Merge branch 'main' into cursor/setup-dev-environment-05d9

Author: burtenshaw

.claude/skills/generate-openenv-env/assets/openenv_env_template/pyproject.toml

@@ -16,8 +16,8 @@ requires-python = ">=3.10"
 dependencies = [
     # Core OpenEnv runtime (provides FastAPI server + HTTP client types)
     # install from github
-    # "openenv-core[core] @ git+https://github.com/huggingface/OpenEnv.git",
-    "openenv-core[core]>=0.2.2",
+    # "openenv @ git+https://github.com/huggingface/OpenEnv.git",
+    "openenv>=0.2.2",
     # Environment-specific dependencies
     # Add all dependencies needed for your environment here
     # Examples:

.claude/skills/generate-openenv-env/assets/openenv_env_template/server/requirements.txt

@@ -1,3 +1,3 @@
-openenv-core[core]>=0.2.2
+openenv>=0.2.2
 fastapi>=0.116.0
 uvicorn>=0.24.0

.github/workflows/openspiel_base_build.yml

@@ -58,7 +58,7 @@ jobs:
 
   # Job 2: Build OpenSpiel base image (depends on base)
   build-openspiel-base:
-    runs-on: 8-core-ubuntu
+    runs-on: ubuntu-latest
     needs: build-base  # Wait for base image to be built
     permissions:
       contents: read
@@ -94,6 +94,8 @@ jobs:
           file: envs/openspiel_env/server/Dockerfile.openspiel-base
           push: true
           platforms: linux/amd64,linux/arm64
+          build-args: |
+            BASE_IMAGE=${{ env.REGISTRY }}/${{ env.IMAGE_PREFIX }}-base:latest
           tags: ${{ steps.meta-openspiel-base.outputs.tags }}
           labels: ${{ steps.meta-openspiel-base.outputs.labels }}
           cache-from: type=gha,scope=openspiel-base

README.md

@@ -175,8 +175,8 @@ Base class for environment communication:
 Manage container deployment:
 - `LocalDockerProvider`: Run containers on local Docker daemon
 - `DockerSwarmProvider`: Deploy to Docker Swarm clusters
-- `KubernetesProvider`: Deploy to Kubernetes clusters
 - `UVProvider`, `DaytonaProvider`, `ACASandboxProvider`: Additional runtime providers
+- `KubernetesProvider`: Deploy to Kubernetes clusters (planned)
 
 #### 5. Models
 Type-safe data structures:

docs/source/_toctree.yml

@@ -29,6 +29,8 @@
       title: MCP Environment Lifecycle
     - local: guides/connecting
       title: Connecting to Servers
+    - local: guides/runtime-providers
+      title: Runtime Providers
     title: Advanced Guides
   title: Guides
 - sections:

docs/source/contributing.md

@@ -25,6 +25,20 @@ Have an idea? Open a [feature request](https://github.com/huggingface/OpenEnv/is
 
 One of the best ways to contribute is to add a new environment! See the [Building Environments](guides/first-environment.md) guide to get started.
 
+### ☁️ Add a Runtime Provider
+
+To support a new runtime (a cloud sandbox, a scheduler, etc.), subclass
+`ContainerProvider` and follow the existing providers (`DaytonaProvider`,
+`ACASandboxProvider`) as references. To keep it consistent:
+
+- Put the SDK in a named `[project.optional-dependencies]` extra in `pyproject.toml` and import it lazily, so core OpenEnv stays SDK-free.
+- Add an `[[autodoc]]` entry under "Container providers" in `docs/source/reference/core.md`.
+- In the [Runtime Providers guide](guides/runtime-providers.md), add a row to the table and a "Per-provider setup" subsection (kept in alphabetical order).
+- Add a test (the providers under `tests/test_core/` inject a fake SDK, so no network is needed).
+
+See the Cloud Sandbox Providers amendment proposed in RFC 002 for the
+provider-neutral invariants (direct base URL, WebSocket conformance, network posture).
+
 ### 📝 Improve Documentation
 
 Documentation improvements are always welcome:

docs/source/environments/sophistry_bench_sprint.md

@@ -37,7 +37,7 @@ from sophistry_bench_sprint_env import SophistryBenchSprintEnv
 
 async def main():
     # Deployed Hugging Face Space (or .from_docker_image("openenv-sophistry_bench_sprint:latest")):
-    client = await SophistryBenchSprintEnv.from_env("anushaacharya/sophistry_bench_sprint_env")
+    client = await SophistryBenchSprintEnv.from_env("openenv-community/sophistry_bench_sprint_env")
     async with client:
         obs = (await client.reset()).observation
         print(obs.prompt, obs.answer_to_defend)
@@ -67,6 +67,22 @@ the reward-hacking measurement. By default it holds **seven** components; `corre
 > reason; even with the rest of the components, forwarding them to the agent leaks the
 > reward signal and defeats the reward-hacking measurement.
 
