Spaces:
Sleeping
Sleeping
File size: 11,847 Bytes
5fde057 | 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 | """
TICKETMELT β The environment.
Four on-call engineers, two production servers, eight minutes until the
traffic wave peaks. One engineer is secretly tracking a high-visibility
session. The trained model plays ONE engineer; three scripted peers play
the rest.
This file implements the OpenEnv Gym-style interface: reset(), step(), state().
"""
from __future__ import annotations
import json
import random
from dataclasses import asdict
from typing import Optional
from .models import (
Action, Observation, State, ServiceState, RoundRecord,
SERVICE_NAMES, VALID_COMMITMENTS, Commitment,
)
from .opponents import DUMB_PEERS, PeerStrategy, run_peer
from .rewards import compute_rewards, RewardBreakdown
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
TOTAL_ROUNDS_DEFAULT = 8
MESSAGE_TOKEN_CAP = 40 # approximate β we count whitespace-separated words
MIN_FIX_ROUNDS = 1
MAX_FIX_ROUNDS = 4
MIN_DEADLINE = 3
MAX_DEADLINE = 8
# ---------------------------------------------------------------------------
# The environment
# ---------------------------------------------------------------------------
class TicketmeltEnv:
"""
Gym-style environment with reset/step/state.
We wrap this in a FastAPI layer (see server.py) for OpenEnv compliance.
"""
def __init__(
self,
total_rounds: int = TOTAL_ROUNDS_DEFAULT,
peer_strategies: dict[str, PeerStrategy] = None,
seed: Optional[int] = None,
):
self.total_rounds = total_rounds
self.peer_strategies = peer_strategies or DUMB_PEERS
self.rng = random.Random(seed)
self._state: Optional[State] = None
self._last_rewards: Optional[RewardBreakdown] = None
# -----------------------------------------------------------------------
# reset()
# -----------------------------------------------------------------------
def reset(self, seed: Optional[int] = None) -> Observation:
"""Start a fresh episode, return initial observation for the trained engineer."""
if seed is not None:
self.rng.seed(seed)
# Name the four engineers β one trained, three scripted peers.
engineer_names = ["trained"] + list(self.peer_strategies.keys())
if len(engineer_names) != 4:
raise ValueError(
f"Expected exactly 3 scripted peers, got {len(self.peer_strategies)}"
)
# Assign services to engineers (one service per engineer)
services_assigned = list(SERVICE_NAMES)
self.rng.shuffle(services_assigned)
# Pick exactly one engineer to be "urgent" (tracking a high-visibility session)
urgent_engineer = self.rng.choice(engineer_names)
# Build ServiceState for each engineer
services = {}
for engineer, service_name in zip(engineer_names, services_assigned):
# Cap fix_rounds so it fits within the episode
max_fix = min(MAX_FIX_ROUNDS, max(MIN_FIX_ROUNDS, self.total_rounds - 1))
fix_rounds = self.rng.randint(MIN_FIX_ROUNDS, max_fix)
# Make deadline reachable-but-tight: fix_rounds + some slack
lo = max(MIN_DEADLINE, fix_rounds + 1)
hi = min(MAX_DEADLINE, self.total_rounds)
if lo > hi:
# Episode is too short for normal slack β use tightest feasible deadline
lo = hi = self.total_rounds
deadline = self.rng.randint(lo, hi)
services[engineer] = ServiceState(
name=service_name,
fix_rounds_total=fix_rounds,
fix_rounds_remaining=fix_rounds,
deadline_round=deadline,
urgency_flag=(engineer == urgent_engineer),
)
self._state = State(
current_round=0,
total_rounds=self.total_rounds,
services=services,
history=[],
trained_engineer="trained",
done=False,
)
self._last_rewards = None
return self._build_observation_for_trained()
# -----------------------------------------------------------------------
# step()
# -----------------------------------------------------------------------
def step(self, action: Action) -> tuple[Observation, float, bool, dict]:
"""
Advance one round. Returns (observation, reward, done, info).
Reward is the binary GRPO reward; the full breakdown is in info.
Reward is 0 every round EXCEPT the final round, where the episode
is scored end-to-end. This is intentional β end-of-episode scoring
is cleaner for GRPO's group comparison mechanic.
"""
if self._state is None:
raise RuntimeError("Must call reset() before step().")
if self._state.done:
raise RuntimeError("Episode already done. Call reset().")
