Upload folder using huggingface_hub

Dockerfile

@@ -11,13 +11,13 @@
 # The build script (openenv build) handles context detection and sets appropriate build args.
 
 ARG BASE_IMAGE=ghcr.io/meta-pytorch/openenv-base:latest
-FROM ${BASE_IMAGE} AS builder
+FROM ghcr.io/meta-pytorch/openenv-base:latest AS builder
 
 WORKDIR /app
 
 # Ensure git is available (required for installing dependencies from VCS)
 RUN apt-get update && \
-    apt-get install -y --no-install-recommends git && \
+    apt-get install -y --no-install-recommends git gcc python3-dev && \
     rm -rf /var/lib/apt/lists/*
 
 # Build argument to control whether we're building standalone or in-repo
@@ -40,22 +40,26 @@ RUN if ! command -v uv >/dev/null 2>&1; then \
     
 # Install dependencies using uv sync
 # If uv.lock exists, use it; otherwise resolve on the fly
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
+    install -m 0755 /root/.local/bin/uv /usr/local/bin/uv && \
+    install -m 0755 /root/.local/bin/uvx /usr/local/bin/uvx
+
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ -f uv.lock ]; then \
-        uv sync --frozen --no-install-project --no-editable; \
+        uv sync --no-install-project --no-editable; \
     else \
         uv sync --no-install-project --no-editable; \
     fi
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ -f uv.lock ]; then \
-        uv sync --frozen --no-editable; \
+        uv sync --no-editable; \
     else \
         uv sync --no-editable; \
     fi
 
 # Final runtime stage
-FROM ${BASE_IMAGE}
+FROM ghcr.io/meta-pytorch/openenv-base:latest
 
 WORKDIR /app
 
@@ -70,6 +74,7 @@ ENV PATH="/app/.venv/bin:$PATH"
 
 # Set PYTHONPATH so imports work correctly
 ENV PYTHONPATH="/app/env:$PYTHONPATH"
+ENV ENABLE_WEB_INTERFACE=true
 
 # Health check
 HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
@@ -77,5 +82,4 @@ HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
 
 # Run the FastAPI server
 # The module path is constructed to work with the /app/env structure
-ENV ENABLE_WEB_INTERFACE=true
 CMD ["sh", "-c", "cd /app/env && uvicorn server.app:app --host 0.0.0.0 --port 8000"]

README.md

@@ -2,17 +2,34 @@
 title: TB2 Environment Server
 emoji: "🧪"
 colorFrom: blue
-colorTo: blue
+colorTo: green
 sdk: docker
 pinned: false
 app_port: 8000
 base_path: /web
 tags:
+  - openenv-0.2.3
   - openenv
   - terminal-bench-2
   - spaces
 ---
 
+## Hugging Face Space Deployment
+
+This Space is built from OpenEnv environment `tbench2_env`.
+
+- Space URL: `https://huggingface.co/spaces/openenv/tbench2`
+- OpenEnv pinned ref: `0.2.3`
+- Hub tag: `openenv`
+
+### Connecting from Code
+
+```python
+from envs.tbench2_env import Env
+
+env = Env(base_url="https://huggingface.co/spaces/openenv/tbench2")
+```
+
 # TB2 Environment (Terminal-Bench 2)
 
 OpenEnv wrapper for [Terminal-Bench 2](https://github.com/laude-institute/terminal-bench-2) tasks. Supports two execution modes:

client.py

@@ -19,12 +19,12 @@ try:
 
     from .models import Tbench2Action, Tbench2Observation, Tbench2State
 except ImportError:
+    from models import Tbench2Action, Tbench2Observation, Tbench2State
+
     # Standalone imports (when environment is standalone with openenv from pip)
     from openenv.core.client_types import StepResult
     from openenv.core.env_client import EnvClient
 
-    from models import Tbench2Action, Tbench2Observation, Tbench2State
-
 
 class Tbench2Env(EnvClient[Tbench2Action, Tbench2Observation, Tbench2State]):
     """HTTP client for the TB2 environment."""

envs/tbench2_env/README.md

@@ -0,0 +1,207 @@
+---
+title: TB2 Environment Server
+emoji: "🧪"
+colorFrom: blue
+colorTo: blue
+sdk: docker
+pinned: false
+app_port: 8000
+base_path: /web
+tags:
+  - openenv
+  - terminal-bench-2
+  - spaces
+---
+
+# TB2 Environment (Terminal-Bench 2)
+
+OpenEnv wrapper for [Terminal-Bench 2](https://github.com/laude-institute/terminal-bench-2) tasks. Supports two execution modes:
+
+| Mode | Description | Use Case |
+|------|-------------|----------|
+| **Local** | Runs commands in the server process (no Docker) | Hugging Face Spaces, environments without Docker access |
+| **Docker** | Runs each task in its own container | Full TB2.0 fidelity with custom task images |
+
+## Quick Start
+
+```python
+from tbench2_env import Tbench2Env, Tbench2Action
+
+env = Tbench2Env(base_url="http://localhost:8000")
+result = env.reset(task_id="headless-terminal")
+print(result.observation.instruction)
+
+result = env.step(Tbench2Action(action_type="exec", command="ls -la"))
+print(result.observation.output)
+
+result = env.step(Tbench2Action(action_type="evaluate"))
+print(result.reward, result.done)
+
+env.close()
+```
+
+## Building the Docker Image
+
+Before using the environment, build the Docker image:
+
+```bash
+# From project root
+docker build -t tbench2-env:latest -f envs/tbench2_env/server/Dockerfile .
+```
+
+## Environment Details
+
+### Action
+**Tbench2Action**: Controls interaction with the TB2 task session
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `action_type` | str | `"exec"` | Action to perform (`exec`, `write`, `view`, `wait`, `kill`, `write_file`, `evaluate`, `close`) |
+| `command` | str | `""` | Shell command or input to send |
+| `session_id` | str \| None | `None` | Session ID for streaming processes |
+| `block` | bool | `True` | Whether to block until command completes |
+| `wait_seconds` | float \| None | `None` | Time to wait (for `wait` action) |
+| `file_path` | str | `""` | File path (for `write_file` action) |
+| `content` | str | `""` | Content to write (for `write_file` action) |
+
+### Observation
+**Tbench2Observation**: Contains the environment response
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `instruction` | str | Task instruction/prompt from the TB2 task |
+| `output` | str | Command output (stdout/stderr) |
+| `success` | bool | Whether the action succeeded |
+| `error` | str | Error message if action failed |
+| `task_id` | str | Current task identifier |
+| `task_path` | str | Path to the task directory |
+| `session_id` | str \| None | Session ID for streaming processes |
+| `action_type` | str | The action type that produced this observation |
+| `info` | dict | Additional metadata |
+
+### State
+**Tbench2State**: Server-side state for the task session
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `task_id` | str | Current task identifier |
+| `task_path` | str | Path to the task directory |
+| `session_id` | str | Active session ID |
+| `terminal_ready` | bool | Whether the terminal is ready for commands |
+| `last_action_type` | str | Last action type executed |
+| `last_command` | str | Last command executed |
+| `last_output` | str | Output from last command |
+
+## Execution Modes
+
+### Local Mode (Default)
+
+Commands execute directly in the server process. Ideal for HF Spaces where Docker-in-Docker is unavailable.
+
+```bash
+# Default - local mode
+python -m tbench2_env.server.app
+
+# Or explicitly set mode
+TB2_MODE=local python -m tbench2_env.server.app
+```
+
+**Note:** Local mode ignores Docker images specified in task.toml. Tasks requiring specific runtime environments may fail.
+
+### Docker Mode
+
+Each task runs in its own Docker container, using the image specified in the task's `task.toml`:
+
+```bash
+# Enable Docker mode
+TB2_MODE=docker python -m tbench2_env.server.app
+```
+
+**Requirements:**
+- Docker socket mounted at `/var/run/docker.sock`
+- Sufficient disk space for container images
+- Network access to pull images if not cached
+
+**Environment Variables for Docker Mode:**
+- `TB2_MODE=docker` - Enable Docker-backed execution
+- Docker socket must be accessible (mounted volume)
+
+## Action Types
+
+| Action | Description | Required Fields |
+|--------|-------------|-----------------|
+| `exec` | Run a shell command | `command`, optionally `block`, `session_id` |
+| `write` | Send input to a running session | `session_id`, `command` |
+| `view` | Read pending output | `session_id` |
+| `wait` | Wait for output | `session_id`, optionally `wait_seconds` |
+| `kill` | Terminate a running session | `session_id` |
+| `write_file` | Write content to a file | `file_path`, `content` |
+| `evaluate` | Run pytest tests, return reward | (none) |
+| `close` | Stop and cleanup | (none) |
+
+## Session IDs (Streaming Processes)
+
+`session_id` is **only** required when you start a non-blocking process and want to interact with it (`write`, `view`, `wait`, `kill`). For plain `exec` commands, you can omit it.
+
+Example (Python):
+```python
+# Start a long-running process
+env.step(Tbench2Action(action_type="exec", command="python -i", block=False, session_id="sess1"))
+
+# Send input to it
+env.step(Tbench2Action(action_type="write", session_id="sess1", command="print(2+2)\n"))
+
+# Read its output
+env.step(Tbench2Action(action_type="view", session_id="sess1"))
+```
+
+## Environment Variables
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `TB2_MODE` | `local` | Execution mode: `local` or `docker` |
+| `TB2_TASKS_DIR` | (auto-download) | Path to local Terminal-Bench-2 repo checkout |
+| `TB2_OUTPUT_DIR` | `/tmp/tbench2_env_runs` | Directory for session logs and cache |
+| `TB2_CACHE_DIR` | `$TB2_OUTPUT_DIR/repo_cache` | Where to extract TB2 repo |
+| `TB2_REPO_URL` | (GitHub main.zip) | Repo zip URL for auto-download |
+
+## Reward
+
+Binary reward on `evaluate` action:
+- `1.0` - All pytest tests pass (exit code 0)
+- `0.0` - Tests fail (non-zero exit code)
+
+Intermediate steps return `reward=None`.
+
+## Running the Server
+
+```bash
+# Install dependencies
+uv sync --all-extras
+
+# Local mode (default, for Spaces)
+python -m tbench2_env.server.app --port 8000
+
+# Docker mode (full TB2.0 compatibility)
+TB2_MODE=docker python -m tbench2_env.server.app --port 8000
+
+# With local TB2 repo
+TB2_TASKS_DIR=/path/to/terminal-bench-2 python -m tbench2_env.server.app
+```
+
+## Project Structure
+
+```
+tbench2_env/
+├── __init__.py              # Module exports (Tbench2Env, Tbench2Action, etc.)
+├── README.md                # This file
+├── client.py                # Tbench2Env client implementation
+├── models.py                # Tbench2Action, Tbench2Observation, Tbench2State
+├── openenv.yaml             # OpenEnv configuration
+├── pyproject.toml           # Package dependencies
+└── server/
+    ├── __init__.py          # Server exports
+    ├── app.py               # FastAPI application
+    ├── tbench2_env_environment.py  # Core environment logic
+    └── Dockerfile           # Container image definition
+```

envs/tbench2_env/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Tbench2 Env Environment."""
+
+from .client import Tbench2Env
+from .models import Tbench2Action, Tbench2Observation, Tbench2State
+
+
+__all__ = [
+    "Tbench2Action",
+    "Tbench2Observation",
+    "Tbench2Env",
+    "Tbench2State",
+]

envs/tbench2_env/client.py

@@ -0,0 +1,75 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""TB2 Environment Client."""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+# Support both in-repo and standalone imports
+try:
+    # In-repo imports (when running from OpenEnv repository)
+    from openenv.core.client_types import StepResult
+    from openenv.core.env_client import EnvClient
+
+    from .models import Tbench2Action, Tbench2Observation, Tbench2State
+except ImportError:
+    from models import Tbench2Action, Tbench2Observation, Tbench2State
+
+    # Standalone imports (when environment is standalone with openenv from pip)
+    from openenv.core.client_types import StepResult
+    from openenv.core.env_client import EnvClient
+
+
+class Tbench2Env(EnvClient[Tbench2Action, Tbench2Observation, Tbench2State]):
+    """HTTP client for the TB2 environment."""
+
+    def _step_payload(self, action: Tbench2Action) -> dict[str, Any]:
+        return {
+            "action_type": action.action_type,
+            "command": action.command,
+            "session_id": action.session_id,
+            "block": action.block,
+            "wait_seconds": action.wait_seconds,
+            "file_path": action.file_path,
+            "content": action.content,
+        }
+
+    def _parse_result(self, payload: dict[str, Any]) -> StepResult[Tbench2Observation]:
+        obs_data = payload.get("observation", {})
+        observation = Tbench2Observation(
+            instruction=obs_data.get("instruction", ""),
+            output=obs_data.get("output", ""),
+            success=obs_data.get("success", True),
+            error=obs_data.get("error", ""),
+            task_id=obs_data.get("task_id", ""),
+            task_path=obs_data.get("task_path", ""),
+            session_id=obs_data.get("session_id"),
+            action_type=obs_data.get("action_type", ""),
+            info=obs_data.get("info", {}),
+            reward=payload.get("reward"),
+            done=payload.get("done", False),
+            metadata=obs_data.get("metadata", {}),
+        )
+        return StepResult(
+            observation=observation,
+            reward=payload.get("reward"),
+            done=payload.get("done", False),
+        )
+
+    def _parse_state(self, payload: dict[str, Any]) -> Tbench2State:
+        return Tbench2State(
+            episode_id=payload.get("episode_id"),
+            step_count=payload.get("step_count", 0),
+            task_id=payload.get("task_id", ""),
+            task_path=payload.get("task_path", ""),
+            terminal_ready=payload.get("terminal_ready", False),
+            last_action_type=payload.get("last_action_type", ""),
+            last_command=payload.get("last_command", ""),
+            last_output=payload.get("last_output", ""),
+        )

envs/tbench2_env/models.py

@@ -0,0 +1,58 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Data models for the TB2 environment.
+"""
+
+from pydantic import Field
+
+
+# Support both in-repo and standalone imports
+try:
+    # In-repo imports (when running from OpenEnv repository)
+    from openenv.core.env_server.types import Action, Observation, State
+except ImportError:
+    # Standalone imports (when environment is standalone with openenv from pip)
+    from openenv.core.env_server.types import Action, Observation, State
+
+
+class Tbench2Action(Action):
+    """Action for interacting with a TB2 task session."""
+
+    action_type: str = Field(default="exec")
+    command: str = Field(default="")
+    session_id: str | None = Field(default=None)
+    block: bool = Field(default=True)
+    wait_seconds: float | None = Field(default=None)
+    file_path: str = Field(default="")
+    content: str = Field(default="")
+
+
+class Tbench2Observation(Observation):
+    """Observation returned from the TB2 environment."""
+
+    instruction: str = Field(default="")
+    output: str = Field(default="")
+    success: bool = Field(default=True)
+    error: str = Field(default="")
+    task_id: str = Field(default="")
+    task_path: str = Field(default="")
+    session_id: str | None = Field(default=None)
+    action_type: str = Field(default="")
+    info: dict = Field(default_factory=dict)
+
+
+class Tbench2State(State):
+    """Server-side state for a TB2 task."""
+
+    task_id: str = Field(default="")
+    task_path: str = Field(default="")
+    session_id: str = Field(default="")
+    terminal_ready: bool = Field(default=False)
+    last_action_type: str = Field(default="")
+    last_command: str = Field(default="")
+    last_output: str = Field(default="")

envs/tbench2_env/openenv.yaml

@@ -0,0 +1,7 @@
+spec_version: 1
+name: tbench2
+type: space
+runtime: fastapi
+app: server.app:app
+port: 8000
+

envs/tbench2_env/openenv_tbench2_env.egg-info/PKG-INFO

@@ -0,0 +1,13 @@
+Metadata-Version: 2.4
+Name: openenv-tbench2_env
+Version: 0.1.0
+Summary: Tbench2 Env environment for OpenEnv
+Requires-Python: >=3.10
+Requires-Dist: openenv-core[core]>=0.2.2
+Requires-Dist: pytest>=8.4.0
+Requires-Dist: camel-ai
+Requires-Dist: docker>=7.0.0
+Requires-Dist: tomli>=2.0.0; python_version < "3.11"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0.0; extra == "dev"

envs/tbench2_env/openenv_tbench2_env.egg-info/SOURCES.txt

@@ -0,0 +1,14 @@
+README.md
+pyproject.toml
+./__init__.py
+./client.py
+./models.py
+openenv_tbench2_env.egg-info/PKG-INFO
+openenv_tbench2_env.egg-info/SOURCES.txt
+openenv_tbench2_env.egg-info/dependency_links.txt
+openenv_tbench2_env.egg-info/entry_points.txt
+openenv_tbench2_env.egg-info/requires.txt
+openenv_tbench2_env.egg-info/top_level.txt
+server/__init__.py
+server/app.py
+server/tbench2_env_environment.py

envs/tbench2_env/openenv_tbench2_env.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

envs/tbench2_env/openenv_tbench2_env.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+server = tbench2_env.server.app:main

envs/tbench2_env/openenv_tbench2_env.egg-info/requires.txt

@@ -0,0 +1,11 @@
+openenv-core[core]>=0.2.2
+pytest>=8.4.0
+camel-ai
+docker>=7.0.0
+
+[:python_version < "3.11"]
+tomli>=2.0.0
+
+[dev]
+pytest>=8.0.0
+pytest-cov>=4.0.0

envs/tbench2_env/openenv_tbench2_env.egg-info/top_level.txt

@@ -0,0 +1 @@
+tbench2_env

envs/tbench2_env/pyproject.toml

@@ -0,0 +1,44 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+[build-system]
+requires = ["setuptools>=45", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "openenv-tbench2_env"
+version = "0.1.0"
+description = "Tbench2 Env environment for OpenEnv"
+requires-python = ">=3.10"
+dependencies = [
+    # Core OpenEnv runtime (provides FastAPI server + HTTP client types)
+    # install from github
+    "openenv-core[core] @ git+https://github.com/meta-pytorch/OpenEnv.git@v0.2.3",
+    "pytest>=8.4.0",
+    # Environment-specific dependencies
+    # Add all dependencies needed for your environment here
+    "camel-ai",
+    # Docker-backed mode (optional, for full TB2.0 compatibility)
+    "docker>=7.0.0",
+     # TOML parsing (tomllib for Python 3.11+, tomli for older versions)
+    "tomli>=2.0.0; python_version < '3.11'",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0.0",
+    "pytest-cov>=4.0.0",
+]
+
+[project.scripts]
+# Server entry point - enables running via: uv run --project . server
+# or: python -m tbench2_env.server.app
+server = "tbench2_env.server.app:main"
+
+[tool.setuptools]
+include-package-data = true
+packages = ["tbench2_env", "tbench2_env.server"]
+package-dir = { "tbench2_env" = ".", "tbench2_env.server" = "server" }

envs/tbench2_env/server/Dockerfile

@@ -0,0 +1,81 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Multi-stage build using openenv-base
+# This Dockerfile is flexible and works for both:
+# - In-repo environments (with local OpenEnv sources)
+# - Standalone environments (with openenv from PyPI/Git)
+# The build script (openenv build) handles context detection and sets appropriate build args.
+
+ARG BASE_IMAGE=ghcr.io/meta-pytorch/openenv-base:latest
+FROM ${BASE_IMAGE} AS builder
+
+WORKDIR /app
+
+# Ensure git is available (required for installing dependencies from VCS)
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends git gcc python3-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+# Build argument to control whether we're building standalone or in-repo
+ARG BUILD_MODE=in-repo
+ARG ENV_NAME=tbench2_env
+
+# Copy environment code (always at root of build context)
+COPY . /app/env
+
+# For in-repo builds, openenv is already vendored in the build context
+# For standalone builds, openenv will be installed via pyproject.toml
+WORKDIR /app/env
+
+# Ensure uv is available (for local builds where base image lacks it)
+RUN if ! command -v uv >/dev/null 2>&1; then \
+        curl -LsSf https://astral.sh/uv/install.sh | sh && \
+        mv /root/.local/bin/uv /usr/local/bin/uv && \
+        mv /root/.local/bin/uvx /usr/local/bin/uvx; \
+    fi
+    
+# Install dependencies using uv sync
+# If uv.lock exists, use it; otherwise resolve on the fly
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-install-project --no-editable; \
+    else \
+        uv sync --no-install-project --no-editable; \
+    fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-editable; \
+    else \
+        uv sync --no-editable; \
+    fi
+
+# Final runtime stage
+FROM ${BASE_IMAGE}
+
+WORKDIR /app
+
+# Copy the virtual environment from builder
+COPY --from=builder /app/env/.venv /app/.venv
+
+# Copy the environment code
+COPY --from=builder /app/env /app/env
+
+# Set PATH to use the virtual environment
+ENV PATH="/app/.venv/bin:$PATH"
+
+# Set PYTHONPATH so imports work correctly
+ENV PYTHONPATH="/app/env:$PYTHONPATH"
+ENV ENABLE_WEB_INTERFACE=true
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost:8000/health || exit 1
+
+# Run the FastAPI server
+# The module path is constructed to work with the /app/env structure
+CMD ["sh", "-c", "cd /app/env && uvicorn server.app:app --host 0.0.0.0 --port 8000"]

envs/tbench2_env/server/__init__.py

@@ -0,0 +1,12 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Tbench2 Env environment server components."""
+
+from .tbench2_env_environment import Tbench2DockerEnvironment, Tbench2Environment
+
+
+__all__ = ["Tbench2Environment", "Tbench2DockerEnvironment"]

envs/tbench2_env/server/app.py

@@ -0,0 +1,107 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+FastAPI application for the Tbench2 Env Environment.
+
+This module creates an HTTP server that exposes the Tbench2Environment
+over HTTP and WebSocket endpoints, compatible with EnvClient.
+
+Endpoints:
+    - POST /reset: Reset the environment
+    - POST /step: Execute an action
+    - GET /state: Get current environment state
+    - GET /schema: Get action/observation schemas
+    - WS /ws: WebSocket endpoint for persistent sessions
+
+Usage:
+    # Development (with auto-reload):
+    uvicorn server.app:app --reload --host 0.0.0.0 --port 8000
+
+    # Production:
+    uvicorn server.app:app --host 0.0.0.0 --port 8000 --workers 4
+
+    # Or run directly:
+    python -m server.app
+"""
+
+import os
+
+
+try:
+    from openenv.core.env_server.http_server import create_app
+
+    # In-repo imports
+    from tbench2_env.models import Tbench2Action, Tbench2Observation
+
+    from .tbench2_env_environment import Tbench2DockerEnvironment, Tbench2Environment
+except Exception as e:  # pragma: no cover
+    from models import Tbench2Action, Tbench2Observation
+
+    # Standalone imports (when environment is standalone with openenv from pip)
+    from openenv.core.env_server.http_server import create_app
+    from server.tbench2_env_environment import (
+        Tbench2DockerEnvironment,
+        Tbench2Environment,
+    )
+
+    _IMPORT_ERROR = e
+
+
+# Determine which environment class to use based on TB2_MODE
+_TB2_MODE = os.getenv("TB2_MODE", "local").lower()
+
+if _TB2_MODE == "docker":
+    _DEFAULT_ENVIRONMENT = Tbench2DockerEnvironment
+    _ENV_SUFFIX = " (Docker mode)"
+elif _TB2_MODE == "auto":
+    # Auto-detect: try Docker, fall back to local
+    _DEFAULT_ENVIRONMENT = Tbench2Environment
+    _ENV_SUFFIX = " (auto-detect mode)"
+else:
+    _DEFAULT_ENVIRONMENT = Tbench2Environment
+    _ENV_SUFFIX = " (local mode)"
+
+
+# Create the app with web interface and README integration
+app = create_app(
+    _DEFAULT_ENVIRONMENT,
+    Tbench2Action,
+    Tbench2Observation,
+    env_name="tbench2_env" + _ENV_SUFFIX,
+    max_concurrent_envs=1,  # increase this number to allow more concurrent WebSocket sessions
+)
+
+
+def main(host: str = "0.0.0.0", port: int = 8000):
+    """
+    Entry point for direct execution via uv run or python -m.
+
+    This function enables running the server without Docker:
+        uv run --project . server
+        uv run --project . server --port 8001
+        python -m tbench2_env.server.app
+
+    Args:
+        host: Host address to bind to (default: "0.0.0.0")
+        port: Port number to listen on (default: 8000)
+
+    For production deployments, consider using uvicorn directly with
+    multiple workers:
+        uvicorn tbench2_env.server.app:app --workers 4
+    """
+    import uvicorn
+
+    uvicorn.run(app, host=host, port=port)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--port", type=int, default=8000)
+    args = parser.parse_args()
+    main(port=args.port)

envs/tbench2_env/server/tbench2_env_environment.py

@@ -0,0 +1,728 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""TB2 environment server implementation (Spaces-compatible local mode)."""
+
+from __future__ import annotations
+
+import logging
+import os
+import sys
+import urllib.request
+import zipfile
+from pathlib import Path
+from typing import Any
+from uuid import uuid4
+
+
+if sys.version_info >= (3, 11):
+    import tomllib
+else:
+    import tomli as tomllib
+
+from openenv.core.env_server.interfaces import Environment
+
+
+# Support both in-repo and standalone imports
+try:
+    # In-repo imports (when running from OpenEnv repository)
+    from tbench2_env.models import Tbench2Action, Tbench2Observation, Tbench2State
+except ImportError:
+    # Standalone imports (when environment is standalone with openenv from pip)
+    from models import Tbench2Action, Tbench2Observation, Tbench2State
+
+_CAMEL_IMPORT_ERROR: Exception | None = None
+
+
+def _require_terminal_toolkit() -> Any:
+    global _CAMEL_IMPORT_ERROR
+    if _CAMEL_IMPORT_ERROR is not None:
+        raise RuntimeError(
+            "camel-ai (TerminalToolkit) is required for TB2. Install from PyPI or from the CAMEL repo."
+        ) from _CAMEL_IMPORT_ERROR
+
+    try:
+        from camel.toolkits import TerminalToolkit
+    except Exception as exc:  # pragma: no cover
+        _CAMEL_IMPORT_ERROR = exc
+        raise RuntimeError(
+            "camel-ai (TerminalToolkit) is required for TB2. Install from PyPI or from the CAMEL repo."
+        ) from exc
+
+    return TerminalToolkit
+
+
+def _download_tb2_repo(cache_dir: Path) -> Path:
+    repo_url = os.getenv(
+        "TB2_REPO_URL",
+        "https://github.com/laude-institute/terminal-bench-2/archive/refs/heads/main.zip",
+    )
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    archive_path = cache_dir / "terminal-bench-2.zip"
+
+    if not archive_path.exists():
+        urllib.request.urlretrieve(repo_url, archive_path)
+
+    with zipfile.ZipFile(archive_path) as zf:
+        root = zf.namelist()[0].split("/")[0]
+        extract_dir = cache_dir / root
+        if not extract_dir.exists():
+            zf.extractall(cache_dir)
+
+    return extract_dir
+
+
+def _read_instruction(task_dir: Path) -> str:
+    instruction_path = task_dir / "instruction.md"
+    if instruction_path.exists():
+        return instruction_path.read_text(encoding="utf-8")
+    return ""
+
+
+def _read_timeout(task_dir: Path, fallback: float) -> float:
+    task_toml = task_dir / "task.toml"
+    if not task_toml.exists():
+        return fallback
+    try:
+        data = tomllib.loads(task_toml.read_text(encoding="utf-8"))
+    except Exception:
+        return fallback
+    verifier = data.get("verifier", {})
+    return float(verifier.get("timeout_sec", fallback))
+
+
+class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2State]):
+    """OpenEnv wrapper around Terminal-Bench 2 tasks (local execution)."""
+
+    SUPPORTS_CONCURRENT_SESSIONS: bool = True
+
+    def __init__(
+        self,
+        tasks_dir: str | None = None,
+        output_dir: str | None = None,
+        command_timeout_s: float = 20.0,
+        safe_mode: bool = False,
+        default_task_id: str | None = None,
+    ) -> None:
+        super().__init__()
+        self.tasks_dir = tasks_dir or os.getenv("TB2_TASKS_DIR", "")
+        self.output_dir = Path(
+            output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs")
+        )
+        self.command_timeout_s = command_timeout_s
+        self.safe_mode = safe_mode
+        self.default_task_id = default_task_id or os.getenv(
+            "TB2_DEFAULT_TASK_ID", "headless-terminal"
+        )
+
+        self._state = Tbench2State()
+        self._task_dir: Path | None = None
+        self._terminal_toolkit = None
+        self._instruction = ""
+
+    def reset(
+        self,
+        seed: int | None = None,
+        episode_id: str | None = None,
+        **kwargs: Any,
+    ) -> Tbench2Observation:
+        del seed
+
+        TerminalToolkit = _require_terminal_toolkit()
+
+        task_id = (
+            kwargs.get("task_id") or kwargs.get("task_name") or self.default_task_id
+        )
+        task_path = kwargs.get("task_path") or kwargs.get("path")
+
+        task_dir = self._resolve_task_path(task_id, task_path)
+        resolved_task_id = task_id or task_dir.name
+
+        self._instruction = _read_instruction(task_dir)
+        self._task_dir = task_dir
+
+        trial_name = f"{resolved_task_id}.{episode_id or uuid4().hex}"
+        session_logs_dir = (
+            self.output_dir / trial_name / "terminal_toolkit_session_logs"
+        )
+        session_logs_dir.mkdir(parents=True, exist_ok=True)
+
+        self._terminal_toolkit = TerminalToolkit(
+            timeout=self.command_timeout_s,
+            working_directory=str(task_dir),
+            use_docker_backend=False,
+            session_logs_dir=session_logs_dir,
+            safe_mode=self.safe_mode,
+        )
+
+        self._state = Tbench2State(
+            episode_id=episode_id or str(uuid4()),
+            step_count=0,
+            task_id=resolved_task_id,
+            task_path=str(task_dir),
+            terminal_ready=True,
+        )
+
+        return Tbench2Observation(
+            instruction=self._instruction,
+            output="",
+            success=True,
+            error="",
+            task_id=resolved_task_id,
+            task_path=str(task_dir),
+            session_id=None,
+            action_type="reset",
+            info={},
+            reward=0.0,
+            done=False,
+        )
+
+    def step(
+        self,
+        action: Tbench2Action,
+        timeout_s: float | None = None,
+        **kwargs: Any,
+    ) -> Tbench2Observation:
+        del timeout_s, kwargs
+
+        if not isinstance(action, Tbench2Action):
+            raise TypeError(f"Expected Tbench2Action, got {type(action)}")
+
+        if self._terminal_toolkit is None or self._task_dir is None:
+            raise RuntimeError("TB2 environment not initialized. Call reset() first.")
+
+        self._state.step_count += 1
+        self._state.last_action_type = action.action_type
+        self._state.last_command = action.command
+
+        output = ""
+        error = ""
+        success = True
+        reward = None
+        done = False
+        info: dict[str, Any] = {}
+        session_id = action.session_id or "tb2-session"
+
+        try:
+            if action.action_type == "exec":
+                output = self._terminal_toolkit.shell_exec(
+                    command=action.command,
+                    block=action.block,
+                    id=session_id,
+                )
+            elif action.action_type == "write":
+                self._ensure_session_id(action.session_id, action.action_type)
+                output = self._terminal_toolkit.shell_write_to_process(
+                    id=action.session_id,
+                    command=action.command,
+                )
+            elif action.action_type == "view":
+                self._ensure_session_id(action.session_id, action.action_type)
+                output = self._terminal_toolkit.shell_view(id=action.session_id)
+            elif action.action_type == "wait":
+                self._ensure_session_id(action.session_id, action.action_type)
+                wait_seconds = action.wait_seconds or 0.0
+                output = self._terminal_toolkit.shell_wait(
+                    id=action.session_id,
+                    wait_seconds=wait_seconds,
+                )
+            elif action.action_type == "kill":
+                self._ensure_session_id(action.session_id, action.action_type)
+                self._terminal_toolkit.shell_kill_process(id=action.session_id)
+                output = f"Killed session {action.session_id}"
+            elif action.action_type == "write_file":
+                self._terminal_toolkit.shell_write_content_to_file(
+                    content=action.content,
+                    file_path=action.file_path,
+                )
+                output = f"Wrote content to {action.file_path}"
+            elif action.action_type == "evaluate":
+                output, reward, info = self._evaluate_task()
+                done = True
+            elif action.action_type == "close":
+                self.close()
+                output = "Closed TB2 environment."
+                done = True
+            else:
+                raise ValueError(f"Unsupported action_type: {action.action_type}")
+        except Exception as exc:  # pragma: no cover
+            success = False
+            error = str(exc)
+
+        self._state.last_output = output
+        self._state.session_id = session_id or ""
+
+        return Tbench2Observation(
+            instruction=self._instruction,
+            output=output,
+            success=success,
+            error=error,
+            task_id=self._state.task_id,
+            task_path=self._state.task_path,
+            session_id=session_id or "",
+            action_type=action.action_type,
+            info=info,
+            reward=reward,
+            done=done,
+        )
+
+    @property
+    def state(self) -> Tbench2State:
+        return self._state
+
+    def close(self) -> None:
+        self._terminal_toolkit = None
+        self._task_dir = None
+        self._instruction = ""
+
+    def _resolve_task_path(self, task_id: str | None, task_path: str | None) -> Path:
+        if task_path:
+            resolved = Path(task_path).expanduser().resolve()
+            if not resolved.exists():
+                raise FileNotFoundError(f"Task path not found: {resolved}")
+            return resolved
+
+        if not task_id:
+            raise ValueError("Provide task_id or task_path to reset TB2 environment.")
+
+        if not self.tasks_dir:
+            cache_dir = Path(
+                os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache"))
+            )
+            repo_dir = _download_tb2_repo(cache_dir)
+            resolved = repo_dir / task_id
+        else:
+            resolved = Path(self.tasks_dir).expanduser().resolve() / task_id
+
+        if not resolved.exists():
+            raise FileNotFoundError(f"Task path not found: {resolved}")
+        return resolved
+
+    def _ensure_session_id(self, session_id: str | None, action_type: str) -> None:
+        if not session_id:
+            raise ValueError(f"session_id is required for action_type='{action_type}'")
+
+    def _evaluate_task(self) -> tuple[str, float, dict[str, Any]]:
+        if self._task_dir is None:
+            raise RuntimeError("TB2 environment not initialized. Call reset() first.")
+        if self._terminal_toolkit is None:
+            raise RuntimeError("Terminal toolkit not initialized.")
+
+        _read_timeout(self._task_dir, fallback=900.0)  # Validate timeout config
+        tests_dir = self._task_dir / "tests"
+        cmd = f"cd {self._task_dir} && python -m pytest -q {tests_dir} -rA; echo __TB2_EXIT_CODE__:$?"
+        output = self._terminal_toolkit.shell_exec(
+            id="tb2-tests",
+            command=cmd,
+            block=True,
+        )
+
+        exit_code = 1
+        marker = "__TB2_EXIT_CODE__"
+        for line in output.splitlines()[::-1]:
+            if marker in line:
+                try:
+                    exit_code = int(line.split(":", 1)[1].strip())
+                except Exception:
+                    exit_code = 1
+                break
+
+        reward = 1.0 if exit_code == 0 else 0.0
+        info = {"tests_passed": exit_code == 0, "exit_code": exit_code}
+        return output, reward, info
+
+
+class Tbench2DockerEnvironment(
+    Environment[Tbench2Action, Tbench2Observation, Tbench2State]
+):
+    """OpenEnv wrapper around Terminal-Bench 2 tasks with Docker isolation.
+
+    This environment runs each task in its own Docker container, reading
+    the image specification from task.toml's [environment] section.
+
+    Requires:
+    - Docker socket mounted (/var/run/docker.sock)
+    - Sufficient disk space for container images
+    """
+
+    SUPPORTS_CONCURRENT_SESSIONS: bool = True
+
+    def __init__(
+        self,
+        tasks_dir: str | None = None,
+        output_dir: str | None = None,
+        command_timeout_s: float = 300.0,
+        safe_mode: bool = True,
+        default_task_id: str | None = None,
+    ) -> None:
+        super().__init__()
+        self.tasks_dir = tasks_dir or os.getenv("TB2_TASKS_DIR", "")
+        self.output_dir = Path(
+            output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs")
+        )
+        self.command_timeout_s = command_timeout_s
+        self.safe_mode = safe_mode
+        self.default_task_id = default_task_id or os.getenv(
+            "TB2_DEFAULT_TASK_ID", "headless-terminal"
+        )
+
+        self._state = Tbench2State()
+        self._task_dir: Path | None = None
+        self._docker_client = None
+        self._container = None
+        self._instruction = ""
+        self._task_image = ""
+        self._task_config: dict[str, Any] = {}
+
+    def _get_docker_client(self) -> Any:
+        """Lazy initialization of Docker client."""
+        if self._docker_client is None:
+            try:
+                import docker
+
+                self._docker_client = docker.from_env()
+            except Exception as exc:
+                raise RuntimeError(
+                    f"Docker client not available. Ensure Docker socket is mounted. Error: {exc}"
+                ) from exc
+        return self._docker_client
+
+    def reset(
+        self,
+        seed: int | None = None,
+        episode_id: str | None = None,
+        **kwargs: Any,
+    ) -> Tbench2Observation:
+        del seed
+
+        task_id = (
+            kwargs.get("task_id") or kwargs.get("task_name") or self.default_task_id
+        )
+        task_path = kwargs.get("task_path") or kwargs.get("path")
+
+        task_dir = self._resolve_task_path(task_id, task_path)
+        resolved_task_id = task_id or task_dir.name
+
+        # Read task configuration including Docker image
+        task_toml_path = task_dir / "task.toml"
+        if task_toml_path.exists():
+            self._task_config = tomllib.loads(
+                task_toml_path.read_text(encoding="utf-8")
+            )
+            self._task_image = self._task_config.get("environment", {}).get(
+                "docker_image", ""
+            )
+        else:
+            self._task_image = ""
+            self._task_config = {}
+
+        self._instruction = _read_instruction(task_dir)
+        self._task_dir = task_dir
+
+        # Create trial directory for logs
+        trial_name = f"{resolved_task_id}.{episode_id or uuid4().hex}"
+        trial_dir = self.output_dir / trial_name
+        trial_dir.mkdir(parents=True, exist_ok=True)
+
+        # Start Docker container if image is specified
+        if self._task_image:
+            self._start_container(task_dir, trial_dir)
+        else:
+            # Fallback to local mode if no image specified
+            self._state = Tbench2State(
+                episode_id=episode_id or str(uuid4()),
+                step_count=0,
+                task_id=resolved_task_id,
+                task_path=str(task_dir),
+                terminal_ready=not self._task_image,  # Ready if no container needed
+            )
+
+        return Tbench2Observation(
+            instruction=self._instruction,
+            output="",
+            success=True,
+            error="",
+            task_id=resolved_task_id,
+            task_path=str(task_dir),
+            session_id=None,
+            action_type="reset",
+            info={"docker_image": self._task_image} if self._task_image else {},
+            reward=0.0,
+            done=False,
+        )
+
+    def _start_container(self, task_dir: Path, trial_dir: Path) -> None:
+        """Start a Docker container for the task.
+
+        Uses file copying instead of bind mounts to support Docker-in-Docker
+        scenarios where the server runs inside a container. Bind mounts reference
+        host paths, which don't exist when the server is containerized.
+        """
+        docker = self._get_docker_client()
+
+        try:
+            # Pull image if needed
+            try:
+                docker.images.get(self._task_image)
+            except Exception:
+                logging.info(f"Pulling image {self._task_image}...")
+                docker.images.pull(self._task_image)
+
+            # Start container WITHOUT bind mounts (for DinD compatibility)
+            self._container = docker.containers.run(
+                image=self._task_image,
+                command="sleep infinity",
+                detach=True,
+                network_mode="host",
+                working_dir="/task",
+                remove=False,
+            )
+
+            # Copy task files into container using tar archive
+            # This works in Docker-in-Docker because we read files from our
+            # filesystem and stream them to the container via the Docker API
+            self._copy_dir_to_container(task_dir, "/task")
+
+            self._state = Tbench2State(
+                episode_id=str(uuid4()),
+                step_count=0,
+                task_id=task_dir.name,
+                task_path=str(task_dir),
+                terminal_ready=True,
+            )
+
+        except Exception as exc:
+            raise RuntimeError(f"Failed to start container: {exc}") from exc
+
+    def _copy_dir_to_container(self, src_dir: Path, dest_path: str) -> None:
+        """Copy a directory into the container using tar archive.
+
+        This method streams files via the Docker API, avoiding bind mount
+        issues in Docker-in-Docker scenarios.
+        """
+        import io
+        import tarfile
+
+        if self._container is None:
+            raise RuntimeError("Container not started")
+
+        # Create tar archive in memory
+        tar_stream = io.BytesIO()
+        with tarfile.open(fileobj=tar_stream, mode="w") as tar:
+            for item in src_dir.rglob("*"):
+                arcname = str(item.relative_to(src_dir))
+                tar.add(str(item), arcname=arcname)
+
+        tar_stream.seek(0)
+
+        # Copy to container
+        self._container.put_archive(dest_path, tar_stream.getvalue())
+
+    def _exec_in_container(
+        self, command: str, workdir: str = "/task"
+    ) -> tuple[int, str]:
+        """Execute a command inside the container."""
+        if self._container is None:
+            raise RuntimeError("Container not started. Call reset() first.")
+
+        exit_code, output = self._container.exec_run(
+            cmd=f"bash -c 'cd {workdir} && {command}'",
+            workdir="/task",
+            stdout=True,
+            stderr=True,
+        )
+        return exit_code, output.decode("utf-8", errors="replace")
+
+    def step(
+        self,
+        action: Tbench2Action,
+        timeout_s: float | None = None,
+        **kwargs: Any,
+    ) -> Tbench2Observation:
+        del timeout_s, kwargs
+
+        if not isinstance(action, Tbench2Action):
+            raise TypeError(f"Expected Tbench2Action, got {type(action)}")
+
+        if self._task_dir is None:
+            raise RuntimeError("TB2 environment not initialized. Call reset() first.")
+
+        self._state.step_count += 1
+        self._state.last_action_type = action.action_type
+        self._state.last_command = action.command
+
+        output = ""
+        error = ""
+        success = True
+        reward = None
+        done = False
+        info: dict[str, Any] = {}
+        session_id = action.session_id or "tb2-session"
+
+        try:
+            if action.action_type == "exec":
+                if self._container:
+                    exit_code, output = self._exec_in_container(action.command)
+                    success = exit_code == 0
+                else:
+                    # Fallback to local execution
+                    import subprocess
+
+                    result = subprocess.run(
+                        action.command,
+                        shell=True,
+                        capture_output=True,
+                        text=True,
+                        timeout=self.command_timeout_s,
+                    )
+                    output = result.stdout + result.stderr
+                    success = result.returncode == 0
+
+            elif action.action_type == "write_file":
+                if self._container:
+                    # Write to container
+                    exit_code, _ = self._exec_in_container(
+                        f"cat > {action.file_path} << 'EOF'\n{action.content}\nEOF"
+                    )
+                    success = exit_code == 0
+                    output = f"Wrote to {action.file_path}"
+                else:
+                    # Local write
+                    Path(action.file_path).write_text(action.content)
+                    output = f"Wrote to {action.file_path}"
+
+            elif action.action_type == "evaluate":
+                if self._container:
+                    output, reward, info = self._evaluate_docker()
+                else:
+                    output, reward, info = self._evaluate_local()
+                done = True
+
+            elif action.action_type == "close":
+                self.close()
+                output = "Closed TB2 environment."
+                done = True
+
+            else:
+                raise ValueError(
+                    f"Unsupported action_type in Docker mode: {action.action_type}"
+                )
+
+        except Exception as exc:
+            success = False
+            error = str(exc)
+
+        self._state.last_output = output
+        self._state.session_id = session_id or ""
+
+        return Tbench2Observation(
+            instruction=self._instruction,
+            output=output,
+            success=success,
+            error=error,
+            task_id=self._state.task_id,
+            task_path=self._state.task_path,
+            session_id=session_id or "",
+            action_type=action.action_type,
+            info=info,
+            reward=reward,
+            done=done,
+        )
+
+    def _evaluate_docker(self) -> tuple[str, float, dict[str, Any]]:
+        """Evaluate task inside Docker container."""
+        if self._container is None:
+            raise RuntimeError("Container not started.")
+        assert self._task_dir is not None, "Task directory not set"
+
+        # Run pytest in the container's /task directory
+        # Use exit code marker for consistency with local mode
+        cmd = "cd /task && python -m pytest -q tests/ -rA; echo __TB2_EXIT_CODE__:$?"
+
+        exit_code, output = self._container.exec_run(
+            cmd=f"bash -c '{cmd}'",
+            workdir="/task",
+            stdout=True,
+            stderr=True,
+        )
+        output_str = output.decode("utf-8", errors="replace")
+
+        # Parse exit code from marker (same logic as local mode)
+        ec = 1
+        marker = "__TB2_EXIT_CODE__"
+        for line in output_str.splitlines()[::-1]:
+            if marker in line:
+                try:
+                    ec = int(line.split(":", 1)[1].strip())
+                except Exception:
+                    ec = 1
+                break
+
+        reward = 1.0 if ec == 0 else 0.0
+        info = {"tests_passed": ec == 0, "exit_code": ec}
+        return output_str, reward, info
+
+    def _evaluate_local(self) -> tuple[str, float, dict[str, Any]]:
+        """Evaluate task locally (fallback)."""
+        if self._task_dir is None:
+            raise RuntimeError("Task not initialized.")
+
+        tests_dir = self._task_dir / "tests"
+        cmd = f"cd {self._task_dir} && python -m pytest -q {tests_dir} -rA; echo __TB2_EXIT_CODE__:$?"
+
+        import subprocess
+
+        result = subprocess.run(
+            cmd,
+            shell=True,
+            capture_output=True,
+            text=True,
+            timeout=900.0,
+        )
+        output = result.stdout + result.stderr
+        exit_code = result.returncode
+
+        reward = 1.0 if exit_code == 0 else 0.0
+        info = {"tests_passed": exit_code == 0, "exit_code": exit_code}
+        return output, reward, info
+
+    @property
+    def state(self) -> Tbench2State:
+        return self._state
+
+    def close(self) -> None:
+        if self._container:
+            try:
+                self._container.stop(timeout=10)
+                self._container.remove(force=True)
+            except Exception:
+                pass
+            self._container = None
+        self._task_dir = None
+        self._instruction = ""
+
+    def _resolve_task_path(self, task_id: str | None, task_path: str | None) -> Path:
+        if task_path:
+            resolved = Path(task_path).expanduser().resolve()
+            if not resolved.exists():
+                raise FileNotFoundError(f"Task path not found: {resolved}")
+            return resolved
+
+        if not task_id:
+            raise ValueError("Provide task_id or task_path to reset TB2 environment.")
+
+        if not self.tasks_dir:
+            cache_dir = Path(
+                os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache"))
+            )
+            repo_dir = _download_tb2_repo(cache_dir)
+            resolved = repo_dir / task_id
+        else:
+            resolved = Path(self.tasks_dir).expanduser().resolve() / task_id
+
+        if not resolved.exists():
+            raise FileNotFoundError(f"Task path not found: {resolved}")
+        return resolved

openenv_tbench2_env.egg-info/PKG-INFO

@@ -0,0 +1,13 @@
+Metadata-Version: 2.4
+Name: openenv-tbench2_env
+Version: 0.1.0
+Summary: Tbench2 Env environment for OpenEnv
+Requires-Python: >=3.10
+Requires-Dist: openenv-core[core]>=0.2.2
+Requires-Dist: pytest>=8.4.0
+Requires-Dist: camel-ai
+Requires-Dist: docker>=7.0.0
+Requires-Dist: tomli>=2.0.0; python_version < "3.11"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0.0; extra == "dev"

openenv_tbench2_env.egg-info/SOURCES.txt

@@ -0,0 +1,14 @@
+README.md
+pyproject.toml
+./__init__.py
+./client.py
+./models.py
+openenv_tbench2_env.egg-info/PKG-INFO
+openenv_tbench2_env.egg-info/SOURCES.txt
+openenv_tbench2_env.egg-info/dependency_links.txt
+openenv_tbench2_env.egg-info/entry_points.txt
+openenv_tbench2_env.egg-info/requires.txt
+openenv_tbench2_env.egg-info/top_level.txt
+server/__init__.py
+server/app.py
+server/tbench2_env_environment.py

openenv_tbench2_env.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

openenv_tbench2_env.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+server = tbench2_env.server.app:main

openenv_tbench2_env.egg-info/requires.txt

@@ -0,0 +1,11 @@
+openenv-core[core]>=0.2.2
+pytest>=8.4.0
+camel-ai
+docker>=7.0.0
+
+[:python_version < "3.11"]
+tomli>=2.0.0
+
+[dev]
+pytest>=8.0.0
+pytest-cov>=4.0.0

openenv_tbench2_env.egg-info/top_level.txt

@@ -0,0 +1 @@
+tbench2_env

pyproject.toml

@@ -16,7 +16,7 @@ requires-python = ">=3.10"
 dependencies = [
     # Core OpenEnv runtime (provides FastAPI server + HTTP client types)
     # install from github
-    "openenv-core @ git+https://github.com/meta-pytorch/OpenEnv.git@v0.2.1",
+    "openenv-core[core] @ git+https://github.com/meta-pytorch/OpenEnv.git@v0.2.3",
     "pytest>=8.4.0",
     # Environment-specific dependencies
     # Add all dependencies needed for your environment here

server/Dockerfile

@@ -0,0 +1,81 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Multi-stage build using openenv-base
+# This Dockerfile is flexible and works for both:
+# - In-repo environments (with local OpenEnv sources)
+# - Standalone environments (with openenv from PyPI/Git)
+# The build script (openenv build) handles context detection and sets appropriate build args.
+
+ARG BASE_IMAGE=ghcr.io/meta-pytorch/openenv-base:latest
+FROM ${BASE_IMAGE} AS builder
+
+WORKDIR /app
+
+# Ensure git is available (required for installing dependencies from VCS)
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends git gcc python3-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+# Build argument to control whether we're building standalone or in-repo
+ARG BUILD_MODE=in-repo
+ARG ENV_NAME=tbench2_env
+
+# Copy environment code (always at root of build context)
+COPY . /app/env
+
+# For in-repo builds, openenv is already vendored in the build context
+# For standalone builds, openenv will be installed via pyproject.toml
+WORKDIR /app/env
+
+# Ensure uv is available (for local builds where base image lacks it)
+RUN if ! command -v uv >/dev/null 2>&1; then \
+        curl -LsSf https://astral.sh/uv/install.sh | sh && \
+        mv /root/.local/bin/uv /usr/local/bin/uv && \
+        mv /root/.local/bin/uvx /usr/local/bin/uvx; \
+    fi
+    
+# Install dependencies using uv sync
+# If uv.lock exists, use it; otherwise resolve on the fly
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-install-project --no-editable; \
+    else \
+        uv sync --no-install-project --no-editable; \
+    fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    if [ -f uv.lock ]; then \
+        uv sync --frozen --no-editable; \
+    else \
+        uv sync --no-editable; \
+    fi
+
+# Final runtime stage
+FROM ${BASE_IMAGE}
+
+WORKDIR /app
+
+# Copy the virtual environment from builder
+COPY --from=builder /app/env/.venv /app/.venv
+
+# Copy the environment code
+COPY --from=builder /app/env /app/env
+
+# Set PATH to use the virtual environment
+ENV PATH="/app/.venv/bin:$PATH"
+
+# Set PYTHONPATH so imports work correctly
+ENV PYTHONPATH="/app/env:$PYTHONPATH"
+ENV ENABLE_WEB_INTERFACE=true
+
+# Health check
+HEALTHCHECK --interval=30s --timeout=3s --start-period=5s --retries=3 \
+    CMD curl -f http://localhost:8000/health || exit 1
+
+# Run the FastAPI server
+# The module path is constructed to work with the /app/env structure
+CMD ["sh", "-c", "cd /app/env && uvicorn server.app:app --host 0.0.0.0 --port 8000"]

server/app.py

@@ -39,11 +39,14 @@ try:
 
     from .tbench2_env_environment import Tbench2DockerEnvironment, Tbench2Environment
 except Exception as e:  # pragma: no cover
+    from models import Tbench2Action, Tbench2Observation
+
     # Standalone imports (when environment is standalone with openenv from pip)
     from openenv.core.env_server.http_server import create_app
-    from server.tbench2_env_environment import Tbench2DockerEnvironment, Tbench2Environment
-
-    from models import Tbench2Action, Tbench2Observation
+    from server.tbench2_env_environment import (
+        Tbench2DockerEnvironment,
+        Tbench2Environment,
+    )
 
     _IMPORT_ERROR = e
 

server/tbench2_env_environment.py

@@ -105,12 +105,18 @@ class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2S
         output_dir: str | None = None,
         command_timeout_s: float = 20.0,
         safe_mode: bool = False,
+        default_task_id: str | None = None,
     ) -> None:
         super().__init__()
         self.tasks_dir = tasks_dir or os.getenv("TB2_TASKS_DIR", "")
-        self.output_dir = Path(output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs"))
+        self.output_dir = Path(
+            output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs")
+        )
         self.command_timeout_s = command_timeout_s
         self.safe_mode = safe_mode
+        self.default_task_id = default_task_id or os.getenv(
+            "TB2_DEFAULT_TASK_ID", "headless-terminal"
+        )
 
         self._state = Tbench2State()
         self._task_dir: Path | None = None
@@ -127,7 +133,9 @@ class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2S
 
         TerminalToolkit = _require_terminal_toolkit()
 
-        task_id = kwargs.get("task_id") or kwargs.get("task_name")
+        task_id = (
+            kwargs.get("task_id") or kwargs.get("task_name") or self.default_task_id
+        )
         task_path = kwargs.get("task_path") or kwargs.get("path")
 
         task_dir = self._resolve_task_path(task_id, task_path)
@@ -137,7 +145,9 @@ class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2S
         self._task_dir = task_dir
 
         trial_name = f"{resolved_task_id}.{episode_id or uuid4().hex}"
-        session_logs_dir = self.output_dir / trial_name / "terminal_toolkit_session_logs"
+        session_logs_dir = (
+            self.output_dir / trial_name / "terminal_toolkit_session_logs"
+        )
         session_logs_dir.mkdir(parents=True, exist_ok=True)
 
         self._terminal_toolkit = TerminalToolkit(
@@ -279,7 +289,9 @@ class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2S
             raise ValueError("Provide task_id or task_path to reset TB2 environment.")
 
         if not self.tasks_dir:
-            cache_dir = Path(os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache")))
+            cache_dir = Path(
+                os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache"))
+            )
             repo_dir = _download_tb2_repo(cache_dir)
             resolved = repo_dir / task_id
         else:
@@ -323,7 +335,9 @@ class Tbench2Environment(Environment[Tbench2Action, Tbench2Observation, Tbench2S
         return output, reward, info
 
 
-class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tbench2State]):
+class Tbench2DockerEnvironment(
+    Environment[Tbench2Action, Tbench2Observation, Tbench2State]
+):
     """OpenEnv wrapper around Terminal-Bench 2 tasks with Docker isolation.
 
     This environment runs each task in its own Docker container, reading
@@ -342,12 +356,18 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
         output_dir: str | None = None,
         command_timeout_s: float = 300.0,
         safe_mode: bool = True,
+        default_task_id: str | None = None,
     ) -> None:
         super().__init__()
         self.tasks_dir = tasks_dir or os.getenv("TB2_TASKS_DIR", "")
-        self.output_dir = Path(output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs"))
+        self.output_dir = Path(
+            output_dir or os.getenv("TB2_OUTPUT_DIR", "/tmp/tbench2_env_runs")
+        )
         self.command_timeout_s = command_timeout_s
         self.safe_mode = safe_mode
+        self.default_task_id = default_task_id or os.getenv(
+            "TB2_DEFAULT_TASK_ID", "headless-terminal"
+        )
 
         self._state = Tbench2State()
         self._task_dir: Path | None = None
@@ -378,7 +398,9 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
     ) -> Tbench2Observation:
         del seed
 
-        task_id = kwargs.get("task_id") or kwargs.get("task_name")
+        task_id = (
+            kwargs.get("task_id") or kwargs.get("task_name") or self.default_task_id
+        )
         task_path = kwargs.get("task_path") or kwargs.get("path")
 
         task_dir = self._resolve_task_path(task_id, task_path)
@@ -387,8 +409,12 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
         # Read task configuration including Docker image
         task_toml_path = task_dir / "task.toml"
         if task_toml_path.exists():
-            self._task_config = tomllib.loads(task_toml_path.read_text(encoding="utf-8"))
-            self._task_image = self._task_config.get("environment", {}).get("docker_image", "")
+            self._task_config = tomllib.loads(
+                task_toml_path.read_text(encoding="utf-8")
+            )
+            self._task_image = self._task_config.get("environment", {}).get(
+                "docker_image", ""
+            )
         else:
             self._task_image = ""
             self._task_config = {}
@@ -495,7 +521,9 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
         # Copy to container
         self._container.put_archive(dest_path, tar_stream.getvalue())
 
-    def _exec_in_container(self, command: str, workdir: str = "/task") -> tuple[int, str]:
+    def _exec_in_container(
+        self, command: str, workdir: str = "/task"
+    ) -> tuple[int, str]:
         """Execute a command inside the container."""
         if self._container is None:
             raise RuntimeError("Container not started. Call reset() first.")
@@ -556,7 +584,9 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
             elif action.action_type == "write_file":
                 if self._container:
                     # Write to container
-                    exit_code, _ = self._exec_in_container(f"cat > {action.file_path} << 'EOF'\n{action.content}\nEOF")
+                    exit_code, _ = self._exec_in_container(
+                        f"cat > {action.file_path} << 'EOF'\n{action.content}\nEOF"
+                    )
                     success = exit_code == 0
                     output = f"Wrote to {action.file_path}"
                 else:
@@ -577,7 +607,9 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
                 done = True
 
             else:
-                raise ValueError(f"Unsupported action_type in Docker mode: {action.action_type}")
+                raise ValueError(
+                    f"Unsupported action_type in Docker mode: {action.action_type}"
+                )
 
         except Exception as exc:
             success = False
@@ -683,7 +715,9 @@ class Tbench2DockerEnvironment(Environment[Tbench2Action, Tbench2Observation, Tb
             raise ValueError("Provide task_id or task_path to reset TB2 environment.")
 
         if not self.tasks_dir:
-            cache_dir = Path(os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache")))
+            cache_dir = Path(
+                os.getenv("TB2_CACHE_DIR", str(self.output_dir / "repo_cache"))
+            )
             repo_dir = _download_tb2_repo(cache_dir)
             resolved = repo_dir / task_id
         else:

src/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""EnvTorch: Standardized agentic execution environments."""

