Phase 7: FastAPI server + Dockerfile + openenv.yaml

Dockerfile

@@ -1,3 +1,14 @@
-# Placeholder — will be filled in Phase 7.
-# Target base: ghcr.io/meta-pytorch/openenv-base:latest
-# Expose port 7860 for HF Spaces Docker SDK.
+FROM ghcr.io/meta-pytorch/openenv-base:latest
+
+WORKDIR /app
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY . .
+
+RUN pip install -e .
+
+EXPOSE 7860
+
+CMD ["uvicorn", "server.app:app", "--host", "0.0.0.0", "--port", "7860"]

__init__.py

@@ -1,3 +0,0 @@
-"""SchemaShift — OpenEnv RL environment for adaptive tool use under API drift."""
-
-__version__ = "0.1.0"

openenv.yaml

@@ -1,7 +1,14 @@
-# Placeholder — will be filled in Phase 7.
-# spec_version: 1
-# name: schemashift
-# type: space
-# runtime: fastapi
-# app: server.app:app
-# port: 7860
+spec_version: 1
+name: schemashift
+type: space
+runtime: fastapi
+app: server.app:app
+port: 7860
+
+tasks:
+  - id: E1_onboard_new_hire
+    difficulty: easy
+  - id: E2_meeting_invite_blast
+    difficulty: easy
+  - id: E3_customer_lookup
+    difficulty: easy

pyproject.toml

@@ -28,8 +28,11 @@ ui = ["gradio>=4.20"]
 eval = ["openai>=1.30"]
 
 [tool.setuptools]
-packages = ["schemashift", "schemashift.tools", "schemashift.server", "schemashift.training"]
-package-dir = {"schemashift" = "."}
+py-modules = ["models", "drift", "scenarios", "graders", "client", "eval"]
+packages = ["tools", "server", "training"]
+
+[tool.setuptools.package-dir]
+"" = "."
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

server/app.py

@@ -1,4 +1,120 @@
-"""FastAPI server + Gradio replay UI. Exposes /health, /reset, /step, /state, /tasks, /grader.
+"""FastAPI server exposing SchemaShiftEnvironment as OpenEnv HTTP service."""
+from __future__ import annotations
 
-Will be filled in Phase 7.
-"""
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel, Field
+
+from models import Action
+from scenarios import SCENARIOS
+from server.environment import SchemaShiftEnvironment
+
+
+app = FastAPI(
+    title="SchemaShift OpenEnv",
+    description="RL environment for training adaptive tool use under schema drift.",
+    version="0.1.0",
+)
+env = SchemaShiftEnvironment()
+
+
+# ─────────────────────────────────────────────────────────────────
+# Request / response models
+# ─────────────────────────────────────────────────────────────────
+
+class ResetRequest(BaseModel):
+    task_id: str
+    seed: int = 0
+
+
+class StepRequest(BaseModel):
+    action: Action
+    tokens_used: int = Field(
+        default=0, ge=0,
+        description="Tokens consumed by the agent on this step",
+    )
+
+
+# ─────────────────────────────────────────────────────────────────
+# Endpoints
+# ─────────────────────────────────────────────────────────────────
+
+@app.get("/")
+def root() -> dict:
+    return {
+        "name": "SchemaShift",
+        "version": "0.1.0",
+        "description": "Adaptive tool use under schema drift",
+        "endpoints": ["/health", "/reset", "/step", "/state", "/tasks", "/grader"],
+    }
+
+
+@app.get("/health")
+def health() -> dict:
+    return {"status": "ok", "version": "0.1.0"}
+
+
+@app.post("/reset")
+def reset(req: ResetRequest) -> dict:
+    """Start new episode. Returns initial Observation as dict."""
+    try:
+        obs = env.reset(req.task_id, req.seed)
+        return obs.model_dump()
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Reset failed: {e}")
+
+
+@app.post("/step")
+def step(req: StepRequest) -> dict:
+    """Submit action, get observation + reward."""
+    try:
+        obs, reward = env.step(req.action, req.tokens_used)
+        return {
+            "observation": obs.model_dump(),
+            "reward": reward.model_dump(),
+        }
+    except RuntimeError as e:
+        raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Step failed: {e}")
+
+
+@app.get("/state")
+def get_state() -> dict:
+    """Return full current episode state for debugging."""
+    if env._state is None:
+        raise HTTPException(status_code=400, detail="No active episode. Call /reset first.")
+    return env._state.model_dump()
+
+
+@app.get("/tasks")
+def get_tasks() -> dict:
+    """List all available scenarios with metadata."""
+    tasks = []
+    for task_id, scenario in SCENARIOS.items():
+        desc = scenario["task_description"]
+        trimmed = desc[:120] + ("..." if len(desc) > 120 else "")
+        tasks.append({
+            "task_id": task_id,
+            "difficulty": scenario["difficulty"],
+            "max_steps": scenario["max_steps"],
+            "required_tools": scenario["required_tools"],
+            "description": trimmed,
+        })
+    return {"tasks": tasks, "count": len(tasks)}
+
+
+@app.get("/grader")
+def get_grader_breakdown() -> dict:
+    """Return grader scoring for current episode state."""
+    if env._state is None:
+        raise HTTPException(status_code=400, detail="No active episode.")
+    reward = env._grader(env._state)
+    return {
+        "cumulative_reward": env._state.cumulative_reward,
+        "current_breakdown": reward.model_dump(),
+        "step": env._state.step,
+        "max_steps": env._state.max_steps,
+        "done": env._state.done,
+    }

