
+# 2. If the environment cannot start or fails inexplicably:
+hud debug test_env:dev # Or your env name that appears when you run hud dev
+# After fixing the error, go back to 1.
+
+# 3. When the environment is in a stable state:
+hud build
+hud push # Requires docker login
+
+# 4. As soon as it's pushed to the newest version, make sure tasks have it updated and run:
+hud rl
+# This is a good test to see if your environment and tasks are high quality!
+
 ## Layout
 ```
 controller/
@@ -83,7 +99,7 @@ save_tasks(tasks, repo_id="your-org/your-dataset")
 hud eval "your-org/your-dataset" --agent claude
 
 # View results at:
-# app.hud.so/leaderboards/your-org/your-dataset
+# hud.so/leaderboards/your-org/your-dataset
 ```
 
 **Note**: Only public HuggingFace datasets appear as leaderboards!

{hud_python-0.4.36 → hud_python-0.4.37}/environments/blank/pyproject.toml

@@ -3,7 +3,7 @@ name = "test_test"
 version = "0.1.0"
 description = "A minimal HUD environment"
 requires-python = ">=3.11"
-dependencies = [ "hud-python==0.4.36", "fastapi", "uvicorn[standard]", "httpx>=0.28.1",]
+dependencies = [ "hud-python==0.4.37", "fastapi", "uvicorn[standard]", "httpx>=0.28.1",]
 
 [build-system]
 requires = [ "hatchling",]

{hud_python-0.4.36 → hud_python-0.4.37}/environments/browser/README.md

@@ -75,7 +75,7 @@ save_tasks(tasks, repo_id="your-org/your-dataset")
 hud eval "your-org/your-dataset" --agent claude
 