src/core/README.md

@@ -0,0 +1,212 @@
+# <img width="35" height="35" alt="image" src="https://github.com/user-attachments/assets/2700a971-e5d6-4036-b03f-2f89c9791609" /> OpenEnv: Agentic Execution Environments
+
+An e2e framework for creating, deploying and using isolated execution environments for agentic RL training, built using Gymnasium style simple APIs. OpenEnv provides a standard for interacting with agentic execution environments via simple Gymnasium style APIs - step(), reset(), state(). Users of agentic execution environments can interact with the environment during RL training loops using these simple APIs.
+
+In addition to making it easier for researchers and RL framework writers, we also provide tools for environment creators making it easier for them to create richer environments and make them available over familiar protocols like HTTP and packaged using canonical technologies like docker. Environment creators can use the OpenEnv framework to create environments that are isolated, secure, and easy to deploy and use.
+
+
+## Overview
+`openenv.core` provides the foundational building blocks for creating and interacting with containerized environments over HTTP. It enables you to build agent environments that can be deployed as Docker containers and accessed via a simple HTTP API.
+
+> ⚠️ **Early Development Warning** OpenEnv is currently in an experimental
+> stage. You should expect bugs, incomplete features, and APIs that may change
+> in future versions. The project welcomes bugfixes, but to make sure things are
+> well coordinated you should discuss any significant change before starting the
+> work. It's recommended that you signal your intention to contribute in the
+> issue tracker, either by filing a new issue or by claiming an existing one.
+
+
+# OpenEnv Core
+
+Core components for OpenEnv - a framework for building HTTP-based agentic environments.
+
+## Features
+
+- **EnvClient**: Async-first client for interacting with remote environments
+- **SyncEnvClient**: Synchronous wrapper via `.sync()` for sync codebases
+- **HTTPEnvServer**: FastAPI-based server wrapper for exposing environments over HTTP/WebSocket
+- **Container Providers**: Pluggable architecture for running containers (Docker, Kubernetes, etc.)
+- **Type System**: Strongly-typed Action/Observation/State interfaces
+- **Web Interface**: Optional web UI for interacting with environments
+
+## Installation
+
+```bash
+pip install "openenv[core]"
+```
+
+For development:
+```bash
+pip install "openenv[core]"
+```
+
+## Quick Start
+
+### Creating an Environment Client
+
+EnvClient is **async by default**. Use `async with` and `await` for all operations:
+
+```python
+import asyncio
+from openenv.core import EnvClient, StepResult
+from dataclasses import dataclass
+from typing import Any
+
+@dataclass
+class MyAction:
+    text: str
+
+@dataclass
+class MyObservation:
+    response: str
+
+class MyEnvClient(EnvClient[MyAction, MyObservation, Any]):
+    def _step_payload(self, action: MyAction) -> dict:
+        return {"text": action.text}
+
+    def _parse_result(self, payload: dict) -> StepResult[MyObservation]:
+        obs_data = payload["observation"]
+        return StepResult(
+            observation=MyObservation(**obs_data),
+            reward=payload.get("reward"),
+            done=payload.get("done", False)
+        )
+
+    def _parse_state(self, payload: dict) -> Any:
+        return payload
+
+# Async usage (recommended)
+async def main():
+    client = await MyEnvClient.from_docker_image("my-env:latest")
+    async with client:
+        result = await client.reset()
+        step_result = await client.step(MyAction(text="hello"))
+
+asyncio.run(main())
+
+# Sync usage (via .sync() wrapper)
+with MyEnvClient(base_url="http://localhost:8000").sync() as client:
+    result = client.reset()
+    step_result = client.step(MyAction(text="hello"))
+```
+
+### Creating an Environment Server
+
+```python
+from openenv.core.env_server import Environment, HTTPEnvServer, create_app
+from dataclasses import dataclass
+
+@dataclass
+class MyAction:
+    text: str
+
+@dataclass
+class MyObservation:
+    response: str
+    reward: float = 0.0
+    done: bool = False
+
+class MyEnvironment(Environment):
+    def reset(self) -> MyObservation:
+        return MyObservation(response="Ready")
+
+    def step(self, action: MyAction) -> MyObservation:
+        return MyObservation(
+            response=f"Echo: {action.text}",
+            reward=1.0,
+            done=False
+        )
+
+# Create FastAPI app
+env = MyEnvironment()
+app = create_app(env, MyAction, MyObservation)
+
+# Run with: uvicorn module:app --host 0.0.0.0 --port 8000
+```
+
+## Container Providers
+
+OpenEnv Core supports multiple container providers:
+
+### Local Docker Provider
+
+```python
+from openenv.core.containers.runtime import LocalDockerProvider
+
+provider = LocalDockerProvider()
+base_url = provider.start_container("my-env:latest")
+provider.wait_for_ready(base_url)
+# Use environment...
+provider.stop_container()
+```
+
+### Kubernetes Provider (Coming Soon)
+
+```python
+from openenv.core.containers.runtime import KubernetesProvider
+
+provider = KubernetesProvider(namespace="envs")
+base_url = provider.start_container("my-env:latest")
+# Use environment...
+provider.stop_container()
+```
+
+
+## API Reference
+
+### EnvClient
+
+Async base class for environment clients. Key methods:
+
+- `async connect()`: Establish WebSocket connection
+- `async reset(**kwargs)`: Reset environment
+- `async step(action)`: Execute action
+- `async state()`: Get current state
+- `async close()`: Close connection and cleanup
+- `sync()`: Return a SyncEnvClient wrapper for synchronous usage
+
+Abstract methods to implement:
+- `_step_payload(action)`: Convert action to JSON
+- `_parse_result(payload)`: Parse response to StepResult
+- `_parse_state(payload)`: Parse state response
+
+### SyncEnvClient
+
+Synchronous wrapper around EnvClient. Use `client.sync()` to get one:
+
+```python
+sync_client = async_client.sync()
+with sync_client:
+    result = sync_client.reset()
+    result = sync_client.step(action)
+```
+
+### HTTPEnvServer
+
+Server wrapper with these methods:
+
+- `register_routes(app)`: Register endpoints on FastAPI app
+- `_deserialize_action(data)`: Convert JSON to Action
+- `_serialize_observation(obs)`: Convert Observation to JSON
+
+### Environment Interface
+
+Base interface for environment implementations:
+
+- `reset()`: Reset environment and return initial observation
+- `step(action)`: Execute action and return observation
+- `state`: Property returning current environment state
+
+## License
+
+This project is licensed under the BSD-3-Clause License - see the LICENSE file for details.
+
+## Contributing
+
+Contributions are welcome! Please see the main OpenEnv repository for contribution guidelines.
+
+## Links
+
+- **Homepage**: https://github.com/meta-pytorch/OpenEnv
+- **Documentation**: https://github.com/meta-pytorch/OpenEnv/blob/main/README.md
+- **Bug Tracker**: https://github.com/meta-pytorch/OpenEnv/issues

src/core/__init__.py

@@ -0,0 +1,81 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Core components for agentic environments."""
+
+from __future__ import annotations
+
+from importlib import import_module
+from typing import TYPE_CHECKING
+
+from . import env_server
+from .env_server import *  # noqa: F403
+
+if TYPE_CHECKING:
+    from .env_client import EnvClient
+    from .generic_client import GenericAction, GenericEnvClient
+    from .llm_client import (
+        AnthropicClient,
+        create_llm_client,
+        LLMClient,
+        LLMResponse,
+        OpenAIClient,
+        ToolCall,
+    )
+    from .mcp_client import MCPClientBase, MCPToolClient
+    from .sync_client import SyncEnvClient
+
+__all__ = [
+    "EnvClient",
+    "SyncEnvClient",
+    "GenericEnvClient",
+    "GenericAction",
+    "MCPClientBase",
+    "MCPToolClient",
+    "AnthropicClient",
+    "LLMClient",
+    "LLMResponse",
+    "OpenAIClient",
+    "ToolCall",
+    "create_llm_client",
+] + env_server.__all__  # type: ignore
+
+
+_LAZY_ATTRS = {
+    "EnvClient": (".env_client", "EnvClient"),
+    "SyncEnvClient": (".sync_client", "SyncEnvClient"),
+    "GenericEnvClient": (".generic_client", "GenericEnvClient"),
+    "GenericAction": (".generic_client", "GenericAction"),
+    "MCPClientBase": (".mcp_client", "MCPClientBase"),
+    "MCPToolClient": (".mcp_client", "MCPToolClient"),
+    "AnthropicClient": (".llm_client", "AnthropicClient"),
+    "LLMClient": (".llm_client", "LLMClient"),
+    "LLMResponse": (".llm_client", "LLMResponse"),
+    "OpenAIClient": (".llm_client", "OpenAIClient"),
+    "ToolCall": (".llm_client", "ToolCall"),
+    "create_llm_client": (".llm_client", "create_llm_client"),
+}
+
+
+def __getattr__(name: str):
+    if name in _LAZY_ATTRS:
+        module_path, attr_name = _LAZY_ATTRS[name]
+        module = import_module(module_path, __name__)
+        value = getattr(module, attr_name)
+        globals()[name] = value
+        return value
+
+    try:
+        value = getattr(env_server, name)
+    except AttributeError as exc:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}") from exc
+
+    globals()[name] = value
+    return value
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals().keys()) | set(__all__))

src/core/client_types.py

@@ -0,0 +1,23 @@
+# Type definitions for EnvTorch
+from dataclasses import dataclass
+from typing import Generic, Optional, TypeVar
+
+# Generic type for observations
+ObsT = TypeVar("ObsT")
+StateT = TypeVar("StateT")
+
+
+@dataclass
+class StepResult(Generic[ObsT]):
+    """
+    Represents the result of one environment step.
+
+    Attributes:
+        observation: The environment's observation after the action.
+        reward: Scalar reward for this step (optional).
+        done: Whether the episode is finished.
+    """
+
+    observation: ObsT
+    reward: Optional[float] = None
+    done: bool = False

src/core/containers/__init__.py

@@ -0,0 +1,7 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Container management for environment servers."""

src/core/containers/images/Dockerfile

@@ -0,0 +1,64 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+#
+# OpenEnv Base Image
+#
+# This is the standard base image for all OpenEnv environment servers.
+# It includes the minimal dependencies needed to run HTTP environment servers
+# and uv for fast dependency management.
+#
+# Build from repo root: docker build -t openenv-base:latest -f src/openenv/core/containers/images/Dockerfile .
+# Tag:   docker tag openenv-base:latest openenv-base:0.2.0
+#
+
+FROM ghcr.io/astral-sh/uv:0.5.27-python3.11-bookworm-slim AS builder
+
+# Set working directory
+WORKDIR /app
+
+# Copy core pyproject.toml and lockfile for dependency installation
+COPY pyproject.toml uv.lock* ./
+
+# Install core dependencies using uv with cache mount
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r pyproject.toml
+
+# Final runtime stage
+FROM python:3.11-slim
+
+# Set metadata
+LABEL maintainer="OpenEnv Team"
+LABEL description="Base image for OpenEnv based environment servers with uv"
+LABEL version="0.2.0"
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    curl \
+    ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy uv from builder
+COPY --from=builder /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
+
+# Copy installed Python packages from builder
+COPY --from=builder /usr/local/lib/python3.11/site-packages /usr/local/lib/python3.11/site-packages
+
+# Copy console scripts installed by pip (uvicorn, fastapi, etc.)
+COPY --from=builder /usr/local/bin/uvicorn /usr/local/bin/fastapi /usr/local/bin/
+
+# Set working directory
+WORKDIR /app
+
+# Default environment variables
+ENV PYTHONPATH=/app/src
+ENV PYTHONUNBUFFERED=1
+ENV UV_SYSTEM_PYTHON=1
+
+# Default expose port (can be overridden)
+EXPOSE 8000
+
+# Note: CMD should be specified in child Dockerfiles

src/core/containers/images/README.md

@@ -0,0 +1,92 @@
+# OpenEnv Base Image
+
+Standard base image for all OpenEnv environment servers.
+
+## What's Included
+
+| Layer | Size | Contents |
+|-------|------|----------|
+| python:3.11-slim | 200 MB  | Base Python runtime |
+| + Dependencies   | 100 MB  | FastAPI, uvicorn, requests |
+| **Total**        | **~300 MB** | Ready for environment servers |
+
+## Image Sizes
+
+```
+openenv-base:latest   300 MB  (python + fastapi + uvicorn)
+```
+echo-env:latest        500 MB  (python + fastapi + uvicorn + app)
+coding-env:latest      520 MB  (python + fastapi + uvicorn + app + tools)
+another-env:latest     510 MB  (python + fastapi + uvicorn + app)
+---
+Total: 1.5 GB (with lots of duplication)
+```
+
+### With Base Images (✅ Solution)
+```
+openenv-base:latest    300 MB  (python + fastapi + uvicorn)
+echo-env:latest         50 MB  (app only, uses base)
+coding-env:latest       70 MB  (app + tools, uses base)
+another-env:latest      45 MB  (app only, uses base)
+---
+Total: 465 MB (base shared, minimal duplication)
+```
+
+## Building the Base Image
+
+```bash
+# From project root
+docker build -t openenv-base:latest -f src/openenv/core/containers/images/Dockerfile .
+```
+
+## Usage in Environment Dockerfiles
+
+Each environment Dockerfile should start with:
+
+```dockerfile
+FROM openenv-base:latest
+
+# Copy only environment-specific files
+COPY src/openenv/core/ /app/src/openenv/core/
+COPY envs/my_env/ /app/envs/my_env/
+
+# Run the server
+CMD ["uvicorn", "envs.my_env.server.app:app", "--host", "0.0.0.0", "--port", "8000"]
+```
+
+## Base Image Contents
+
+- Python 3.11-slim
+- FastAPI >= 0.104.0
+- Uvicorn >= 0.24.0
+- Requests >= 2.25.0
+- curl (for health checks)
+
+## Example: Building Echo Environment
+
+```bash
+# Step 1: Build base image (do this once)
+docker build -t openenv-base:latest -f src/openenv/core/containers/images/Dockerfile .
+
+# Step 2: Build echo environment (uses base)
+docker build -t echo-env:latest -f envs/echo_env/server/Dockerfile .
+
+# Step 3: Run echo environment
+docker run -p 8000:8000 echo-env:latest
+```
+
+## Updating the Base
+
+When dependencies need updating:
+
+1. Update `src/openenv/core/containers/images/Dockerfile`
+2. Rebuild base image
+3. Rebuild all environment images (they'll use new base)
+
+```bash
+# Update base
+docker build -t openenv-base:latest -f src/openenv/core/containers/images/Dockerfile .
+
+# Rebuild environments (they automatically use new base)
+docker build -t echo-env:latest -f envs/echo_env/server/Dockerfile .
+```

src/core/containers/runtime/__init__.py

@@ -0,0 +1,25 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Container runtime providers."""
+
+from .providers import (
+    ContainerProvider,
+    DockerSwarmProvider,
+    KubernetesProvider,
+    LocalDockerProvider,
+    RuntimeProvider,
+)
+from .uv_provider import UVProvider
+
+__all__ = [
+    "ContainerProvider",
+    "DockerSwarmProvider",
+    "LocalDockerProvider",
+    "KubernetesProvider",
+    "RuntimeProvider",
+    "UVProvider",
+]

src/core/containers/runtime/daytona_provider.py

@@ -0,0 +1,572 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Daytona container provider for running OpenEnv environments in Daytona cloud sandboxes.
+
+Requires the ``daytona`` SDK: ``pip install daytona>=0.10``
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import shlex
+import time
+from typing import Any, Callable, Dict, Optional
+
+import yaml
+
+from .providers import ContainerProvider
+
+
+class DaytonaProvider(ContainerProvider):
+    """
+    Container provider that runs environments in Daytona cloud sandboxes.
+
+    Example:
+        >>> provider = DaytonaProvider(api_key="your-key")
+        >>> image = DaytonaProvider.image_from_dockerfile("envs/echo_env/server/Dockerfile")
+        >>> base_url = provider.start_container(image)
+        >>> provider.wait_for_ready(base_url)
+        >>> provider.stop_container()
+    """
+
+    _dockerfile_registry: Dict[str, Dict[str, Any]] = {}
+
+    def __init__(
+        self,
+        *,
+        api_key: Optional[str] = None,
+        public: bool = False,
+        resources: Optional[Any] = None,
+        auto_stop_interval: int = 15,
+        target: Optional[str] = None,
+        on_snapshot_create_logs: Optional[Callable[[str], None]] = None,
+        cmd: Optional[str] = None,
+        create_timeout: float = 300,
+    ):
+        """
+        Args:
+            api_key: Daytona API key. Falls back to ``DAYTONA_API_KEY`` env var.
+            public: If True, the sandbox preview is publicly accessible.
+            resources: Optional ``daytona.Resources`` instance for CPU/memory.
+            auto_stop_interval: Minutes of inactivity before auto-stop (0 disables).
+            target: Daytona target region (e.g. "us").
+            on_snapshot_create_logs: Callback for snapshot build log lines.
+            cmd: Shell command to start the server inside the sandbox.
+            create_timeout: Seconds to wait for sandbox creation (default 300).
+                Heavy images (e.g. with Playwright/Chromium) may need more.
+        """
+        from daytona import Daytona, DaytonaConfig
+
+        config_kwargs: Dict[str, Any] = {}
+        resolved_key = api_key or os.environ.get("DAYTONA_API_KEY")
+        if resolved_key:
+            config_kwargs["api_key"] = resolved_key
+        if target:
+            config_kwargs["target"] = target
+
+        self._daytona = Daytona(DaytonaConfig(**config_kwargs))
+        self._public = public
+        self._resources = resources
+        self._auto_stop_interval = auto_stop_interval
+        self._on_snapshot_create_logs = on_snapshot_create_logs
+        self._cmd = cmd
+        self._create_timeout = create_timeout
+        self._sandbox: Any = None
+        self._preview_url: Optional[str] = None
+
+    def _discover_server_cmd(self, sandbox: Any, port: int = 8000) -> str:
+        """Discover the server command from ``openenv.yaml`` inside *sandbox*.
+
+        Finds the file, reads the ``app`` field, and constructs a command
+        of the form ``cd <env_root> && python -m uvicorn <app> --host 0.0.0.0 --port <port>``.
+
+        Raises:
+            ValueError: If ``openenv.yaml`` is not found or lacks an ``app`` field.
+        """
+        yaml_path = self._find_openenv_yaml(sandbox)
+        if yaml_path is None:
+            raise ValueError(
+                "Could not find openenv.yaml inside the sandbox. "
+                "Pass an explicit cmd= to DaytonaProvider or start_container()."
+            )
+
+        cat_resp = sandbox.process.exec(f"cat {shlex.quote(yaml_path)}", timeout=10)
+        content = cat_resp.result if hasattr(cat_resp, "result") else str(cat_resp)
+        app = self._parse_app_field(content)
+        if app is None:
+            raise ValueError(
+                f"openenv.yaml at {yaml_path} does not contain an 'app' field. "
+                "Pass an explicit cmd= to DaytonaProvider or start_container()."
+            )
+
+        # The directory containing openenv.yaml is the env root
+        env_root = yaml_path.rsplit("/", 1)[0]
+        return (
+            f"cd {shlex.quote(env_root)} && "
+            f"python -m uvicorn {shlex.quote(app)} --host 0.0.0.0 --port {port}"
+        )
+
+    def _find_openenv_yaml(self, sandbox: Any) -> Optional[str]:
+        """Locate ``openenv.yaml`` inside the sandbox.
+
+        Tries the modern layout path ``/app/env/openenv.yaml`` first,
+        then falls back to a ``find`` command for the old layout.
+        """
+        # Fast path: modern Dockerfile layout
+        resp = sandbox.process.exec(
+            "test -f /app/env/openenv.yaml && echo found", timeout=10
+        )
+        out = resp.result if hasattr(resp, "result") else str(resp)
+        if "found" in (out or ""):
+            return "/app/env/openenv.yaml"
+
+        # Fallback: search for it (redirect stderr so error messages
+        # like "No such file or directory" don't get mistaken for paths).
+        resp = sandbox.process.exec(
+            "find /app -maxdepth 4 -name openenv.yaml -print -quit 2>/dev/null",
+            timeout=10,
+        )
+        path = (resp.result if hasattr(resp, "result") else str(resp) or "").strip()
+        if path and path.startswith("/"):
+            return path
+
+        return None
+
+    @staticmethod
+    def _parse_app_field(yaml_content: str) -> Optional[str]:
+        """Extract the ``app`` value from raw openenv.yaml content.
+
+        Uses PyYAML to handle comments, quotes, and nested keys correctly.
+        """
+        try:
+            data = yaml.safe_load(yaml_content) or {}
+        except Exception:
+            return None
+
+        if not isinstance(data, dict):
+            return None
+
+        value = data.get("app")
+        if isinstance(value, str):
+            value = value.strip()
+            return value if value else None
+        return None
+
+    @staticmethod
+    def _parse_dockerfile_cmd(dockerfile_content: str) -> Optional[str]:
+        """Extract the server command from the last ``CMD`` in a Dockerfile.
+
+        Handles exec form (``CMD ["prog", "arg"]``) and shell form
+        (``CMD prog arg``).  When a Dockerfile has multiple ``CMD``
+        instructions (e.g. multi-stage builds), the last one wins - same
+        semantics as Docker itself.  Lines where ``CMD`` appears inside a
+        comment are ignored.
+
+        Returns:
+            The command as a single string, or ``None`` if no ``CMD`` found.
+        """
+        import re
+
+        last_cmd: Optional[str] = None
+        for line in dockerfile_content.splitlines():
+            stripped = line.strip()
+            # Skip comments
+            if stripped.startswith("#"):
+                continue
+            match = re.match(r"CMD\s+(.+)", stripped, flags=re.IGNORECASE)
+            if match:
+                last_cmd = match.group(1).strip()
+
+        if last_cmd is None:
+            return None
+
+        # Exec form: CMD ["executable", "param1", ...]
+        if last_cmd.startswith("["):
+            try:
+                parts = json.loads(last_cmd)
+                if isinstance(parts, list) and all(isinstance(p, str) for p in parts):
+                    return " ".join(parts)
+            except (json.JSONDecodeError, TypeError):
+                pass
+
+        # Shell form: CMD executable param1 ...
+        return last_cmd if last_cmd else None
+
+    @staticmethod
+    def strip_buildkit_syntax(dockerfile_content: str) -> str:
+        """Remove BuildKit ``--mount=...`` flags from ``RUN`` instructions.
+
+        Handles single-line flags, multi-line continuations, and multiple
+        ``--mount`` flags spread across continuation lines. Only leading
+        ``--mount`` flags are removed (before the actual command starts).
+
+        Daytona's ``Image.from_dockerfile`` does not support BuildKit
+        ``--mount`` syntax.  This helper strips the flags so that standard
+        Dockerfiles (like the ones generated by ``openenv build``) can
+        be used directly.
+        """
+        import re
+
+        def strip_leading_mounts(text: str) -> str:
+            remaining = text
+            while True:
+                match = re.match(r"\s*--mount=\S+\s*", remaining)
+                if not match:
+                    return remaining
+                remaining = remaining[match.end() :]
+
+        lines = dockerfile_content.split("\n")
+        result: list[str] = []
+        in_run = False
+        in_mount_prefix = False
+
+        for line in lines:
+            line_out = line
+            run_start = False
+            if re.match(r"\s*RUN(\s+|$)", line, flags=re.IGNORECASE):
+                in_run = True
+                in_mount_prefix = True
+                run_start = True
+
+            if in_run and in_mount_prefix:
+                original_ends_with_slash = line_out.rstrip().endswith("\\")
+                if run_start:
+                    match = re.match(r"(\s*RUN\s+)(.*)$", line_out, flags=re.IGNORECASE)
+                    if match:
+                        run_prefix, remainder = match.group(1), match.group(2)
+                    else:
+                        run_prefix, remainder = line_out, ""
+                    new_remainder = strip_leading_mounts(remainder)
+                    line_out = run_prefix + new_remainder
+                    content_for_check = new_remainder
+                else:
+                    new_remainder = strip_leading_mounts(line_out)
+                    line_out = new_remainder
+                    content_for_check = new_remainder
+
+                if original_ends_with_slash and not line_out.rstrip().endswith("\\"):
+                    line_out = line_out.rstrip() + " \\"
+
+                if content_for_check.strip() not in ("", "\\"):
+                    in_mount_prefix = False
+
+            if in_run and not line_out.rstrip().endswith("\\"):
+                in_run = False
+                in_mount_prefix = False
+
+            result.append(line_out)
+
+        return "\n".join(result)
+
+    @classmethod
+    def image_from_dockerfile(
+        cls,
+        dockerfile_path: str,
+        context_dir: str | None = None,
+    ) -> str:
+        """Validate a Dockerfile and return a ``dockerfile:`` URI for
+        :meth:`start_container`.
+
+        Eagerly validates the Dockerfile (existence, COPY sources,
+        BuildKit stripping) and stores the processed content in an
+        internal registry.  The actual ``daytona.Image`` is created
+        later inside ``start_container``.
+
+        Args:
+            dockerfile_path: Path to the Dockerfile on disk.
+            context_dir: Build context directory.  Defaults to the
+                Dockerfile's grandparent directory, matching the
+                ``openenv init`` convention where Dockerfiles live in
+                ``<env>/server/Dockerfile`` and the build context is
+                ``<env>/``.  Pass explicitly for non-standard layouts
+                (e.g. ``context_dir="."`` for repo-root contexts).
+
+        Returns:
+            A ``"dockerfile:<abs_path>"`` string to pass to
+            ``start_container``.
+
+        Raises:
+            FileNotFoundError: If *dockerfile_path* does not exist.
+            ValueError: If *context_dir* is given but does not exist,
+                or if COPY sources in the Dockerfile cannot be found
+                under the resolved context directory.
+        """
+        import pathlib
+        import re
+
+        src = pathlib.Path(dockerfile_path).resolve()
+        if not src.is_file():
+            raise FileNotFoundError(f"Dockerfile not found: {dockerfile_path}")
+
+        if context_dir is not None:
+            ctx = pathlib.Path(context_dir)
+            if not ctx.is_dir():
+                raise ValueError(f"context_dir does not exist: {context_dir}")
+        else:
+            # Default: grandparent of the Dockerfile, matching the
+            # openenv init layout (<env>/server/Dockerfile -> <env>/).
+            ctx = src.parent.parent
+
+        content = src.read_text()
+        stripped = cls.strip_buildkit_syntax(content)
+
+        # Validate that COPY sources exist under the context directory.
+        # This catches mismatches early (e.g. a Dockerfile expecting repo
+        # root as context when we defaulted to the env directory).
+        for line in stripped.splitlines():
+            m = re.match(r"^\s*COPY\s+(?!--from=)(\S+)\s+", line, re.IGNORECASE)
+            if not m:
+                continue
+            copy_src = m.group(1)
+            if copy_src.startswith("/"):
+                continue
+            resolved = ctx / copy_src
+            if not resolved.exists() and not any(ctx.glob(copy_src)):
+                raise ValueError(
+                    f"Dockerfile COPY source '{copy_src}' not found "
+                    f"under context_dir '{ctx}'. This Dockerfile may "
+                    f"expect a different build context (e.g. the repo "
+                    f"root). Pass context_dir explicitly."
+                )
+
+        # Parse CMD from the original Dockerfile so start_container can
+        # use it as a fallback when openenv.yaml is unavailable.
+        parsed_cmd = cls._parse_dockerfile_cmd(content)
+
+        cls._dockerfile_registry[str(src)] = {
+            "stripped_content": stripped,
+            "context_dir": str(ctx),
+            "server_cmd": parsed_cmd,
+        }
+
+        return f"dockerfile:{src}"
+
+    def start_container(
+        self,
+        image: str,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        **kwargs: Any,
+    ) -> str:
+        """
+        Create a Daytona sandbox from a Docker image or snapshot.
+
+        Daytona does not execute the image's CMD (known bug — ENTRYPOINT
+        runs, CMD does not).  The server command is resolved in order:
+
+        1. Explicit ``cmd`` passed to the constructor.
+        2. ``cmd`` key in ``**kwargs`` (popped before forwarding).
+        3. Auto-discovered from ``openenv.yaml`` inside the sandbox.
+        4. ``CMD`` parsed from the Dockerfile (when *image* came from
+           ``image_from_dockerfile``).
+
+        Args:
+            image: Docker image name (e.g. ``"echo-env:latest"``),
+                   ``"snapshot:<name>"`` to create from a pre-built snapshot,
+                   or ``"dockerfile:<path>"`` returned by
+                   :meth:`image_from_dockerfile`.
+            port: Must be ``None`` or ``8000``. Daytona exposes port 8000
+                  via its preview proxy; other ports raise ``ValueError``.
+            env_vars: Environment variables forwarded to the sandbox.
+            **kwargs: ``cmd`` (str) to override the server command;
+                remaining kwargs passed through to ``Daytona.create()``.
+
+        Returns:
+            HTTPS preview URL for the sandbox (base_url).
+        """
+        if port is not None and port != 8000:
+            raise ValueError(
+                f"DaytonaProvider only supports port 8000 (got {port}). "
+                "The Daytona preview proxy routes to port 8000 inside the sandbox."
+            )
+
+        # Resolve the server command (may be None; discovery happens after
+        # sandbox creation when we can inspect the filesystem).
+        cmd = kwargs.pop("cmd", None) or self._cmd
+
+        # CMD parsed from Dockerfile (populated for "dockerfile:" images).
+        parsed_cmd: Optional[str] = None
+
+        # Build creation params
+        create_kwargs: Dict[str, Any] = {}
+        if env_vars:
+            create_kwargs["env_vars"] = env_vars
+        if self._public:
+            create_kwargs["public"] = True
+        if self._auto_stop_interval != 15:
+            create_kwargs["auto_stop_interval"] = self._auto_stop_interval
+
+        if image.startswith("snapshot:"):
+            from daytona import CreateSandboxFromSnapshotParams
+
+            snapshot_name = image[len("snapshot:") :]
+            params = CreateSandboxFromSnapshotParams(
+                snapshot=snapshot_name, **create_kwargs
+            )
+        elif image.startswith("dockerfile:"):
+            from daytona import CreateSandboxFromImageParams, Image
+
+            dockerfile_path = image[len("dockerfile:") :]
+            meta = self._dockerfile_registry.get(dockerfile_path)
+            if meta is None:
+                raise ValueError(
+                    f"No registered Dockerfile metadata for {dockerfile_path}. "
+                    "Call DaytonaProvider.image_from_dockerfile() first."
+                )
+
+            parsed_cmd = meta.get("server_cmd")
+
+            # Build the daytona Image from the pre-stripped content.
+            import pathlib
+            import uuid
+
+            ctx = pathlib.Path(meta["context_dir"])
+            tmp_name = f".daytona-{uuid.uuid4().hex[:8]}.dockerfile"
+            tmp_path = ctx / tmp_name
+            try:
+                tmp_path.write_text(meta["stripped_content"])
+                daytona_image = Image.from_dockerfile(str(tmp_path))
+            finally:
+                tmp_path.unlink(missing_ok=True)
+
+            img_kwargs: Dict[str, Any] = {
+                "image": daytona_image,
+                **create_kwargs,
+            }
+            if self._resources is not None:
+                img_kwargs["resources"] = self._resources
+            params = CreateSandboxFromImageParams(**img_kwargs)
+        else:
+            from daytona import CreateSandboxFromImageParams
+
+            img_kwargs = {"image": image, **create_kwargs}
+            if self._resources is not None:
+                img_kwargs["resources"] = self._resources
+            params = CreateSandboxFromImageParams(**img_kwargs)
+
+        # Create sandbox
+        extra: Dict[str, Any] = dict(kwargs)
+        if self._on_snapshot_create_logs is not None:
+            extra["on_snapshot_create_logs"] = self._on_snapshot_create_logs
+
+        self._sandbox = self._daytona.create(
+            params, timeout=self._create_timeout, **extra
+        )
+
+        try:
+            # Discover server command from openenv.yaml if not explicitly set.
+            if cmd is None:
+                try:
+                    cmd = self._discover_server_cmd(self._sandbox)
+                except ValueError:
+                    # Fall back to CMD parsed from Dockerfile (if available).
+                    if parsed_cmd:
+                        cmd = parsed_cmd
+                    else:
+                        raise
+
+            # Wrap in bash -c so compound commands (cd ... && uvicorn ...)
+            # are handled correctly by nohup.  Write PID so we can check
+            # if the process crashed later in wait_for_ready().
+            escaped_cmd = shlex.quote(cmd)
+            self._sandbox.process.exec(
+                f"nohup bash -c {escaped_cmd} > /tmp/openenv-server.log 2>&1 &"
+                " echo $! > /tmp/openenv-server.pid",
+                timeout=10,
+            )
+
+            # Get a signed preview URL for port 8000.  The token is
+            # embedded in the URL itself so no extra headers are needed.
+            signed = self._sandbox.create_signed_preview_url(
+                8000, expires_in_seconds=86400
+            )
+            self._preview_url = signed.url
+        except Exception:
+            self.stop_container()
+            raise
+
+        return self._preview_url
+
+    def refresh_preview_url(self) -> str:
+        """Get a fresh signed preview URL (valid for 24h).
+
+        Daytona signed URLs expire after at most 24 hours.  Call this to
+        get a new one for long-running sessions.  The returned URL points
+        to the same sandbox — clients will need to reconnect using it.
+        """
+        if self._sandbox is None:
+            raise RuntimeError("No active sandbox to refresh URL for.")
+        signed = self._sandbox.create_signed_preview_url(8000, expires_in_seconds=86400)
+        self._preview_url = signed.url
+        return self._preview_url
+
+    def stop_container(self) -> None:
+        """Delete the Daytona sandbox."""
+        if self._sandbox is None:
+            return
+
+        try:
+            self._daytona.delete(self._sandbox)
+        finally:
+            self._sandbox = None
+            self._preview_url = None
+
+    def wait_for_ready(self, base_url: str, timeout_s: float = 120.0) -> None:
+        """
+        Poll the /health endpoint until the sandbox is ready.
+
+        Uses a longer default timeout (120s) than Docker providers because
+        Daytona sandboxes may have cold-start latency.
+
+        Args:
+            base_url: Preview URL returned by ``start_container()``.
+            timeout_s: Maximum seconds to wait.
+
+        Raises:
+            TimeoutError: If the sandbox doesn't become ready in time.
+            RuntimeError: If the server process died (detected via PID check).
+        """
+        import requests
+
+        health_url = f"{base_url}/health"
+
+        deadline = time.time() + timeout_s
+        while time.time() < deadline:
+            try:
+                response = requests.get(health_url, timeout=5.0)
+                if response.status_code == 200:
+                    return
+            except requests.RequestException:
+                pass
+
+            # Early exit: if the server process died, raise immediately
+            # instead of waiting for the full health-check timeout.
+            if self._sandbox is not None:
+                resp = self._sandbox.process.exec(
+                    "kill -0 $(cat /tmp/openenv-server.pid) 2>/dev/null"
+                    " && echo RUNNING || echo DEAD",
+                    timeout=10,
+                )
+                out = resp.result if hasattr(resp, "result") else str(resp)
+                if "DEAD" in (out or ""):
+                    log_resp = self._sandbox.process.exec(
+                        "cat /tmp/openenv-server.log 2>/dev/null", timeout=10
+                    )
+                    log = (
+                        log_resp.result
+                        if hasattr(log_resp, "result")
+                        else str(log_resp)
+                    )
+                    raise RuntimeError(f"Server process died.\nLog:\n{log}")
+
+            time.sleep(1.0)
+
+        raise TimeoutError(
+            f"Daytona sandbox at {base_url} did not become ready within {timeout_s}s"
+        )

