Spaces:
Sleeping
Sleeping
File size: 9,816 Bytes
9b47159 d192bd6 9b47159 cd482eb 9b47159 cd482eb d192bd6 9b47159 7e60e2c 9b47159 7e60e2c 9b47159 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 | """
Core RL environment implementing the OpenEnv interface.
Provides reset(), step(), and state for the Bug Triage environment.
Thread-safe episode store for concurrent HTTP sessions.
"""
from __future__ import annotations
import json
import logging
import random
import threading
import uuid
from pathlib import Path
from typing import Any, Dict, List, Optional
try:
from core.env_server import Environment as _OpenEnvBase
except ImportError:
class _OpenEnvBase: # type: ignore[no-redef]
pass
from ..graders import GRADERS
from ..models import (
DEVELOPERS,
BugReport,
BugTriageAction,
BugTriageObservation,
BugTriageState,
)
logger = logging.getLogger(__name__)
DATA_PATH = Path(__file__).parent.parent / "data" / "bugs.json"
class BugTriageEnvironment(_OpenEnvBase):
"""
Bug Triage RL Environment.
Each episode:
1. reset(task_id) - selects a random bug report
2. step(action) - agent triages the bug; grader scores; done=True
3. state - returns episode metadata
Episode store is thread-safe for concurrent HTTP sessions.
"""
def __init__(self, data_path: Path = DATA_PATH) -> None:
self._bugs: List[Dict[str, Any]] = self._load_data(data_path)
self._episodes: Dict[str, Dict[str, Any]] = {}
self._lock = threading.Lock()
self._current_episode_id: Optional[str] = None
logger.info("Loaded %d bug reports.", len(self._bugs))
# -- OpenEnv interface -----------------------------------------------------
def reset(self, task_id: str = "task_1") -> BugTriageObservation:
"""Start a new episode. Returns initial observation."""
bug = random.choice(self._bugs)
episode_id = uuid.uuid4().hex[:12]
episode: Dict[str, Any] = {
"episode_id": episode_id,
"task_id": task_id,
"bug": bug,
"ground_truth": bug["ground_truth"],
"actions": [],
"done": False,
}
with self._lock:
self._episodes[episode_id] = episode
self._current_episode_id = episode_id
bug_report = self._make_bug_report(bug)
logger.info(
"Episode %s | task=%s | bug=%s", episode_id, task_id, bug["bug_id"]
)
return BugTriageObservation(
done=False,
reward=0.01,
task_id=task_id,
bug_report=bug_report,
available_developers=list(DEVELOPERS),
step_number=0,
feedback="New bug report received. Please triage.",
episode_id=episode_id,
)
def step(
self,
action: BugTriageAction,
episode_id: Optional[str] = None,
) -> BugTriageObservation:
"""Execute one triage action. Episode terminates immediately."""
ep_id = episode_id or self._current_episode_id
if ep_id is None:
raise ValueError("No active episode. Call reset() first.")
with self._lock:
ep = self._episodes.get(ep_id)
if ep is None:
raise ValueError(f"Unknown episode_id: {ep_id}")
if ep["done"]:
raise ValueError(f"Episode {ep_id} is already done.")
ep["actions"].append(action.model_dump())
ep["done"] = True
# Score via task-specific grader
grader_fn = GRADERS.get(ep["task_id"], GRADERS["task_1"])
grader_score = grader_fn(ep["actions"], ep["ground_truth"])
# Shaped reward: map [0, 1] to [-0.5, 1.0] for GRPO training
reward = (grader_score * 1.5) - 0.5
# Confidence calibration bonus/penalty
calibration_bonus = self._compute_calibration_bonus(
action.confidence, grader_score
)
reward += calibration_bonus
reward = max(0.01, min(0.99, reward))
# Build feedback string
feedback = self._build_feedback(
action, ep["ground_truth"], ep["task_id"], grader_score
)
if action.confidence is not None:
feedback += (
f" | confidence={action.confidence:.2f}"
f" (calibration={calibration_bonus:+.2f})"
)
bug_report = self._make_bug_report(ep["bug"])
log_msg = "Episode %s | score=%.3f | reward=%.3f"
log_args: list[Any] = [ep_id, grader_score, reward]
if action.confidence is not None:
log_msg += " | cal=%+.2f"
log_args.append(calibration_bonus)
logger.info(log_msg, *log_args)
return BugTriageObservation(
done=True,
reward=round(reward, 4),
task_id=ep["task_id"],
bug_report=bug_report,
available_developers=list(DEVELOPERS),
step_number=1,
feedback=feedback,
grader_score=round(grader_score, 4),
episode_id=ep_id,
)
@property
def state(self) -> BugTriageState:
"""Return current episode metadata."""
