Spaces:
Sleeping
Sleeping
File size: 10,895 Bytes
e2eb9d7 56724ad e2eb9d7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 | """
AgentOps Gym β Core Environment class.
Implements the OpenEnv Environment interface: reset(), step(), state.
Orchestrates tool execution, reward shaping, and episode grading.
Each episode is fully deterministic given a task_id:
- Snapshot is restored from PROJECT_SNAPSHOTS on reset
- All tool calls operate on the in-memory snapshot
- No real filesystem, no real subprocess
"""
import copy
import logging
import uuid
from typing import Optional, Any
from openenv.core.env_server.interfaces import Environment
try:
from agentops_gym.models import ToolCall, AgentObservation, AgentState
from agentops_gym.server.tools import run_tool, PROJECT_SNAPSHOTS, AVAILABLE_TOOLS
from agentops_gym.server.tasks import (
TASK_REGISTRY,
get_task,
list_task_ids,
compute_step_reward,
grade_episode,
)
except (ModuleNotFoundError, ImportError):
from models import ToolCall, AgentObservation, AgentState
from server.tools import run_tool, PROJECT_SNAPSHOTS, AVAILABLE_TOOLS
from server.tasks import (
TASK_REGISTRY,
get_task,
list_task_ids,
compute_step_reward,
grade_episode,
)
logger = logging.getLogger(__name__)
_last_grader_result: Optional[dict] = None
class AgentOpsEnvironment(Environment[ToolCall, AgentObservation, AgentState]):
"""Tool-use efficiency training environment.
Each episode:
1. reset() selects a task, initialises the in-memory snapshot, returns initial obs
2. step() executes a tool call, computes reward, checks completion
3. state property returns current episode metadata
"""
def __init__(self):
super().__init__()
self._episode_id: str = ""
self._task_id: str = ""
self._task: dict = {}
self._snapshot: dict = {}
self._visible_files: list = []
self._discovered_files: list = []
self._action_history: list = []
self._step_count: int = 0
self._max_steps: int = 10
self._done: bool = True
self._cumulative_reward: float = 0.0
self._grader_score: Optional[float] = None
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
**kwargs: Any,
) -> AgentObservation:
"""Start a new episode.
kwargs may include 'task_id' to select a specific task.
If not given, defaults to task_1 (can be cycled externally).
"""
task_id = kwargs.get("task_id", "task_1")
if task_id not in TASK_REGISTRY:
task_id = "task_1"
self._episode_id = episode_id or str(uuid.uuid4())
self._task_id = task_id
self._task = get_task(task_id)
self._max_steps = self._task["max_steps"]
self._snapshot = copy.deepcopy(PROJECT_SNAPSHOTS.get(task_id, {}))
self._visible_files = list(self._task["initial_visible_files"])
self._discovered_files = list(self._visible_files)
self._action_history = []
self._step_count = 0
self._done = False
self._cumulative_reward = 0.0
self._grader_score = None
logger.info("Episode %s started: task=%s", self._episode_id, task_id)
return AgentObservation(
visible_files=list(self._visible_files),
last_tool_result=None,
action_history=[],
step_count=0,
task_description=self._task["description"],
message=f"Episode started. Available tools: {', '.join(AVAILABLE_TOOLS.keys())}",
done=False,
reward=0.0,
metadata={
"task_id": task_id,
"difficulty": self._task["difficulty"],
"max_steps": self._max_steps,
"available_tools": list(AVAILABLE_TOOLS.keys()),
},
)
def step(
self,
action: ToolCall,
**kwargs: Any,
) -> AgentObservation:
"""Execute one tool call and return updated observation."""
if self._done:
return self._terminal_obs("Episode already done. Call reset() first.")
self._step_count += 1
tool = action.tool
params = action.parameters
tool_result, self._snapshot, self._discovered_files = run_tool(
tool=tool,
parameters=params,
snapshot=self._snapshot,
discovered_files=self._discovered_files,
)
history_before = list(self._action_history)
action_str = f"{tool}({params})"
self._action_history.append(action_str)
for f in self._discovered_files:
if f not in self._visible_files:
self._visible_files.append(f)
step_reward, reward_breakdown = compute_step_reward(
task_id=self._task_id,
tool=tool,
parameters=params,
tool_result=tool_result,
action_history=history_before,
discovered_files=self._discovered_files,
snapshot=self._snapshot,
)
self._cumulative_reward += step_reward
self._cumulative_reward = max(0.0, min(1.0, self._cumulative_reward))
done = False
message = None
if self._step_count >= self._max_steps:
done = True
message = f"Max steps ({self._max_steps}) reached."