src/core/containers/runtime/providers.py

@@ -0,0 +1,669 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Container provider abstractions for running environment servers.
+
+This module provides a pluggable architecture for different container providers
+(local Docker, Kubernetes, cloud providers, etc.) to be used with EnvClient.
+"""
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any, Dict, Optional, Sequence
+
+
+class ContainerProvider(ABC):
+    """
+    Abstract base class for container providers.
+
+    Providers implement this interface to support different container platforms:
+    - LocalDockerProvider: Runs containers on local Docker daemon
+    - KubernetesProvider: Runs containers in Kubernetes cluster
+    - FargateProvider: Runs containers on AWS Fargate
+    - CloudRunProvider: Runs containers on Google Cloud Run
+
+    The provider manages a single container lifecycle and provides the base URL
+    for connecting to it.
+
+    Example:
+        >>> provider = LocalDockerProvider()
+        >>> base_url = provider.start_container("echo-env:latest")
+        >>> print(base_url)  # http://localhost:8000
+        >>> # Use the environment via base_url
+        >>> provider.stop_container()
+    """
+
+    @abstractmethod
+    def start_container(
+        self,
+        image: str,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        **kwargs: Any,
+    ) -> str:
+        """
+        Start a container from the specified image.
+
+        Args:
+            image: Container image name (e.g., "echo-env:latest")
+            port: Port to expose (if None, provider chooses)
+            env_vars: Environment variables to pass to container
+            **kwargs: Provider-specific options
+
+        Returns:
+            Base URL to connect to the container (e.g., "http://localhost:8000")
+
+        Raises:
+            RuntimeError: If container fails to start
+        """
+        pass
+
+    @abstractmethod
+    def stop_container(self) -> None:
+        """
+        Stop and remove the running container.
+
+        This cleans up the container that was started by start_container().
+        """
+        pass
+
+    @abstractmethod
+    def wait_for_ready(self, base_url: str, timeout_s: float = 30.0) -> None:
+        """
+        Wait for the container to be ready to accept requests.
+
+        This typically polls the /health endpoint until it returns 200.
+
+        Args:
+            base_url: Base URL of the container
+            timeout_s: Maximum time to wait
+
+        Raises:
+            TimeoutError: If container doesn't become ready in time
+        """
+        pass
+
+
+class LocalDockerProvider(ContainerProvider):
+    """
+    Container provider for local Docker daemon.
+
+    This provider runs containers on the local machine using Docker.
+    Useful for development and testing.
+
+    Example:
+        >>> provider = LocalDockerProvider()
+        >>> base_url = provider.start_container("echo-env:latest")
+        >>> # Container running on http://localhost:<random-port>
+        >>> provider.stop_container()
+    """
+
+    def __init__(self):
+        """Initialize the local Docker provider."""
+        self._container_id: Optional[str] = None
+        self._container_name: Optional[str] = None
+
+        # Check if Docker is available
+        import subprocess
+
+        try:
+            subprocess.run(
+                ["docker", "version"],
+                check=True,
+                capture_output=True,
+                timeout=5,
+            )
+        except (
+            subprocess.CalledProcessError,
+            FileNotFoundError,
+            subprocess.TimeoutExpired,
+        ):
+            raise RuntimeError(
+                "Docker is not available. Please install Docker Desktop or Docker Engine."
+            )
+
+    def start_container(
+        self,
+        image: str,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        **kwargs: Any,
+    ) -> str:
+        """
+        Start a Docker container locally.
+
+        Args:
+            image: Docker image name
+            port: Port to expose (if None, finds available port)
+            env_vars: Environment variables for the container
+            **kwargs: Additional Docker run options
+
+        Returns:
+            Base URL to connect to the container
+        """
+        import subprocess
+        import time
+
+        # Find available port if not specified
+        if port is None:
+            port = self._find_available_port()
+
+        # Generate container name
+        self._container_name = self._generate_container_name(image)
+
+        # Build docker run command
+        cmd = [
+            "docker",
+            "run",
+            "-d",  # Detached
+            "--name",
+            self._container_name,
+            "-p",
+            f"{port}:8000",  # Map port
+        ]
+
+        # Add environment variables
+        if env_vars:
+            for key, value in env_vars.items():
+                cmd.extend(["-e", f"{key}={value}"])
+
+        # Add image
+        cmd.append(image)
+
+        # Run container
+        try:
+            result = subprocess.run(cmd, capture_output=True, text=True, check=True)
+            self._container_id = result.stdout.strip()
+        except subprocess.CalledProcessError as e:
+            error_msg = f"Failed to start Docker container.\nCommand: {' '.join(cmd)}\nExit code: {e.returncode}\nStderr: {e.stderr}\nStdout: {e.stdout}"
+            raise RuntimeError(error_msg) from e
+
+        # Wait a moment for container to start
+        time.sleep(1)
+
+        base_url = f"http://localhost:{port}"
+        return base_url
+
+    def stop_container(self) -> None:
+        """
+        Stop and remove the Docker container.
+        """
+        if self._container_id is None:
+            return
+
+        import subprocess
+
+        try:
+            # Stop container
+            subprocess.run(
+                ["docker", "stop", self._container_id],
+                capture_output=True,
+                check=True,
+                timeout=10,
+            )
+
+            # Remove container
+            subprocess.run(
+                ["docker", "rm", self._container_id],
+                capture_output=True,
+                check=True,
+                timeout=10,
+            )
+        except subprocess.CalledProcessError:
+            # Container might already be stopped/removed
+            pass
+        finally:
+            self._container_id = None
+            self._container_name = None
+
+    def wait_for_ready(self, base_url: str, timeout_s: float = 30.0) -> None:
+        """
+        Wait for container to be ready by polling /health endpoint.
+
+        Args:
+            base_url: Base URL of the container
+            timeout_s: Maximum time to wait
+
+        Raises:
+            TimeoutError: If container doesn't become ready
+        """
+        import time
+
+        import requests
+
+        start_time = time.time()
+        health_url = f"{base_url}/health"
+
+        # Bypass proxy for localhost to avoid proxy issues
+        proxies = {"http": None, "https": None}
+
+        while time.time() - start_time < timeout_s:
+            try:
+                response = requests.get(health_url, timeout=2.0, proxies=proxies)
+                if response.status_code == 200:
+                    return
+            except requests.RequestException:
+                pass
+
+            time.sleep(0.5)
+
+        raise TimeoutError(
+            f"Container at {base_url} did not become ready within {timeout_s}s"
+        )
+
+    def _find_available_port(self) -> int:
+        """
+        Find an available port on localhost.
+
+        Returns:
+            An available port number
+        """
+        import socket
+
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+            s.bind(("", 0))
+            s.listen(1)
+            port = s.getsockname()[1]
+        return port
+
+    def _generate_container_name(self, image: str) -> str:
+        """
+        Generate a unique container name based on image name and timestamp.
+
+        Args:
+            image: Docker image name
+
+        Returns:
+            A unique container name
+        """
+        import time
+
+        clean_image = image.split("/")[-1].split(":")[0]
+        timestamp = int(time.time() * 1000)
+        return f"{clean_image}-{timestamp}"
+
+
+class DockerSwarmProvider(ContainerProvider):
+    """
+    Container provider that uses Docker Swarm services for local concurrency.
+
+    This provider creates a replicated Swarm service backed by the local Docker
+    engine. The built-in load-balancer fans requests across the replicas,
+    allowing multiple container instances to run concurrently on the developer
+    workstation (mirroring the workflow described in the Docker stack docs).
+    """
+
+    def __init__(
+        self,
+        *,
+        auto_init_swarm: bool = True,
+        overlay_network: Optional[str] = None,
+    ):
+        """
+        Args:
+            auto_init_swarm: Whether to call ``docker swarm init`` when Swarm
+                is not active. Otherwise, user must manually initialize Swarm.
+            overlay_network: Optional overlay network name for the service.
+                When provided, the network is created with
+                ``docker network create --driver overlay --attachable`` if it
+                does not already exist.
+        """
+        self._service_name: Optional[str] = None
+        self._service_id: Optional[str] = None
+        self._published_port: Optional[int] = None
+        self._overlay_network = overlay_network
+        self._auto_init_swarm = auto_init_swarm
+
+        self._ensure_docker_available()
+        self._ensure_swarm_initialized()
+        if self._overlay_network:
+            self._ensure_overlay_network(self._overlay_network)
+
+    def start_container(
+        self,
+        image: str,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        **kwargs: Any,
+    ) -> str:
+        """
+        Start (or scale) a Swarm service for the given image.
+
+        Supported kwargs:
+            replicas (int): Number of container replicas (default: 2).
+            cpu_limit (float | str): CPU limit passed to ``--limit-cpu``.
+            memory_limit (str): Memory limit passed to ``--limit-memory``.
+            constraints (Sequence[str]): Placement constraints.
+            labels (Dict[str, str]): Service labels.
+            command (Sequence[str] | str): Override container command.
+        """
+        import shlex
+        import subprocess
+        import time
+
+        allowed_kwargs = {
+            "replicas",
+            "cpu_limit",
+            "memory_limit",
+            "constraints",
+            "labels",
+            "command",
+        }
+        unknown = set(kwargs) - allowed_kwargs
+        if unknown:
+            raise ValueError(f"Unsupported kwargs for DockerSwarmProvider: {unknown}")
+
+        replicas = int(kwargs.get("replicas", 2))
+        cpu_limit = kwargs.get("cpu_limit")
+        memory_limit = kwargs.get("memory_limit")
+        constraints: Optional[Sequence[str]] = kwargs.get("constraints")
+        labels: Optional[Dict[str, str]] = kwargs.get("labels")
+        command_override = kwargs.get("command")
+
+        if port is None:
+            port = self._find_available_port()
+
+        self._service_name = self._generate_service_name(image)
+        self._published_port = port
+
+        cmd = [
+            "docker",
+            "service",
+            "create",
+            "--detach",
+            "--name",
+            self._service_name,
+            "--replicas",
+            str(max(1, replicas)),
+            "--publish",
+            f"{port}:8000",
+        ]
+
+        if self._overlay_network:
+            cmd.extend(["--network", self._overlay_network])
+
+        if env_vars:
+            for key, value in env_vars.items():
+                cmd.extend(["--env", f"{key}={value}"])
+
+        if cpu_limit is not None:
+            cmd.extend(["--limit-cpu", str(cpu_limit)])
+
+        if memory_limit is not None:
+            cmd.extend(["--limit-memory", str(memory_limit)])
+
+        if constraints:
+            for constraint in constraints:
+                cmd.extend(["--constraint", constraint])
+
+        if labels:
+            for key, value in labels.items():
+                cmd.extend(["--label", f"{key}={value}"])
+
+        cmd.append(image)
+
+        if command_override:
+            if isinstance(command_override, str):
+                cmd.extend(shlex.split(command_override))
+            else:
+                cmd.extend(command_override)
+
+        try:
+            result = subprocess.run(
+                cmd,
+                capture_output=True,
+                text=True,
+                check=True,
+            )
+            self._service_id = result.stdout.strip()
+        except subprocess.CalledProcessError as e:
+            error_msg = (
+                "Failed to start Docker Swarm service.\n"
+                f"Command: {' '.join(cmd)}\n"
+                f"Exit code: {e.returncode}\n"
+                f"Stdout: {e.stdout}\n"
+                f"Stderr: {e.stderr}"
+            )
+            raise RuntimeError(error_msg) from e
+
+        # Give Swarm a brief moment to schedule the tasks.
+        time.sleep(1.0)
+
+        return f"http://localhost:{port}"
+
+    def stop_container(self) -> None:
+        """
+        Remove the Swarm service (and keep the Swarm manager running).
+        """
+        if not self._service_name:
+            return
+
+        import subprocess
+
+        try:
+            subprocess.run(
+                ["docker", "service", "rm", self._service_name],
+                capture_output=True,
+                check=True,
+                timeout=10,
+            )
+        except subprocess.CalledProcessError:
+            # Service may already be gone; ignore.
+            pass
+        finally:
+            self._service_name = None
+            self._service_id = None
+            self._published_port = None
+
+    def wait_for_ready(self, base_url: str, timeout_s: float = 30.0) -> None:
+        """
+        Wait for at least one replica to become healthy by polling /health.
+
+        Note: With Swarm's load balancer, requests round-robin across replicas,
+        so this only verifies that at least one replica is responding. Some
+        replicas may still be starting when this returns.
+        """
+        import time
+
+        import requests
+
+        deadline = time.time() + timeout_s
+        health_url = f"{base_url}/health"
+
+        # Bypass proxy for localhost to avoid proxy issues
+        proxies = {"http": None, "https": None}
+
+        while time.time() < deadline:
+            try:
+                response = requests.get(health_url, timeout=2.0, proxies=proxies)
+                if response.status_code == 200:
+                    return
+            except requests.RequestException:
+                pass
+
+            time.sleep(0.5)
+
+        raise TimeoutError(
+            f"Swarm service at {base_url} did not become ready within {timeout_s}s"
+        )
+
+    def _ensure_docker_available(self) -> None:
+        import subprocess
+
+        try:
+            subprocess.run(
+                ["docker", "version"],
+                check=True,
+                capture_output=True,
+                timeout=5,
+            )
+        except (
+            subprocess.CalledProcessError,
+            FileNotFoundError,
+            subprocess.TimeoutExpired,
+        ) as exc:
+            raise RuntimeError(
+                "Docker is not available. Please install Docker Desktop or Docker Engine."
+            ) from exc
+
+    def _ensure_swarm_initialized(self) -> None:
+        import subprocess
+
+        try:
+            result = subprocess.run(
+                ["docker", "info", "--format", "{{.Swarm.LocalNodeState}}"],
+                capture_output=True,
+                text=True,
+                check=True,
+                timeout=5,
+            )
+            state = result.stdout.strip().lower()
+            if state == "active":
+                return
+        except subprocess.CalledProcessError:
+            state = "unknown"
+
+        if not self._auto_init_swarm:
+            raise RuntimeError(
+                f"Docker Swarm is not active (state={state}). Enable Swarm manually or pass auto_init_swarm=True."
+            )
+
+        try:
+            subprocess.run(
+                ["docker", "swarm", "init"],
+                check=True,
+                capture_output=True,
+                timeout=10,
+            )
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError("Failed to initialize Docker Swarm") from e
+
+    def _ensure_overlay_network(self, network: str) -> None:
+        import subprocess
+
+        inspect = subprocess.run(
+            ["docker", "network", "inspect", network],
+            capture_output=True,
+            text=True,
+            check=False,
+        )
+        if inspect.returncode == 0:
+            return
+
+        try:
+            subprocess.run(
+                [
+                    "docker",
+                    "network",
+                    "create",
+                    "--driver",
+                    "overlay",
+                    "--attachable",
+                    network,
+                ],
+                check=True,
+                capture_output=True,
+                timeout=10,
+            )
+        except subprocess.CalledProcessError as e:
+            raise RuntimeError(f"Failed to create overlay network '{network}'") from e
+
+    def _find_available_port(self) -> int:
+        import socket
+
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+            s.bind(("", 0))
+            s.listen(1)
+            port = s.getsockname()[1]
+        return port
+
+    def _generate_service_name(self, image: str) -> str:
+        import time
+
+        clean_image = image.split("/")[-1].split(":")[0]
+        timestamp = int(time.time() * 1000)
+        return f"{clean_image}-swarm-{timestamp}"
+
+
+class KubernetesProvider(ContainerProvider):
+    """
+    Container provider for Kubernetes clusters.
+
+    This provider creates pods in a Kubernetes cluster and exposes them
+    via services or port-forwarding.
+
+    Example:
+        >>> provider = KubernetesProvider(namespace="envtorch-dev")
+        >>> base_url = provider.start_container("echo-env:latest")
+        >>> # Pod running in k8s, accessible via service or port-forward
+        >>> provider.stop_container()
+    """
+
+    pass
+
+
+class RuntimeProvider(ABC):
+    """
+    Abstract base class for runtime providers that are not container providers.
+    Providers implement this interface to support different runtime platforms:
+    - UVProvider: Runs environments via `uv run`
+
+    The provider manages a single runtime lifecycle and provides the base URL
+    for connecting to it.
+
+    Example:
+        >>> provider = UVProvider(project_path="/path/to/env")
+        >>> base_url = provider.start()
+        >>> print(base_url)  # http://localhost:8000
+        >>> provider.stop()
+    """
+
+    @abstractmethod
+    def start(
+        self,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        **kwargs: Any,
+    ) -> str:
+        """
+        Start a runtime from the specified image.
+
+        Args:
+            image: Runtime image name
+            port: Port to expose (if None, provider chooses)
+            env_vars: Environment variables for the runtime
+            **kwargs: Additional runtime options
+        """
+
+    @abstractmethod
+    def stop(self) -> None:
+        """
+        Stop the runtime.
+        """
+        pass
+
+    @abstractmethod
+    def wait_for_ready(self, timeout_s: float = 30.0) -> None:
+        """
+        Wait for the runtime to be ready to accept requests.
+        """
+        pass
+
+    def __enter__(self) -> "RuntimeProvider":
+        """
+        Enter the runtime provider.
+        """
+        self.start()
+        return self
+
+    def __exit__(self, exc_type, exc, tb) -> None:
+        """
+        Exit the runtime provider.
+        """
+        self.stop()
+        return False

src/core/containers/runtime/uv_provider.py

@@ -0,0 +1,224 @@
+"""Providers for launching ASGI applications via ``uv run``."""
+
+from __future__ import annotations
+
+import os
+import socket
+import subprocess
+import time
+from typing import Dict, Optional
+
+import requests
+
+from .providers import RuntimeProvider
+
+
+def _check_uv_installed() -> None:
+    try:
+        subprocess.check_output(["uv", "--version"])
+    except FileNotFoundError as exc:
+        raise RuntimeError(
+            "`uv` executable not found. Install uv from https://docs.astral.sh and ensure it is on PATH."
+        ) from exc
+
+
+def _find_free_port() -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
+        sock.bind(("", 0))
+        sock.listen(1)
+        return sock.getsockname()[1]
+
+
+def _create_uv_command(
+    *,
+    host: str,
+    port: int,
+    reload: bool,
+    workers: int,
+    app: str,
+    project_path: str,
+) -> list[str]:
+    command: list[str] = ["uv", "run", "--isolated", "--project", project_path]
+
+    command.append("--")
+    command.extend(
+        [
+            "uvicorn",
+            app,
+            "--host",
+            host,
+            "--port",
+            str(port),
+            "--workers",
+            str(workers),
+        ]
+    )
+
+    if reload:
+        command.append("--reload")
+
+    return command
+
+
+def _poll_health(health_url: str, timeout_s: float) -> None:
+    """Poll a health endpoint until it returns HTTP 200 or times out."""
+
+    deadline = time.time() + timeout_s
+    while time.time() < deadline:
+        try:
+            timeout = max(0.0001, min(deadline - time.time(), 2.0))
+            response = requests.get(health_url, timeout=timeout)
+            if response.status_code == 200:
+                return
+        except requests.RequestException:
+            continue
+
+        time.sleep(0.5)
+
+    raise TimeoutError(f"Server did not become ready within {timeout_s:.1f} seconds")
+
+
+class UVProvider(RuntimeProvider):
+    """
+    RuntimeProvider implementation backed by ``uv run``.
+
+    Args:
+        project_path: Local path to a uv project (passed to ``uv run --project``)
+        app: ASGI application path for uvicorn (defaults to ``server.app:app``)
+        host: Host interface to bind to (defaults to ``0.0.0.0``)
+        reload: Whether to enable uvicorn's reload mode
+        env_vars: Environment variables to pass through to the spawned process
+        context_timeout_s: How long to wait for the environment to become ready
+
+    Example:
+        >>> provider = UVProvider(project_path="/path/to/env")
+        >>> base_url = provider.start()
+        >>> print(base_url)  # http://localhost:8000
+        >>> # Use the environment via base_url
+        >>> provider.stop()
+    """
+
+    def __init__(
+        self,
+        *,
+        project_path: str,
+        app: str = "server.app:app",
+        host: str = "0.0.0.0",
+        reload: bool = False,
+        env_vars: Optional[Dict[str, str]] = None,
+        context_timeout_s: float = 60.0,
+    ):
+        """Initialize the UVProvider."""
+        self.project_path = os.path.abspath(project_path)
+        self.app = app
+        self.host = host
+        self.reload = reload
+        self.env_vars = env_vars
+        self.context_timeout_s = context_timeout_s
+        _check_uv_installed()
+        self._process = None
+        self._base_url = None
+
+    def start(
+        self,
+        port: Optional[int] = None,
+        env_vars: Optional[Dict[str, str]] = None,
+        workers: int = 1,
+        **_: Dict[str, str],
+    ) -> str:
+        """
+        Start the environment via `uv run`.
+
+        Args:
+            port: The port to bind the environment to
+            env_vars: Environment variables to pass to the environment
+            workers: The number of workers to use
+
+        Returns:
+            The base URL of the environment
+
+        Raises:
+            RuntimeError: If the environment is already running
+        """
+        if self._process is not None and self._process.poll() is None:
+            raise RuntimeError("UVProvider is already running")
+
+        bind_port = port or _find_free_port()
+
+        command = _create_uv_command(
+            host=self.host,
+            port=bind_port,
+            reload=self.reload,
+            workers=workers,
+            app=self.app,
+            project_path=self.project_path,
+        )
+
+        env = os.environ.copy()
+
+        if self.env_vars:
+            env.update(self.env_vars)
+        if env_vars:
+            env.update(env_vars)
+
+        try:
+            self._process = subprocess.Popen(command, env=env)
+        except OSError as exc:
+            raise RuntimeError(f"Failed to launch `uv run`: {exc}") from exc
+
+        client_host = "127.0.0.1" if self.host in {"0.0.0.0", "::"} else self.host
+        self._base_url = f"http://{client_host}:{bind_port}"
+        return self._base_url
+
+    def wait_for_ready(self, timeout_s: float = 60.0) -> None:
+        """
+        Wait for the environment to become ready.
+
+        Args:
+            timeout_s: The timeout to wait for the environment to become ready
+
+        Raises:
+            RuntimeError: If the environment is not running
+            TimeoutError: If the environment does not become ready within the timeout
+        """
+        if self._process and self._process.poll() is not None:
+            code = self._process.returncode
+            raise RuntimeError(f"uv process exited prematurely with code {code}")
+
+        _poll_health(f"{self._base_url}/health", timeout_s=timeout_s)
+
+    def stop(self) -> None:
+        """
+        Stop the environment.
+
+        Raises:
+            RuntimeError: If the environment is not running
+        """
+        if self._process is None:
+            return
+
+        if self._process.poll() is None:
+            self._process.terminate()
+            try:
+                self._process.wait(timeout=10.0)
+            except subprocess.TimeoutExpired:
+                self._process.kill()
+                self._process.wait(timeout=5.0)
+
+        self._process = None
+        self._base_url = None
+
+    @property
+    def base_url(self) -> str:
+        """
+        The base URL of the environment.
+
+        Returns:
+            The base URL of the environment
+
+        Raises:
+            RuntimeError: If the environment is not running
+        """
+        if self._base_url is None:
+            raise RuntimeError("UVProvider has not been started")
+        return self._base_url