 # View results at:
-# app.hud.so/leaderboards/your-org/your-dataset
+# hud.so/leaderboards/your-org/your-dataset
 ```
 
 **Note**: Only public HuggingFace datasets appear as leaderboards!

{hud_python-0.4.36 → hud_python-0.4.37}/environments/browser/pyproject.toml

@@ -3,7 +3,7 @@ name = "hud-browser-controller"
 version = "0.1.0"
 description = "HUD Browser Controller - MCP interface for browser environments"
 requires-python = ">=3.11,<3.14"
-dependencies = [ "pydantic>=2.6,<3", "pydantic-settings>=2.2,<3", "hud-python@git+https://github.com/hud-evals/hud-python@env-cli-improvements", "playwright", "pyautogui", "httpx", "typer", "fastapi", "uvicorn",]
+dependencies = [ "pydantic>=2.6,<3", "pydantic-settings>=2.2,<3", "hud-python@git+https://github.com/hud-evals/hud-python@env-cli-improvements", "playwright", "pyautogui", "httpx", "typer", "fastapi>=0.104.1", "uvicorn[standard]>=0.24.0", "python-multipart>=0.0.6",]
 
 [build-system]
 requires = [ "hatchling",]
@@ -19,4 +19,4 @@ image = "hud-browser:dev"
 allow-direct-references = true
 
 [tool.hatch.build.targets.wheel]
-packages = [ "controller", "problems",]
+packages = [ "controller", "environment",]

{hud_python-0.4.36 → hud_python-0.4.37}/environments/deepresearch/pyproject.toml

@@ -3,7 +3,7 @@ name = "deepresearch"
 version = "0.1.0"
 description = "DeepResearch HUD environment with HTTP backend (EXA on server)"
 requires-python = ">=3.11"
-dependencies = [ "hud-python==0.4.36", "fastapi>=0.104.1", "uvicorn[standard]>=0.24.0", "httpx>=0.24.0",]
+dependencies = [ "hud-python==0.4.37", "fastapi>=0.104.1", "uvicorn[standard]>=0.24.0", "httpx>=0.24.0",]
 
 [build-system]
 requires = [ "hatchling",]

{hud_python-0.4.36 → hud_python-0.4.37}/hud/agents/__init__.py

@@ -2,12 +2,14 @@ from __future__ import annotations
 
 from .base import MCPAgent
 from .claude import ClaudeAgent
+from .lite_llm import LiteAgent
 from .openai import OperatorAgent
 from .openai_chat_generic import GenericOpenAIChatAgent
 
 __all__ = [
     "ClaudeAgent",
     "GenericOpenAIChatAgent",
+    "LiteAgent",
     "MCPAgent",
     "OperatorAgent",
 ]

hud_python-0.4.37/hud/agents/lite_llm.py

@@ -0,0 +1,72 @@
+"""LiteLLM MCP Agent implementation.
+
+Same OpenAI chat-completions shape + MCP tool plumbing,
+but transport is LiteLLM and (optionally) tools are shaped by LiteLLM's MCP transformer.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any, ClassVar
+
+import litellm
+
+from .openai_chat_generic import GenericOpenAIChatAgent
+
+logger = logging.getLogger(__name__)
+
+# Prefer LiteLLM's built-in MCP -> OpenAI tool transformer (handles Bedrock nuances)
+try:
+    from litellm.experimental_mcp_client.tools import (
+        transform_mcp_tool_to_openai_tool,
+    )
+except Exception:  # pragma: no cover - optional dependency
+    transform_mcp_tool_to_openai_tool = None  # type: ignore
+
+
+class LiteAgent(GenericOpenAIChatAgent):
+    """
+    Same OpenAI chat-completions shape + MCP tool plumbing,
+    but transport is LiteLLM and (optionally) tools are shaped by LiteLLM's MCP transformer.
+    """
+
+    metadata: ClassVar[dict[str, Any]] = {}
+
+    def __init__(
+        self,
+        *,
+        model_name: str = "gpt-4o-mini",
+        completion_kwargs: dict[str, Any] | None = None,
+        **agent_kwargs: Any,
+    ) -> None:
+        # We don't need an OpenAI client; pass None
+        super().__init__(
+            openai_client=None,
+            model_name=model_name,
+            completion_kwargs=completion_kwargs,
+            **agent_kwargs,
+        )
+
+    def get_tool_schemas(self) -> list[dict]:
+        # Prefer LiteLLM's stricter transformer (handles Bedrock & friends)
+        if transform_mcp_tool_to_openai_tool is not None:
+            return [
+                transform_mcp_tool_to_openai_tool(t)  # returns ChatCompletionToolParam-like dict
+                for t in self.get_available_tools()
+            ]
+        # Fallback to the generic OpenAI sanitizer
+        return GenericOpenAIChatAgent.get_tool_schemas(self)
+
+    async def _invoke_chat_completion(
+        self,
+        *,
+        messages: list[Any],
+        tools: list[dict] | None,
+        extra: dict[str, Any],
+    ):
+        return await litellm.acompletion(
+            model=self.model_name,
+            messages=messages,
+            tools=tools or None,  # LiteLLM tolerates None better than []
+            **extra,
+        )

{hud_python-0.4.36 → hud_python-0.4.37}/hud/agents/openai_chat_generic.py

@@ -42,7 +42,7 @@ class GenericOpenAIChatAgent(MCPAgent):
     def __init__(
         self,
         *,
-        openai_client: AsyncOpenAI,
+        openai_client: AsyncOpenAI | None,
         model_name: str = "gpt-4o-mini",
         completion_kwargs: dict[str, Any] | None = None,
         **agent_kwargs: Any,
@@ -171,6 +171,23 @@ class GenericOpenAIChatAgent(MCPAgent):
             openai_tools.append(openai_tool)
         return openai_tools
 
+    async def _invoke_chat_completion(
+        self,
+        *,
+        messages: list[Any],
+        tools: list[dict] | None,
+        extra: dict[str, Any],
+    ):
+        if self.oai is None:
+            raise ValueError("openai_client is required for GenericOpenAIChatAgent")
+        # default transport = OpenAI SDK
+        return await self.oai.chat.completions.create(
+            model=self.model_name,
+            messages=messages,
+            tools=tools,  # already ChatCompletionToolParam-shaped
+            **extra,
+        )
+
     @instrument(
         span_type="agent",
         record_args=False,
@@ -180,17 +197,14 @@ class GenericOpenAIChatAgent(MCPAgent):
         """Send chat request to OpenAI and convert the response."""
 
         # Convert MCP tool schemas to OpenAI format
-        mcp_schemas = self.get_tool_schemas()
+        tools = cast("list[ChatCompletionToolParam]", self.get_tool_schemas())
 
         protected_keys = {"model", "messages", "tools"}
         extra = {k: v for k, v in (self.completion_kwargs or {}).items() if k not in protected_keys}
 
         try:
-            response = await self.oai.chat.completions.create(
