Spaces:
Sleeping
Sleeping
File size: 11,742 Bytes
e5310eb | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 | """
multi_agent_env.py — MultiAgentDevOpsEnv
Two agents collaborate on the same DevOps incident episode:
Investigator — gathers evidence and diagnoses
Responder — applies remediation actions
Both share one underlying DevOpsIncidentEnv. The Investigator's evidence
is visible to the Responder through the shared observation's evidence_log.
"""
from __future__ import annotations
import copy
import uuid
from dataclasses import dataclass, field
from typing import Optional
from env import DevOpsIncidentEnv
from models import Action, ActionType, Observation
# ---------------------------------------------------------------------------
# Role action sets
# ---------------------------------------------------------------------------
INVESTIGATOR_ACTIONS: frozenset[ActionType] = frozenset({
ActionType.READ_LOGS,
ActionType.READ_METRICS,
ActionType.READ_RUNBOOK,
ActionType.SEARCH_LOGS,
ActionType.DIAGNOSE,
ActionType.ACKNOWLEDGE,
ActionType.NOOP,
})
RESPONDER_ACTIONS: frozenset[ActionType] = frozenset({
ActionType.RESTART_SERVICE,
ActionType.ROLLBACK,
ActionType.BLOCK_IP_RANGE,
ActionType.CREATE_INDEX,
ActionType.FAILOVER,
ActionType.SCALE_UP,
ActionType.ALERT_ONCALL,
ActionType.NOOP,
})
# ---------------------------------------------------------------------------
# Joint observation container
# ---------------------------------------------------------------------------
@dataclass
class MultiAgentObservation:
"""Holds both agents' views of the current environment state."""
investigator_obs: Observation
responder_obs: Observation
step: int = 0
done: bool = False
def to_dict(self) -> dict:
return {
"investigator_obs": (
self.investigator_obs.model_dump()
if hasattr(self.investigator_obs, "model_dump")
else self.investigator_obs.dict()
),
"responder_obs": (
self.responder_obs.model_dump()
if hasattr(self.responder_obs, "model_dump")
else self.responder_obs.dict()
),
"step": self.step,
"done": self.done,
}
# ---------------------------------------------------------------------------
# Main class
# ---------------------------------------------------------------------------
class MultiAgentDevOpsEnv:
"""
Multi-agent wrapper around DevOpsIncidentEnv.
Two agents — Investigator and Responder — collaborate on the same
incident episode. They share one underlying environment (and thus
one step counter) but are restricted to disjoint action sets.
Usage::
env = MultiAgentDevOpsEnv(task_id="bonus", seed=42)
obs = env.reset()
# obs is a MultiAgentObservation
result = env.step_investigator(
Action(action_type=ActionType.READ_LOGS, service="log-aggregator")
)
# result["investigator_reward"] > 0
result = env.step_responder(
Action(action_type=ActionType.ROLLBACK, service="ml-inference-service")
)
# result["responder_reward"] > 0
Both agents may NOOP on any turn. The episode ends when the
underlying environment signals done (max_steps or resolution).
"""
LEVELS = {
"investigator": INVESTIGATOR_ACTIONS,
"responder": RESPONDER_ACTIONS,
}
def __init__(self, task_id: str = "easy", seed: int = 42) -> None:
self.task_id = task_id
self.seed = seed
self.session_id = str(uuid.uuid4())
self._env = DevOpsIncidentEnv(task_id=task_id, seed=seed)
self._obs: Optional[Observation] = None
self._done: bool = False
self._step: int = 0
# Per-agent cumulative rewards
self._investigator_reward: float = 0.0
self._responder_reward: float = 0.0
# History for debugging / logging
self._investigator_actions: list[dict] = []
self._responder_actions: list[dict] = []
# -----------------------------------------------------------------------
# Public API
# -----------------------------------------------------------------------
def reset(self) -> MultiAgentObservation:
"""Reset the environment. Both agent views are returned."""
self._obs = self._env.reset(seed=self.seed)
self._done = False
self._step = 0
self._investigator_reward = 0.0
self._responder_reward = 0.0
self._investigator_actions.clear()
self._responder_actions.clear()
return self._make_obs()
def step_investigator(self, action: Action) -> dict:
"""
Investigator takes one step.
Returns a dict with keys:
investigator_obs — updated Observation (after action)
investigator_reward — float reward for this step
done — bool, episode over
error — str (only present if action was invalid)
"""
if self._done:
return {"error": "episode complete", "investigator_reward": 0.0, "done": True}
if action.action_type not in INVESTIGATOR_ACTIONS:
return {
"error": (
f"Action '{action.action_type.value}' is not allowed for Investigator. "
f"Allowed: {[a.value for a in INVESTIGATOR_ACTIONS]}"
),
"investigator_reward": 0.0,
"done": False,
}
result = self._env.step(action)
self._obs = result.observation
self._step += 1
if result.done:
self._done = True
# Investigator-specific reward scaling
reward = self._compute_investigator_reward(action, result.reward)
self._investigator_reward += reward
self._investigator_actions.append({
"step": self._step,
"action_type": action.action_type.value,
"service": action.service,
"reward": reward,
})
return {
"investigator_obs": (
result.observation.model_dump()
if hasattr(result.observation, "model_dump")
else result.observation.dict()
),
"investigator_reward": reward,
"done": result.done,
"info": result.info,
}
def step_responder(self, action: Action) -> dict:
"""
Responder takes one step.
Returns a dict with keys:
responder_obs — updated Observation (after action)
responder_reward — float reward for this step
done — bool, episode over
error — str (only present if action was invalid)
"""
if self._done:
return {"error": "episode complete", "responder_reward": 0.0, "done": True}
if action.action_type not in RESPONDER_ACTIONS:
return {
"error": (
f"Action '{action.action_type.value}' is not allowed for Responder. "
f"Allowed: {[a.value for a in RESPONDER_ACTIONS]}"
),
"responder_reward": 0.0,
"done": False,
}
result = self._env.step(action)
self._obs = result.observation
self._step += 1
if result.done:
self._done = True
# Responder-specific reward scaling
reward = self._compute_responder_reward(action, result.reward)
self._responder_reward += reward
self._responder_actions.append({
"step": self._step,
"action_type": action.action_type.value,
"service": action.service,
"reward": reward,
})
return {
"responder_obs": (
result.observation.model_dump()
if hasattr(result.observation, "model_dump")
else result.observation.dict()
),
"responder_reward": reward,
"done": result.done,
"info": result.info,
}
# -----------------------------------------------------------------------
# Reward accessors
# -----------------------------------------------------------------------
def get_investigator_reward(self) -> float:
"""Cumulative reward earned by the Investigator so far."""
return self._investigator_reward
def get_responder_reward(self) -> float:
"""Cumulative reward earned by the Responder so far."""
return self._responder_reward
def get_joint_reward(self) -> float:
"""Mean of both agents' cumulative rewards."""
return (self._investigator_reward + self._responder_reward) / 2.0
def get_state(self) -> dict:
"""Full environment state for logging / debugging."""
s = self._env.state()
return {
"session_id": self.session_id,
"task_id": self.task_id,
"seed": self.seed,
"step": self._step,
"done": self._done,
"investigator_reward": self._investigator_reward,
"responder_reward": self._responder_reward,
"joint_reward": self.get_joint_reward(),
"investigator_actions": self._investigator_actions,
"responder_actions": self._responder_actions,
"ground_truth_root_cause": s.ground_truth_root_cause,
"ground_truth_fix": s.ground_truth_fix,
"incident_resolved": s.incident_resolved,
}
# -----------------------------------------------------------------------
# Internal helpers
# -----------------------------------------------------------------------
def _make_obs(self) -> MultiAgentObservation:
"""Build a MultiAgentObservation from the current env state."""
assert self._obs is not None
# Both agents see the full observation; role is distinguished by
# which step_*() method they call, not by observation filtering.
return MultiAgentObservation(
investigator_obs=self._obs,
responder_obs=self._obs,
step=self._step,
done=self._done,
)
@staticmethod
def _compute_investigator_reward(action: Action, raw: float) -> float:
"""
Scale raw env reward for the Investigator role.
The Investigator is credited for gathering evidence and diagnosing.
Remediation rewards are not credited to them (they'll be ≤0 anyway
since the env penalises remediation without diagnosis).
"""
if action.action_type == ActionType.READ_LOGS:
return max(raw, 0.0) # keep positive signal
if action.action_type == ActionType.READ_METRICS:
return max(raw, 0.0)
if action.action_type == ActionType.READ_RUNBOOK:
return max(raw * 0.5, 0.0) # lower weight for runbook reads
if action.action_type == ActionType.DIAGNOSE:
return raw # full credit — can be negative too
if action.action_type == ActionType.NOOP:
return min(raw, 0.0) # noop penalised as normal
return max(raw, 0.0)
@staticmethod
def _compute_responder_reward(action: Action, raw: float) -> float:
"""
Scale raw env reward for the Responder role.
The Responder is credited for correct remediation and penalised for
collateral damage.
"""
if action.action_type == ActionType.NOOP:
return min(raw, 0.0)
# Pass through raw reward (includes collateral damage penalties)
return raw
|