src/core/containers/test_local_docker_provider.py

@@ -0,0 +1,260 @@
+#!/usr/bin/env python3
+"""
+End-to-end test for LocalDockerProvider.
+
+This script tests the complete flow:
+1. Start a container using LocalDockerProvider
+2. Wait for it to be ready
+3. Make HTTP requests to test the environment
+4. Clean up the container
+"""
+
+import sys
+from pathlib import Path
+
+# Add src to path
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+import requests
+from openenv.core.containers.runtime import LocalDockerProvider
+
+
+# TODO: Remove this test or make it a functional test sicne this will be tested in e2e test for echo env
+def test_local_docker_provider():
+    """Test LocalDockerProvider end-to-end."""
+    print("=" * 60)
+    print("LocalDockerProvider End-to-End Test")
+    print("=" * 60)
+    print()
+
+    provider = None
+
+    try:
+        # Step 1: Create provider
+        print("Step 1: Creating LocalDockerProvider...")
+        provider = LocalDockerProvider()
+        print("✓ Provider created\n")
+
+        # Step 2: Start container
+        print("Step 2: Starting echo-env container...")
+        base_url = provider.start_container("echo-env:latest")
+        print(f"✓ Container started at: {base_url}")
+        if provider._container_id:
+            print(f"  Container ID: {provider._container_id[:12]}...")
+        if provider._container_name:
+            print(f"  Container name: {provider._container_name}\n")
+
+        # Step 3: Wait for ready
+        print("Step 3: Waiting for container to be ready...")
+        provider.wait_for_ready(base_url, timeout_s=30.0)
+        print("✓ Container is ready!\n")
+
+        # Step 4: Test health endpoint
+        print("Step 4: Testing /health endpoint...")
+        response = requests.get(f"{base_url}/health")
+        print(f"  Status: {response.status_code}")
+        print(f"  Response: {response.json()}")
+        assert response.status_code == 200
+        assert response.json()["status"] == "healthy"
+        print("✓ Health check passed\n")
+
+        # Step 5: Test reset endpoint
+        print("Step 5: Testing /reset endpoint...")
+        response = requests.post(
+            f"{base_url}/reset",
+            json={},
+            headers={"Content-Type": "application/json"},
+        )
+        print(f"  Status: {response.status_code}")
+        data = response.json()
+        print(f"  Message: {data['observation']['echoed_message']}")
+        print(f"  Reward: {data['reward']}")
+        print(f"  Done: {data['done']}")
+        assert response.status_code == 200
+        assert data["observation"]["echoed_message"] == "Echo environment ready!"
+        print("✓ Reset test passed\n")
+
+        # Step 6: Test step endpoint
+        print("Step 6: Testing /step endpoint...")
+        response = requests.post(
+            f"{base_url}/step",
+            json={"action": {"message": "Hello from LocalDockerProvider!"}},
+            headers={"Content-Type": "application/json"},
+        )
+        print(f"  Status: {response.status_code}")
+        data = response.json()
+        print(f"  Echoed: {data['observation']['echoed_message']}")
+        print(f"  Length: {data['observation']['message_length']}")
+        print(f"  Reward: {data['reward']}")
+        assert response.status_code == 200
+        assert (
+            data["observation"]["echoed_message"] == "Hello from LocalDockerProvider!"
+        )
+        assert data["observation"]["message_length"] == 31
+        print("✓ Step test passed\n")
+
+        # Step 7: Test state endpoint
+        print("Step 7: Testing /state endpoint...")
+        response = requests.get(f"{base_url}/state")
+        print(f"  Status: {response.status_code}")
+        data = response.json()
+        print(f"  Episode ID: {data['episode_id']}")
+        print(f"  Step count: {data['step_count']}")
+        assert response.status_code == 200
+        assert data["step_count"] == 1  # One step from above
+        print("✓ State test passed\n")
+
+        # Step 8: Multiple steps
+        print("Step 8: Testing multiple steps...")
+        for i in range(3):
+            response = requests.post(
+                f"{base_url}/step",
+                json={"action": {"message": f"Message {i + 1}"}},
+                headers={"Content-Type": "application/json"},
+            )
+            assert response.status_code == 200
+            print(f"  Step {i + 1}: ✓")
+
+        # Check state updated
+        response = requests.get(f"{base_url}/state")
+        data = response.json()
+        assert data["step_count"] == 4  # 1 + 3 more steps
+        print(f"  Final step count: {data['step_count']}")
+        print("✓ Multiple steps test passed\n")
+
+        print("=" * 60)
+        print("✓ All tests passed!")
+        print("=" * 60)
+        print()
+
+        return True
+
+    except Exception as e:
+        print(f"\n❌ Test failed: {e}")
+        import traceback
+
+        traceback.print_exc()
+        return False
+
+    finally:
+        # Step 9: Cleanup
+        if provider is not None:
+            print("\nStep 9: Cleaning up container...")
+            try:
+                provider.stop_container()
+                print("✓ Container stopped and removed\n")
+            except Exception as e:
+                print(f"⚠️  Cleanup warning: {e}\n")
+
+
+def test_provider_with_custom_port():
+    """Test provider with custom port."""
+    print("=" * 60)
+    print("LocalDockerProvider with Custom Port Test")
+    print("=" * 60)
+    print()
+
+    provider = None
+
+    try:
+        provider = LocalDockerProvider()
+
+        print("Starting container on custom port 8123...")
+        base_url = provider.start_container("echo-env:latest", port=8123)
+        print(f"✓ Started at: {base_url}")
+        assert ":8123" in base_url
+
+        print("Waiting for ready...")
+        provider.wait_for_ready(base_url)
+        print("✓ Ready!")
+
+        print("Testing health...")
+        response = requests.get(f"{base_url}/health")
+        assert response.status_code == 200
+        print("✓ Health check passed")
+
+        print("\n✓ Custom port test passed!\n")
+        return True
+
+    except Exception as e:
+        print(f"\n❌ Test failed: {e}")
+        return False
+
+    finally:
+        if provider is not None:
+            provider.stop_container()
+            print("✓ Cleaned up\n")
+
+
+def test_provider_with_env_vars():
+    """Test provider with environment variables."""
+    print("=" * 60)
+    print("LocalDockerProvider with Environment Variables Test")
+    print("=" * 60)
+    print()
+
+    provider = None
+
+    try:
+        provider = LocalDockerProvider()
+
+        print("Starting container with environment variables...")
+        base_url = provider.start_container(
+            "echo-env:latest", env_vars={"DEBUG": "true", "LOG_LEVEL": "info"}
+        )
+        print(f"✓ Started at: {base_url}")
+
+        print("Waiting for ready...")
+        provider.wait_for_ready(base_url)
+        print("✓ Ready!")
+
+        print("Testing health...")
+        response = requests.get(f"{base_url}/health")
+        assert response.status_code == 200
+        print("✓ Health check passed")
+
+        print("\n✓ Environment variables test passed!\n")
+        return True
+
+    except Exception as e:
+        print(f"\n❌ Test failed: {e}")
+        return False
+
+    finally:
+        if provider is not None:
+            provider.stop_container()
+            print("✓ Cleaned up\n")
+
+
+if __name__ == "__main__":
+    print()
+    print("🐳 LocalDockerProvider Test Suite")
+    print()
+
+    results = []
+
+    # Run basic test
+    results.append(("Basic End-to-End", test_local_docker_provider()))
+
+    # Run custom port test
+    results.append(("Custom Port", test_provider_with_custom_port()))
+
+    # Run environment variables test
+    results.append(("Environment Variables", test_provider_with_env_vars()))
+
+    # Summary
+    print("=" * 60)
+    print("Test Summary")
+    print("=" * 60)
+    for name, passed in results:
+        status = "✓ PASSED" if passed else "✗ FAILED"
+        print(f"{name:25} {status}")
+    print("=" * 60)
+
+    all_passed = all(result for _, result in results)
+    if all_passed:
+        print("\n🎉 All tests passed!")
+        exit(0)
+    else:
+        print("\n❌ Some tests failed")
+        exit(1)

src/core/env_client.py

@@ -0,0 +1,484 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Environment client for persistent sessions.
+
+This module provides a WebSocket-based client that maintains a persistent connection
+to an environment server, enabling efficient multi-step interactions without
+the overhead of HTTP request/response cycles.
+
+The client is async by default. For synchronous usage, use the `.sync()` method
+to get a `SyncEnvClient` wrapper.
+
+Example (async):
+    >>> async with GenericEnvClient(base_url="ws://localhost:8000") as env:
+    ...     result = await env.reset()
+    ...     result = await env.step({"code": "print('hello')"})
+
+Example (sync wrapper):
+    >>> env = GenericEnvClient(base_url="ws://localhost:8000").sync()
+    >>> with env:
+    ...     result = env.reset()
+    ...     result = env.step({"code": "print('hello')"})
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import os
+from abc import ABC, abstractmethod
+from typing import Any, Dict, Generic, Optional, Type, TYPE_CHECKING, TypeVar
+
+from .client_types import StateT, StepResult
+from .containers.runtime import LocalDockerProvider, UVProvider
+from .utils import convert_to_ws_url
+
+if TYPE_CHECKING:
+    from websockets.asyncio.client import ClientConnection
+
+    from .containers.runtime import ContainerProvider, RuntimeProvider
+    from .sync_client import SyncEnvClient
+
+from websockets.asyncio.client import connect as ws_connect
+
+ActT = TypeVar("ActT")
+ObsT = TypeVar("ObsT")
+EnvClientT = TypeVar("EnvClientT", bound="EnvClient")
+
+
+class EnvClient(ABC, Generic[ActT, ObsT, StateT]):
+    """
+    Async environment client for persistent sessions.
+
+    This client maintains a persistent WebSocket connection to an environment
+    server, enabling efficient multi-step interactions. Each client instance
+    corresponds to a dedicated environment session on the server.
+
+    The client is async by default. For synchronous usage, use the `.sync()`
+    method to get a `SyncEnvClient` wrapper.
+
+    Features:
+    - Lower latency for sequential interactions
+    - Session state is maintained server-side
+    - Better suited for long-running episodes
+    - Async by default for modern Python async/await patterns
+
+    Example (async):
+        >>> from envs.coding_env.client import CodingEnv
+        >>>
+        >>> # Connect to a server using async context manager
+        >>> async with CodingEnv(base_url="ws://localhost:8000") as env:
+        ...     result = await env.reset(seed=42)
+        ...     while not result.done:
+        ...         action = agent.predict(result.observation)
+        ...         result = await env.step(action)
+
+    Example (sync wrapper):
+        >>> env = CodingEnv(base_url="ws://localhost:8000").sync()
+        >>> with env:
+        ...     result = env.reset(seed=42)
+        ...     result = env.step(action)
+    """
+
+    def __init__(
+        self,
+        base_url: str,
+        connect_timeout_s: float = 10.0,
+        message_timeout_s: float = 60.0,
+        max_message_size_mb: float = 100.0,
+        provider: Optional["ContainerProvider | RuntimeProvider"] = None,
+        mode: Optional[str] = None,
+    ):
+        """
+        Initialize environment client.
+
+        Args:
+            base_url: Base URL of the environment server (http:// or ws://).
+                     Will be converted to ws:// if http:// is provided.
+            connect_timeout_s: Timeout for establishing WebSocket connection
+            message_timeout_s: Timeout for receiving responses to messages
+            max_message_size_mb: Maximum WebSocket message size in megabytes.
+                                Default 100MB to handle large observations (screenshots, DOM, etc.)
+            provider: Optional container/runtime provider for lifecycle management.
+                     Can be a ContainerProvider (Docker) or RuntimeProvider (UV).
+            mode: Communication mode: 'simulation' for Gym-style API (default) or
+                 'production' for MCP JSON-RPC protocol. Can also be set via the
+                 OPENENV_CLIENT_MODE environment variable. Constructor parameter
+                 takes precedence over environment variable. Case-insensitive.
+        """
+        # Determine mode (constructor > env var > default)
+        if mode is None:
+            mode = os.environ.get("OPENENV_CLIENT_MODE", "simulation")
+
+        # Normalize and validate mode
+        mode = mode.lower()
+        if mode not in ("simulation", "production"):
+            raise ValueError(
+                f"Invalid mode: '{mode}'. Must be 'simulation' or 'production'. "
+                f"Set via constructor parameter or OPENENV_CLIENT_MODE environment variable."
+            )
+
+        # Store mode (use object.__setattr__ to bypass immutability)
+        object.__setattr__(self, "_mode", mode)
+
+        # Convert HTTP URL to WebSocket URL
+        ws_url = convert_to_ws_url(base_url)
+
+        self._ws_url = f"{ws_url}/ws"
+        self._connect_timeout = connect_timeout_s
+        self._message_timeout = message_timeout_s
+        self._max_message_size = int(
+            max_message_size_mb * 1024 * 1024
+        )  # Convert MB to bytes
+        self._provider = provider
+        self._ws: Optional[ClientConnection] = None
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        """Prevent modification of _mode after initialization."""
+        if name == "_mode" and hasattr(self, "_mode"):
+            raise AttributeError("Cannot modify mode after initialization")
+        super().__setattr__(name, value)
+
+    async def connect(self) -> "EnvClient":
+        """
+        Establish WebSocket connection to the server.
+
+        Returns:
+            self for method chaining
+
+        Raises:
+            ConnectionError: If connection cannot be established
+        """
+        if self._ws is not None:
+            return self
+
+        # Bypass proxy for localhost connections
+        ws_url_lower = self._ws_url.lower()
+        is_localhost = "localhost" in ws_url_lower or "127.0.0.1" in ws_url_lower
+
+        old_no_proxy = os.environ.get("NO_PROXY")
+        if is_localhost:
+            # Set NO_PROXY to bypass proxy for localhost
+            current_no_proxy = old_no_proxy or ""
+            if "localhost" not in current_no_proxy.lower():
+                os.environ["NO_PROXY"] = (
+                    f"{current_no_proxy},localhost,127.0.0.1"
+                    if current_no_proxy
+                    else "localhost,127.0.0.1"
+                )
+
+        try:
+            self._ws = await ws_connect(
+                self._ws_url,
+                open_timeout=self._connect_timeout,
+                max_size=self._max_message_size,
+            )
+        except Exception as e:
+            raise ConnectionError(f"Failed to connect to {self._ws_url}: {e}") from e
+        finally:
+            # Restore original NO_PROXY value
+            if is_localhost:
+                if old_no_proxy is None:
+                    os.environ.pop("NO_PROXY", None)
+                else:
+                    os.environ["NO_PROXY"] = old_no_proxy
+
+        return self
+
+    async def disconnect(self) -> None:
+        """Close the WebSocket connection."""
+        if self._ws is not None:
+            try:
+                # Send close message
+                await self._send({"type": "close"})
+            except Exception:
+                pass  # Best effort
+            try:
+                await self._ws.close()
+            except Exception:
+                pass
+            self._ws = None
+
+    async def _ensure_connected(self) -> None:
+        """Ensure WebSocket connection is established."""
+        if self._ws is None:
+            await self.connect()
+
+    async def _send(self, message: Dict[str, Any]) -> None:
+        """Send a message over the WebSocket."""
+        await self._ensure_connected()
+        assert self._ws is not None
+        await self._ws.send(json.dumps(message))
+
+    async def _receive(self) -> Dict[str, Any]:
+        """Receive and parse a message from the WebSocket."""
+        assert self._ws is not None
+        raw = await asyncio.wait_for(self._ws.recv(), timeout=self._message_timeout)
+        return json.loads(raw)
+
+    async def _send_and_receive(self, message: Dict[str, Any]) -> Dict[str, Any]:
+        """Send a message and wait for response."""
+        await self._send(message)
+        response = await self._receive()
+
+        # Check for error response
+        if response.get("type") == "error":
+            error_data = response.get("data", {})
+            raise RuntimeError(
+                f"Server error: {error_data.get('message', 'Unknown error')} "
+                f"(code: {error_data.get('code', 'UNKNOWN')})"
+            )
+
+        return response
+
+    @classmethod
+    async def from_docker_image(
+        cls: Type[EnvClientT],
+        image: str,
+        provider: Optional["ContainerProvider"] = None,
+        **kwargs: Any,
+    ) -> EnvClientT:
+        """
+        Create an environment client by spinning up a Docker container.
+
+        Args:
+            image: Docker image name to run (e.g., "coding-env:latest")
+            provider: Container provider to use (defaults to LocalDockerProvider)
+            **kwargs: Additional arguments to pass to provider.start_container()
+
+        Returns:
+            Connected client instance
+        """
+        if provider is None:
+            provider = LocalDockerProvider()
+
+        # Start container
+        base_url = provider.start_container(image, **kwargs)
+
+        # Wait for server to be ready
+        provider.wait_for_ready(base_url)
+
+        # Create and connect client
+        client = cls(base_url=base_url, provider=provider)
+        await client.connect()
+
+        return client
+
+    @classmethod
+    async def from_env(
+        cls: Type[EnvClientT],
+        repo_id: str,
+        *,
+        use_docker: bool = True,
+        provider: Optional["ContainerProvider | RuntimeProvider"] = None,
+        **provider_kwargs: Any,
+    ) -> EnvClientT:
+        """
+        Create a client from a Hugging Face Space.
+
+        Args:
+            repo_id: Hugging Face space identifier ``{org}/{space}``.
+            use_docker: When ``True`` (default) pull from the HF registry and
+                launch via :class:`LocalDockerProvider`. When ``False`` run the
+                space locally with :class:`UVProvider`.
+            provider: Optional provider instance to reuse. Must be a
+                :class:`ContainerProvider` when ``use_docker=True`` and a
+                :class:`RuntimeProvider` otherwise.
+            provider_kwargs: Additional keyword arguments forwarded to
+                either the container provider's ``start_container`` (docker)
+                or to the ``UVProvider`` constructor/start (uv). When
+                ``use_docker=False``, the ``project_path`` argument can be
+                used to override the default git URL
+                (``git+https://huggingface.co/spaces/{repo_id}``).
+
+        Returns:
+            Connected client instance
+
+        Examples:
+            >>> # Pull and run from HF Docker registry
+            >>> env = await MyEnv.from_env("openenv/echo-env")
+            >>>
+            >>> # Run locally with UV (clones the space)
+            >>> env = await MyEnv.from_env("openenv/echo-env", use_docker=False)
+            >>>
+            >>> # Run from a local checkout
+            >>> env = await MyEnv.from_env(
+            ...     "openenv/echo-env",
+            ...     use_docker=False,
+            ...     project_path="/path/to/local/checkout"
+            ... )
+        """
+        # Extract start args that apply to both providers
+        start_args = {}
+        for key in ("port", "env_vars", "workers"):
+            if key in provider_kwargs:
+                start_args[key] = provider_kwargs.pop(key)
+
+        if use_docker:
+            # Docker mode: pull from HF registry
+            docker_provider = provider or LocalDockerProvider()
+            tag = provider_kwargs.pop("tag", "latest")
+            image = f"registry.hf.space/{repo_id.replace('/', '-')}:{tag}"
+            base_url = docker_provider.start_container(
+                image, **start_args, **provider_kwargs
+            )
+            docker_provider.wait_for_ready(base_url)
+
+            client = cls(base_url=base_url, provider=docker_provider)
+            await client.connect()
+            return client
+        else:
+            # UV mode: clone and run with uv
+            if provider is None:
+                uv_kwargs = dict(provider_kwargs)
+                project_path = uv_kwargs.pop("project_path", None)
+                if project_path is None:
+                    project_path = f"git+https://huggingface.co/spaces/{repo_id}"
+
+                provider = UVProvider(project_path=project_path, **uv_kwargs)
+            else:
+                if provider_kwargs:
+                    raise ValueError(
+                        "provider_kwargs cannot be used when supplying a provider instance"
+                    )
+
+            base_url = provider.start(**start_args)
+            provider.wait_for_ready()
+
+            client = cls(base_url=base_url, provider=provider)
+            await client.connect()
+            return client
+
+    @abstractmethod
+    def _step_payload(self, action: ActT) -> Dict[str, Any]:
+        """Convert an Action object to the JSON data expected by the env server."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def _parse_result(self, payload: Dict[str, Any]) -> StepResult[ObsT]:
+        """Convert a JSON response from the env server to StepResult[ObsT]."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def _parse_state(self, payload: Dict[str, Any]) -> StateT:
+        """Convert a JSON response from the state endpoint to a State object."""
+        raise NotImplementedError
+
+    async def reset(self, **kwargs: Any) -> StepResult[ObsT]:
+        """
+        Reset the environment with optional parameters.
+
+        Args:
+            **kwargs: Optional parameters passed to the environment's reset method.
+                     Common parameters include:
+                     - seed: Random seed for reproducibility
+                     - episode_id: Custom episode identifier
+
+        Returns:
+            StepResult containing initial observation
+        """
+        message = {
+            "type": "reset",
+            "data": kwargs,
+        }
+        response = await self._send_and_receive(message)
+        return self._parse_result(response.get("data", {}))
+
+    async def step(self, action: ActT, **kwargs: Any) -> StepResult[ObsT]:
+        """
+        Execute an action in the environment.
+
+        Args:
+            action: The action to execute
+            **kwargs: Optional parameters (currently ignored)
+
+        Returns:
+            StepResult containing observation, reward, and done status
+        """
+        message = {
+            "type": "step",
+            "data": self._step_payload(action),
+        }
+        response = await self._send_and_receive(message)
+        return self._parse_result(response.get("data", {}))
+
+    async def state(self) -> StateT:
+        """
+        Get the current environment state from the server.
+
+        Returns:
+            State object with environment state information
+        """
+        message = {"type": "state"}
+        response = await self._send_and_receive(message)
+        return self._parse_state(response.get("data", {}))
+
+    async def close(self) -> None:
+        """
+        Close the WebSocket connection and clean up resources.
+
+        If this client was created via from_docker_image() or from_env(),
+        this will also stop and remove the associated container/process.
+        """
+        await self.disconnect()
+
+        if self._provider is not None:
+            # Handle both ContainerProvider and RuntimeProvider
+            if hasattr(self._provider, "stop_container"):
+                self._provider.stop_container()
+            elif hasattr(self._provider, "stop"):
+                self._provider.stop()
+
+    async def __aenter__(self) -> "EnvClient":
+        """Enter async context manager, ensuring connection is established."""
+        await self.connect()
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb) -> None:
+        """Exit async context manager, closing connection."""
+        await self.close()
+
+    def __enter__(self) -> "EnvClient":
+        """Sync context manager entry - raises error suggesting async usage."""
+        raise TypeError(
+            "EnvClient is async by default. Use 'async with' instead of 'with', "
+            "or call .sync() to get a synchronous wrapper:\n"
+            "  async with client:  # async usage\n"
+            "  with client.sync():  # sync wrapper"
+        )
+
+    def __exit__(self, exc_type, exc_val, exc_tb) -> None:
+        """Sync context manager exit - should not be reached."""
+        pass  # pragma: no cover
+
+    def sync(self) -> "SyncEnvClient":
+        """
+        Return a synchronous wrapper around this async client.
+
+        Use this method when you need synchronous access to the environment
+        without async/await syntax. This is useful for:
+        - Integration with synchronous codebases
+        - Interactive/REPL usage
+        - Stopping async from "infecting" the call stack
+
+        Returns:
+            SyncEnvClient wrapper that provides synchronous methods
+
+        Example:
+            >>> # Create async client and get sync wrapper
+            >>> async_client = GenericEnvClient(base_url="http://localhost:8000")
+            >>> sync_client = async_client.sync()
+            >>>
+            >>> # Use synchronous API
+            >>> with sync_client:
+            ...     result = sync_client.reset()
+            ...     result = sync_client.step({"code": "print('hello')"})
+        """
+        from .sync_client import SyncEnvClient
+
+        return SyncEnvClient(self)

src/core/env_server/__init__.py

@@ -0,0 +1,150 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Core environment interfaces and types."""
+
+from .base_transforms import CompositeTransform, NullTransform
+from .exceptions import (
+    ConcurrencyConfigurationError,
+    EnvironmentFactoryError,
+    OpenEnvError,
+    SessionCapacityError,
+    SessionCreationError,
+    SessionNotFoundError,
+)
+from .http_server import create_app, create_fastapi_app, HTTPEnvServer
+from .interfaces import Environment, Message, ModelTokenizer, Transform
+
+try:
+    from .mcp_environment import MCPEnvironment
+except ModuleNotFoundError:
+    MCPEnvironment = None  # type: ignore[assignment]
+
+from .mcp_types import (
+    CallToolAction,
+    CallToolObservation,
+    JsonRpcError,
+    # JSON-RPC types
+    JsonRpcErrorCode,
+    JsonRpcRequest,
+    JsonRpcResponse,
+    ListToolsAction,
+    ListToolsObservation,
+    McpMethod,
+    RESERVED_TOOL_NAMES,
+    Tool,
+    ToolError,
+    ToolErrorType,
+    WSMCPMessage,
+    WSMCPResponse,
+)
+from .route_config import GetEndpointConfig
+from .serialization import (
+    deserialize_action,
+    deserialize_action_with_preprocessing,
+    serialize_observation,
+)
+from .types import (
+    Action,
+    BaseMessage,
+    ConcurrencyConfig,
+    HealthResponse,
+    HealthStatus,
+    Observation,
+    SchemaResponse,
+    ServerCapacityStatus,
+    ServerMode,
+    SessionInfo,
+    State,
+    WSCloseMessage,
+    WSErrorCode,
+    WSErrorResponse,
+    WSIncomingMessage,
+    WSObservationResponse,
+    WSResetMessage,
+    WSStateMessage,
+    WSStateResponse,
+    WSStepMessage,
+)
+
+try:
+    from .web_interface import create_web_interface_app, WebInterfaceManager
+except ModuleNotFoundError:
+    create_web_interface_app = None  # type: ignore[assignment]
+    WebInterfaceManager = None  # type: ignore[assignment]
+
+__all__ = [
+    # Core interfaces
+    "Environment",
+    "Transform",
+    "Message",
+    "ModelTokenizer",
+    # Types
+    "Action",
+    "Observation",
+    "State",
+    "SchemaResponse",
+    "HealthResponse",
+    # Enums
+    "HealthStatus",
+    "ServerMode",
+    "WSErrorCode",
+    # WebSocket message types
+    "BaseMessage",
+    "WSIncomingMessage",
+    "WSResetMessage",
+    "WSStepMessage",
+    "WSStateMessage",
+    "WSCloseMessage",
+    "WSObservationResponse",
+    "WSStateResponse",
+    "WSErrorResponse",
+    # Concurrency types
+    "ConcurrencyConfig",
+    "ServerCapacityStatus",
+    "SessionInfo",
+    # Exceptions
+    "OpenEnvError",
+    "ConcurrencyConfigurationError",
+    "SessionCapacityError",
+    "SessionNotFoundError",
+    "SessionCreationError",
+    "EnvironmentFactoryError",
+    # Base transforms
+    "CompositeTransform",
+    "NullTransform",
+    # HTTP Server
+    "HTTPEnvServer",
+    "create_app",
+    "create_fastapi_app",
+    # Web Interface
+    "create_web_interface_app",
+    "WebInterfaceManager",
+    # Serialization utilities
+    "deserialize_action",
+    "deserialize_action_with_preprocessing",
+    "serialize_observation",
+    # Route configuration
+    "GetEndpointConfig",
+    # MCP types
+    "Tool",
+    "ToolError",
+    "ToolErrorType",
+    "ListToolsAction",
+    "CallToolAction",
+    "ListToolsObservation",
+    "CallToolObservation",
+    "WSMCPMessage",
+    "WSMCPResponse",
+    "RESERVED_TOOL_NAMES",
+    "MCPEnvironment",
+    # JSON-RPC types
+    "JsonRpcErrorCode",
+    "JsonRpcError",
+    "JsonRpcRequest",
+    "JsonRpcResponse",
+    "McpMethod",
+]

src/core/env_server/base_transforms.py

@@ -0,0 +1,29 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Base transform implementations for composing environment-specific transforms."""
+
+from .interfaces import Transform
+from .types import Observation
+
+
+class CompositeTransform(Transform):
+    """Combines multiple transforms into a single transform."""
+
+    def __init__(self, transforms: list[Transform]):
+        self.transforms = transforms
+
+    def __call__(self, observation: Observation) -> Observation:
+        for transform in self.transforms:
+            observation = transform(observation)
+        return observation
+
+
+class NullTransform(Transform):
+    """Default transform that passes through unchanged."""
+
+    def __call__(self, observation: Observation) -> Observation:
+        return observation

src/core/env_server/exceptions.py

@@ -0,0 +1,105 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Custom exceptions for environment server operations."""
+
+from typing import Optional
+
+
+class OpenEnvError(Exception):
+    """Base exception for all OpenEnv errors."""
+
+    pass
+
+
+class ConcurrencyConfigurationError(OpenEnvError):
+    """
+    Raised when an environment is misconfigured for concurrent sessions.
+
+    This error is raised during server startup when max_concurrent_envs > 1
+    is specified for an environment that is not marked as SUPPORTS_CONCURRENT_SESSIONS.
+    """
+
+    def __init__(
+        self,
+        environment_name: str,
+        max_concurrent_envs: int,
+        message: Optional[str] = None,
+    ):
+        self.environment_name = environment_name
+        self.max_concurrent_envs = max_concurrent_envs
+
+        if message is None:
+            message = (
+                f"Environment '{environment_name}' is not marked as SUPPORTS_CONCURRENT_SESSIONS. "
+                f"Cannot run with max_concurrent_envs={max_concurrent_envs}. "
+                f"Either set max_concurrent_envs=1 or ensure the environment "
+                f"properly isolates session state and set SUPPORTS_CONCURRENT_SESSIONS=True."
+            )
+
+        super().__init__(message)
+
+
+class SessionCapacityError(OpenEnvError):
+    """
+    Raised when the server cannot accept new sessions due to capacity limits.
+
+    This error is raised when a new WebSocket connection is attempted but
+    the server has already reached max_concurrent_envs active sessions.
+    """
+
+    def __init__(
+        self,
+        active_sessions: int,
+        max_sessions: int,
+        message: Optional[str] = None,
+    ):
+        self.active_sessions = active_sessions
+        self.max_sessions = max_sessions
+
+        if message is None:
+            message = (
+                f"Server at capacity: {active_sessions}/{max_sessions} sessions active. "
+                f"Cannot accept new connections."
+            )
+
+        super().__init__(message)
+
+
+class SessionNotFoundError(OpenEnvError):
+    """Raised when attempting to access a session that does not exist."""
+
+    def __init__(self, session_id: str, message: Optional[str] = None):
+        self.session_id = session_id
+
+        if message is None:
+            message = f"Session '{session_id}' not found."
+
+        super().__init__(message)
+
+
+class SessionCreationError(OpenEnvError):
+    """Raised when a session cannot be created."""
+
+    def __init__(self, reason: str, message: Optional[str] = None):
+        self.reason = reason
+
+        if message is None:
+            message = f"Failed to create session: {reason}"
+
+        super().__init__(message)
+
+
+class EnvironmentFactoryError(OpenEnvError):
+    """Raised when the environment factory fails to create an instance."""
+
+    def __init__(self, factory_name: str, message: Optional[str] = None):
+        self.factory_name = factory_name
+
+        if message is None:
+            message = f"Environment factory '{factory_name}' failed to create instance."
+
+        super().__init__(message)