-                model=self.model_name,
-                messages=messages,
-                tools=cast("list[ChatCompletionToolParam]", mcp_schemas),
-                **extra,
+            response = await self._invoke_chat_completion(
+                messages=messages, tools=tools, extra=extra
             )
         except Exception as e:
             error_content = f"Error getting response {e}"

{hud_python-0.4.36 → hud_python-0.4.37}/hud/cli/__init__.py

@@ -912,7 +912,7 @@ def eval(
     agent: str | None = typer.Argument(
         None,
         help=(
-            "Agent backend to use (claude, openai, or vllm). If not provided, will prompt interactively."  # noqa: E501
+            "Agent backend to use (claude, openai, vllm, or litellm). If not provided, will prompt interactively."  # noqa: E501
         ),
     ),
     full: bool = typer.Option(
@@ -960,6 +960,12 @@ def eval(
         "--verbose",
         help="Enable verbose output from the agent",
     ),
+    very_verbose: bool = typer.Option(
+        False,
+        "--very-verbose",
+        "-vv",
+        help="Enable debug-level logs for maximum visibility",
+    ),
     vllm_base_url: str | None = typer.Option(
         None,
         "--vllm-base-url",
@@ -1025,13 +1031,14 @@ def eval(
                 {"name": "Claude 4 Sonnet", "value": "claude"},
                 {"name": "OpenAI Computer Use", "value": "openai"},
                 {"name": "vLLM (Local Server)", "value": "vllm"},
+                {"name": "LiteLLM (Multi-provider)", "value": "litellm"},
             ]
         )
 
         agent = hud_console.select("Select an agent to use:", choices=choices, default=0)
 
     # Handle HUD model selection
-    if agent and agent not in ["claude", "openai", "vllm"]:
+    if agent and agent not in ["claude", "openai", "vllm", "litellm"]:
         # Find remote model name
         model = agent
         if not vllm_base_url:
@@ -1052,7 +1059,7 @@ def eval(
         hud_console.info(f"Using HUD model: {model} (trained on {base_model})")
 
     # Validate agent choice
-    valid_agents = ["claude", "openai", "vllm"]
+    valid_agents = ["claude", "openai", "vllm", "litellm"]
     if agent not in valid_agents:
         hud_console.error(f"Invalid agent: {agent}. Must be one of: {', '.join(valid_agents)}")
         raise typer.Exit(1)
@@ -1070,6 +1077,7 @@ def eval(
         max_workers=max_workers,
         max_concurrent_per_worker=max_concurrent_per_worker,
         verbose=verbose,
+        very_verbose=very_verbose,
         vllm_base_url=vllm_base_url,
         group_size=group_size,
     )
@@ -1119,7 +1127,7 @@ def rl(
     ),
     model: str | None = typer.Argument(
         None,
-        help="Model to train (default: interactive selection)",
+        help="Model to train from https://hud.so/models (default: interactive selection)",
     ),
     config_file: Path | None = typer.Option(  # noqa: B008
         None,
@@ -1159,6 +1167,12 @@ def rl(
         "--ddp-gpus",
         help="Specific GPUs for DDP (e.g., '0,1,2,3')",
     ),
+    yes: bool = typer.Option(
+        False,
+        "--yes",
+        "-y",
+        help="Auto-accept all prompts and use defaults (lazy mode)",
+    ),
     vllm_gpu: int | None = typer.Option(
         None,
         "--vllm-gpu",
@@ -1180,6 +1194,7 @@ def rl(
         no_ddp=no_ddp,
         ddp_gpus=ddp_gpus,
         vllm_gpu=vllm_gpu,
+        yes=yes,
     )
 
 

{hud_python-0.4.36 → hud_python-0.4.37}/hud/cli/build.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import asyncio
+import contextlib
 import hashlib
 import subprocess
 import time
@@ -13,6 +14,7 @@ from typing import Any
 import typer
 import yaml
 
+from hud.cli.utils.source_hash import compute_source_hash, list_source_files
 from hud.clients import MCPClient
 from hud.utils.hud_console import HUDConsole
 from hud.version import __version__ as hud_version
@@ -341,10 +343,11 @@ def build_environment(
     required_env, optional_env = extract_env_vars_from_dockerfile(dockerfile_path)
 
     # Merge user-provided env vars with detected ones
-    provided_env_vars = {}
+    provided_env_vars: dict[str, str] = {}
     missing_required = []
     if env_vars:
-        provided_env_vars = env_vars.copy()
+        # Use placeholders in lock file for any provided values to avoid storing secrets
+        provided_env_vars = {k: f"${{{k}}}" for k in env_vars}
         # Track which required vars are still missing
         missing_required = [e for e in required_env if e not in env_vars]
 
@@ -384,6 +387,8 @@ def build_environment(
             "hudVersion": hud_version,
             "directory": str(env_dir.name),
             "version": new_version,  # Internal environment version
+            # Fast source fingerprint for change detection
+            "sourceHash": compute_source_hash(env_dir),
         },
         "environment": {
             "initializeMs": analysis["initializeMs"],
@@ -424,6 +429,16 @@ def build_environment(
     with open(lock_path, "w") as f:
         yaml.dump(lock_content, f, default_flow_style=False, sort_keys=False)
 
+    # Also write the file list we hashed for transparency (non-essential)
+    with contextlib.suppress(Exception):
+        files = [
+            str(p.resolve().relative_to(env_dir)).replace("\\", "/")
+            for p in list_source_files(env_dir)
+        ]
+        lock_content["build"]["sourceFiles"] = files
+        with open(lock_path, "w") as f:
+            yaml.dump(lock_content, f, default_flow_style=False, sort_keys=False)
+
     hud_console.success("Created lock file: hud.lock.yaml")
 
     # Calculate lock file hash

{hud_python-0.4.36 → hud_python-0.4.37}/hud/cli/dev.py

@@ -530,7 +530,7 @@ async def start_mcp_proxy(
                     stderr=asyncio.subprocess.DEVNULL,
                 )
                 await stop_result.communicate()
-                hud_console.success("✅ Container stopped successfully")
+                hud_console.success("Container stopped successfully")
                 container_stopped = True
         except Exception as e:
             hud_console.warning(f"Failed to stop container: {e}")