Add HFSandboxBackend for the OpenCode harness

.github/workflows/docker-build.yml

@@ -109,6 +109,9 @@ jobs:
           - name: opencode-env
             dockerfile: envs/opencode_env/server/Dockerfile
             context: envs/opencode_env
+          - name: opencode-sandbox
+            dockerfile: envs/opencode_env/sandbox/hf_image/Dockerfile
+            context: envs/opencode_env
           - name: openapp-env
             dockerfile: envs/openapp_env/server/Dockerfile
             context: envs/openapp_env

docs/source/environments/opencode.md

@@ -127,6 +127,32 @@ turns = session.fetch_proxy_trace()             # per-turn (tokens, logprobs)
 session.close()
 ```
 
+#### Sandbox backends
+
+`E2BSandboxBackend` is the default. To run the agent in a **Hugging Face sandbox**
+instead, swap the backend (`huggingface_hub>=1.22` ships with the package):
+
+```python
+from opencode_env.sandbox import HFSandboxBackend
+
+factory = OpenCodeSessionFactory(
+    config=OpenCodeConfig(..., sandbox_home="/root"),   # HF sandbox execs as root
+    sandbox_backend=HFSandboxBackend(image="python:3.12"),
+    mode="transparent_proxy",
+)
+```
+
+`image="python:3.12"` cold-installs opencode + the proxy deps on every rollout. For faster
+rollouts use the pre-baked image (opencode + proxy deps already installed), built by CI from
+`sandbox/hf_image/Dockerfile`:
+
+```python
+sandbox_backend=HFSandboxBackend(image="ghcr.io/huggingface/openenv-opencode-sandbox:latest")
+```
+
+Any backend satisfying the `SandboxBackend` / `SandboxHandle` / `BgJob` protocols in
+`opencode_env.sandbox.base` can be plugged in the same way.
+
 ## Building the Docker Image
 
 The Dockerfile lives at `server/Dockerfile`. Use the `openenv` CLI from

envs/opencode_env/README.md

@@ -140,6 +140,32 @@ turns = session.fetch_proxy_trace()             # per-turn (tokens, logprobs)
 session.close()
 ```
 
+#### Sandbox backends
+
+`E2BSandboxBackend` is the default. To run the agent in a **Hugging Face sandbox**
+instead, swap the backend (`huggingface_hub>=1.22` ships with the package):
+
+```python
+from opencode_env.sandbox import HFSandboxBackend
+
+factory = OpenCodeSessionFactory(
+    config=OpenCodeConfig(..., sandbox_home="/root"),   # HF sandbox execs as root
+    sandbox_backend=HFSandboxBackend(image="python:3.12"),
+    mode="transparent_proxy",
+)
+```
+
+`image="python:3.12"` cold-installs opencode + the proxy deps on every rollout. For faster
+rollouts use the pre-baked image (opencode + proxy deps already installed), built by CI from
+`sandbox/hf_image/Dockerfile`:
+
+```python
+sandbox_backend=HFSandboxBackend(image="ghcr.io/huggingface/openenv-opencode-sandbox:latest")
+```
+
+Any backend satisfying the `SandboxBackend` / `SandboxHandle` / `BgJob` protocols in
+`opencode_env.sandbox.base` can be plugged in the same way.
+
 ## Building the Docker Image
 
 The Dockerfile lives at `server/Dockerfile`. Use the `openenv` CLI from

envs/opencode_env/harness.py

@@ -49,21 +49,22 @@
     build_opencode_json,
     build_run_cmd,
     instruction_path,
+    opencode_bin_path,
     opencode_config_path,
+    proxy_dir,
+    proxy_log_path,
+    proxy_source_path,
+    proxy_trace_path,
     system_prompt_path,
 )
 from .sandbox.base import BgJob, SandboxBackend, SandboxHandle
 from .task import OpenCodeTask
 
 
-# Inside-sandbox proxy paths (Mode B).
+# Mode B proxy port. In-sandbox paths derive from config.sandbox_home (opencode_runtime).
 _PROXY_PORT = 7000
-_PROXY_TRACE_PATH = "/home/user/logs/agent/proxy_trace.jsonl"
-_PROXY_LOG_PATH = "/home/user/logs/agent/proxy.log"
 
-# Where the proxy source lives on disk (in this repo). Uploaded into the
-# sandbox at /home/user/proxy/interception.py before each rollout, unless
-# the sandbox was created from a template that already has it baked in.
+# Local proxy source, uploaded to proxy_source_path(config) unless already baked in.
 _PROXY_SOURCE_PATH = Path(__file__).parent / "sandbox" / "interception.py"
 