src/core/env_server/gradio_theme.py

@@ -0,0 +1,128 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Unified terminal-style theme for OpenEnv Gradio UI (light/dark)."""
+
+from __future__ import annotations
+
+import gradio as gr
+
+_MONO_FONTS = (
+    "JetBrains Mono",
+    "Fira Code",
+    "Cascadia Code",
+    "Consolas",
+    "ui-monospace",
+    "monospace",
+)
+
+_CORE_FONT = (
+    "Lato",
+    "Inter",
+    "Arial",
+    "Helvetica",
+    "sans-serif",
+)
+
+_ZERO_RADIUS = gr.themes.Size(
+    xxs="0px",
+    xs="0px",
+    sm="0px",
+    md="0px",
+    lg="0px",
+    xl="0px",
+    xxl="0px",
+)
+
+_GREEN_HUE = gr.themes.Color(
+    c50="#e6f4ea",
+    c100="#ceead6",
+    c200="#a8dab5",
+    c300="#6fcc8b",
+    c400="#3fb950",
+    c500="#238636",
+    c600="#1a7f37",
+    c700="#116329",
+    c800="#0a4620",
+    c900="#033a16",
+    c950="#04200d",
+)
+
+_NEUTRAL_HUE = gr.themes.Color(
+    c50="#f6f8fa",
+    c100="#eaeef2",
+    c200="#d0d7de",
+    c300="#afb8c1",
+    c400="#8c959f",
+    c500="#6e7781",
+    c600="#57606a",
+    c700="#424a53",
+    c800="#32383f",
+    c900="#24292f",
+    c950="#1b1f24",
+)
+
+OPENENV_GRADIO_THEME = gr.themes.Base(
+    primary_hue=_GREEN_HUE,
+    secondary_hue=_NEUTRAL_HUE,
+    neutral_hue=_NEUTRAL_HUE,
+    font=_CORE_FONT,
+    font_mono=_MONO_FONTS,
+    radius_size=_ZERO_RADIUS,
+).set(
+    body_background_fill="#ffffff",
+    background_fill_primary="#ffffff",
+    background_fill_secondary="#f6f8fa",
+    block_background_fill="#ffffff",
+    block_border_color="#ffffff",
+    block_label_text_color="#57606a",
+    block_title_text_color="#24292f",
+    border_color_primary="#d0d7de",
+    input_background_fill="#ffffff",
+    input_border_color="#d0d7de",
+    button_primary_background_fill="#1a7f37",
+    button_primary_background_fill_hover="#116329",
+    button_primary_text_color="#ffffff",
+    button_secondary_background_fill="#f6f8fa",
+    button_secondary_background_fill_hover="#eaeef2",
+    button_secondary_text_color="#24292f",
+    button_secondary_border_color="#d0d7de",
+    body_background_fill_dark="#0d1117",
+    background_fill_primary_dark="#0d1117",
+    background_fill_secondary_dark="#0d1117",
+    block_background_fill_dark="#0d1117",
+    block_border_color_dark="#0d1117",
+    block_label_text_color_dark="#8b949e",
+    block_title_text_color_dark="#c9d1d9",
+    border_color_primary_dark="#30363d",
+    input_background_fill_dark="#0d1117",
+    input_border_color_dark="#30363d",
+    button_primary_background_fill_dark="#30363d",
+    button_primary_background_fill_hover_dark="#484f58",
+    button_primary_text_color_dark="#c9d1d9",
+    button_secondary_background_fill_dark="#21262d",
+    button_secondary_background_fill_hover_dark="#30363d",
+    button_secondary_text_color_dark="#c9d1d9",
+    button_secondary_border_color_dark="#30363d",
+)
+
+OPENENV_GRADIO_CSS = """
+* { border-radius: 0 !important; }
+.col-left { padding: 16px !important; }
+.col-right { padding: 16px !important; }
+.prose, .markdown-text, .md,
+.prose > *, .markdown-text > * {
+    background: transparent !important;
+    border: none !important;
+    box-shadow: none !important;
+}
+.dark .col-left {
+    border-left-color: rgba(139, 148, 158, 0.4) !important;
+}
+.dark .col-right {
+    border-left-color: rgba(201, 209, 217, 0.3) !important;
+}
+"""

src/core/env_server/gradio_ui.py

@@ -0,0 +1,240 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Gradio-based web UI for OpenEnv environments.
+
+Replaces the legacy HTML/JavaScript interface when ENABLE_WEB_INTERFACE is set.
+Mount at /web via gr.mount_gradio_app() from create_web_interface_app().
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from typing import Any, Dict, List, Optional
+
+import gradio as gr
+
+from .types import EnvironmentMetadata
+
+
+def _escape_md(text: str) -> str:
+    """Escape Markdown special characters in user-controlled content."""
+    return re.sub(r"([\\`*_\{\}\[\]()#+\-.!|~>])", r"\\\1", str(text))
+
+
+def _format_observation(data: Dict[str, Any]) -> str:
+    """Format reset/step response for Markdown display."""
+    lines: List[str] = []
+    obs = data.get("observation", {})
+    if isinstance(obs, dict):
+        if obs.get("prompt"):
+            lines.append(f"**Prompt:**\n\n{_escape_md(obs['prompt'])}\n")
+        messages = obs.get("messages", [])
+        if messages:
+            lines.append("**Messages:**\n")
+            for msg in messages:
+                sender = _escape_md(str(msg.get("sender_id", "?")))
+                content = _escape_md(str(msg.get("content", "")))
+                cat = _escape_md(str(msg.get("category", "")))
+                lines.append(f"- `[{cat}]` Player {sender}: {content}")
+            lines.append("")
+    reward = data.get("reward")
+    done = data.get("done")
+    if reward is not None:
+        lines.append(f"**Reward:** `{reward}`")
+    if done is not None:
+        lines.append(f"**Done:** `{done}`")
+    return "\n".join(lines) if lines else "*No observation data*"
+
+
+def _readme_section(metadata: Optional[EnvironmentMetadata]) -> str:
+    """README content for the left panel."""
+    if not metadata or not metadata.readme_content:
+        return "*No README available.*"
+    return metadata.readme_content
+
+
+def get_gradio_display_title(
+    metadata: Optional[EnvironmentMetadata],
+    fallback: str = "OpenEnv Environment",
+) -> str:
+    """Return the title used for the Gradio app (browser tab and Blocks)."""
+    name = metadata.name if metadata else fallback
+    return f"OpenEnv Agentic Environment: {name}"
+
+
+def build_gradio_app(
+    web_manager: Any,
+    action_fields: List[Dict[str, Any]],
+    metadata: Optional[EnvironmentMetadata],
+    is_chat_env: bool,
+    title: str = "OpenEnv Environment",
+    quick_start_md: Optional[str] = None,
+) -> gr.Blocks:
+    """
+    Build a Gradio Blocks app for the OpenEnv web interface.
+
+    Args:
+        web_manager: WebInterfaceManager (reset/step_environment, get_state).
+        action_fields: Field dicts from _extract_action_fields(action_cls).
+        metadata: Environment metadata for README/name.
+        is_chat_env: If True, single message textbox; else form from action_fields.
+        title: App title (overridden by metadata.name when present; see get_gradio_display_title).
+        quick_start_md: Optional Quick Start markdown (class names already replaced).
+
+    Returns:
+        gr.Blocks to mount with gr.mount_gradio_app(app, blocks, path="/web").
+    """
+    readme_content = _readme_section(metadata)
+    display_title = get_gradio_display_title(metadata, fallback=title)
+
+    async def reset_env():
+        try:
+            data = await web_manager.reset_environment()
+            obs_md = _format_observation(data)
+            return (
+                obs_md,
+                json.dumps(data, indent=2),
+                "Environment reset successfully.",
+            )
+        except Exception as e:
+            return ("", "", f"Error: {e}")
+
+    def _step_with_action(action_data: Dict[str, Any]):
+        async def _run():
+            try:
+                data = await web_manager.step_environment(action_data)
+                obs_md = _format_observation(data)
+                return (
+                    obs_md,
+                    json.dumps(data, indent=2),
+                    "Step complete.",
+                )
+            except Exception as e:
+                return ("", "", f"Error: {e}")
+
+        return _run
+
+    async def step_chat(message: str):
+        if not (message or str(message).strip()):
+            return ("", "", "Please enter an action message.")
+        action = {"message": str(message).strip()}
+        return await _step_with_action(action)()
+
+    def get_state_sync():
+        try:
+            data = web_manager.get_state()
+            return json.dumps(data, indent=2)
+        except Exception as e:
+            return f"Error: {e}"
+
+    with gr.Blocks(title=display_title) as demo:
+        with gr.Row():
+            with gr.Column(scale=1, elem_classes="col-left"):
+                if quick_start_md:
+                    with gr.Accordion("Quick Start", open=True):
+                        gr.Markdown(quick_start_md)
+                with gr.Accordion("README", open=False):
+                    gr.Markdown(readme_content)
+
+            with gr.Column(scale=2, elem_classes="col-right"):
+                obs_display = gr.Markdown(
+                    value=("# Playground\n\nClick **Reset** to start a new episode."),
+                )
+                with gr.Group():
+                    if is_chat_env:
+                        action_input = gr.Textbox(
+                            label="Action message",
+                            placeholder="e.g. Enter your message...",
+                        )
+                        step_inputs = [action_input]
+                        step_fn = step_chat
+                    else:
+                        step_inputs = []
+                        for field in action_fields:
+                            name = field["name"]
+                            field_type = field.get("type", "text")
+                            label = name.replace("_", " ").title()
+                            placeholder = field.get("placeholder", "")
+                            if field_type == "checkbox":
+                                inp = gr.Checkbox(label=label)
+                            elif field_type == "number":
+                                inp = gr.Number(label=label)
+                            elif field_type == "select":
+                                choices = field.get("choices") or []
+                                inp = gr.Dropdown(
+                                    choices=choices,
+                                    label=label,
+                                    allow_custom_value=False,
+                                )
+                            elif field_type in ("textarea", "tensor"):
+                                inp = gr.Textbox(
+                                    label=label,
+                                    placeholder=placeholder,
+                                    lines=3,
+                                )
+                            else:
+                                inp = gr.Textbox(
+                                    label=label,
+                                    placeholder=placeholder,
+                                )
+                            step_inputs.append(inp)
+
+                        async def step_form(*values):
+                            if not action_fields:
+                                return await _step_with_action({})()
+                            action_data = {}
+                            for i, field in enumerate(action_fields):
+                                if i >= len(values):
+                                    break
+                                name = field["name"]
+                                val = values[i]
+                                if field.get("type") == "checkbox":
+                                    action_data[name] = bool(val)
+                                elif val is not None and val != "":
+                                    action_data[name] = val
+                            return await _step_with_action(action_data)()
+
+                        step_fn = step_form
+
+                    with gr.Row():
+                        step_btn = gr.Button("Step", variant="primary")
+                        reset_btn = gr.Button("Reset", variant="secondary")
+                        state_btn = gr.Button("Get state", variant="secondary")
+                    with gr.Row():
+                        status = gr.Textbox(
+                            label="Status",
+                            interactive=False,
+                        )
+                    raw_json = gr.Code(
+                        label="Raw JSON response",
+                        language="json",
+                        interactive=False,
+                    )
+
+        reset_btn.click(
+            fn=reset_env,
+            outputs=[obs_display, raw_json, status],
+        )
+        step_btn.click(
+            fn=step_fn,
+            inputs=step_inputs,
+            outputs=[obs_display, raw_json, status],
+        )
+        if is_chat_env:
+            action_input.submit(
+                fn=step_fn,
+                inputs=step_inputs,
+                outputs=[obs_display, raw_json, status],
+            )
+        state_btn.click(
+            fn=get_state_sync,
+            outputs=[raw_json],
+        )
+
+    return demo

