Spaces:
Sleeping
Sleeping
File size: 11,632 Bytes
922c4d1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 | """
IncidentOps Core Environment.
Implements the OpenEnv Environment base class for the IncidentOps SRE
incident-response training environment.
Episode lifecycle:
1. reset(task_name=...) β loads a scenario and returns the initial briefing
2. step(action) β executes a command, returns observation + reward
3. state() β returns current State (episode_id, step_count)
The environment terminates (done=True) when:
- The agent calls 'resolve'
- step_count reaches scenario.max_steps (episode timeout)
"""
from __future__ import annotations
import os
from datetime import datetime, timezone
from typing import Any, Optional
from uuid import uuid4
try:
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import Action, Observation, State
except ImportError:
from openenv.core.env_server.interfaces import Environment # type: ignore[no-redef]
from openenv.core.env_server.types import Action, Observation, State
try:
from models import IncidentAction, IncidentObservation
from server.graders import grade
from server.scenarios import ALL_TASK_NAMES, Scenario, get_scenario
from server.simulation import SimulationEngine
except ImportError:
from ..models import IncidentAction, IncidentObservation
from .graders import grade
from .scenarios import ALL_TASK_NAMES, Scenario, get_scenario
from .simulation import SimulationEngine
# Default task if none specified at reset
DEFAULT_TASK = os.getenv("INCIDENT_TASK", "service-restart")
# ---------------------------------------------------------
# Strict open-interval clamp for ALL rewards/scores.
# Duplicated here as a second defense layer.
# ---------------------------------------------------------
def _safe_reward(value: float) -> float:
"""
Clamp reward to the OPEN interval (0.01, 0.99).
The OpenEnv Phase 2 validator rejects scores that are exactly 0.0 or 1.0.
We never allow a reward of 0 or 1 to leave this module.
"""
v = float(value)
if v <= 0.0:
return 0.01
if v >= 1.0:
return 0.99
if v < 0.01:
return 0.01
if v > 0.99:
return 0.99
return v
class IncidentOpsEnvironment(Environment):
"""
IncidentOps: AI Incident Response Training Environment.
The agent acts as an on-call SRE engineer, interacting with a simulated
production system through a text-based terminal interface.
Three tasks of increasing difficulty are supported:
- service-restart (easy)
- config-drift (medium)
- cascading-failure (hard)
Each task is self-contained and can be selected at reset time by passing
task_name in the reset request body:
{"task_name": "config-drift"}
Attributes:
SUPPORTS_CONCURRENT_SESSIONS: False β environment is stateful per session.
"""
SUPPORTS_CONCURRENT_SESSIONS = False
def __init__(self) -> None:
self._state: State = State(episode_id=str(uuid4()), step_count=0)
self._scenario: Optional[Scenario] = None
self._engine: Optional[SimulationEngine] = None
self._done: bool = False
self._final_score: float = 0.5
self._task_name: str = DEFAULT_TASK
self._cumulative_reward: float = 0.0
# --- Core API -------------------------------------------------
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
task_name: Optional[str] = None,
**kwargs: Any,
) -> IncidentObservation:
"""
Reset the environment for a new episode.
Args:
seed: Optional seed for reproducible scenario generation.
episode_id: Optional custom episode identifier.
task_name: Which task to load. One of:
"service-restart" | "config-drift" | "cascading-failure"
Defaults to the INCIDENT_TASK env var or "service-restart".
**kwargs: Ignored extra keyword arguments.
Returns:
IncidentObservation with initial briefing text and alert counts.
"""
self._task_name = task_name or DEFAULT_TASK
if self._task_name not in ALL_TASK_NAMES:
valid = ", ".join(ALL_TASK_NAMES)
self._task_name = DEFAULT_TASK
briefing = (
f" [WARNING] Unknown task_name. Valid options: {valid}\n"
f"Loading default task: '{self._task_name}'.\n"
)
else:
briefing = ""
self._scenario = get_scenario(self._task_name, seed=seed)
self._engine = SimulationEngine(self._scenario)
self._done = False
self._cumulative_reward = 0.0
self._final_score = 0.5
self._state = State(
episode_id=episode_id or str(uuid4()),
step_count=0,
)
briefing += self._build_briefing()
# Reset reward must be returned as a valid open interval float.
# The Phase 1 validator crashes if it receives null/None here.
return IncidentObservation(
output=briefing,
timestamp=self._sim_time(),
alert_count=len(self._engine.active_alerts),
severity=self._engine.current_severity,
affected_services=self._engine.affected_services,
done=False,
reward=_safe_reward(0.5),
metadata={
"task_name": self._task_name,
"episode_id": self._state.episode_id,
"difficulty": self._scenario.difficulty,
"max_steps": self._scenario.max_steps,
},
)
def step(
self,
action: Action,
timeout_s: Optional[float] = None,
**kwargs: Any,
) -> IncidentObservation:
"""
Execute one agent action.
Args:
action: An IncidentAction with a text command string.
timeout_s: Ignored (commands are synchronous and fast).
**kwargs: Extra keyword arguments are ignored.
Returns:
IncidentObservation with command output, updated metrics, and reward.
"""
if self._done:
return self._terminal_observation("Episode already finished. Call reset() to start a new one.")
if self._scenario is None or self._engine is None:
return self._terminal_observation("Environment not initialised. Call reset() first.")
# Extract command string from action
if isinstance(action, IncidentAction):
command = action.command
elif isinstance(action, dict):
command = action.get("command", "")
else:
command = str(getattr(action, "command", action))
self._state.step_count += 1
step_num = self._state.step_count
# Execute command in simulation
output, incremental_reward, done_by_resolve = self._engine.execute(command)
# Check episode termination
timeout_reached = step_num >= self._scenario.max_steps
self._done = done_by_resolve or timeout_reached
# Compute reward β always pass through _safe_reward
if self._done:
self._final_score = grade(
scenario=self._scenario,
engine=self._engine,
step_count=step_num,
)
# _safe_reward is defense-in-depth (grader already clamps)
reward = _safe_reward(self._final_score)
else:
# Incremental reward centred at 0.5, clamped to (0.01, 0.99)
reward = _safe_reward(0.5 + incremental_reward)
self._cumulative_reward += incremental_reward
# Append timeout notice if applicable
if timeout_reached and not done_by_resolve:
output += (
f"\n\n [TIMEOUT] Episode timeout reached ({self._scenario.max_steps} steps).\n"
f" Final score computed based on actions taken so far.\n"
)
return IncidentObservation(
output=output,
timestamp=self._sim_time(offset_minutes=step_num),
alert_count=len(self._engine.active_alerts),
severity=self._engine.current_severity,
affected_services=self._engine.affected_services,
done=self._done,
reward=reward,
metadata={
"task_name": self._task_name,
"episode_id": self._state.episode_id,
"step": step_num,
"command": command,
"final_score": _safe_reward(self._final_score) if self._done else None,
},
)
@property
def state(self) -> State:
"""Return the current episode state."""
return self._state
def close(self) -> None:
"""Release any resources (nothing to release in this implementation)."""
pass
# --- Helpers ---------------------------------------------
def _build_briefing(self) -> str:
"""Build the initial incident briefing shown to the agent."""
s = self._scenario
e = self._engine
lines = [
"",
"+------------------------------------------------------------------β",
"| IncidentOps β Incident Response Environment |",
"+------------------------------------------------------------------β",
"",
f" [Task] {s.display_name}",
f" [Difficulty] {s.difficulty.upper()}",
f" [Sim-Time] {s.start_time}",
f" [Alerts] {len(e.active_alerts)} active",
f" [Severity] {e.current_severity.upper()}",
"",
f" DESCRIPTION:",
]
# Wrap description text
import textwrap
for line in textwrap.wrap(s.description, width=66):
lines.append(f" {line}")
lines += [
"",
" -------------------------------------------------------------",
" You are the on-call engineer. Triage the incident, find the",
" root cause, remediate it, and type 'resolve' when done.",
"",
" Type 'help' to see all available commands.",
" Type 'alerts' to see active alerts.",
" Type 'status' for a system-wide dashboard.",
" -------------------------------------------------------------",
"",
]
return "\n".join(lines)
def _sim_time(self, offset_minutes: int = 0) -> str:
"""Return a simulated ISO-8601 timestamp."""
if self._scenario:
base = self._scenario.start_time # e.g. "2026-04-10T04:00:00Z"
try:
from datetime import timedelta
dt = datetime.fromisoformat(base.replace("Z", "+00:00"))
dt += timedelta(minutes=offset_minutes)
return dt.strftime("%Y-%m-%dT%H:%M:%SZ")
except Exception:
pass
return datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ")
def _terminal_observation(self, message: str) -> IncidentObservation:
"""Return a terminal observation for error/edge cases."""
return IncidentObservation(
output=f"\n {message}\n",
timestamp=self._sim_time(),
alert_count=0,
severity="none",
affected_services=[],
done=True,
reward=_safe_reward(0.5), # Always use _safe_reward
metadata={},
)
|