 
@@ -240,53 +241,57 @@ def create(
             or getattr(getattr(sandbox, "raw", None), "sandbox_id", "?")
         )
         _log.info("factory.create: sandbox=%s — bootstrapping…", sid)
+        # Any failure past here (bootstrap/proxy/agent) must tear the sandbox down.
         try:
             self._bootstrap_sandbox(sandbox, oc_task)
-        except Exception as exc:
-            _log.error("factory.create: bootstrap failed: %r", exc)
-            sandbox.kill()
-            raise
 
-        base_url_override: str | None = None
-        proxy_trace_path: str | None = None
-        proxy_bg_job: BgJob | None = None
-        if self._mode == "transparent_proxy":
-            _log.info(
-                "factory.create: starting interception proxy on :%d → %s",
-                _PROXY_PORT, self._config.base_url,
-            )
-            proxy_bg_job, base_url_override, proxy_trace_path = self._start_proxy(
-                sandbox
-            )
-            _log.info("factory.create: proxy up at %s", base_url_override)
-            # Rewrite opencode.json so opencode points at the proxy. Force
-            # ``openai_compatible`` so opencode hits ``/v1/chat/completions``
-            # (which the proxy serves) rather than provider-specific paths.
-            from .config import OpenCodeConfig as _OCC
-
-            proxy_cfg = _OCC(
-                **{
-                    **self._config.model_dump(),
-                    "provider": "openai_compatible",
-                    "base_url": base_url_override,
-                }
-            )
-            sandbox.write_text(
-                opencode_config_path(self._config),
-                build_opencode_json(proxy_cfg),
-            )
+            base_url_override: str | None = None
+            proxy_trace_path: str | None = None
+            proxy_bg_job: BgJob | None = None
+            if self._mode == "transparent_proxy":
+                _log.info(
+                    "factory.create: starting interception proxy on :%d → %s",
+                    _PROXY_PORT, self._config.base_url,
+                )
+                proxy_bg_job, base_url_override, proxy_trace_path = self._start_proxy(
+                    sandbox
+                )
+                _log.info("factory.create: proxy up at %s", base_url_override)
+                # Rewrite opencode.json so opencode points at the proxy. Force
+                # ``openai_compatible`` so opencode hits ``/v1/chat/completions``
+                # (which the proxy serves) rather than provider-specific paths.
+                from .config import OpenCodeConfig as _OCC
+
+                proxy_cfg = _OCC(
+                    **{
+                        **self._config.model_dump(),
+                        "provider": "openai_compatible",
+                        "base_url": base_url_override,
+                    }
+                )
+                sandbox.write_text(
+                    opencode_config_path(self._config),
+                    build_opencode_json(proxy_cfg),
+                )
 
-        session = OpenCodeSession(
-            sandbox=sandbox,
-            config=self._config,
-            task=oc_task,
-            verifier=self._verifier,
-            base_url_override=base_url_override,
-            proxy_trace_path=proxy_trace_path,
-            proxy_bg_job=proxy_bg_job,
-        )
-        session.start_agent()
-        return session
+            session = OpenCodeSession(
+                sandbox=sandbox,
+                config=self._config,
+                task=oc_task,
+                verifier=self._verifier,
+                base_url_override=base_url_override,
+                proxy_trace_path=proxy_trace_path,
+                proxy_bg_job=proxy_bg_job,
+            )
+            session.start_agent()
+            return session
+        except Exception as exc:
+            _log.error("factory.create: setup failed, killing sandbox: %r", exc)
+            try:
+                sandbox.kill()  # best-effort: don't let a cleanup failure mask the root cause
+            except Exception:
+                _log.exception("factory.create: sandbox.kill() during cleanup also failed")
+            raise
 
     # ------------------------------------------------------------------
     def _wait_for_sandbox_ready(
@@ -369,7 +374,7 @@ def _opencode_already_installed(self, sandbox: SandboxHandle) -> bool:
         """
         try:
             r = sandbox.exec(
-                "/home/user/.opencode/bin/opencode --version",
+                f"{opencode_bin_path(self._config)} --version",
                 timeout=10,
             )
             return r.exit_code == 0
@@ -440,9 +445,7 @@ def _start_proxy(
         Skips the pip install + source-upload steps when the prebaked
         template already has them in place.
         """
-        proxy_already_present = sandbox.exists(
-            "/home/user/proxy/interception.py"
-        )
+        proxy_already_present = sandbox.exists(proxy_source_path(self._config))
 
         if not proxy_already_present:
             # Install proxy deps (idempotent on retries).
@@ -457,18 +460,18 @@ def _start_proxy(
             )
             # Upload the proxy module into the sandbox.
             sandbox.write_text(
-                "/home/user/proxy/interception.py",
+                proxy_source_path(self._config),
                 _PROXY_SOURCE_PATH.read_text(),
             )
-            sandbox.write_text("/home/user/proxy/__init__.py", "")
+            sandbox.write_text(f"{proxy_dir(self._config)}/__init__.py", "")
 
         proxy_args = [
             "python",
             "interception.py",
             "--upstream-url",
             self._config.base_url,
             "--trace",
-            _PROXY_TRACE_PATH,
+            proxy_trace_path(self._config),
             "--port",
             str(_PROXY_PORT),
             "--top-logprobs",
@@ -487,9 +490,9 @@ def _start_proxy(
 
         quoted_proxy_args = " ".join(shlex.quote(arg) for arg in proxy_args)
         proxy_cmd = (
-            "cd /home/user/proxy && "
+            f"cd {shlex.quote(proxy_dir(self._config))} && "
             f"{quoted_proxy_args} "
-            f"> {shlex.quote(_PROXY_LOG_PATH)} 2>&1"
+            f"> {shlex.quote(proxy_log_path(self._config))} 2>&1"
         )
         proxy_env = {"OPENCODE_UPSTREAM_API_KEY": self._config.api_key}
         proxy_job = sandbox.start_bg(proxy_cmd, envs=proxy_env)
@@ -511,7 +514,7 @@ def _start_proxy(
         else:
             log = ""
             try:
-                log = sandbox.read_text(_PROXY_LOG_PATH)
+                log = sandbox.read_text(proxy_log_path(self._config))
             except Exception:
                 pass
             proxy_job.kill()
@@ -521,7 +524,7 @@ def _start_proxy(
             )
 
         base_url_override = f"http://127.0.0.1:{_PROXY_PORT}/v1"
-        return proxy_job, base_url_override, _PROXY_TRACE_PATH
+        return proxy_job, base_url_override, proxy_trace_path(self._config)
 
 
 __all__ = [

envs/opencode_env/opencode_runtime.py

@@ -43,6 +43,26 @@ def workdir_path(config: OpenCodeConfig) -> str:
     return f"{config.sandbox_home}/workdir"
 
 
+def opencode_bin_path(config: OpenCodeConfig) -> str:
+    return f"{config.sandbox_home}/.opencode/bin/opencode"
+
+
+def proxy_dir(config: OpenCodeConfig) -> str:
+    return f"{config.sandbox_home}/proxy"
+
+
+def proxy_source_path(config: OpenCodeConfig) -> str:
+    return f"{proxy_dir(config)}/interception.py"
+
+
+def proxy_trace_path(config: OpenCodeConfig) -> str:
+    return f"{config.sandbox_home}/logs/agent/proxy_trace.jsonl"
+
+
+def proxy_log_path(config: OpenCodeConfig) -> str:
+    return f"{config.sandbox_home}/logs/agent/proxy.log"
+
+
 def build_opencode_json(config: OpenCodeConfig) -> str:
     """Return the serialized ``opencode.json`` the sandbox should install.
 

envs/opencode_env/pyproject.toml

@@ -29,6 +29,8 @@ dependencies = [
     # OpenCode harness primitive — sandbox + proxy + agent driver
     "httpx>=0.27.0",
     "e2b>=1.0.0",
+    # >=1.22 ships huggingface_hub.Sandbox (HFSandboxBackend); core only floors 0.20.
+    "huggingface_hub>=1.22",
 ]
 
 [project.optional-dependencies]

envs/opencode_env/sandbox/__init__.py

@@ -41,6 +41,29 @@ def __init__(self, *_args, **_kwargs):
 
     E2BBgJob = E2BSandboxBackend = E2BSandboxHandle = _RequiresE2B  # type: ignore[assignment]
 
+try:
+    from .hf import HFBgJob, HFSandboxBackend, HFSandboxHandle  # noqa: F401
+except ImportError as _hf_err:  # pragma: no cover
+
+    class _RequiresHFSandbox:
+        """Stub raised when a recent enough ``huggingface_hub`` (with ``Sandbox``) is missing.
+
+        Lets the package import cleanly so unit tests, ``openenv validate``, and the
+        docs build can run. Actually constructing one of these classes raises a clear
+        ImportError. ``Sandbox`` ships in ``huggingface_hub>=1.22``.
+        """
+
+        _hf_import_error = _hf_err
+
+        def __init__(self, *_args, **_kwargs):
+            raise ImportError(
+                "huggingface_hub>=1.22 (with `Sandbox`) is required to use "
+                "HFSandboxBackend; upgrade via `pip install -U huggingface_hub`. "
+                f"Original import error: {self._hf_import_error}"
+            )
+
+    HFBgJob = HFSandboxBackend = HFSandboxHandle = _RequiresHFSandbox  # type: ignore[assignment]
+
 
 __all__ = [
     "BgJob",
@@ -50,4 +73,7 @@ def __init__(self, *_args, **_kwargs):
     "E2BBgJob",
     "E2BSandboxBackend",
     "E2BSandboxHandle",
+    "HFBgJob",
+    "HFSandboxBackend",
+    "HFSandboxHandle",
 ]