ep_id = self._current_episode_id
if ep_id is None:
return BugTriageState()
with self._lock:
ep = self._episodes.get(ep_id, {})
return BugTriageState(
episode_id=ep_id,
step_count=len(ep.get("actions", [])),
task_id=ep.get("task_id", ""),
bug_id=ep.get("bug", {}).get("bug_id"),
)
# -- Grading ---------------------------------------------------------------
def grade_episode(self, episode_id: str, task_id: str) -> Dict[str, Any]:
"""Grade a completed episode. Used by the /grader endpoint."""
with self._lock:
ep = self._episodes.get(episode_id)
if ep is None:
return {"score": 0.01, "breakdown": {}, "error": "episode_not_found"}
if not ep["actions"]:
return {"score": 0.01, "breakdown": {}, "error": "no_actions"}
grader_fn = GRADERS.get(task_id, GRADERS["task_1"])
score = grader_fn(ep["actions"], ep["ground_truth"])
last_action = ep["actions"][-1]
gt = ep["ground_truth"]
breakdown: Dict[str, float] = {
"bug_type_match": float(
last_action.get("bug_type") == gt.get("bug_type")
),
"priority_match": float(
last_action.get("priority") == gt.get("priority")
),
}
if task_id == "task_3":
breakdown["developer_match"] = float(
last_action.get("assigned_developer")
== gt.get("assigned_developer")
)
breakdown["action_match"] = float(
last_action.get("suggested_action")
== gt.get("suggested_action")
)
return {"score": round(score, 4), "breakdown": breakdown}
# -- Private helpers -------------------------------------------------------
@staticmethod
def _load_data(path: Path) -> List[Dict[str, Any]]:
with open(path, "r", encoding="utf-8") as fh:
return json.load(fh)
@staticmethod
def _make_bug_report(bug: Dict[str, Any]) -> BugReport:
return BugReport(
bug_id=bug["bug_id"],
title=bug["title"],
description=bug["description"],
logs=bug.get("logs"),
environment=bug.get("environment"),
reporter=bug.get("reporter"),
created_at=bug.get("created_at"),
metadata=bug.get("metadata", {}),
)
@staticmethod
def _compute_calibration_bonus(
confidence: Optional[float],
grader_score: float,
) -> float:
"""
Compute reward adjustment based on confidence calibration.
Returns a bonus in [-0.15, +0.10] or 0.0 if confidence is not provided.
"""
if confidence is None:
return 0.0
calibration_error = abs(confidence - grader_score)
if grader_score >= 0.8 and confidence >= 0.8:
return 0.10 # correct and confident
if grader_score < 0.5 and confidence >= 0.8:
return -0.15 # wrong and overconfident
if calibration_error < 0.2:
return 0.05 # well-calibrated
return -0.05 # poorly calibrated
@staticmethod
def _build_feedback(
action: BugTriageAction,
ground_truth: Dict[str, Any],
task_id: str,
score: float,
) -> str:
parts: list[str] = [f"score={score:.2f}"]
if task_id in ("task_1", "task_3"):
match = action.bug_type == ground_truth.get("bug_type")
parts.append(
f"bug_type={'ok' if match else 'wrong'}"
f" (pred={action.bug_type}, expected={ground_truth.get('bug_type')})"
)
if task_id in ("task_2", "task_3"):
match = action.priority == ground_truth.get("priority")
parts.append(
f"priority={'ok' if match else 'wrong'}"
f" (pred={action.priority}, expected={ground_truth.get('priority')})"
)
if task_id == "task_3":
match = action.assigned_developer == ground_truth.get(
"assigned_developer"
)
parts.append(
f"developer={'ok' if match else 'wrong'}"
f" (pred={action.assigned_developer},"
f" expected={ground_truth.get('assigned_developer')})"
)
match = action.suggested_action == ground_truth.get(
"suggested_action"
)
parts.append(
f"action={'ok' if match else 'wrong'}"
f" (pred={action.suggested_action},"
f" expected={ground_truth.get('suggested_action')})"
)
return " | ".join(parts)
|