src/core/env_server/http_server.py

@@ -0,0 +1,1646 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+HTTP server wrapper for Environment instances.
+
+This module provides utilities to wrap any Environment subclass and expose it
+over HTTP and WebSocket endpoints that EnvClient can consume.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import inspect
+import json
+import logging
+import os
+import time
+import uuid
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import AsyncExitStack
+from typing import Any, AsyncContextManager, Callable, cast, Dict, Optional, Type
+
+_MISSING = object()
+
+from fastapi import (
+    Body,
+    FastAPI,
+    HTTPException,
+    Request,
+    status,
+    WebSocket,
+    WebSocketDisconnect,
+)
+from pydantic import ValidationError
+
+from .interfaces import Environment
+from .mcp_environment import get_server_tools
+from .mcp_types import (
+    JsonRpcErrorCode,
+    JsonRpcRequest,
+    JsonRpcResponse,
+    McpMethod,
+    WSMCPMessage,
+    WSMCPResponse,
+)
+from .route_config import GetEndpointConfig, register_get_endpoints
+from .serialization import deserialize_action, serialize_observation
+from .types import (
+    Action,
+    ConcurrencyConfig,
+    EnvironmentMetadata,
+    HealthResponse,
+    HealthStatus,
+    Observation,
+    ResetRequest,
+    ResetResponse,
+    SchemaResponse,
+    ServerCapacityStatus,
+    ServerMode,
+    SessionInfo,
+    State,
+    StepRequest,
+    StepResponse,
+    WSCloseMessage,
+    WSErrorCode,
+    WSErrorResponse,
+    WSObservationResponse,
+    WSResetMessage,
+    WSStateMessage,
+    WSStateResponse,
+    WSStepMessage,
+)
+
+
+def _make_json_serializable(obj: Any) -> Any:
+    """
+    Convert an object to a JSON-serializable form.
+
+    Handles Pydantic models, dataclasses, and other common types.
+
+    Args:
+        obj: The object to convert
+
+    Returns:
+        A JSON-serializable representation of the object
+    """
+    if obj is None:
+        return None
+    if isinstance(obj, (str, int, float, bool)):
+        return obj
+    if isinstance(obj, (list, tuple)):
+        return [_make_json_serializable(item) for item in obj]
+    if isinstance(obj, dict):
+        return {k: _make_json_serializable(v) for k, v in obj.items()}
+    if hasattr(obj, "model_dump"):
+        # Pydantic model
+        return obj.model_dump()
+    if hasattr(obj, "__dict__"):
+        # Object with __dict__
+        return {k: _make_json_serializable(v) for k, v in obj.__dict__.items()}
+    # Fallback to string representation
+    return str(obj)
+
+
+from .exceptions import (
+    ConcurrencyConfigurationError,
+    EnvironmentFactoryError,
+    SessionCapacityError,
+)
+
+
+class HTTPEnvServer:
+    """
+    HTTP server wrapper for Environment instances.
+
+    This class wraps an Environment and exposes its reset(), step(), and state
+    methods as HTTP and WebSocket endpoints compatible with EnvClient.
+
+    The server expects:
+    - Action deserialization: Converts JSON dict to Action subclass
+    - Observation serialization: Converts Observation subclass to JSON dict
+
+    Example:
+        >>> from core.env_server import HTTPEnvServer
+        >>> from envs.coding_env.server import CodeExecutionEnvironment
+        >>> from envs.coding_env.models import CodeAction, CodeObservation
+        >>>
+        >>> # Pass environment class (factory pattern)
+        >>> server = HTTPEnvServer(
+        ...     env=CodeExecutionEnvironment,
+        ...     action_cls=CodeAction,
+        ...     observation_cls=CodeObservation,
+        ...     max_concurrent_envs=4,
+        ... )
+        >>>
+        >>> # Register routes with FastAPI
+        >>> from fastapi import FastAPI
+        >>> app = FastAPI()
+        >>> server.register_routes(app)
+    """
+
+    def __init__(
+        self,
+        env: Callable[[], Environment],
+        action_cls: Type[Action],
+        observation_cls: Type[Observation],
+        max_concurrent_envs: Optional[int] = None,
+        concurrency_config: Optional[ConcurrencyConfig] = None,
+    ):
+        """
+        Initialize HTTP server wrapper.
+
+        Args:
+            env: Environment factory (callable) that creates new instances.
+                 Will be called to create a new environment for each WebSocket session.
+            action_cls: The Action subclass this environment expects
+            observation_cls: The Observation subclass this environment returns
+            max_concurrent_envs: Maximum number of concurrent WebSocket sessions.
+                                 Mutually exclusive with concurrency_config.
+            concurrency_config: Optional ConcurrencyConfig for advanced concurrency settings.
+                                Mutually exclusive with max_concurrent_envs.
+
+        Raises:
+            ValueError: If both max_concurrent_envs and concurrency_config are provided.
+            ConcurrencyConfigurationError: If max_concurrent_envs > 1 for an
+                environment that is not marked as SUPPORTS_CONCURRENT_SESSIONS.
+        """
+        # Validate that env is callable
+        if not callable(env):
+            raise TypeError(
+                f"env must be a callable (class or factory function), got {type(env)}. "
+                f"Pass the environment class (e.g., MyEnvironment) not an instance (e.g., MyEnvironment())."
+            )
+
+        self._env_factory: Callable[[], Environment] = env
+
+        # Handle concurrency configuration
+        if max_concurrent_envs is not None and concurrency_config is not None:
+            raise ValueError(
+                "Cannot specify both 'max_concurrent_envs' and 'concurrency_config'. "
+                "Please use only one method to configure concurrency."
+            )
+
+        if concurrency_config is not None:
+            self._concurrency_config = concurrency_config
+        elif max_concurrent_envs is not None:
+            self._concurrency_config = ConcurrencyConfig(
+                max_concurrent_envs=max_concurrent_envs,
+                session_timeout=None,
+            )
+        else:
+            # Default configuration
+            self._concurrency_config = ConcurrencyConfig(
+                max_concurrent_envs=1,
+                session_timeout=None,
+            )
+
+        self._max_concurrent_envs = self._concurrency_config.max_concurrent_envs
+
+        # Validate concurrency configuration
+        self._validate_concurrency_safety()
+
+        self.action_cls = action_cls
+        self.observation_cls = observation_cls
+
+        # Session management for WebSocket connections
+        self._sessions: Dict[str, Optional[Environment]] = {}
+        self._session_executors: Dict[str, ThreadPoolExecutor] = {}
+        self._session_stacks: Dict[str, AsyncExitStack] = {}
+        self._session_info: Dict[str, SessionInfo] = {}
+        self._session_lock = asyncio.Lock()
+
+        # Create thread pool for running sync code in async context
+        # This is needed for environments using sync libraries (e.g., Playwright)
+        self._executor = ThreadPoolExecutor(max_workers=32)
+
+        # Idle session reaper configuration.
+        # Timeout is taken from ConcurrencyConfig.session_timeout;
+        # None means no timeout (default — reaper is a no-op).
+        self._session_idle_timeout_s: Optional[float] = (
+            self._concurrency_config.session_timeout
+        )
+        self._reaper_task: Optional[asyncio.Task[None]] = None
+
+    def _validate_concurrency_safety(self) -> None:
+        """
+        Validate that the environment supports the configured concurrency level.
+
+        Raises:
+            ConcurrencyConfigurationError: If max_concurrent_envs > 1 for an
+                environment that is not marked as SUPPORTS_CONCURRENT_SESSIONS.
+        """
+        if self._max_concurrent_envs <= 1:
+            return
+
+        if inspect.isclass(self._env_factory):
+            env_cls = self._env_factory
+        else:
+            _temp_env = self._env_factory()
+            env_cls = type(_temp_env)
+            _temp_env.close()
+            del _temp_env
+
+        if not getattr(env_cls, "SUPPORTS_CONCURRENT_SESSIONS", False):
+            raise ConcurrencyConfigurationError(
+                environment_name=env_cls.__name__,
+                max_concurrent_envs=self._max_concurrent_envs,
+            )
+
+    def get_capacity_status(self) -> ServerCapacityStatus:
+        """
+        Get the current capacity status of the server.
+
+        Returns:
+            ServerCapacityStatus with current session counts and availability.
+        """
+        return ServerCapacityStatus.from_counts(
+            active=len(self._sessions),
+            max_sessions=self._max_concurrent_envs,
+        )
+
+    async def _run_sync_in_thread_pool(
+        self, func: Callable[..., Observation], *args, **kwargs
+    ) -> Observation:
+        """Run a synchronous function in the thread pool executor."""
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(self._executor, lambda: func(*args, **kwargs))
+
+    def _get_valid_kwargs(
+        self,
+        sig: inspect.Signature,
+        kwargs: Dict[str, Any],
+        skip_params: Optional[set[str]] = None,
+    ) -> Dict[str, Any]:
+        """Filter kwargs to only include parameters accepted by the function signature."""
+        if skip_params is None:
+            skip_params = set()
+
+        valid_kwargs = {}
+
+        has_kwargs = any(
+            p.kind == inspect.Parameter.VAR_KEYWORD for p in sig.parameters.values()
+        )
+
+        for k, v in kwargs.items():
+            if k in sig.parameters or has_kwargs:
+                if k not in skip_params:
+                    valid_kwargs[k] = v
+
+        return valid_kwargs
+
+    async def _create_session(self) -> tuple[str, Environment]:
+        """
+        Create a new WebSocket session with its own environment instance.
+
+        Returns:
+            Tuple of (session_id, environment)
+
+        Raises:
+            SessionCapacityError: If max concurrent sessions reached
+            EnvironmentFactoryError: If the factory fails to create an environment
+        """
+        async with self._session_lock:
+            if len(self._sessions) >= self._max_concurrent_envs:
+                raise SessionCapacityError(
+                    active_sessions=len(self._sessions),
+                    max_sessions=self._max_concurrent_envs,
+                )
+
+            session_id = str(uuid.uuid4())
+            current_time = time.time()
+
+            # Create executor and reserve slot so capacity is not exceeded while
+            # we create the env outside the lock (avoids blocking other sessions)
+            executor = ThreadPoolExecutor(max_workers=1)
+            self._session_executors[session_id] = executor
+            self._sessions[session_id] = None  # placeholder until env is ready
+
+        try:
+            # Create environment in the executor thread (outside lock)
+            loop = asyncio.get_event_loop()
+            env = await loop.run_in_executor(executor, self._env_factory)
+        except Exception as e:
+            async with self._session_lock:
+                executor.shutdown(wait=False)
+                self._session_executors.pop(session_id, None)
+                self._sessions.pop(session_id, None)
+            factory_name = getattr(
+                self._env_factory, "__name__", str(self._env_factory)
+            )
+            raise EnvironmentFactoryError(factory_name) from e
+
+        # Hold the MCP session open for the lifetime of this session,
+        # matching the WebSocket path's AsyncExitStack pattern.  This
+        # prevents per-request MCP transport teardown/reconnection and
+        # preserves FastMCP session state (ctx.set_state / ctx.get_state)
+        # across HTTP calls within the same OpenEnv session.
+        stack = AsyncExitStack()
+        try:
+            mcp_session_factory = getattr(env, "mcp_session", None)
+            if callable(mcp_session_factory):
+                mcp_session_cm = cast(AsyncContextManager[Any], mcp_session_factory())
+                await stack.enter_async_context(mcp_session_cm)
+        except Exception:
+            # MCP transport failed to start — clean up the reserved slot,
+            # the env, and the executor so they don't leak permanently
+            # against _max_concurrent_envs.
+            await stack.aclose()  # best-effort
+            async with self._session_lock:
+                self._sessions.pop(session_id, None)
+                self._session_executors.pop(session_id, None)
+                self._session_info.pop(session_id, None)
+            await self._cleanup_session_resources(env, executor)
+            raise
+
+        async with self._session_lock:
+            self._sessions[session_id] = env
+            self._session_stacks[session_id] = stack
+            now = time.time()
+            self._session_info[session_id] = SessionInfo(
+                session_id=session_id,
+                created_at=current_time,
+                last_activity_at=now,
+                step_count=0,
+                environment_type=type(env).__name__,
+            )
+
+        return session_id, env
+
+    async def _destroy_session(self, session_id: str) -> None:
+        """
+        Destroy a WebSocket session and cleanup resources.
+
+        Args:
+            session_id: The session ID to destroy
+        """
+        async with self._session_lock:
+            env = self._sessions.pop(session_id, None)
+            executor = self._session_executors.pop(session_id, None)
+            stack = self._session_stacks.pop(session_id, None)
+            self._session_info.pop(session_id, None)
+
+        await self._cleanup_session_resources(env, executor, stack)
+
+    async def _cleanup_session_resources(
+        self,
+        env: Optional[Environment],
+        executor: Optional[ThreadPoolExecutor],
+        stack: Optional[AsyncExitStack] = None,
+    ) -> None:
+        """Close an environment and shut down its executor (best-effort)."""
+        # Close the MCP session stack first — this gracefully exits the
+        # mcp_session() context (and the underlying FastMCP Client session)
+        # before we tear down the environment references.
+        if stack is not None:
+            try:
+                await stack.aclose()
+            except Exception:
+                pass  # Best effort cleanup
+
+        # Run close() in the same executor where the env was created
+        # This is required for thread-sensitive libraries like Playwright/greenlet
+        if env is not None:
+            if executor is not None:
+                try:
+                    loop = asyncio.get_event_loop()
+                    await loop.run_in_executor(executor, env.close)
+                except Exception:
+                    # If executor close fails, try direct close as fallback
+                    try:
+                        env.close()
+                    except Exception:
+                        pass  # Best effort cleanup
+            else:
+                try:
+                    env.close()
+                except Exception:
+                    pass  # Best effort cleanup
+
+        # Shutdown executor after close is done
+        if executor is not None:
+            executor.shutdown(wait=False)
+
+    def _update_session_activity(
+        self, session_id: str, increment_step: bool = False
+    ) -> None:
+        """
+        Update session activity timestamp and optionally increment step count.
+
+        Args:
+            session_id: The session ID to update
+            increment_step: If True, increment the step count
+        """
+        if session_id in self._session_info:
+            self._session_info[session_id].last_activity_at = time.time()
+            if increment_step:
+                self._session_info[session_id].step_count += 1
+
+    async def _reap_idle_sessions(self) -> None:
+        """Background task that periodically destroys sessions idle beyond the timeout."""
+        timeout = self._session_idle_timeout_s
+        if timeout is None:
+            return  # no timeout configured — noop
+        interval = max(timeout / 4, 5.0)  # check frequently enough
+        while True:
+            try:
+                await asyncio.sleep(interval)
+                now = time.time()
+                stale_ids: list[str] = []
+                async with self._session_lock:
+                    for sid, info in self._session_info.items():
+                        if now - info.last_activity_at > timeout:
+                            stale_ids.append(sid)
+                for sid in stale_ids:
+                    # Re-check under lock: activity may have arrived since
+                    # the snapshot was taken, making this session active again.
+                    # Refresh `now` so slow _destroy_session calls don't cause
+                    # subsequent entries to be validated against a stale clock.
+                    now = time.time()
+                    async with self._session_lock:
+                        info = self._session_info.get(sid)
+                        if info is None or (now - info.last_activity_at) <= timeout:
+                            continue
+                    await self._destroy_session(sid)
+            except asyncio.CancelledError:
+                break
+            except Exception as exc:
+                logging.getLogger(__name__).warning(
+                    "Idle-session reaper encountered an error (will retry): %s",
+                    exc,
+                )
+
+    def _start_reaper(self) -> None:
+        """Start the idle-session reaper if a timeout is configured."""
+        if self._session_idle_timeout_s is not None and self._reaper_task is None:
+            self._reaper_task = asyncio.create_task(self._reap_idle_sessions())
+
+    def _stop_reaper(self) -> None:
+        """Cancel the reaper background task."""
+        if self._reaper_task is not None:
+            self._reaper_task.cancel()
+            self._reaper_task = None
+
+    def get_session_info(self, session_id: str) -> Optional[SessionInfo]:
+        """
+        Get information about a specific session.
+
+        Args:
+            session_id: The session ID to query
+
+        Returns:
+            SessionInfo if the session exists, None otherwise
+        """
+        return self._session_info.get(session_id)
+
+    async def _run_in_session_executor(
+        self, session_id: str, func: Callable[..., Observation], *args, **kwargs
+    ) -> Observation:
+        """Run a synchronous function in the session's thread pool executor."""
+        executor = self._session_executors.get(session_id, self._executor)
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(executor, lambda: func(*args, **kwargs))
+
+    @property
+    def active_sessions(self) -> int:
+        """Return the number of active WebSocket sessions."""
+        return len(self._sessions)
+
+    @property
+    def max_concurrent_envs(self) -> int:
+        """Return the maximum number of concurrent environments."""
+        return self._max_concurrent_envs
+
+    @property
+    def is_concurrency_safe(self) -> bool:
+        """Return whether the environment is marked as concurrency safe."""
+        import inspect
+
+        if inspect.isclass(self._env_factory):
+            return getattr(self._env_factory, "SUPPORTS_CONCURRENT_SESSIONS", False)
+        else:
+            _temp_env = self._env_factory()
+            result = getattr(_temp_env, "SUPPORTS_CONCURRENT_SESSIONS", False)
+            _temp_env.close()
+            del _temp_env
+            return result
+
+    @property
+    def concurrency_config(self) -> ConcurrencyConfig:
+        """Return the concurrency configuration."""
+        return self._concurrency_config
+
+    def register_routes(
+        self, app: FastAPI, mode: ServerMode | str = ServerMode.SIMULATION
+    ) -> None:
+        """
+        Register HTTP routes on a FastAPI application.
+
+        Args:
+            app: FastAPI application instance
+            mode: Server mode - either SIMULATION or PRODUCTION (or string equivalents).
+                  In production mode, simulation control endpoints (/reset, /step, /state)
+                  are NOT registered. Only safe endpoints (/health, /schema, /metadata, /ws)
+                  are available. Defaults to SIMULATION for backwards compatibility.
+
+        Raises:
+            ValueError: If mode is not a valid ServerMode or string equivalent.
+        """
+        # Convert string to ServerMode enum for backwards compatibility
+        if isinstance(mode, str):
+            try:
+                mode = ServerMode(mode.lower())
+            except ValueError:
+                valid_modes = [m.value for m in ServerMode]
+                raise ValueError(
+                    f"Invalid mode: '{mode}'. Must be one of: {valid_modes}"
+                )
+
+        # Wire up idle-session reaper lifecycle via app events
+        server_ref = self
+
+        async def _start_session_reaper() -> None:
+            server_ref._start_reaper()
+
+        async def _stop_session_reaper() -> None:
+            server_ref._stop_reaper()
+
+        if not getattr(app.router, "_openenv_reaper_registered", False):
+            app.router.on_startup.append(_start_session_reaper)
+            app.router.on_shutdown.append(_stop_session_reaper)
+            app.router._openenv_reaper_registered = True  # type: ignore[attr-defined]
+
+        # Helper function to handle reset endpoint
+        async def reset_handler(
+            request: ResetRequest = Body(default_factory=ResetRequest),
+        ) -> ResetResponse:
+            """Reset endpoint - returns initial observation."""
+            _env = self._env_factory()
+
+            try:
+                kwargs = request.model_dump(exclude_unset=True)
+
+                is_async = _env.reset_async.__func__ is not Environment.reset_async
+
+                if is_async:
+                    sig = inspect.signature(_env.reset_async)
+                else:
+                    sig = inspect.signature(_env.reset)
+                valid_kwargs = self._get_valid_kwargs(sig, kwargs)
+
+                if is_async:
+                    observation = await _env.reset_async(**valid_kwargs)
+                else:
+                    observation = await self._run_sync_in_thread_pool(
+                        _env.reset, **valid_kwargs
+                    )
+                return ResetResponse(**serialize_observation(observation))
+            finally:
+                _env.close()
+
+        # Helper function to handle step endpoint
+        async def step_handler(request: StepRequest) -> StepResponse:
+            """Step endpoint - executes action and returns observation."""
+            action_data = request.action
+
+            try:
+                action = deserialize_action(action_data, self.action_cls)
+            except ValidationError as e:
+                raise HTTPException(
+                    status_code=status.HTTP_422_UNPROCESSABLE_CONTENT, detail=e.errors()
+                )
+
+            _env = self._env_factory()
+
+            try:
+                kwargs = request.model_dump(exclude_unset=True, exclude={"action"})
+
+                is_async = _env.step_async.__func__ is not Environment.step_async
+
+                if is_async:
+                    sig = inspect.signature(_env.step_async)
+                else:
+                    sig = inspect.signature(_env.step)
+                valid_kwargs = self._get_valid_kwargs(
+                    sig, kwargs, skip_params={"action"}
+                )
+
+                if is_async:
+                    observation = await _env.step_async(action, **valid_kwargs)
+                else:
+                    observation = await self._run_sync_in_thread_pool(
+                        _env.step, action, **valid_kwargs
+                    )
+
+                return StepResponse(**serialize_observation(observation))
+            finally:
+                _env.close()
+
+        # Helper function to handle MCP endpoint
+        async def mcp_handler(
+            request: JsonRpcRequest,
+            session_env: Optional[Environment] = None,
+            session_id: Optional[str] = None,
+        ) -> JsonRpcResponse:
+            """
+            Handle MCP JSON-RPC requests.
+
+            Supports tools/list and tools/call methods in JSON-RPC 2.0 format,
+            plus OpenEnv session lifecycle methods for HTTP MCP:
+            - openenv/session/create
+            - openenv/session/close
+            """
+            method = request.method
+            request_id = request.id
+            params = request.params
+            if not isinstance(params, dict):
+                return JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.INVALID_PARAMS,
+                    "Params must be an object",
+                    request_id=request_id,
+                )
+
+            # OpenEnv extension methods for explicit MCP session management.
+            # This enables persistent MCP lifecycles over HTTP /mcp, matching WebSocket semantics.
+            if method == "openenv/session/create":
+                if session_env is not None and session_id is not None:
+                    return JsonRpcResponse.success(
+                        result={"session_id": session_id},
+                        request_id=request_id,
+                    )
+                try:
+                    created_session_id, _ = await self._create_session()
+                except SessionCapacityError as e:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.SERVER_ERROR,
+                        str(e),
+                        request_id=request_id,
+                        data={
+                            "active_sessions": e.active_sessions,
+                            "max_sessions": e.max_sessions,
+                        },
+                    )
+                except EnvironmentFactoryError as e:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.SERVER_ERROR,
+                        str(e),
+                        request_id=request_id,
+                        data={"factory_name": e.factory_name},
+                    )
+                return JsonRpcResponse.success(
+                    result={"session_id": created_session_id},
+                    request_id=request_id,
+                )
+
+            if method == "openenv/session/close":
+                target_session_id = params.get("session_id")
+                if not target_session_id:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_PARAMS,
+                        "Invalid params - 'session_id' is required",
+                        request_id=request_id,
+                    )
+
+                if session_id is not None and target_session_id == session_id:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_REQUEST,
+                        "Cannot close active WebSocket-managed session via MCP method",
+                        request_id=request_id,
+                    )
+
+                async with self._session_lock:
+                    env = self._sessions.pop(target_session_id, _MISSING)
+                    if env is not _MISSING:
+                        executor = self._session_executors.pop(target_session_id, None)
+                        stack = self._session_stacks.pop(target_session_id, None)
+                        self._session_info.pop(target_session_id, None)
+                    else:
+                        executor = None
+                        stack = None
+
+                if env is _MISSING:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_PARAMS,
+                        f"Unknown session_id: {target_session_id}",
+                        request_id=request_id,
+                    )
+
+                if env is None:
+                    # Session slot reserved but env factory still running;
+                    # re-insert the placeholder AND the executor so
+                    # _create_session can finish and the executor remains
+                    # tracked for eventual shutdown.
+                    async with self._session_lock:
+                        self._sessions[target_session_id] = None
+                        if executor is not None:
+                            self._session_executors[target_session_id] = executor
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_REQUEST,
+                        f"Session {target_session_id} is still initializing; retry shortly",
+                        request_id=request_id,
+                    )
+
+                # env/executor/stack cleanup outside the lock
+                await self._cleanup_session_resources(env, executor, stack)
+                return JsonRpcResponse.success(
+                    result={"session_id": target_session_id, "closed": True},
+                    request_id=request_id,
+                )
+
+            requested_session_id = params.get("session_id")
+            managed_session_id = session_id
+
+            # Use provided session environment or create temporary one
+            if session_env is not None:
+                _env = session_env
+                should_close = False
+            elif requested_session_id:
+                async with self._session_lock:
+                    _env = self._sessions.get(requested_session_id, _MISSING)
+
+                if _env is _MISSING:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_PARAMS,
+                        f"Unknown session_id: {requested_session_id}",
+                        request_id=request_id,
+                    )
+
+                if _env is None:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INVALID_REQUEST,
+                        f"Session {requested_session_id} is still initializing; retry shortly",
+                        request_id=request_id,
+                    )
+
+                should_close = False
+                managed_session_id = requested_session_id
+            else:
+                _env = self._env_factory()
+                should_close = True
+            try:
+                mcp_client = getattr(_env, "mcp_client", None)
+                mcp_server = getattr(_env, "mcp_server", None)
+                mcp_session_factory = getattr(_env, "mcp_session", None)
+
+                if method == McpMethod.TOOLS_LIST:
+                    # Check if environment is MCP-enabled
+                    if mcp_client is None and mcp_server is None:
+                        return JsonRpcResponse.error_response(
+                            JsonRpcErrorCode.INTERNAL_ERROR,
+                            "Environment does not support MCP",
+                            request_id=request_id,
+                        )
+
+                    if mcp_client:
+                        if managed_session_id and mcp_client.is_connected():
+                            # Session-managed with live transport — call
+                            # directly, no redundant re-entry.
+                            tools = await mcp_client.list_tools()
+                        elif callable(mcp_session_factory):
+                            # Stateless request, or session-managed but the
+                            # background transport was lost: (re-)open.
+                            mcp_session_cm = cast(
+                                AsyncContextManager[Any], mcp_session_factory()
+                            )
+                            async with mcp_session_cm:
+                                tools = await mcp_client.list_tools()
+                        else:
+                            async with mcp_client:
+                                tools = await mcp_client.list_tools()
+
+                        return JsonRpcResponse.success(
+                            result={
+                                "tools": [
+                                    t.model_dump()
+                                    if hasattr(t, "model_dump")
+                                    else dict(t)
+                                    for t in tools
+                                ]
+                            },
+                            request_id=request_id,
+                        )
+
+                    if mcp_server:
+                        tools = []
+                        for _tool_name, tool in get_server_tools(mcp_server).items():
+                            tools.append(
+                                {
+                                    "name": tool.name,
+                                    "description": tool.description or "",
+                                    "inputSchema": tool.parameters or {},
+                                }
+                            )
+                        return JsonRpcResponse.success(
+                            result={"tools": tools},
+                            request_id=request_id,
+                        )
+
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.INTERNAL_ERROR,
+                        "MCP server not available",
+                        request_id=request_id,
+                    )
+
+                elif method == McpMethod.TOOLS_CALL:
+                    tool_name = params.get("name")
+                    arguments = params.get("arguments", {})
+
+                    if mcp_client is None and mcp_server is None:
+                        return JsonRpcResponse.error_response(
+                            JsonRpcErrorCode.INTERNAL_ERROR,
+                            "Environment does not support MCP",
+                            request_id=request_id,
+                        )
+
+                    if not tool_name:
+                        return JsonRpcResponse.error_response(
+                            JsonRpcErrorCode.INVALID_PARAMS,
+                            "Missing 'name' in params",
+                            request_id=request_id,
+                        )
+
+                    if mcp_client:
+                        if managed_session_id and mcp_client.is_connected():
+                            # Session-managed with live transport.
+                            result = await mcp_client.call_tool(
+                                name=tool_name, arguments=arguments
+                            )
+                        elif callable(mcp_session_factory):
+                            # Stateless request, or session-managed but the
+                            # background transport was lost: (re-)open.
+                            mcp_session_cm = cast(
+                                AsyncContextManager[Any], mcp_session_factory()
+                            )
+                            async with mcp_session_cm:
+                                result = await mcp_client.call_tool(
+                                    name=tool_name, arguments=arguments
+                                )
+                        else:
+                            async with mcp_client:
+                                result = await mcp_client.call_tool(
+                                    name=tool_name, arguments=arguments
+                                )
+                    elif mcp_server:
+                        server_tools = get_server_tools(mcp_server)
+                        if tool_name in server_tools:
+                            tool = server_tools[tool_name]
+                            if inspect.iscoroutinefunction(tool.fn):
+                                result = await tool.fn(**arguments)
+                            else:
+                                result = tool.fn(**arguments)
+                        else:
+                            return JsonRpcResponse.error_response(
+                                JsonRpcErrorCode.INVALID_PARAMS,
+                                f"Tool not found: {tool_name}",
+                                request_id=request_id,
+                            )
+                    else:
+                        return JsonRpcResponse.error_response(
+                            JsonRpcErrorCode.INTERNAL_ERROR,
+                            "MCP server not available",
+                            request_id=request_id,
+                        )
+
+                    # Ensure result is JSON serializable
+                    serializable_result = _make_json_serializable(result)
+
+                    return JsonRpcResponse.success(
+                        result=serializable_result,
+                        request_id=request_id,
+                    )
+
+                else:
+                    return JsonRpcResponse.error_response(
+                        JsonRpcErrorCode.METHOD_NOT_FOUND,
+                        f"Method not found: {method}",
+                        request_id=request_id,
+                    )
+
+            except Exception as e:
+                return JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.INTERNAL_ERROR,
+                    str(e),
+                    request_id=request_id,
+                )
+            finally:
+                if managed_session_id:
+                    self._update_session_activity(
+                        managed_session_id,
+                        increment_step=(method == McpMethod.TOOLS_CALL),
+                    )
+                if should_close:
+                    _env.close()
+
+        # Register MCP WebSocket endpoint (available in both production and simulation modes)
+        @app.websocket("/mcp")
+        async def mcp_websocket_endpoint(websocket: WebSocket):
+            """
+            WebSocket endpoint for MCP JSON-RPC requests.
+
+            Each WebSocket connection gets its own environment instance for MCP operations.
+
+            Message Protocol:
+            - Client sends: JSON-RPC 2.0 request (tools/list, tools/call)
+            - Server responds: JSON-RPC 2.0 response (result or error)
+            """
+            await websocket.accept()
+
+            session_id = None
+            session_env = None
+
+            try:
+                # Create session with dedicated environment
+                session_id, session_env = await self._create_session()
+                if session_env is None:
+                    raise RuntimeError(
+                        "Session environment not initialized for MCP websocket"
+                    )
+
+                # If environment has an mcp_session context manager, hold it open
+                # for the lifetime of the websocket connection
+
+                async with AsyncExitStack() as stack:
+                    mcp_session_factory = getattr(session_env, "mcp_session", None)
+                    if callable(mcp_session_factory):
+                        mcp_session_cm = cast(
+                            AsyncContextManager[Any], mcp_session_factory()
+                        )
+                        await stack.enter_async_context(mcp_session_cm)
+
+                    while True:
+                        # Receive message from client
+                        raw_message = await websocket.receive_text()
+
+                        try:
+                            jsonrpc_dict = json.loads(raw_message)
+                            jsonrpc_request = JsonRpcRequest(**jsonrpc_dict)
+                        except json.JSONDecodeError as e:
+                            error_resp = JsonRpcResponse.error_response(
+                                JsonRpcErrorCode.PARSE_ERROR,
+                                f"Parse error: {e}",
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+                            continue
+                        except ValidationError as e:
+                            error_resp = JsonRpcResponse.error_response(
+                                JsonRpcErrorCode.INVALID_REQUEST,
+                                f"Invalid request: {e}",
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+                            continue
+
+                        try:
+                            # Call mcp_handler with session environment
+                            response = await mcp_handler(
+                                jsonrpc_request,
+                                session_env=session_env,
+                                session_id=session_id,
+                            )
+                            await websocket.send_text(response.model_dump_json())
+                        except Exception as e:
+                            error_resp = JsonRpcResponse.error_response(
+                                JsonRpcErrorCode.INTERNAL_ERROR,
+                                str(e),
+                                request_id=jsonrpc_request.id,
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+
+            except WebSocketDisconnect:
+                pass
+            except SessionCapacityError as e:
+                error_resp = JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.SERVER_ERROR,
+                    str(e),
+                    data={
+                        "active_sessions": e.active_sessions,
+                        "max_sessions": e.max_sessions,
+                    },
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            except EnvironmentFactoryError as e:
+                error_resp = JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.SERVER_ERROR,
+                    str(e),
+                    data={"factory_name": e.factory_name},
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            except Exception as e:
+                error_resp = JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.SERVER_ERROR,
+                    str(e),
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            finally:
+                if session_id:
+                    await self._destroy_session(session_id)
+                try:
+                    await websocket.close()
+                except RuntimeError:
+                    pass
+
+        # Register simulation control routes only in simulation mode
+        if mode == ServerMode.SIMULATION:
+
+            @app.post(
+                "/reset",
+                response_model=ResetResponse,
+                tags=["Environment Control"],
+                summary="Reset the environment",
+                description="""
+Reset the environment to its initial state and return the first observation.
+
+You can optionally provide a seed for reproducibility and an episode_id for tracking.
+                """,
+                responses={
+                    200: {
+                        "description": "Environment reset successfully",
+                        "content": {
+                            "application/json": {
+                                "example": {
+                                    "observation": {"status": "ready", "data": {}},
+                                    "reward": None,
+                                    "done": False,
+                                }
+                            }
+                        },
+                    }
+                },
+            )
+            async def reset(
+                request: ResetRequest = Body(default_factory=ResetRequest),
+            ) -> ResetResponse:
+                return await reset_handler(request)
+
+            @app.post(
+                "/step",
+                response_model=StepResponse,
+                tags=["Environment Control"],
+                summary="Execute an action in the environment",
+                description="""
+Execute an action in the environment and receive the resulting observation.
+
+The action must conform to the environment's action schema, which can be
+retrieved from the `/schema` endpoint. If the action is invalid,
+the endpoint will return HTTP 422 with detailed validation errors.
+
+The response includes:
+- **observation**: The environment's response to the action
+- **reward**: Optional reward signal (float or None)
+- **done**: Boolean indicating if the episode has terminated
+                """,
+                responses={
+                    200: {
+                        "description": "Action executed successfully",
+                        "content": {
+                            "application/json": {
+                                "example": {
+                                    "observation": {"status": "success", "data": {}},
+                                    "reward": 1.0,
+                                    "done": False,
+                                }
+                            }
+                        },
+                    },
+                    422: {
+                        "description": "Validation error - invalid action format or values",
+                        "content": {
+                            "application/json": {
+                                "example": {
+                                    "detail": [
+                                        {
+                                            "type": "string_too_short",
+                                            "loc": ["body", "action", "message"],
+                                            "msg": "String should have at least 1 character",
+                                            "input": "",
+                                        }
+                                    ]
+                                }
+                            }
+                        },
+                    },
+                    500: {
+                        "description": "Internal server error during action execution"
+                    },
+                },
+            )
+            async def step(request: StepRequest) -> StepResponse:
+                return await step_handler(request)
+
+        def get_state_handler() -> State:
+            _env = self._env_factory()
+            try:
+                return _env.state
+            finally:
+                _env.close()
+
+        def get_metadata_handler() -> EnvironmentMetadata:
+            _env = self._env_factory()
+            try:
+                return _env.get_metadata()
+            finally:
+                _env.close()
+
+        # Build list of GET endpoints based on mode
+        get_endpoints = [
+            GetEndpointConfig(
+                path="/metadata",
+                handler=get_metadata_handler,
+                response_model=EnvironmentMetadata,
+                tag="Environment Info",
+                summary="Get environment metadata",
+                description="""
+Get metadata about this environment.
+
+Returns information about the environment including name, description,
+version, author, and documentation links.
+                """,
+            ),
+            GetEndpointConfig(
+                path="/health",
+                handler=lambda: HealthResponse(status=HealthStatus.HEALTHY),
+                response_model=HealthResponse,
+                tag="Health",
+                summary="Health check",
+                description="Check if the environment server is running and healthy.",
+            ),
+        ]
+
+        # Only register /state endpoint in simulation mode
+        if mode == ServerMode.SIMULATION:
+            get_endpoints.insert(
+                0,
+                GetEndpointConfig(
+                    path="/state",
+                    handler=get_state_handler,
+                    response_model=State,
+                    tag="State Management",
+                    summary="Get current environment state",
+                    description="""
+Retrieve the current internal state of the environment.
+
+The structure of the state object is defined by the environment's State model.
+                    """,
+                ),
+            )
+
+        register_get_endpoints(app, get_endpoints)
+
+        # Register combined schema endpoint
+        @app.get(
+            "/schema",
+            response_model=SchemaResponse,
+            tags=["Schema"],
+            summary="Get all JSON schemas",
+            description="""
+Get JSON schemas for actions, observations, and state in a single response.
+
+Returns a combined schema object containing:
+- **action**: JSON schema for actions accepted by this environment
+- **observation**: JSON schema for observations returned by this environment
+- **state**: JSON schema for environment state objects
+
+This is more efficient than calling individual schema endpoints and provides
+all schema information needed to interact with the environment.
+            """,
+            responses={
+                200: {
+                    "description": "Combined schemas retrieved successfully",
+                    "content": {
+                        "application/json": {
+                            "example": {
+                                "action": {
+                                    "type": "object",
+                                    "properties": {"message": {"type": "string"}},
+                                },
+                                "observation": {
+                                    "type": "object",
+                                    "properties": {"response": {"type": "string"}},
+                                },
+                                "state": {
+                                    "type": "object",
+                                    "properties": {"step_count": {"type": "integer"}},
+                                },
+                            }
+                        }
+                    },
+                }
+            },
+        )
+        async def get_schemas() -> SchemaResponse:
+            """Return all schemas in one response."""
+            return SchemaResponse(
+                action=self.action_cls.model_json_schema(),
+                observation=self.observation_cls.model_json_schema(),
+                state=State.model_json_schema(),
+            )
+
+        # Register MCP endpoint for production mode (direct MCP access)
+        @app.post("/mcp")
+        async def mcp_endpoint(request_raw: Request) -> Dict[str, Any]:
+            """
+            MCP JSON-RPC endpoint for production mode.
+
+            Bypasses step() overhead and provides direct access to MCP tools.
+            Supports tools/list and tools/call methods.
+            """
+            # Parse JSON manually to handle parse errors gracefully
+            try:
+                body = await request_raw.body()
+                request_dict = json.loads(body)
+                request = JsonRpcRequest(**request_dict)
+            except json.JSONDecodeError:
+                return JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.PARSE_ERROR
+                ).model_dump()
+            except ValidationError as e:
+                return JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.INVALID_REQUEST,
+                    f"Invalid request: {e}",
+                ).model_dump()
+            except Exception:
+                return JsonRpcResponse.error_response(
+                    JsonRpcErrorCode.PARSE_ERROR
+                ).model_dump()
+
+            response = await mcp_handler(request)
+            return response.model_dump()
+
+        # Register WebSocket endpoint for persistent sessions
+        @app.websocket("/ws")
+        async def websocket_endpoint(websocket: WebSocket):
+            """
+            WebSocket endpoint for persistent environment sessions.
+
+            Each WebSocket connection gets its own environment instance.
+
+            Message Protocol:
+            - Client sends: WSResetMessage | WSStepMessage | WSStateMessage | WSCloseMessage
+            - Server responds: WSObservationResponse | WSStateResponse | WSErrorResponse
+            """
+            await websocket.accept()
+
+            session_id = None
+            session_env = None
+
+            try:
+                # Create session with dedicated environment
+                session_id, session_env = await self._create_session()
+                if session_env is None:
+                    raise RuntimeError(
+                        "Session environment not initialized for websocket"
+                    )
+
+                # Keep MCP session open for entire websocket lifetime
+                # (avoids reconnect overhead on every message)
+
+                async with AsyncExitStack() as stack:
+                    mcp_session_factory = getattr(session_env, "mcp_session", None)
+                    if callable(mcp_session_factory):
+                        mcp_session_cm = cast(
+                            AsyncContextManager[Any], mcp_session_factory()
+                        )
+                        await stack.enter_async_context(mcp_session_cm)
+
+                    while True:
+                        # Receive message from client
+                        raw_message = await websocket.receive_text()
+
+                        try:
+                            message_dict = json.loads(raw_message)
+                        except json.JSONDecodeError as e:
+                            error_resp = WSErrorResponse(
+                                data={
+                                    "message": f"Invalid JSON: {e}",
+                                    "code": WSErrorCode.INVALID_JSON,
+                                }
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+                            continue
+
+                        msg_type = message_dict.get("type", "")
+
+                        try:
+                            match msg_type:
+                                case "reset":
+                                    msg = WSResetMessage(**message_dict)
+
+                                    is_async = (
+                                        session_env.reset_async.__func__
+                                        is not Environment.reset_async
+                                    )
+
+                                    if is_async:
+                                        sig = inspect.signature(session_env.reset_async)
+                                        valid_kwargs = self._get_valid_kwargs(
+                                            sig, msg.data
+                                        )
+                                        observation = await session_env.reset_async(
+                                            **valid_kwargs
+                                        )
+                                    else:
+                                        sig = inspect.signature(session_env.reset)
+                                        valid_kwargs = self._get_valid_kwargs(
+                                            sig, msg.data
+                                        )
+                                        observation = (
+                                            await self._run_in_session_executor(
+                                                session_id,
+                                                session_env.reset,
+                                                **valid_kwargs,
+                                            )
+                                        )
+
+                                    self._update_session_activity(session_id)
+
+                                    response = WSObservationResponse(
+                                        data=serialize_observation(observation),
+                                    )
+
+                                case "step":
+                                    msg = WSStepMessage(**message_dict)
+                                    action = deserialize_action(
+                                        msg.data, self.action_cls
+                                    )
+
+                                    is_async = (
+                                        session_env.step_async.__func__
+                                        is not Environment.step_async
+                                    )
+
+                                    if is_async:
+                                        observation = await session_env.step_async(
+                                            action
+                                        )
+                                    else:
+                                        observation = (
+                                            await self._run_in_session_executor(
+                                                session_id, session_env.step, action
+                                            )
+                                        )
+
+                                    self._update_session_activity(
+                                        session_id, increment_step=True
+                                    )
+
+                                    response = WSObservationResponse(
+                                        data=serialize_observation(observation)
+                                    )
+
+                                case "state":
+                                    msg = WSStateMessage(**message_dict)
+                                    state = session_env.state
+                                    if hasattr(state, "model_dump"):
+                                        state_data = state.model_dump()
+                                    else:
+                                        state_data = dict(state) if state else {}
+
+                                    response = WSStateResponse(data=state_data)
+
+                                case "close":
+                                    msg = WSCloseMessage(**message_dict)
+                                    break
+
+                                case "mcp":
+                                    msg = WSMCPMessage(**message_dict)
+                                    try:
+                                        rpc_request = JsonRpcRequest(**msg.data)
+                                    except (ValidationError, Exception) as e:
+                                        rpc_response = JsonRpcResponse.error_response(
+                                            JsonRpcErrorCode.INVALID_REQUEST,
+                                            f"Invalid request: {e}",
+                                        )
+                                    else:
+                                        rpc_response = await mcp_handler(
+                                            rpc_request,
+                                            session_env=session_env,
+                                            session_id=session_id,
+                                        )
+                                    response = WSMCPResponse(
+                                        data=rpc_response.model_dump()
+                                    )
+
+                                case _:
+                                    response = WSErrorResponse(
+                                        data={
+                                            "message": f"Unknown message type: {msg_type}",
+                                            "code": WSErrorCode.UNKNOWN_TYPE,
+                                        }
+                                    )
+
+                            await websocket.send_text(response.model_dump_json())
+
+                        except ValidationError as e:
+                            error_resp = WSErrorResponse(
+                                data={
+                                    "message": "Invalid message",
+                                    "code": WSErrorCode.VALIDATION_ERROR,
+                                    "errors": e.errors(),
+                                }
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+                        except Exception as e:
+                            error_resp = WSErrorResponse(
+                                data={
+                                    "message": str(e),
+                                    "code": WSErrorCode.EXECUTION_ERROR,
+                                }
+                            )
+                            await websocket.send_text(error_resp.model_dump_json())
+
+            except WebSocketDisconnect:
+                pass
+            except SessionCapacityError as e:
+                error_resp = WSErrorResponse(
+                    data={
+                        "message": str(e),
+                        "code": WSErrorCode.CAPACITY_REACHED,
+                        "active_sessions": e.active_sessions,
+                        "max_sessions": e.max_sessions,
+                    }
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            except EnvironmentFactoryError as e:
+                error_resp = WSErrorResponse(
+                    data={
+                        "message": str(e),
+                        "code": WSErrorCode.FACTORY_ERROR,
+                        "factory_name": e.factory_name,
+                    }
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            except Exception as e:
+                error_resp = WSErrorResponse(
+                    data={"message": str(e), "code": WSErrorCode.SESSION_ERROR}
+                )
+                await websocket.send_text(error_resp.model_dump_json())
+            finally:
+                if session_id:
+                    await self._destroy_session(session_id)
+                try:
+                    await websocket.close()
+                except RuntimeError:
+                    pass
+
+
+def create_app(
+    env: Callable[[], Environment],
+    action_cls: Type[Action],
+    observation_cls: Type[Observation],
+    env_name: Optional[str] = None,
+    max_concurrent_envs: Optional[int] = None,
+    concurrency_config: Optional[ConcurrencyConfig] = None,
+    gradio_builder: Optional[Callable[..., Any]] = None,
+) -> FastAPI:
+    """
+    Create a FastAPI application with or without web interface.
+
+    This function creates a FastAPI app with the web interface enabled by default,
+    including README integration for better user experience.
+
+    Args:
+        env: Environment factory (callable) that creates new instances
+        action_cls: The Action subclass this environment expects
+        observation_cls: The Observation subclass this environment returns
+        env_name: Optional environment name for README loading
+        max_concurrent_envs: Maximum concurrent WebSocket sessions.
+                             Mutually exclusive with concurrency_config.
+        concurrency_config: Optional ConcurrencyConfig for advanced concurrency settings.
+                            Mutually exclusive with max_concurrent_envs.
+        gradio_builder: Optional callable to build a custom Gradio UI at /web.
+            Signature: (web_manager, action_fields, metadata, is_chat_env, title,
+            quick_start_md) -> gr.Blocks. When None, the default Gradio app is used.
+            See docs/customizing-web-ui.md.
+
+    Returns:
+        FastAPI application instance with or without web interface and README integration
+    """
+    # Check if web interface should be enabled
+    # This can be controlled via environment variable or build argument
+    enable_web = os.getenv("ENABLE_WEB_INTERFACE", "false").lower() in (
+        "true",
+        "1",
+        "yes",
+    )
+
+    if enable_web:
+        # Gradio-based web UI (gradio is a core dependency)
+        from .web_interface import create_web_interface_app
+
+        return create_web_interface_app(
+            cast(Any, env),
+            action_cls,
+            observation_cls,
+            env_name,
+            max_concurrent_envs,
+            concurrency_config,
+            gradio_builder=gradio_builder,
+        )
+    else:
+        # Use standard FastAPI app without web interface
+        return create_fastapi_app(
+            env, action_cls, observation_cls, max_concurrent_envs, concurrency_config
+        )
+
+
+def create_fastapi_app(
+    env: Callable[[], Environment],
+    action_cls: Type[Action],
+    observation_cls: Type[Observation],
+    max_concurrent_envs: Optional[int] = None,
+    concurrency_config: Optional[ConcurrencyConfig] = None,
+) -> FastAPI:
+    """
+    Create a FastAPI application with comprehensive documentation.
+
+    Args:
+        env: Environment factory (callable) that creates new instances
+        action_cls: The Action subclass this environment expects
+        observation_cls: The Observation subclass this environment returns
+        max_concurrent_envs: Maximum concurrent WebSocket sessions.
+                             Mutually exclusive with concurrency_config.
+        concurrency_config: Optional ConcurrencyConfig for advanced concurrency settings.
+                            Mutually exclusive with max_concurrent_envs.
+
+    Returns:
+        FastAPI application instance
+    """
+    try:
+        from fastapi import FastAPI
+    except ImportError:
+        raise ImportError(
+            "FastAPI is required. Install with: pip install fastapi uvicorn"
+        )
+
+    app = FastAPI(
+        title="OpenEnv Environment HTTP API",
+        version="1.0.0",
+        description="""
+# OpenEnv Environment HTTP API
+
+HTTP API for interacting with OpenEnv environments through a standardized interface.
+
+## Features
+
+* **Environment Reset**: Initialize or restart episodes
+* **Action Execution**: Send actions and receive observations
+* **State Inspection**: Query current environment state
+* **Schema Access**: Retrieve JSON schemas for actions and observations
+
+## Workflow
+
+1. Call `/reset` to start a new episode and get initial observation
+2. Call `/step` repeatedly with actions to interact with environment
+3. Episode ends when observation returns `done: true`
+4. Call `/state` anytime to inspect current environment state
+
+## Documentation
+
+* **Swagger UI**: Available at `/docs`
+* **ReDoc**: Available at `/redoc`
+* **OpenAPI Schema**: Available at `/openapi.json`
+        """,
+        openapi_tags=[
+            {
+                "name": "Environment Control",
+                "description": "Core operations for environment interaction (reset, step)",
+            },
+            {
+                "name": "State Management",
+                "description": "Operations for inspecting environment state",
+            },
+            {
+                "name": "Environment Info",
+                "description": "Information about the environment",
+            },
+            {
+                "name": "Schema",
+                "description": "JSON Schema endpoints for actions, observations, and state",
+            },
+            {"name": "Health", "description": "Service health and status checks"},
+        ],
+        docs_url="/docs",
+        redoc_url="/redoc",
+        openapi_url="/openapi.json",
+        contact={
+            "name": "OpenEnv Team",
+            "url": "https://github.com/meta-pytorch/OpenEnv",
+        },
+        license_info={
+            "name": "BSD-3-Clause",
+            "url": "https://github.com/meta-pytorch/OpenEnv/blob/main/LICENSE",
+        },
+    )
+
+    server = HTTPEnvServer(
+        env,
+        action_cls,
+        observation_cls,
+        max_concurrent_envs,
+        concurrency_config=concurrency_config,
+    )
+    server.register_routes(app)
+    return app

