Spaces:
Running
Running
Commit ·
b4b210e
1
Parent(s): 24f0bf0
fix: add OpenEnv root reset and step aliases
Browse filesExpose OpenEnv-compatible root endpoints (/reset, /step, /state/{episode_id}) and API aliases (/api/reset, /api/step, /api/state/{episode_id}) to prevent 405 on evaluator probes.
Add endpoint coercion for relaxed action payloads and tests for alias behavior.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
backend/app/api/routes/__init__.py
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
"""API routes package."""
|
| 2 |
|
| 3 |
-
from app.api.routes import agents, episode, health, memory, sites, tasks, tools
|
| 4 |
|
| 5 |
-
__all__ = ["agents", "episode", "health", "memory", "sites", "tasks", "tools"]
|
|
|
|
| 1 |
"""API routes package."""
|
| 2 |
|
| 3 |
+
from app.api.routes import agents, episode, health, memory, openenv, sites, tasks, tools
|
| 4 |
|
| 5 |
+
__all__ = ["agents", "episode", "health", "memory", "openenv", "sites", "tasks", "tools"]
|
backend/app/api/routes/openenv.py
ADDED
|
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""OpenEnv compatibility endpoints exposed at root-level paths."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
from typing import Any
|
| 6 |
+
|
| 7 |
+
from fastapi import APIRouter, Body, HTTPException, status
|
| 8 |
+
from pydantic import BaseModel, Field
|
| 9 |
+
|
| 10 |
+
from app.api.deps import SettingsDep
|
| 11 |
+
from app.api.routes.episode import (
|
| 12 |
+
EpisodeState,
|
| 13 |
+
ResetRequest,
|
| 14 |
+
ResetResponse,
|
| 15 |
+
StepRequest,
|
| 16 |
+
get_episode_state,
|
| 17 |
+
reset_episode,
|
| 18 |
+
step_episode,
|
| 19 |
+
)
|
| 20 |
+
from app.core.action import Action, ActionType
|
| 21 |
+
|
| 22 |
+
router = APIRouter(tags=["OpenEnv"])
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class OpenEnvResetRequest(BaseModel):
|
| 26 |
+
"""Lenient reset request supporting common OpenEnv field aliases."""
|
| 27 |
+
|
| 28 |
+
task_id: str | None = Field(default=None)
|
| 29 |
+
task: str | None = Field(default=None)
|
| 30 |
+
task_name: str | None = Field(default=None)
|
| 31 |
+
seed: int | None = Field(default=None)
|
| 32 |
+
config: dict[str, Any] | None = Field(default=None)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class OpenEnvStepRequest(BaseModel):
|
| 36 |
+
"""Lenient step request supporting common OpenEnv field aliases."""
|
| 37 |
+
|
| 38 |
+
episode_id: str | None = Field(default=None)
|
| 39 |
+
episode: str | None = Field(default=None)
|
| 40 |
+
session_id: str | None = Field(default=None)
|
| 41 |
+
action: Any = Field(default_factory=dict)
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
def _coerce_action(action_payload: Any) -> Action:
|
| 45 |
+
"""Coerce OpenEnv-style actions into internal Action model."""
