Spaces:
Sleeping
Sleeping
File size: 10,503 Bytes
543a85f 63c7e0c 543a85f 4fa79b5 b4cadb7 4fa79b5 543a85f 4fa79b5 543a85f c1f8942 b4cadb7 543a85f c1f8942 543a85f 63c7e0c 543a85f 63c7e0c 543a85f 63c7e0c 543a85f 63c7e0c 543a85f c1f8942 b4cadb7 543a85f c1f8942 543a85f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 | """
FastAPI application for the PrivilegeDesk Environment.
Endpoints (OpenEnv standard):
POST /reset - start a new episode
POST /step β execute a tool call
GET /state β current episode state
GET /schema β action/observation JSON schemas
WS /ws β WebSocket for persistent sessions
Additional endpoints (hackathon required):
GET /tasks β list the 3 tasks with action schemas
POST /grader β return episode score breakdown (0.0β1.0)
POST /baseline β run baseline agent and return scores
NOTE on HTTP statefulness
--------------------------
OpenEnv's create_app() HTTP /reset and /step handlers are intentionally
stateless β each request creates a new throwaway env instance. State is
only preserved over WebSocket.
We remove those stateless routes and replace them with our own singleton-
backed versions so that curl-style HTTP testing works correctly. WebSocket
(/ws) is untouched and still creates proper per-session environments.
"""
from typing import Any, Dict, Optional, List
from fastapi import HTTPException
from pydantic import BaseModel
from starlette.routing import Route
try:
from openenv.core.env_server.http_server import create_app
except Exception as e:
raise ImportError("openenv is required. Install with: uv sync") from e
try:
from ..models import PrivilegeDeskAction, PrivilegeDeskObservation
from .privilege_desk_environment import PrivilegeDeskEnvironment
except (ImportError, ModuleNotFoundError):
from models import PrivilegeDeskAction, PrivilegeDeskObservation
from server.privilege_desk_environment import PrivilegeDeskEnvironment
from env.world_state import WorldState
from reward.grader import grade
from pipeline.task_templates import TASK_TEMPLATES
# ββ Module-level singleton β shared across /reset, /step, /grader βββββββββββββ
# Mimics what OpenEnv's WebSocket session manager does per-connection, but
# over plain HTTP for easy curl / script testing.
_world: WorldState = WorldState()
# ββ Build base app from OpenEnv (registers /ws, /schema, /health, /state) ββββ
app = create_app(
PrivilegeDeskEnvironment,
PrivilegeDeskAction,
PrivilegeDeskObservation,
env_name="privilege_desk",
max_concurrent_envs=1,
)
# Remove the stateless /reset and /step routes that create_app registered.
# FastAPI uses first-match routing, so if we don't remove them our overrides
# (registered below) would never be reached.
_OVERRIDE_PATHS = {"/reset", "/step"}
app.routes[:] = [
r for r in app.routes
if not (
isinstance(r, Route)
and r.path in _OVERRIDE_PATHS
and "POST" in (r.methods or set())
)
]
# ββ Pydantic request models βββββββββββββββββββββββββββββββββββββββββββββββββββ
class ResetRequest(BaseModel):
task_id: str = "access_decision"
seed: Optional[int] = None
difficulty_level: int = 1
class StepRequest(BaseModel):
action: Dict[str, Any] # {"tool_name": "...", "arguments": {...}}
# ββ /reset ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.post("/reset")
def reset_episode(body: ResetRequest = None) -> Dict[str, Any]:
"""Reset to a new episode and return the initial observation.
The world state is stored in a module-level singleton so that subsequent
/step and /grader calls operate on the same episode.
