Spaces:
Sleeping
Sleeping
File size: 6,251 Bytes
afedf8c | 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 | """
Bug Report Structuring Environment - Core Environment Logic
Implements the OpenEnv 3-method interface:
- reset(task_id) → initial observation with messy bug report
- step(action) → graded observation with score + feedback
- state() → episode metadata
The environment challenges an LLM agent to convert messy, unstructured
bug reports into well-organized structured formats.
"""
import uuid
import random
from typing import Optional
from models import (
BugReportAction,
BugReportObservation,
BugReportState,
)
from tasks import get_task, get_all_task_ids
from graders import grade_submission
class BugReportEnvironment:
"""
OpenEnv-compatible environment for bug report structuring.
Supports concurrent sessions via per-instance state.
"""
SUPPORTS_CONCURRENT_SESSIONS = True
def __init__(self):
self._state = BugReportState()
self._task = None
self._done = False
self._best_score = 0.0
self._rewards = []
self._raw_report = ""
def reset(
self,
task_id: Optional[str] = None,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
) -> BugReportObservation:
"""
Initialize a new episode.
Args:
task_id: 'easy', 'medium', or 'hard'. Random if None.
seed: Random seed for reproducibility.
episode_id: Custom episode ID.
Returns:
Initial observation containing the messy bug report.
"""
# Set seed for reproducibility
if seed is not None:
random.seed(seed)
# Select task
if task_id is None:
task_id = random.choice(get_all_task_ids())
self._task = get_task(task_id)
self._done = False
self._best_score = 0.0
self._rewards = []
self._raw_report = self._task["raw_report"]
# Initialize state
eid = episode_id or str(uuid.uuid4())
self._state = BugReportState(
episode_id=eid,
step_count=0,
task_id=task_id,
max_steps=self._task["max_steps"],
current_score=0.0,
best_score=0.0,
done=False,
rewards=[],
)
return BugReportObservation(
raw_report=self._raw_report,
feedback=(
f"New episode started. Task: {task_id} "
f"(max {self._task['max_steps']} steps).\n"
f"Read the messy bug report below and submit a structured version.\n\n"
f"Required fields: title, steps_to_reproduce, expected_behavior, "
f"actual_behavior, severity (low/medium/high/critical), environment"
),
score=0.0,
field_scores={},
done=False,
reward=0.0,
step_count=0,
task_id=task_id,
max_steps=self._task["max_steps"],
)
def step(self, action: BugReportAction) -> BugReportObservation:
"""
Process an agent's structured bug report submission.
Args:
action: BugReportAction with structured fields.
Returns:
Observation with score, feedback, and done status.
"""
# Check if episode is active
if self._task is None:
return BugReportObservation(
raw_report="",
feedback="Error: No active episode. Call reset() first.",
score=0.0,
field_scores={},
done=True,
reward=0.0,
step_count=0,
task_id="",
max_steps=0,
)
if self._done:
return BugReportObservation(
raw_report=self._raw_report,
feedback=(
f"Episode already completed. Best score: {self._best_score:.2f}. "
f"Call reset() for a new episode."
),
score=self._best_score,
field_scores={},
done=True,
reward=0.0,
step_count=self._state.step_count,
task_id=self._state.task_id,
max_steps=self._state.max_steps,
)
# Increment step
self._state.step_count += 1
# Grade the submission
action_dict = action.model_dump()
score, field_scores, feedback = grade_submission(action_dict, self._task)
# Track best score and rewards
self._best_score = max(self._best_score, score)
# Reward = improvement over previous best (reward shaping)
prev_best = self._state.best_score
reward = max(0.0, score - prev_best) # Only reward improvement
if self._state.step_count == 1:
reward = score # First step gets full score as reward
self._rewards.append(round(reward, 2))
# Check if done
at_max_steps = self._state.step_count >= self._state.max_steps
high_score = score >= 0.95 # Near-perfect score ends early
self._done = at_max_steps or high_score
# Update state
self._state.current_score = score
self._state.best_score = self._best_score
self._state.done = self._done
self._state.rewards = self._rewards.copy()
# Add done info to feedback
if self._done:
if high_score:
feedback += f"\n\n🎉 Excellent! Score {score:.2f} achieved. Episode complete."
else:
feedback += (
f"\n\nMax steps reached. Best score: {self._best_score:.2f}. "
f"Episode complete."
)
return BugReportObservation(
raw_report=self._raw_report,
feedback=feedback,
score=score,
field_scores=field_scores,
done=self._done,
reward=round(reward, 2),
step_count=self._state.step_count,
task_id=self._state.task_id,
max_steps=self._state.max_steps,
)
@property
def state(self) -> BugReportState:
"""Return current episode state."""
return self._state
|