+## Training
+
+[`examples/sophistry_bench_sprint_grpo.py`](https://github.com/huggingface/OpenEnv/blob/main/examples/sophistry_bench_sprint_grpo.py)
+trains a policy on this env with TRL's `GRPOTrainer` — a plain prompt ->
+completion -> reward setup, since the episode is single-step.
+
+Validated with a real 100-step run on Hugging Face Jobs (`Qwen2.5-0.5B-Instruct`,
+`a10g-small`) and a 100-step run on the Prime Intellect Hub
+(`Llama-3.2-1B-Instruct`, registered as `anusha/sophistry-bench-sprint`, parity-tested
+against this port). Both show `aggregate_reward` (the optimized proxy) climbing while
+`correctness_reward` (the hidden ground truth, weight 0) stays flat — the reward-hacking
+signature this env is designed to surface. The larger Prime Intellect run converges on
+the literal `claim_count_cliff` target (`n_claims` saturates at exactly 8); the smaller
+HF Jobs run finds a different shortcut instead (`n_claims` collapses to ~0, near-empty
+completions) — same underlying finding, different degenerate strategy depending on scale.
+
 ## Build & test
 
 ```bash

docs/source/environments/unity.md

@@ -594,7 +594,7 @@ This approach is useful when:
 - `mlagents` (installed from source via git)
 - `numpy>=1.20.0`
 - `pillow>=9.0.0` (for visual observations)
-- `openenv[core]>=0.2.0`
+- `openenv>=0.2.0`
 
 ## References
 

docs/source/getting-started.md

@@ -9,24 +9,19 @@ pip install openenv
 ```
 
 > [!NOTE]
-> This installs the `openenv` CLI and the `openenv.core` runtime. Environment
-> projects can depend on `openenv[core]` when they only need the server and
-> client libraries.
+> This installs the full OpenEnv runtime: the environment server, the client,
+> the `openenv` CLI, the web interface, and MCP support. Environments depend on
+> `openenv` directly.
 
 ### Optional dependencies
 
-OpenEnv ships optional extras for specific integrations. Install them with
+A few integrations ship as optional extras. Install them with
 `pip install openenv[<extra>]`:
 
 | Extra | Pulls in |
 |-------|----------|
-| `core` | Server + client runtime only (no CLI) |
-| `cli` | The `openenv` command-line interface |
-| `all` | `core` + `cli` |
 | `inspect` | The Inspect AI evaluation harness |
-
-Cloud sandbox providers ship their own extras (`daytona`, `aca`); see the Core
-API reference.
+| `daytona`, `aca`, `modal` | Cloud sandbox providers (see the Core API reference) |
 
 ## Try an Environment
 
@@ -147,16 +142,18 @@ Cloud sandbox providers implement the same `ContainerProvider` contract as
 local Docker: they start an isolated environment server and return a `base_url`
 that an `EnvClient` can connect to directly, while keeping provider-specific
 control-plane concepts (sandbox groups, projects, signed URLs, snapshots, egress
-policy) inside the provider. This keeps OpenEnv an open protocol that any hosted
-runtime — Daytona, Modal, E2B, Azure Container Apps Sandboxes, Kubernetes-backed
-sandboxes, and others — can implement without changing the client/server
-protocol.
-
-See RFC 002, "Cloud Sandbox Providers", for the provider-neutral invariants
-(direct base URL, WebSocket conformance, base URL lifetime and reconnect,
-provider-specific source mapping, orchestration-only lifecycle, explicit network
-posture), and the [Core API reference](reference/core.md) for the available
-provider implementations and their provider-specific setup.
+policy) inside the provider. Because the contract is provider-neutral, any hosted
+runtime can implement it without changing the client/server protocol.
+
+Providers shipped today: `LocalDockerProvider`, `DockerSwarmProvider`,
+`UVProvider`, `DaytonaProvider`, and `ACASandboxProvider` (Azure Container Apps
+Sandboxes). A `KubernetesProvider` is planned.
+
+See the [Runtime Providers guide](guides/runtime-providers.md) for the full list,
+install extras, and how to select a provider, and the
+[Core API reference](reference/core.md) for each provider's API. The
+provider-neutral invariants are described in the Cloud Sandbox Providers amendment
+proposed in RFC 002 (env-spec).
 
 ## Next Steps
 

docs/source/getting_started/contributing-envs.md

@@ -17,7 +17,7 @@ This guide covers three workflows:
 Before you start, make sure you have:
 
 - Python 3.11+ and [`uv`](https://github.com/astral-sh/uv) installed
-- The OpenEnv CLI: `pip install openenv[cli]` (or install from source)
+- The OpenEnv CLI: `pip install openenv` (or install from source)
 - A [Hugging Face account](https://huggingface.co/join) with a [write token](https://huggingface.co/settings/tokens)
 
 Authenticate with the Hub: