Merge pull request #4 from facebookexternal/env_code

Add BaseHTTPClient to talk to Envs via JSON over RPC.

Author: pankit-eng

src/core/base.py

@@ -43,3 +43,7 @@ def step(self, action: ActT) -> StepResult[ObsT]:
             StepResult[ObsT]: The resulting observation, reward, done flag, and info.
         """
         raise NotImplementedError
+
+    def close(self) -> None:
+        """Release resources (containers, sessions, etc.)."""
+        pass

src/core/base_env_client.py

@@ -0,0 +1,82 @@
+"""
+core/runner_env.py
+Minimal HTTP-based environment client.
+- Talks to a single env worker exposing: POST /reset, POST /step
+
+Future hooks (commented below) for:
+- episode_id, seed on reset
+- request_id on step
+- custom headers (auth/trace)
+"""
+
+from __future__ import annotations
+
+from abc import abstractmethod
+from typing import Any, Dict, Generic, Optional, TypeVar
+
+import requests
+
+from .base import BaseEnv
+from .types import StepResult
+
+ActT = TypeVar("ActT")
+ObsT = TypeVar("ObsT")
+
+
+class HTTPEnvClient(BaseEnv[ActT, ObsT], Generic[ActT, ObsT]):
+    def __init__(
+        self,
+        base_url: str,
+        request_timeout_s: float = 15.0,
+        default_headers: Optional[Dict[str, str]] = None,
+    ):
+        self._base = base_url.rstrip("/")
+        self._timeout = float(request_timeout_s)
+        self._http = requests.Session()
+        self._headers = default_headers or {}
+
+    @abstractmethod
+    def _step_payload(self, action: ActT) -> dict:
+        """Convert an Action object to the JSON body expected by the env server."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def _parse_result(self, payload: dict) -> StepResult[ObsT]:
+        """Convert a JSON response from the env server to StepResult[ObsT]."""
+        raise NotImplementedError
+
+    # ---------- BaseEnv ----------
+    def reset(self) -> ObsT:
+        body: Dict[str, Any] = {}
+        # TODO: later:
+        # body["seed"] = seed
+        # body["episode_id"] = episode_id
+        r = self._http.post(
+            f"{self._base}/reset",
+            json=body,
+            headers=self._headers,
+            timeout=self._timeout,
+        )
+        r.raise_for_status()
+        return self._parse_result(r.json()).observation
+
+    def step(self, action: ActT) -> StepResult[ObsT]:
+        body: Dict[str, Any] = {
+            "action": self._step_payload(action),
+            "timeout_s": int(self._timeout),
+        }
+        # TODO: later:
+        # body["request_id"] = str(uuid.uuid4())
+        # body["episode_id"] = current_episode_id
+        r = self._http.post(
+            f"{self._base}/step",
+            json=body,
+            headers=self._headers,
+            timeout=self._timeout,
+        )
+        r.raise_for_status()
+        return self._parse_result(r.json())
+
+    def close(self) -> None:
+        # nothing to close; higher-level libraries own lifecycles of the endpoints
+        pass

src/envs/coding_env/env.py

@@ -1,72 +1,50 @@
 """
-envs/coding_env/env.py
---------------------------------
-Concrete environment implementation using the core BaseEnv.
-POC implementation runs code locally via subprocess that can be changed later.
+CodingEnv
+---------
+Client-side wrapper for the Coding environment server.
+Talks HTTP to a single base_url exposing: /reset and /step.
+
+- users instantiate CodingEnv with a base_url provided by the higher-level
+  vector/orchestration layer.
+- Environment authors ship the Docker image that serves the HTTP API.
+
+(Seeds, episode IDs, request IDs, capabilities can be added later in the payloads.)
 """
 
 from __future__ import annotations
 
-import subprocess
 from typing import Optional
 
-from core.base import BaseEnv
+from core.base_env_client import HTTPEnvClient
 from core.types import StepResult
 
 from .models import CodeAction, CodeObservation
 
 
-class CodingEnv(BaseEnv[CodeAction, CodeObservation]):
-    """
-    Minimal Coding Environment.
-
-    POC behavior:
-      - reset(): returns a fresh, empty observation (no persistent state).
-      - step(action): runs Python code with `python -c` and returns stdout/stderr/exit_code.
-
-    Future swap:
-      Replace _run_code_locally() with a call to your Docker/gateway backend without
-      changing the public API.
-    """
-
+class CodingEnv(HTTPEnvClient[CodeAction, CodeObservation]):
     def __init__(
         self,
-        default_timeout_s: float = 10.0,
-        python_executable: str = "python",
+        base_url: str,
+        request_timeout_s: float = 15.0,
     ):
-        """
-        Args:
-            default_timeout_s: Max seconds to allow code execution before timing out.
-            python_executable: Interpreter to run (e.g., "python3", a venv path, etc.).
-        """
-        self._default_timeout_s = float(default_timeout_s)
-        self._python = python_executable
-
-    # --- BaseEnv interface ---
-
-    def reset(self) -> CodeObservation:
-        # No state to clear in this POC; return an initial observation.
-        return CodeObservation(stdout="", stderr="", exit_code=0)
-
-    def step(self, action: CodeAction) -> StepResult[CodeObservation]:
-        if not isinstance(action, CodeAction):
-            raise TypeError(f"Expected CodeAction, got {type(action)!r}")
-
-        # TODO: replace dummy response with the call to the code executor inside the container
-        obs, timed_out = CodeObservation(stderr="", stdout="", exit_code=0), False
-
-        # Simple reward heuristic: success and no stderr -> 1.0 else 0.0
-        reward: Optional[float] = (
-            1.0 if (obs.exit_code == 0 and not obs.stderr) else 0.0
+        super().__init__(
+            base_url=base_url,
+            request_timeout_s=request_timeout_s,
         )
 
-        info = {
-            "timed_out": timed_out,
-            "interpreter": self._python,
+    # --- HTTPEnvClient abstract hooks ---
+
+    def _step_payload(self, action: CodeAction) -> dict:
+        # Shape expected by the server's /step endpoint under "action"
+        return {
+            "code": action.code,
         }
 
+    def _parse_result(self, payload: dict) -> StepResult[CodeObservation]:
+        # Expecting: { "observation": {...}, "reward": <float|null>, "done": <bool>, "info": {...} }
+        obs = CodeObservation(**payload["observation"])
         return StepResult(
             observation=obs,
-            reward=reward,
-            done=False,  # Coding env is not episodic by default
+            reward=payload.get("reward"),
+            done=bool(payload.get("done", False)),
         )

src/envs/coding_env/example_usage.py

@@ -0,0 +1,7 @@
+from envs.coding_env.env import CodingEnv
+from envs.coding_env.models import CodeAction
+
+env = CodingEnv(base_url="http://localhost:8080")
+obs0 = env.reset()
+result = env.step(CodeAction(code="print('hi')"))
+print(result.observation.stdout.strip(), result.reward)