File size: 8,902 Bytes
83136ac | 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 | """OpenEnv-compatible wrapper around :class:`WorldSim`.
Every step advances exactly one agent's turn in a round-robin schedule:
``SRE -> DEV -> MANAGER -> OVERSIGHT -> SRE -> ...``
This keeps the standard single-agent OpenEnv API while letting us train a
multi-agent pipeline: TRL sees each turn as an independent (obs, action)
pair with a per-turn reward drawn from the decomposed reward function.
Self-play across roles is handled by conditioning the same policy on the
role field in the observation.
``openenv-core`` is an optional dependency. If it isn't installed, this
module falls back to a minimal base class so unit tests, dashboard runs,
and offline baselines still work. The OpenEnv server entry-point in
``server/app.py`` requires the real package.
"""
from __future__ import annotations
from typing import Any
from chaosops.env.metrics import MetricsRecorder, MetricsSnapshot
from chaosops.env.models import (
AgentRole,
ChaosOpsAction,
ChaosOpsObservation,
ChaosOpsState,
)
from chaosops.env.world_sim import Scenario, WorldSim
from chaosops.rewards.reward_fn import (
StepRewardBreakdown,
compute_step_reward,
terminal_penalty_if_unresolved,
)
try: # pragma: no cover β only relevant when openenv-core is installed.
from openenv.core.env_server import Environment as _OpenEnvBase
_HAS_OPENENV = True
except ImportError: # pragma: no cover
_HAS_OPENENV = False
class _OpenEnvBase: # type: ignore[no-redef]
"""Stand-in so ``chaosops.env.environment`` is importable without deps."""
SUPPORTS_CONCURRENT_SESSIONS: bool = False
DEFAULT_TURN_ORDER: tuple[AgentRole, ...] = (
# SRE observes first (metrics + logs). Oversight runs second so it has a
# chance to flag rogue fleet activity BEFORE the Dev agent remediates β
# otherwise a fast fix ends the episode without anyone collecting the
# rogue-catch bonus, which creates a degenerate training signal.
AgentRole.SRE,
AgentRole.OVERSIGHT,
AgentRole.DEV,
AgentRole.MANAGER,
)
class ChaosOpsEnvironment(_OpenEnvBase):
"""Round-robin multi-agent OpenEnv environment."""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(
self,
turn_order: tuple[AgentRole, ...] = DEFAULT_TURN_ORDER,
*,
metrics_capacity: int = 512,
) -> None:
self._sim = WorldSim()
self._turn_order = turn_order
self._turn_index = 0
self._last_breakdown: StepRewardBreakdown | None = None
# Allow construction without a scenario for OpenEnv's reflection step.
self._default_scenario: Scenario | None = None
# Persistent metrics recorder β replaces synthetic dashboard series.
self._metrics = MetricsRecorder(capacity=metrics_capacity)
# ------------------------------------------------------------------
# OpenEnv interface
# ------------------------------------------------------------------
def reset(
self,
seed: int | None = None,
episode_id: str | None = None,
scenario: Scenario | None = None,
**_kwargs: Any,
) -> ChaosOpsObservation:
"""Start a new episode.
``scenario`` wins over ``seed``. When neither is supplied we fall
back to a deterministic DB_DEADLOCK scenario so the server endpoint
``/reset`` always returns something sensible.
"""
if scenario is None:
if self._default_scenario is not None:
scenario = self._default_scenario
else:
from chaosops.env.models import DifficultyTier, FailureType
scenario = Scenario.from_type(
FailureType.DB_DEADLOCK,
seed=seed if seed is not None else 0,
difficulty=DifficultyTier.EASY,
)
self._sim.reset(scenario)
if episode_id:
self._sim.state.episode_id = episode_id
self._turn_index = 0
self._last_breakdown = None
self._metrics.reset()
role = self._current_role()
view = self._sim.project_view(role)
return ChaosOpsObservation(
done=False,
reward=0.0,
view=view,
step=self._sim.state.step_count,
turn_role=role,
message=(
f"Incident detected. Turn 1: {role.value.upper()}. "
"Diagnose, coordinate, and resolve."