src/core/env_server/interfaces.py

@@ -0,0 +1,297 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import inspect
+from abc import ABC, abstractmethod
+from typing import Any, Generic, Optional, Protocol, TYPE_CHECKING, TypeVar
+
+from typing_extensions import TypedDict
+
+from .types import Action, EnvironmentMetadata, Observation, State
+
+if TYPE_CHECKING:
+    from openenv.core.rubrics import Rubric
+
+ActT = TypeVar("ActT", bound=Action)
+ObsT = TypeVar("ObsT", bound=Observation)
+StateT = TypeVar("StateT", bound=State)
+
+
+class Message(TypedDict):
+    """A message in a conversation.
+
+    Compatible with Huggingface chat template format.
+    """
+
+    role: str
+    content: str
+
+
+class ModelTokenizer(Protocol):
+    """Protocol for tokenizers that support chat templates.
+
+    This protocol defines the interface that tokenizers must implement
+    to work with chat-based environments. It's compatible with
+    Huggingface transformers tokenizers.
+    """
+
+    def apply_chat_template(
+        self,
+        conversation: list[Message],
+        tokenize: bool = True,
+        return_tensors: str | None = None,
+        **kwargs: Any,
+    ) -> Any:
+        """Apply a chat template to format and optionally tokenize a conversation.
+
+        Args:
+            conversation: List of message dictionaries with 'role' and 'content'
+            tokenize: Whether to tokenize the output
+            return_tensors: Format for returned tensors ('pt' for PyTorch)
+            **kwargs: Additional arguments
+
+        Returns:
+            Formatted and optionally tokenized conversation
+        """
+        ...
+
+    def decode(
+        self, token_ids: Any, skip_special_tokens: bool = False, **kwargs: Any
+    ) -> str:
+        """Decode token IDs back to text.
+
+        Args:
+            token_ids: Token IDs to decode
+            skip_special_tokens: Whether to skip special tokens in output
+            **kwargs: Additional arguments
+
+        Returns:
+            Decoded text string
+        """
+        ...
+
+
+class Transform(ABC, Generic[ObsT]):
+    """Transform observations to add rewards, metrics, or other modifications.
+
+    Transforms follow the TorchRL pattern where they take an observation
+    and return a (potentially modified) observation. This allows for
+    flexible reward computation and observation augmentation.
+    """
+
+    @abstractmethod
+    def __call__(self, observation: ObsT) -> ObsT:
+        """Transform an observation.
+
+        Args:
+            observation: The input observation
+
+        Returns:
+            The transformed observation
+        """
+        pass
+
+
+class Environment(ABC, Generic[ActT, ObsT, StateT]):
+    """Base class for all environment servers following Gym/Gymnasium API.
+
+    Args:
+        transform: Optional transform to apply to observations
+        rubric: Optional rubric for reward computation. When provided, the
+            rubric's output can be used to set the observation's reward in step().
+
+    Class Attributes:
+        SUPPORTS_CONCURRENT_SESSIONS: Whether this environment supports concurrent sessions.
+            When True, multiple WebSocket connections can each have their own
+            environment instance (up to max_concurrent_envs). When False (default),
+            the environment should only be used with a single session at a time.
+
+            Set this to True in your Environment subclass if:
+            - The environment uses proper session isolation (e.g., unique working dirs)
+            - No shared mutable state exists between instances
+            - External resources (databases, APIs) can handle concurrent access
+
+    Attributes:
+        rubric: Optional rubric for computing rewards. Environments can set this
+            in __init__ and use it in step() to compute observation rewards.
+            Training infrastructure can access it for introspection:
+                for name, r in env.rubric.named_rubrics():
+                    print(f"{name}: {r.last_score}")
+
+    See RFC 004 for rubric design: rfcs/004-rubrics.md
+    """
+
+    # Class-level flag indicating whether this environment supports concurrent sessions
+    SUPPORTS_CONCURRENT_SESSIONS: bool = False
+
+    # Optional rubric for reward computation
+    rubric: Optional["Rubric"]
+
+    def __init__(
+        self,
+        transform: Optional[Transform[ObsT]] = None,
+        rubric: Optional["Rubric"] = None,
+    ):
+        self.transform = transform
+        self.rubric = rubric
+
+    @abstractmethod
+    def reset(
+        self,
+        seed: Optional[int] = None,
+        episode_id: Optional[str] = None,
+        **kwargs: Any,
+    ) -> ObsT:
+        """Reset the environment and return initial observation."""
+        pass
+
+    async def reset_async(
+        self,
+        seed: Optional[int] = None,
+        episode_id: Optional[str] = None,
+        **kwargs: Any,
+    ) -> ObsT:
+        """Async version of reset. Default implementation calls sync reset.
+
+        Override to provide true async implementation.
+        """
+        return self.reset(seed=seed, episode_id=episode_id, **kwargs)
+
+    @abstractmethod
+    def step(
+        self,
+        action: ActT,
+        timeout_s: Optional[float] = None,
+        **kwargs: Any,
+    ) -> ObsT:
+        """Take a step in the environment."""
+        pass
+
+    async def step_async(
+        self,
+        action: ActT,
+        timeout_s: Optional[float] = None,
+        **kwargs: Any,
+    ) -> ObsT:
+        """Async version of step. Default implementation calls sync step.
+
+        Override to provide true async implementation.
+        """
+        return self.step(action, timeout_s=timeout_s, **kwargs)
+
+    @property
+    @abstractmethod
+    def state(self) -> StateT:
+        """Get the current environment state."""
+        pass
+
+    def get_metadata(self) -> EnvironmentMetadata:
+        """
+        Get metadata about this environment.
+
+        Override this method to provide custom metadata for the environment.
+        Default implementation returns basic metadata derived from class name.
+
+        Returns:
+            EnvironmentMetadata with environment information
+        """
+        return EnvironmentMetadata(
+            name=self.__class__.__name__,
+            description=f"{self.__class__.__name__} environment",
+            version="1.0.0",
+        )
+
+    def _apply_transform(self, observation: ObsT) -> ObsT:
+        """Apply transform if one is provided."""
+        if self.transform is not None:
+            return self.transform(observation)
+        return observation
+
+    def _apply_rubric(self, action: ActT, observation: ObsT) -> float:
+        """Apply rubric if one is provided.
+
+        Args:
+            action: The action taken by the agent.
+            observation: The resulting observation.
+
+        Returns:
+            Reward value from the rubric, or 0.0 if no rubric is set.
+
+        Usage in step():
+            def step(self, action: MyAction, ...) -> MyObservation:
+                # ... execute action and create observation ...
+                observation.reward = self._apply_rubric(action, observation)
+                return observation
+        """
+        if self.rubric is not None:
+            return self.rubric(action, observation)
+        return 0.0
+
+    async def _apply_rubric_async(self, action: ActT, observation: ObsT) -> float:
+        """Apply rubric asynchronously if one is provided.
+
+        Args:
+            action: The action taken by the agent.
+            observation: The resulting observation.
+
+        Returns:
+            Reward value from the rubric, or 0.0 if no rubric is set.
+
+        Usage in step_async():
+            async def step_async(self, action: MyAction, ...) -> MyObservation:
+                # ... execute action and create observation ...
+                observation.reward = await self._apply_rubric_async(action, observation)
+                return observation
+        """
+        if self.rubric is not None:
+            result = self.rubric(action, observation)
+            # If rubric returns a coroutine, await it
+            if inspect.iscoroutine(result):
+                return await result
+            return result
+        return 0.0
+
+    def _reset_rubric(self) -> None:
+        """Reset the rubric state if one is provided.
+
+        Call this in reset() to clear any trajectory state in the rubric.
+
+        Usage in reset():
+            def reset(self, ...) -> MyObservation:
+                self._reset_rubric()
+                # ... create initial observation ...
+                return observation
+        """
+        if self.rubric is not None:
+            self.rubric.reset()
+
+    async def _reset_rubric_async(self) -> None:
+        """Reset the rubric state asynchronously if one is provided.
+
+        Call this in reset_async() to clear any trajectory state in the rubric.
+
+        Usage in reset_async():
+            async def reset_async(self, ...) -> MyObservation:
+                await self._reset_rubric_async()
+                # ... create initial observation ...
+                return observation
+        """
+        if self.rubric is not None:
+            # Check if rubric has async reset method
+            if hasattr(self.rubric, "reset_async"):
+                result = self.rubric.reset_async()
+                if inspect.iscoroutine(result):
+                    await result
+            else:
+                self.rubric.reset()
+
+    def close(self) -> None:
+        """Clean up resources used by the environment.
+
+        Override this method to implement custom cleanup logic.
+        Called when the environment is being destroyed or reset.
+        """
+        pass

src/core/env_server/mcp_environment.py

@@ -0,0 +1,645 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+MCP Environment base class for OpenEnv.
+
+This module provides the MCPEnvironment base class that integrates FastMCP servers
+with OpenEnv's Gym-style Environment interface. It handles MCP tool discovery
+and invocation through the step() API, following RFC 003.
+
+Key features:
+- Automatic routing of ListToolsAction and CallToolAction to MCP server
+- Reserved tool name validation (reset, step, state, close are protected)
+- Timeout handling for tool calls
+- Proper error categorization (tool not found, execution errors, timeouts)
+- Mode-aware tool registration (production vs simulation)
+- Code mode support via get_callables() and execute_code()
+
+Usage:
+    from fastmcp import FastMCP
+    from openenv.core.env_server.mcp_environment import MCPEnvironment
+
+    class MyMCPEnv(MCPEnvironment):
+        def __init__(self):
+            mcp = FastMCP("my-server")
+
+            # Register mode-specific tools
+            @self.tool(mode="production")
+            def my_tool(arg: str) -> str:
+                return f"Production: {arg}"
+
+            @self.tool(mode="simulation")
+            def my_tool(arg: str) -> str:
+                return f"Simulation: {arg}"
+
+            super().__init__(mcp)
+
+        def reset(self, seed=None, episode_id=None, **kwargs):
+            # Reset logic here
+            ...
+
+        def _step_impl(self, action):
+            # Handle non-MCP actions
+            ...
+
+        @property
+        def state(self):
+            # Return current state
+            ...
+"""
+
+import asyncio
+import inspect
+from abc import abstractmethod
+from collections import defaultdict
+from contextlib import asynccontextmanager
+from typing import Any, Callable, Dict, Optional
+
+from fastmcp import Client
+from fastmcp.client.client import CallToolResult
+from mcp.types import TextContent
+
+from ..utils import run_async_safely
+from .interfaces import Environment
+from .mcp_types import (
+    CallToolAction,
+    CallToolObservation,
+    ListToolsAction,
+    ListToolsObservation,
+    RESERVED_TOOL_NAMES,
+    Tool,
+    ToolError,
+    ToolErrorType,
+)
+from .types import Action, Observation
+
+
+# Default timeout for MCP tool calls in seconds
+MCP_TOOL_CALL_TIMEOUT = 30.0
+
+# Valid modes for tool registration
+VALID_MODES = {"production", "simulation"}
+
+
+def get_server_tools(mcp_server: Any) -> Dict[str, Any]:
+    """
+    Get tools from a FastMCP server, compatible with both 2.x and 3.x.
+
+    Returns:
+        Dictionary mapping tool names to tool objects.
+    """
+    # FastMCP 2.x: get_tools() returns dict {name: Tool}
+    if hasattr(mcp_server, "get_tools"):
+        result = run_async_safely(mcp_server.get_tools())
+        if isinstance(result, dict):
+            return result
+    # FastMCP 3.x: list_tools() returns list of Tool objects
+    if hasattr(mcp_server, "list_tools"):
+        tools_list = run_async_safely(mcp_server.list_tools())
+        return {t.name: t for t in tools_list}
+    return {}
+
+
+class MCPEnvironment(Environment):
+    """
+    Base class for environments that expose tools via MCP (Model Context Protocol).
+
+    MCPEnvironment bridges FastMCP servers with OpenEnv's Gym-style API, allowing
+    agents to discover and invoke MCP tools through the standard step() interface.
+
+    The class automatically handles:
+    - ListToolsAction: Returns available tools from the MCP server
+    - CallToolAction: Invokes a specific tool with arguments
+
+    All other actions are delegated to the abstract _step_impl() method,
+    which subclasses must implement.
+
+    Args:
+        mcp_server: A FastMCP server instance containing tool definitions.
+            The server's tools will be validated against reserved names.
+        transform: Optional transform to apply to observations (inherited from Environment).
+
+    Raises:
+        ValueError: If any tool in the MCP server uses a reserved name
+            (reset, step, state, close).
+
+    Example:
+        >>> from fastmcp import FastMCP
+        >>> mcp = FastMCP("calculator")
+        >>> @mcp.tool()
+        ... def add(a: int, b: int) -> int:
+        ...     return a + b
+        >>> env = MyMCPEnvironment(mcp)
+        >>> obs = env.step(ListToolsAction())
+        >>> obs.tools[0].name
+        'add'
+    """
+
+    def __init__(self, mcp_server: Any, transform: Optional[Any] = None) -> None:
+        """
+        Initialize the MCP environment.
+
+        Args:
+            mcp_server: A FastMCP server instance with tool definitions.
+            transform: Optional transform to apply to observations.
+
+        Raises:
+            ValueError: If any tool uses a reserved name (reset, step, state, close).
+        """
+        super().__init__(transform=transform)
+
+        # Validate tool names before storing
+        self._validate_tool_names(mcp_server)
+
+        self.mcp_server = mcp_server
+        self.mcp_client = Client(mcp_server)
+
+        # Track mode-specific tools: {tool_name: {mode: func}}
+        # mode can be "production", "simulation", or None (available in all modes)
+        self._mode_tools = defaultdict(dict)
+
+        # Track tool schemas for list_tools: {tool_name: {mode: schema}}
+        self._mode_tool_schemas = defaultdict(dict)
+
+    def _require_mcp_client(self) -> Any:
+        """Return MCP client or raise if environment has been closed."""
+        if self.mcp_client is None:
+            raise RuntimeError("MCP client is not available; environment is closed")
+        return self.mcp_client
+
+    def _require_mcp_server(self) -> Any:
+        """Return MCP server or raise if environment has been closed."""
+        if self.mcp_server is None:
+            raise RuntimeError("MCP server is not available; environment is closed")
+        return self.mcp_server
+
+    @asynccontextmanager
+    async def mcp_session(self):
+        """
+        Context manager for MCP client sessions.
+
+        This wrapper serves two purposes:
+
+        1. **Null guard** — raises a clear error if ``close()`` has already
+           been called (``mcp_client`` is ``None``).
+
+        2. **AsyncExitStack adapter** — FastMCP's ``Client.__aenter__``
+           creates a background ``asyncio.Task`` for session management.
+           When entered directly via ``AsyncExitStack`` in the HTTP session
+           path (``_create_session``), this task can be cancelled by ASGI
+           harnesses (e.g. Starlette ``TestClient``) between requests,
+           corrupting session state.  Wrapping in an ``asynccontextmanager``
+           generator isolates the task lifecycle: the generator frame keeps
+           ``async with client:`` suspended at ``yield``, so cleanup only
+           runs when the stack explicitly closes the generator — not when
+           the event loop cancels orphaned tasks.
+
+        Delegates to FastMCP's ``Client`` context manager which is
+        reentrant: the first entry opens the transport and subsequent
+        (nested) entries simply increment an internal reference counter.
+        The transport is closed only when the outermost context exits.
+
+        No external lock is needed because ``Client._connect`` /
+        ``Client._disconnect`` already serialise connection state changes
+        through their own ``anyio.Lock``.
+        """
+        client = self._require_mcp_client()
+        async with client:
+            yield client
+
+    @property
+    def supports_code_mode(self) -> bool:
+        """Check if this environment supports code mode (execute_code)."""
+        return True
+
+    def _get_server_tools(self, mcp_server: Any) -> Dict[str, Any]:
+        """
+        Get tools from a FastMCP server, compatible with both 2.x and 3.x.
+
+        Returns:
+            Dictionary mapping tool names to tool objects.
+        """
+        return get_server_tools(mcp_server)
+
+    def get_callables(self) -> Dict[str, Callable]:
+        """
+        Get callable functions for code mode.
+
+        Returns tool functions as direct Python callables, enabling code mode
+        where agents write Python code that calls tools directly (no JSON-RPC
+        overhead). Mode-specific tools are filtered by the current mode.
+
+        Returns:
+            Dictionary mapping tool names to callables.
+        """
+        callables: Dict[str, Callable] = {}
+        current_mode = getattr(self, "_mode", None)
+
+        # Extract callables from FastMCP server using public API
+        for tool_name, tool in self._get_server_tools(self.mcp_server).items():
+            if hasattr(tool, "fn") and callable(tool.fn):
+                callables[tool_name] = tool.fn
+
+        # Add mode-specific tools available in current mode
+        for tool_name, mode_funcs in self._mode_tools.items():
+            if None in mode_funcs:
+                # Tool available in all modes (already in FastMCP if registered there)
+                if tool_name not in callables:
+                    callables[tool_name] = mode_funcs[None]
+            elif current_mode in mode_funcs:
+                # Tool available in current mode only
+                callables[tool_name] = mode_funcs[current_mode]
+
+        return callables
+
+    def execute_code(self, code: str) -> Observation:
+        """
+        Execute Python code with tools available as callables.
+
+        This enables the CodeAct pattern where agents write Python code
+        that calls tools directly as functions, avoiding JSON-RPC overhead.
+
+        Args:
+            code: Python code to execute. Tools are available as functions
+                in the execution namespace. Set a variable named 'result'
+                to capture the return value.
+
+        Returns:
+            Observation with result in metadata["result"] or error in
+            metadata["error"].
+        """
+        namespace = self.get_callables()
+
+        result_dict: Dict[str, Any] = {}
+        try:
+            exec(code, namespace, result_dict)
+            result = result_dict.get("result")
+            return Observation(done=False, reward=0.0, metadata={"result": result})
+        except SyntaxError as e:
+            return Observation(
+                done=False, reward=0.0, metadata={"error": f"Syntax error: {str(e)}"}
+            )
+        except Exception as e:
+            return Observation(done=False, reward=0.0, metadata={"error": str(e)})
+
+    def _validate_tool_names(self, mcp_server: Any) -> None:
+        """
+        Validate that no tools use reserved names.
+
+        Reserved names (reset, step, state, close) are protected to maintain
+        the dual API boundary between infrastructure and agent APIs.
+
+        Args:
+            mcp_server: The FastMCP server to validate.
+
+        Raises:
+            ValueError: If any tool uses a reserved name.
+        """
+        tools_dict = self._get_server_tools(mcp_server)
+        if tools_dict:
+            tool_names = set(tools_dict.keys())
+            conflicts = tool_names & RESERVED_TOOL_NAMES
+            if conflicts:
+                raise ValueError(
+                    f"MCP tools cannot use reserved names: {sorted(conflicts)}. "
+                    f"Reserved names are: {sorted(RESERVED_TOOL_NAMES)}"
+                )
+
+    def tool(self, mode: Optional[str] = None) -> Callable:
+        """
+        Decorator for registering mode-aware tools.
+
+        Args:
+            mode: Optional mode for the tool ("production" or "simulation").
+                If None, tool is available in all modes.
+
+        Returns:
+            A decorator function for registering tools.
+
+        Raises:
+            ValueError: If mode is not None, "production", or "simulation".
+        """
+        if mode is not None and mode not in VALID_MODES:
+            raise ValueError(
+                f"Invalid mode '{mode}'. Mode must be 'production', 'simulation', or None."
+            )
+
+        def decorator(func: Callable) -> Callable:
+            tool_name = func.__name__
+            # Validate tool name is not reserved
+            if tool_name in RESERVED_TOOL_NAMES:
+                raise ValueError(
+                    f"Tool name '{tool_name}' is reserved and cannot be used. "
+                    f"Reserved names are: {sorted(RESERVED_TOOL_NAMES)}"
+                )
+
+            # If mode is None, register with FastMCP as usual
+            if mode is None:
+                mcp_server = self._require_mcp_server()
+                decorated_func = mcp_server.tool()(func)
+                self._mode_tools[tool_name][None] = func
+                return decorated_func
+
+            # For mode-specific tools, don't register with FastMCP
+            # Instead, track them ourselves
+            self._mode_tools[tool_name][mode] = func
+
+            # Extract schema information from function signature
+            sig = inspect.signature(func)
+            schema = {
+                "type": "object",
+                "properties": {},
+                "required": [],
+            }
+
+            for param_name, param in sig.parameters.items():
+                # Get type annotation
+                param_type = param.annotation
+                json_type = "string"  # default
+                if param_type in (int, "int"):
+                    json_type = "integer"
+                elif param_type in (float, "float"):
+                    json_type = "number"
+                elif param_type in (bool, "bool"):
+                    json_type = "boolean"
+
+                schema["properties"][param_name] = {"type": json_type}
+
+                # If no default value, it's required
+                if param.default == inspect.Parameter.empty:
+                    schema["required"].append(param_name)
+
+            # Store the schema for this mode-specific tool
+            self._mode_tool_schemas[tool_name][mode] = {
+                "name": tool_name,
+                "description": func.__doc__ or "",
+                "input_schema": schema,
+            }
+
+            return func
+
+        return decorator
+
+    def step(
+        self,
+        action: Action,
+        timeout_s: Optional[float] = None,
+        **kwargs: Any,
+    ) -> Observation:
+        """
+        Execute an action in the environment.
+
+        This method routes MCP-specific actions (ListToolsAction, CallToolAction)
+        to the appropriate handlers, while delegating all other actions to
+        the subclass's _step_impl() method.
+
+        Args:
+            action: The action to execute. Can be:
+                - ListToolsAction: Returns available MCP tools
+                - CallToolAction: Invokes a specific MCP tool
+                - Any other Action: Delegated to _step_impl()
+            timeout_s: Optional timeout in seconds for the action.
+                Defaults to MCP_TOOL_CALL_TIMEOUT (30s) for MCP actions.
+            **kwargs: Additional arguments passed to handlers.
+
+        Returns:
+            Observation appropriate to the action type:
+                - ListToolsObservation for ListToolsAction
+                - CallToolObservation for CallToolAction
+                - Subclass-defined Observation for other actions
+        """
+        if isinstance(action, ListToolsAction):
+            return self._handle_list_tools()
+        elif isinstance(action, CallToolAction):
+            return self._handle_call_tool(action, timeout_s=timeout_s)
+        else:
+            return self._step_impl(action, timeout_s=timeout_s, **kwargs)
+
+    def _handle_list_tools(self) -> ListToolsObservation:
+        """Sync wrapper — delegates to the canonical async implementation."""
+        return run_async_safely(self._async_handle_list_tools())
+
+    async def _async_list_tools(self) -> list:
+        """
+        Async helper to list tools from the MCP client.
+
+        Returns:
+            List of tool objects from the MCP server.
+        """
+        async with self.mcp_session() as client:
+            return await client.list_tools()
+
+    def _handle_call_tool(
+        self,
+        action: CallToolAction,
+        timeout_s: Optional[float] = None,
+    ) -> CallToolObservation:
+        """Sync wrapper — delegates to the canonical async implementation."""
+        return run_async_safely(
+            self._async_handle_call_tool(action, timeout_s=timeout_s)
+        )
+
+    async def _async_call_tool(self, tool_name: str, arguments: dict) -> Any:
+        """
+        Async helper to call a tool on the MCP server.
+
+        Args:
+            tool_name: Name of the tool to invoke.
+            arguments: Dictionary of arguments to pass to the tool.
+
+        Returns:
+            The result from the tool execution.
+        """
+        async with self.mcp_session() as client:
+            return await client.call_tool(tool_name, arguments)
+
+    async def _async_handle_list_tools(self) -> ListToolsObservation:
+        """Async version of _handle_list_tools — avoids run_async_safely."""
+        try:
+            current_mode = getattr(self, "_mode", None)
+            tools_result = await self._async_list_tools()
+            tools = []
+            for tool in tools_result:
+                if tool.name not in self._mode_tool_schemas:
+                    tools.append(
+                        Tool(
+                            name=tool.name,
+                            description=tool.description or "",
+                            input_schema=tool.inputSchema
+                            if hasattr(tool, "inputSchema")
+                            else {},
+                        )
+                    )
+            for tool_name, mode_schemas in self._mode_tool_schemas.items():
+                if None in mode_schemas:
+                    schema = mode_schemas[None]
+                    tools.append(
+                        Tool(
+                            name=schema["name"],
+                            description=schema["description"],
+                            input_schema=schema["input_schema"],
+                        )
+                    )
+                elif current_mode in mode_schemas:
+                    schema = mode_schemas[current_mode]
+                    tools.append(
+                        Tool(
+                            name=schema["name"],
+                            description=schema["description"],
+                            input_schema=schema["input_schema"],
+                        )
+                    )
+            return ListToolsObservation(tools=tools)
+        except Exception as e:
+            return ListToolsObservation(
+                tools=[],
+                metadata={"error": str(e), "error_type": "list_tools_failed"},
+            )
+
+    async def _async_handle_call_tool(
+        self,
+        action: CallToolAction,
+        timeout_s: Optional[float] = None,
+    ) -> CallToolObservation:
+        """Async version of _handle_call_tool — avoids run_async_safely."""
+        timeout = timeout_s if timeout_s is not None else MCP_TOOL_CALL_TIMEOUT
+        tool_name = action.tool_name
+        current_mode = getattr(self, "_mode", None)
+
+        if tool_name in self._mode_tools:
+            mode_info = self._mode_tools[tool_name]
+            if None in mode_info:
+                func = mode_info[None]
+            elif current_mode in mode_info:
+                func = mode_info[current_mode]
+            else:
+                return CallToolObservation(
+                    tool_name=tool_name,
+                    result=None,
+                    error=ToolError(
+                        error_type=ToolErrorType.TOOL_NOT_FOUND,
+                        message=f"Tool '{tool_name}' not available in {current_mode} mode",
+                    ),
+                )
+            try:
+                if inspect.iscoroutinefunction(func):
+                    result = await func(**action.arguments)
+                else:
+                    result = func(**action.arguments)
+                return CallToolObservation(
+                    tool_name=tool_name,
+                    result=CallToolResult(
+                        content=[TextContent(type="text", text=str(result))],
+                        structured_content={"result": result},
+                        meta=None,
+                        data=result,
+                        is_error=False,
+                    ),
+                )
+            except Exception as e:
+                return CallToolObservation(
+                    tool_name=tool_name,
+                    result=None,
+                    error=ToolError(
+                        error_type=ToolErrorType.EXECUTION_ERROR,
+                        message=str(e),
+                    ),
+                )
+
+        try:
+            result = await asyncio.wait_for(
+                self._async_call_tool(action.tool_name, action.arguments),
+                timeout=timeout,
+            )
+            return CallToolObservation(tool_name=action.tool_name, result=result)
+        except asyncio.TimeoutError:
+            return CallToolObservation(
+                tool_name=action.tool_name,
+                result=None,
+                error=ToolError(
+                    error_type=ToolErrorType.TIMEOUT,
+                    message=f"Tool '{action.tool_name}' timed out after {timeout} seconds",
+                ),
+            )
+        except Exception as e:
+            error_message = str(e)
+            if (
+                "not found" in error_message.lower()
+                or "unknown tool" in error_message.lower()
+            ):
+                error_type = ToolErrorType.TOOL_NOT_FOUND
+            elif (
+                "invalid" in error_message.lower()
+                or "argument" in error_message.lower()
+            ):
+                error_type = ToolErrorType.INVALID_ARGS
+            else:
+                error_type = ToolErrorType.EXECUTION_ERROR
+            return CallToolObservation(
+                tool_name=action.tool_name,
+                result=None,
+                error=ToolError(error_type=error_type, message=error_message),
+            )
+
+    async def step_async(
+        self,
+        action: Action,
+        timeout_s: Optional[float] = None,
+        **kwargs: Any,
+    ) -> Observation:
+        """
+        Async step that routes MCP actions without going through run_async_safely.
+
+        The WebSocket handler calls this directly on the outer event loop, where
+        the MCP session is already open, avoiding the thread/event-loop deadlock
+        that occurs when the sync step() path is used via run_in_executor.
+        """
+        if isinstance(action, ListToolsAction):
+            return await self._async_handle_list_tools()
+        elif isinstance(action, CallToolAction):
+            return await self._async_handle_call_tool(action, timeout_s=timeout_s)
+        else:
+            loop = asyncio.get_event_loop()
+            return await loop.run_in_executor(
+                None, lambda: self._step_impl(action, timeout_s=timeout_s, **kwargs)
+            )
+
+    @abstractmethod
+    def _step_impl(
+        self,
+        action: Action,
+        timeout_s: Optional[float] = None,
+        **kwargs: Any,
+    ) -> Observation:
+        """
+        Handle non-MCP actions in the environment.
+
+        Subclasses must implement this method to handle any actions that are
+        not ListToolsAction or CallToolAction. This is where environment-specific
+        action processing should occur.
+
+        Args:
+            action: The action to execute (guaranteed not to be an MCP action).
+            timeout_s: Optional timeout in seconds.
+            **kwargs: Additional arguments.
+
+        Returns:
+            An Observation appropriate for the action.
+        """
+        pass
+
+    def close(self) -> None:
+        """
+        Clean up resources used by the environment.
+
+        This method cleans up the MCP client and any other resources.
+        Subclasses should call super().close() if they override this method.
+        """
+        # The MCP client uses async context manager, so cleanup happens
+        # automatically when the context exits. We just clear references.
+        self.mcp_client = None
+        self.mcp_server = None