tests/test_server.py

@@ -0,0 +1,129 @@
+"""FastAPI server acceptance tests — Phase 7.
+
+Each test gets a fresh SchemaShiftEnvironment via monkeypatch so state doesn't
+leak across tests (the default module-level env would pollute test ordering).
+"""
+from __future__ import annotations
+
+import pytest
+from fastapi.testclient import TestClient
+
+from server import app as app_module
+from server.environment import SchemaShiftEnvironment
+
+
+@pytest.fixture
+def client(monkeypatch):
+    """Fresh env per test for isolation."""
+    fresh_env = SchemaShiftEnvironment()
+    monkeypatch.setattr(app_module, "env", fresh_env)
+    return TestClient(app_module.app)
+
+
+def test_root_endpoint(client) -> None:
+    r = client.get("/")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["name"] == "SchemaShift"
+    assert "endpoints" in body
+
+
+def test_health_endpoint(client) -> None:
+    r = client.get("/health")
+    assert r.status_code == 200
+    assert r.json() == {"status": "ok", "version": "0.1.0"}
+
+
+def test_tasks_endpoint(client) -> None:
+    r = client.get("/tasks")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["count"] == 3
+    task_ids = {t["task_id"] for t in body["tasks"]}
+    assert task_ids == {
+        "E1_onboard_new_hire",
+        "E2_meeting_invite_blast",
+        "E3_customer_lookup",
+    }
+    for t in body["tasks"]:
+        assert t["difficulty"] == "easy"
+        assert isinstance(t["required_tools"], list)
+
+
+def test_reset_valid_task(client) -> None:
+    r = client.post("/reset", json={"task_id": "E1_onboard_new_hire"})
+    assert r.status_code == 200
+    body = r.json()
+    assert body["task_id"] == "E1_onboard_new_hire"
+    assert body["step"] == 0
+    assert body["done"] is False
+    assert "mail" in body["tool_schemas"]
+    assert "calendar" in body["tool_schemas"]
+
+
+def test_reset_invalid_task(client) -> None:
+    r = client.post("/reset", json={"task_id": "nonexistent_task"})
+    assert r.status_code == 400
+    body = r.json()
+    assert "detail" in body
+    assert "nonexistent_task" in body["detail"]
+
+
+def test_step_before_reset_returns_400(client) -> None:
+    action_payload = {
+        "action": {
+            "type": "inspect_schema",
+            "inspect": {"tool": "mail"},
+        },
+        "tokens_used": 0,
+    }
+    r = client.post("/step", json=action_payload)
+    assert r.status_code == 400
+    assert "reset" in r.json()["detail"].lower()
+
+
+def test_step_valid_action_after_reset(client) -> None:
+    client.post("/reset", json={"task_id": "E1_onboard_new_hire"})
+    action_payload = {
+        "action": {
+            "type": "inspect_schema",
+            "inspect": {"tool": "mail"},
+        },
+        "tokens_used": 0,
+    }
+    r = client.post("/step", json=action_payload)
+    assert r.status_code == 200
+    body = r.json()
+    assert "observation" in body
+    assert "reward" in body
+    assert body["observation"]["step"] == 1
+    assert body["observation"]["last_response"]["ok"] is True
+
+
+def test_state_endpoint_before_reset_returns_400(client) -> None:
+    r = client.get("/state")
+    assert r.status_code == 400
+
+
+def test_state_endpoint_after_reset(client) -> None:
+    client.post("/reset", json={"task_id": "E1_onboard_new_hire"})
+    r = client.get("/state")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["task_id"] == "E1_onboard_new_hire"
+    assert body["step"] == 0
+    assert body["max_steps"] == 8
+    assert body["done"] is False
+    assert isinstance(body["drift_plan"], list)
+
+
+def test_grader_endpoint_after_reset(client) -> None:
+    client.post("/reset", json={"task_id": "E1_onboard_new_hire"})
+    r = client.get("/grader")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["cumulative_reward"] == 0.0
+    assert body["step"] == 0
+    assert body["max_steps"] == 8
+    assert body["done"] is False
+    assert "current_breakdown" in body