),
)
def step(
self,
action: ChaosOpsAction,
timeout_s: float | None = None,
**_kwargs: Any,
) -> ChaosOpsObservation:
"""Advance one agent's turn.
The action is applied, physics ticks, the reward is computed, and
the next role's observation is returned. On terminal steps we
include the terminal penalty so the final reward fully accounts for
unresolved episodes.
"""
role = self._current_role()
# Enforce role field β the TRL adapter sets it, but humans forget.
if action.role != role:
action = action.model_copy(update={"role": role})
outcome_flags = self._sim.apply_action(action)
self._sim.tick()
breakdown = compute_step_reward(state=self._sim.state, outcome_flags=outcome_flags)
self._last_breakdown = breakdown
self._sim.state.cumulative_reward += breakdown.total
if outcome_flags.get("wrong_fix"):
self._sim.state.wrong_fixes = max(self._sim.state.wrong_fixes, 1)
if outcome_flags.get("miscommunication"):
self._sim.state.miscommunications = max(self._sim.state.miscommunications, 1)
terminal = self._sim.is_terminal()
if terminal:
terminal_bonus = terminal_penalty_if_unresolved(self._sim.state)
self._sim.state.cumulative_reward += terminal_bonus
# Record a real telemetry snapshot after every step β used by the
# dashboard sparkline panels and written to training logs.
self._metrics.on_step(self._sim.state, action)
self._turn_index += 1
next_role = self._current_role()
view = self._sim.project_view(next_role)
message = self._format_message(role, action, outcome_flags, terminal)
return ChaosOpsObservation(
done=terminal,
reward=breakdown.total,
view=view,
step=self._sim.state.step_count,
turn_role=next_role,
message=message,
)
@property
def state(self) -> ChaosOpsState:
return self._sim.state
# ------------------------------------------------------------------
# Introspection helpers used by the dashboard / tests
# ------------------------------------------------------------------
@property
def last_breakdown(self) -> StepRewardBreakdown | None:
return self._last_breakdown
@property
def current_role(self) -> AgentRole:
return self._current_role()
@property
def turn_order(self) -> tuple[AgentRole, ...]:
return self._turn_order
def set_default_scenario(self, scenario: Scenario) -> None:
"""Allow the server to pre-configure the next ``reset`` call."""
self._default_scenario = scenario
@property
def metrics(self) -> MetricsRecorder:
"""Expose the ring-buffer recorder for dashboards and tests."""
return self._metrics
def latest_metrics(self) -> MetricsSnapshot | None:
return self._metrics.latest()
# ------------------------------------------------------------------
# Internals
# ------------------------------------------------------------------
def _current_role(self) -> AgentRole:
return self._turn_order[self._turn_index % len(self._turn_order)]
@staticmethod
def _format_message(
role: AgentRole,
action: ChaosOpsAction,
flags: dict[str, bool],
terminal: bool,
) -> str:
parts = [f"{role.value.upper()}: {action.action_type.value}"]
if action.target:
parts.append(f"-> {action.target}")
if flags.get("resolved"):
parts.append("β incident resolved")
if flags.get("wrong_fix"):
parts.append("β wrong fix")
if flags.get("cascade_triggered"):
parts.append("β cascade triggered")
if flags.get("rogue_flagged_correctly"):
parts.append("β rogue agent caught")
if flags.get("rogue_flagged_incorrectly"):
parts.append("β false rogue flag")
if flags.get("root_cause_correct"):
parts.append("β root cause identified")
if terminal and not flags.get("resolved"):
parts.append("Γ episode ended unresolved")
return " | ".join(parts)
__all__ = ["ChaosOpsEnvironment", "DEFAULT_TURN_ORDER"]
|