# Sanitize the trained engineer's action
action = self._sanitize_action(action)
# Collect all four actions for this round (trained + 3 peers)
round_actions: dict[str, Action] = {"trained": action}
for peer_name, strategy in self.peer_strategies.items():
peer_service = self._state.services[peer_name]
peer_action = run_peer(strategy, peer_service, self._state.history)
peer_action = self._sanitize_action(peer_action)
round_actions[peer_name] = peer_action
# Resolve collisions and advance services
record = self._resolve_round(round_actions)
self._state.history.append(record)
self._state.current_round += 1
# Check episode termination
all_done = all(s.completed for s in self._state.services.values())
out_of_rounds = self._state.current_round >= self._state.total_rounds
self._state.done = all_done or out_of_rounds
# Compute reward only at end of episode (zero otherwise)
if self._state.done:
self._finalize_service_status()
self._last_rewards = compute_rewards(self._state)
reward_for_training = self._last_rewards.binary_grpo_reward
else:
reward_for_training = 0.0
obs = self._build_observation_for_trained()
info = self._build_info()
return obs, reward_for_training, self._state.done, info
# -----------------------------------------------------------------------
# state()
# -----------------------------------------------------------------------
def state(self) -> dict:
"""Return full state for inspection. Includes everything β peers' urgency, all history."""
if self._state is None:
return {"error": "not initialized; call reset() first"}
out = {
"current_round": self._state.current_round,
"total_rounds": self._state.total_rounds,
"done": self._state.done,
"trained_engineer": self._state.trained_engineer,
"services": {
name: asdict(svc) for name, svc in self._state.services.items()
},
"history": [asdict(r) for r in self._state.history],
}
if self._last_rewards is not None:
out["rewards"] = asdict(self._last_rewards)
out["episode_summary"] = self._state.episode_summary()
return out
# -----------------------------------------------------------------------
# Internal helpers
# -----------------------------------------------------------------------
def _sanitize_action(self, action: Action) -> Action:
"""Enforce format safety: invalid commitment β MONITOR, long message β truncated."""
commitment = action.commitment if action.commitment in VALID_COMMITMENTS else "MONITOR"
msg = action.channel_msg or ""
# Token cap: crude whitespace split
tokens = msg.split()
if len(tokens) > MESSAGE_TOKEN_CAP:
msg = " ".join(tokens[:MESSAGE_TOKEN_CAP])
return Action(commitment=commitment, channel_msg=msg)
def _resolve_round(self, round_actions: dict[str, Action]) -> RoundRecord:
"""Determine who successfully deployed, who collided, who just monitored."""
record = RoundRecord(round_number=self._state.current_round + 1)
for name, a in round_actions.items():
record.messages[name] = a.channel_msg
record.commitments[name] = a.commitment
# Count deploys per server
deploys_a = [n for n, a in round_actions.items() if a.commitment == "DEPLOY_PROD_A"]
deploys_b = [n for n, a in round_actions.items() if a.commitment == "DEPLOY_PROD_B"]
# Collisions: 2+ on same server
if len(deploys_a) >= 2:
record.collisions.append("PROD_A")
else:
for name in deploys_a:
self._advance_service(name)
record.successful_deploys[name] = "PROD_A"
if len(deploys_b) >= 2:
record.collisions.append("PROD_B")
else:
for name in deploys_b:
self._advance_service(name)
record.successful_deploys[name] = "PROD_B"
return record
def _advance_service(self, engineer_name: str):
"""One successful solo deploy β service progresses by one round."""
svc = self._state.services[engineer_name]
if svc.completed:
return
svc.fix_rounds_remaining = max(0, svc.fix_rounds_remaining - 1)
if svc.fix_rounds_remaining == 0:
svc.completed = True
svc.completion_round = self._state.current_round + 1
svc.completed_on_time = svc.completion_round <= svc.deadline_round
def _finalize_service_status(self):
"""At episode end, mark uncompleted services with completed=False but correct flags."""
for svc in self._state.services.values():
if not svc.completed:
svc.completed_on_time = False
def _build_observation_for_trained(self) -> Observation:
"""Return the observation visible to the trained engineer."""
me = self._state.trained_engineer
my_service = self._state.services[me]
peer_progress = {}
for name, svc in self._state.services.items():
if name == me:
continue
peer_progress[name] = {
"service": svc.name,
"rounds_remaining": svc.fix_rounds_remaining,
"deadline": svc.deadline_round,
"completed": svc.completed,
}
return Observation(
current_round=self._state.current_round,
total_rounds=self._state.total_rounds,
my_service=my_service,
my_engineer_name=me,
peer_progress=peer_progress,
history=list(self._state.history),
done=self._state.done,
)
def _build_info(self) -> dict:
"""Extra info for logging β includes continuous reward breakdown if episode ended."""
info = {
"round": self._state.current_round,
"done": self._state.done,
}
if self._last_rewards is not None:
info["reward_breakdown"] = {
"r1_service_restored": self._last_rewards.r1,
"r2_site_uptime": self._last_rewards.r2,
"r3_clean_deploys": self._last_rewards.r3,
"r4_yield_to_critical": self._last_rewards.r4,
"weighted_sum": self._last_rewards.weighted_sum,
"binary_grpo_reward": self._last_rewards.binary_grpo_reward,
}
info["episode_summary"] = self._state.episode_summary()
return info
|