"""
global _world
req = body or ResetRequest()
_world = WorldState()
obs = _world.reset(
seed=req.seed,
task_id=req.task_id,
difficulty_level=req.difficulty_level,
)
return {
"observation": obs,
"reward": 0.0,
"done": False,
"info": {
"task_id": req.task_id,
"seed": req.seed,
"episode": "started",
},
}
# ββ /step βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.post("/step")
def step_episode(body: StepRequest) -> Dict[str, Any]:
"""Execute one tool call on the current episode.
body.action must be a dict with:
tool_name (str) β e.g. "request.view", "access.decide"
arguments (dict) β tool-specific kwargs (can be empty {})
"""
global _world
if _world._router is None:
raise HTTPException(
status_code=409,
detail="No active episode. Call POST /reset first.",
)
obs, reward, terminated, truncated, info = _world.step(body.action)
done = terminated or truncated
# Clamp step reward strictly to (0.01, 0.99) β Phase 2 requirement
clamped_reward = min(max(round(reward, 4), 0.10), 0.90)
return {
"observation": obs,
"reward": clamped_reward,
"done": done,
"info": info,
}
# ββ /grader βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.post("/grader")
def grade_episode(body: Dict[str, Any] = None) -> Dict[str, Any]:
"""Return the grading breakdown for the current live episode.
Reads the world state already mutated by /reset + /step calls and
scores it β no re-execution, no fake episode.
"""
global _world
if _world._router is None or not _world._raw:
raise HTTPException(
status_code=409,
detail="No active episode. Call POST /reset first, then POST /step.",
)
score = grade(_world._raw)
# Ensure score is strictly in (0, 1) β never exactly 0.0 or 1.0
raw_score = score.get("score", 0.10)
final_score = max(0.10, min(0.90, raw_score))
return {
"task_id": _world._raw.get("task_id", "unknown"),
"score": round(final_score, 4),
"breakdown": score.get("breakdown", {}),
"weights": score.get("weights", {}),
"details": score.get("details", {}),
"steps_taken": _world.step_count,
"episode_done": _world.done,
}
# ββ /tasks ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.get("/tasks")
def list_tasks() -> List[Dict[str, Any]]:
"""List all available tasks with their schemas."""
tasks = []
for task_id, template in TASK_TEMPLATES.items():
tasks.append({
"id": task_id,
"name": task_id.replace("_", " ").title(),
"description": template["task_goal"],
"difficulty": template["difficulty"],
"grader": f"graders.{task_id}_grader",
"time_limit_seconds": 600,
"max_steps": template["max_steps"],
"available_tools": template["available_tools"],
"grading_weights": template["grading_weights"],
"action_schema": {
"type": "object",
"properties": {
"tool_name": {
"type": "string",
"enum": template["available_tools"],
"description": "Tool to call",
},
"arguments": {
"type": "object",
"description": "Tool-specific arguments",
},
},
"required": ["tool_name"],
},
})
return tasks
# ββ /baseline βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.post("/baseline")
def run_baseline() -> Dict[str, Any]:
"""Run a naive baseline agent on all 3 tasks and return scores."""
results = []
for task_id in TASK_TEMPLATES:
ws = WorldState()
ws.reset(seed=42, task_id=task_id)
total_reward = 0.0
steps = 0
if task_id == "access_decision":
_, r, _, _, _ = ws.step({
"tool_name": "access.decide",
"arguments": {"decision": "approve", "role": "viewer",
"ttl_hours": 4, "justification_category": "operational"},
})
total_reward += r
steps = 1
elif task_id == "access_review":
_, r, _, _, _ = ws.step({
"tool_name": "review.submit",
"arguments": {"summary": "baseline review"},
})
total_reward += r
steps = 1
elif task_id == "jit_escalation":
req_id = next(iter(ws._raw.get("pending_requests", {})), None)
if req_id:
_, r, _, _, _ = ws.step({
"tool_name": "access.grant",
"arguments": {"request_id": req_id},
})
total_reward += r
steps = 1
score_info = grade(ws._raw)
# Ensure score is strictly in (0, 1) β never exactly 0.0 or 1.0
raw_score = score_info.get("score", 0.10)
episode_score = max(0.10, min(0.90, raw_score))
results.append({
"task_id": task_id,
"steps": steps,
"total_step_reward": round(total_reward, 4),
"episode_score": round(episode_score, 4),
})
avg_score = sum(r["episode_score"] for r in results) / len(results) if results else 0.0
return {
"baseline_agent": "naive_terminal_tool",
"results": results,
"average_episode_score": round(avg_score, 4),
"note": "Naive baseline β hits the terminal tool immediately with default args.",
}
# ββ Server entry point ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main(host: str = "0.0.0.0", port: int = 8000):
import uvicorn
uvicorn.run(app, host=host, port=port)
if __name__ == '__main__':
main()
|