|
| 46 |
+
if isinstance(action_payload, Action):
|
| 47 |
+
return action_payload
|
| 48 |
+
|
| 49 |
+
if isinstance(action_payload, str):
|
| 50 |
+
action_type = action_payload.strip().lower()
|
| 51 |
+
try:
|
| 52 |
+
return Action(action_type=ActionType(action_type), parameters={})
|
| 53 |
+
except ValueError:
|
| 54 |
+
return Action.wait()
|
| 55 |
+
|
| 56 |
+
if isinstance(action_payload, dict):
|
| 57 |
+
payload = dict(action_payload)
|
| 58 |
+
|
| 59 |
+
if "action_type" not in payload:
|
| 60 |
+
for alias in ("action", "type", "name"):
|
| 61 |
+
alias_value = payload.get(alias)
|
| 62 |
+
if isinstance(alias_value, str) and alias_value.strip():
|
| 63 |
+
payload["action_type"] = alias_value.strip().lower()
|
| 64 |
+
break
|
| 65 |
+
|
| 66 |
+
if "parameters" not in payload:
|
| 67 |
+
params = payload.get("params")
|
| 68 |
+
payload["parameters"] = params if isinstance(params, dict) else {}
|
| 69 |
+
|
| 70 |
+
if "reasoning" not in payload and isinstance(payload.get("thought"), str):
|
| 71 |
+
payload["reasoning"] = payload["thought"]
|
| 72 |
+
|
| 73 |
+
action_type = payload.get("action_type")
|
| 74 |
+
if not isinstance(action_type, str):
|
| 75 |
+
payload["action_type"] = ActionType.WAIT.value
|
| 76 |
+
payload["parameters"] = {}
|
| 77 |
+
else:
|
| 78 |
+
normalized = action_type.strip().lower()
|
| 79 |
+
try:
|
| 80 |
+
ActionType(normalized)
|
| 81 |
+
payload["action_type"] = normalized
|
| 82 |
+
except ValueError:
|
| 83 |
+
payload["action_type"] = ActionType.WAIT.value
|
| 84 |
+
payload["parameters"] = {}
|
| 85 |
+
|
| 86 |
+
try:
|
| 87 |
+
return Action.model_validate(payload)
|
| 88 |
+
except Exception:
|
| 89 |
+
return Action.wait()
|
| 90 |
+
|
| 91 |
+
return Action.wait()
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
@router.post(
|
| 95 |
+
"/reset",
|
| 96 |
+
response_model=ResetResponse,
|
| 97 |
+
status_code=status.HTTP_200_OK,
|
| 98 |
+
summary="OpenEnv-compatible reset endpoint",
|
| 99 |
+
)
|
| 100 |
+
@router.post(
|
| 101 |
+
"/api/reset",
|
| 102 |
+
response_model=ResetResponse,
|
| 103 |
+
status_code=status.HTTP_200_OK,
|
| 104 |
+
include_in_schema=False,
|
| 105 |
+
)
|
| 106 |
+
async def openenv_reset(
|
| 107 |
+
settings: SettingsDep,
|
| 108 |
+
request: OpenEnvResetRequest | None = Body(default=None),
|
| 109 |
+
) -> ResetResponse:
|
| 110 |
+
"""
|
| 111 |
+
Root-level reset alias used by OpenEnv evaluators.
|
| 112 |
+
|
| 113 |
+
Defaults to `task_001` when no explicit task identifier is provided.
|
| 114 |
+
"""
|
| 115 |
+
payload = request or OpenEnvResetRequest()
|
| 116 |
+
task_id = payload.task_id or payload.task or payload.task_name or "task_001"
|
| 117 |
+
normalized_request = ResetRequest(task_id=task_id, seed=payload.seed, config=payload.config)
|
| 118 |
+
return await reset_episode(normalized_request, settings)
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
@router.post(
|
| 122 |
+
"/step",
|
| 123 |
+
status_code=status.HTTP_200_OK,
|
| 124 |
+
summary="OpenEnv-compatible step endpoint",
|
| 125 |
+
)
|
| 126 |
+
@router.post(
|
| 127 |
+
"/api/step",
|
| 128 |
+
status_code=status.HTTP_200_OK,
|
| 129 |
+
include_in_schema=False,
|
| 130 |
+
)
|
| 131 |
+
async def openenv_step(
|
| 132 |
+
request: OpenEnvStepRequest = Body(default_factory=OpenEnvStepRequest),
|
| 133 |
+
) -> dict[str, Any]:
|
| 134 |
+
"""Root-level step alias used by OpenEnv evaluators."""
|
| 135 |
+
episode_id = request.episode_id or request.episode or request.session_id
|
| 136 |
+
if not episode_id:
|
| 137 |
+
raise HTTPException(
|
| 138 |
+
status_code=status.HTTP_400_BAD_REQUEST,
|
| 139 |
+
detail="Missing episode_id",
|
| 140 |
+
)
|
| 141 |
+
|
| 142 |
+
result = await step_episode(
|
| 143 |
+
StepRequest(
|
| 144 |
+
episode_id=episode_id,
|
| 145 |
+
action=_coerce_action(request.action),
|
| 146 |
+
)
|
| 147 |
+
)
|
| 148 |
+
payload = result.model_dump()
|
| 149 |
+
payload["done"] = bool(result.terminated or result.truncated)
|
| 150 |
+
return payload
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
@router.get(
|
| 154 |
+
"/state/{episode_id}",
|
| 155 |
+
response_model=EpisodeState,
|
| 156 |
+
status_code=status.HTTP_200_OK,
|
| 157 |
+
summary="OpenEnv-compatible state endpoint",
|
| 158 |
+
)
|
| 159 |
+
@router.get(
|
| 160 |
+
"/api/state/{episode_id}",
|
| 161 |
+
response_model=EpisodeState,
|
| 162 |
+
status_code=status.HTTP_200_OK,
|
| 163 |
+
include_in_schema=False,
|
| 164 |
+
)
|
| 165 |
+
async def openenv_state(episode_id: str) -> EpisodeState:
|
| 166 |
+
"""Root-level state alias used by OpenEnv evaluators."""
|
| 167 |
+
return await get_episode_state(episode_id)
|
| 168 |
+
|
backend/app/main.py
CHANGED
|
@@ -11,7 +11,7 @@ from fastapi.middleware.cors import CORSMiddleware
|
|
| 11 |
from fastapi.responses import FileResponse, HTMLResponse
|
| 12 |
from fastapi.staticfiles import StaticFiles
|
| 13 |
|
| 14 |
-
from app.api.routes import agents, episode, health, memory, plugins, scrape, sites, tasks, tools
|
| 15 |
from app.api.routes import settings as settings_routes
|
| 16 |
from app.config import get_settings
|
| 17 |
from app.memory.manager import MemoryManager
|
|
@@ -135,6 +135,7 @@ def create_app() -> FastAPI:
|
|
| 135 |
app.include_router(plugins.router, prefix=api_prefix, tags=["Plugins"])
|
| 136 |
app.include_router(sites.router, prefix=api_prefix, tags=["Sites"])
|
| 137 |
app.include_router(scrape.router, prefix=api_prefix, tags=["Scraping"])
|
|
|
|
| 138 |
|
| 139 |
# Import and include providers router
|
| 140 |
from app.api.routes import providers
|
|
|
|
| 11 |
from fastapi.responses import FileResponse, HTMLResponse
|
| 12 |
from fastapi.staticfiles import StaticFiles
|
| 13 |
|
| 14 |
+
from app.api.routes import agents, episode, health, memory, openenv, plugins, scrape, sites, tasks, tools
|
| 15 |
from app.api.routes import settings as settings_routes
|
| 16 |
from app.config import get_settings
|
| 17 |
from app.memory.manager import MemoryManager
|
|
|
|
| 135 |
app.include_router(plugins.router, prefix=api_prefix, tags=["Plugins"])
|
| 136 |
app.include_router(sites.router, prefix=api_prefix, tags=["Sites"])
|
| 137 |
app.include_router(scrape.router, prefix=api_prefix, tags=["Scraping"])
|
| 138 |
+
app.include_router(openenv.router, tags=["OpenEnv"])
|
| 139 |
|
| 140 |
# Import and include providers router
|
| 141 |
from app.api.routes import providers
|
backend/tests/test_api/test_episode.py
CHANGED
|
@@ -46,3 +46,30 @@ def test_get_state(client: TestClient, sample_task: dict) -> None:
|
|
| 46 |
assert response.status_code == 200
|
| 47 |
data = response.json()
|
| 48 |
assert data["episode_id"] == episode_id
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
assert response.status_code == 200
|
| 47 |
data = response.json()
|
| 48 |
assert data["episode_id"] == episode_id
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def test_openenv_reset_alias(client: TestClient, sample_task: dict) -> None:
|
| 52 |
+
"""Test OpenEnv-compatible reset alias at root path."""
|
| 53 |
+
response = client.post("/reset", json={"task": sample_task["task_id"]})
|
| 54 |
+
assert response.status_code == 200
|
| 55 |
+
data = response.json()
|
| 56 |
+
assert "episode_id" in data
|
| 57 |
+
assert data["task_id"] == sample_task["task_id"]
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def test_openenv_step_alias_with_string_action(client: TestClient, sample_task: dict) -> None:
|
| 61 |
+
"""Test OpenEnv-compatible step alias accepts string action payloads."""
|
| 62 |
+
reset_response = client.post("/reset", json={"task_id": sample_task["task_id"]})
|
| 63 |
+
assert reset_response.status_code == 200
|
| 64 |
+
episode_id = reset_response.json()["episode_id"]
|
| 65 |
+
|
| 66 |
+
step_response = client.post(
|
| 67 |
+
"/step",
|
| 68 |
+
json={
|
| 69 |
+
"episode_id": episode_id,
|
| 70 |
+
"action": "done",
|
| 71 |
+
},
|
| 72 |
+
)
|
| 73 |
+
assert step_response.status_code == 200
|
| 74 |
+
data = step_response.json()
|
| 75 |
+
assert "done" in data
|