# Hard cap for task_3
if self._task_id == "task_3" and self._step_count > 8:
done = True
message = "Hard step cap (8) exceeded. Score capped at 0.3."
# ββ Task completion detection ββββββββββββββββββββββββββββββββββ
# task_1: linter ran and found the bug (or agent read main.py + grepped json)
if self._task_id == "task_1":
linted = any("BASH" in h.upper() and "LINT" in h.upper() for h in self._action_history)
read_main = any("FILEREAD" in h.upper() and "MAIN.PY" in h.upper() for h in self._action_history)
found_json = any("GREP" in h.upper() and "JSON" in h.upper() for h in self._action_history)
if linted or (read_main and found_json):
done = True
message = "Bug identified β grading episode."
# task_2: config.json was written with timeout=10
elif self._task_id == "task_2":
import json as _json
try:
cfg = _json.loads(self._snapshot.get("config.json", "{}"))
if cfg.get("timeout") == 10:
done = True
message = "Config patched successfully β grading episode."
except Exception:
pass
# task_3: main.py now contains a cache mechanism
elif self._task_id == "task_3":
main_src = self._snapshot.get("main.py", "")
if "lru_cache" in main_src or "_cache" in main_src:
done = True
message = "Caching implemented β grading episode."
# task_4: .env contains API_KEY and main.py uses os.getenv
elif self._task_id == "task_4":
main_src = self._snapshot.get("main.py", "")
env_src = self._snapshot.get(".env", "")
if "API_KEY=SECRET_TOKEN_XYZ" in env_src.replace(" ", "") and \
"os.getenv" in main_src and \
"SECRET_TOKEN_XYZ" not in main_src:
done = True
message = "Secret migrated successfully β grading episode."
# Redundant call message (non-terminating)
if len(self._action_history) >= 2 and self._action_history[-1] == self._action_history[-2]:
message = (message or "") + " Redundant call detected."
self._done = done
# Compute final grader score at episode end
grader_score = None
if done:
grader_score, breakdown = grade_episode(
task_id=self._task_id,
snapshot=self._snapshot,
action_history=self._action_history,
steps_used=self._step_count,
)
self._grader_score = grader_score
# Store globally for /grader endpoint
global _last_grader_result
_last_grader_result = {
"task_id": self._task_id,
"episode_id": self._episode_id,
"score": grader_score,
"breakdown": breakdown,
"steps_used": self._step_count,
}
# Add completion bonus proportional to grader score
step_reward += grader_score * 0.5
logger.info(
"Episode %s done: task=%s score=%.3f steps=%d",
self._episode_id, self._task_id, grader_score, self._step_count,
)
return AgentObservation(
visible_files=list(self._visible_files),
last_tool_result=tool_result,
action_history=list(self._action_history),
step_count=self._step_count,
task_description=self._task["description"],
message=message,
done=done,
reward=round(step_reward, 4),
metadata={
"task_id": self._task_id,
"difficulty": self._task["difficulty"],
"cumulative_reward": round(self._cumulative_reward, 4),
"grader_score": grader_score,
"reward_breakdown": reward_breakdown,
"steps_remaining": self._max_steps - self._step_count,
},
)
@property
def state(self) -> AgentState:
return AgentState(
episode_id=self._episode_id,
step_count=self._step_count,
task_id=self._task_id,
task_description=self._task.get("description", ""),
difficulty=self._task.get("difficulty", ""),
max_steps=self._max_steps,
visible_files=list(self._visible_files),
discovered_files=list(self._discovered_files),
action_history=list(self._action_history),
current_reward=round(self._cumulative_reward, 4),
completed=self._done,
grader_score=self._grader_score,
)
def close(self) -> None:
pass
def _terminal_obs(self, msg: str) -> AgentObservation:
return AgentObservation(
visible_files=list(self._visible_files),
last_tool_result=msg,
action_history=list(self._action_history),
step_count=self._step_count,
task_description=self._task.get("description", ""),
message=msg,
done=True,
reward=0.0,
metadata={"task_id": self._task_id, "grader_score": self._grader_score},
)
def get_last_grader_result() -> Optional[dict]:
return _last_grader_result |