File size: 21,514 Bytes
698f4d8 81b4b70 698f4d8 2568517 698f4d8 80b3b2e 698f4d8 2568517 698f4d8 2568517 698f4d8 2568517 698f4d8 2568517 698f4d8 80b3b2e 698f4d8 2568517 698f4d8 2568517 698f4d8 9d82eed 80b3b2e 9d82eed 698f4d8 80b3b2e 698f4d8 2568517 698f4d8 2568517 698f4d8 2568517 698f4d8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 | """
Parlay MCP tools — universal protocol (works with any MCP client).
8 tools covering negotiation lifecycle, game state, and leaderboard.
Tool names intentionally avoid exactly: reset, step, state, close — those
identifiers are reserved by OpenEnv / client connection helpers.
"""
import logging
import uuid
from typing import Optional
import numpy as np
from fastmcp import FastMCP
from parlay_env.models import (
PersonaType, TacticalMove, BeliefState, HiddenState, ParlayState,
)
from parlay_env.game_theory import compute_zopa, compute_nash_bargaining_solution
from parlay_env.grader import grade_episode
from game.scenarios import SCENARIOS, get_scenario
from game.tactical_cards import TACTICAL_CARDS, draw_hand, get_card
from game.leaderboard import Leaderboard
from agent.personas import PERSONAS, build_system_prompt
from agent.gemini_client import MODEL_ID_DEMO, call_gemini, validate_ai_offer_direction
from agent.tom_tracker import ToMTracker
logger = logging.getLogger(__name__)
mcp = FastMCP(
"Parlay Negotiation Environment",
description=(
"An RL negotiation environment. Train agents, play scenarios, "
"access game state and leaderboards via MCP tools. "
"Any MCP-compatible client can use these tools."
),
)
# In-memory session store for MCP sessions
_sessions: dict[str, dict] = {}
_leaderboard = Leaderboard()
CP_START = 100
CP_REGEN = 5
MAX_TURNS = 20
# CP costs per tactical move
_CP_COSTS: dict[TacticalMove, int] = {
TacticalMove.ANCHOR_HIGH: 0,
TacticalMove.BATNA_REVEAL: 20,
TacticalMove.SILENCE: 5,
}
def _get_hidden_state(scenario_id: str, seed: int = 42) -> HiddenState:
"""Build a HiddenState for the given scenario."""
rng = np.random.default_rng(seed)
scenario = get_scenario(scenario_id)
noise = float(rng.uniform(0.95, 1.05))
return HiddenState(
budget_ceiling=round(scenario.batna_buyer * noise, 2),
walk_away_price=round(scenario.batna_seller * noise, 2),
urgency_score=float(np.clip(0.5 + rng.uniform(-0.15, 0.15), 0.0, 1.0)),
has_alternative=scenario.id in ("saas_enterprise", "acquisition_term_sheet"),
persona_drifted=False,
)
@mcp.tool()
async def start_negotiation(
scenario_id: str,
persona: str,
player_name: str = "Agent",
) -> dict:
"""
Start a new negotiation episode.
Args:
scenario_id: One of: saas_enterprise, hiring_package, acquisition_term_sheet
persona: One of: shark, diplomat, veteran
player_name: Display name for the leaderboard (default: "Agent")
Returns:
session_id: Unique session identifier for subsequent calls.
observation: Initial game state including ZOPA, Nash point, and opening message.
scenario: Scenario context (title, description, currency, unit).
available_cards: List of tactical cards available to the player.
opening_message: The AI opponent's opening statement.
"""
valid_scenarios = list(SCENARIOS.keys())
if scenario_id not in valid_scenarios:
return {"error": f"Invalid scenario_id. Valid options: {valid_scenarios}"}
try:
persona_type = PersonaType(persona)
except ValueError:
return {"error": f"Invalid persona. Valid options: {[p.value for p in PersonaType]}"}
scenario = get_scenario(scenario_id)
session_id = str(uuid.uuid4())
hidden = _get_hidden_state(scenario_id, seed=hash(session_id) % 10000)
initial_belief = BeliefState(
est_budget=hidden.budget_ceiling * 0.80,
est_walk_away=hidden.walk_away_price * 1.15,
est_urgency=0.50,
est_has_alternative=False,
confidence=0.30,
)
tom = ToMTracker(initial_belief, persona_type)
hand = draw_hand(3, rng_seed=hash(session_id) % 9999)
system_prompt = build_system_prompt(
persona=persona_type,
scenario_id=scenario_id,
scenario_title=scenario.title,
scenario_description=scenario.description,
batna=hidden.walk_away_price,
budget=hidden.budget_ceiling,
urgency=hidden.urgency_score,
)
persona_cfg = PERSONAS[persona_type]
opening_message = persona_cfg.opening_line
zopa = compute_zopa(hidden.budget_ceiling, hidden.walk_away_price)
nash = compute_nash_bargaining_solution(hidden.budget_ceiling, hidden.walk_away_price)
_sessions[session_id] = {
"session_id": session_id,
"player_name": player_name,
"scenario_id": scenario_id,
"persona": persona,
"persona_type": persona_type,
"hidden": hidden,
"tom": tom,
"system_prompt": system_prompt,
"conversation": [{"role": "opponent", "content": opening_message, "turn": 0}],
"offer_history": [],
"step_count": 0,
"cumulative_reward": 0.0,
"credibility_points": CP_START,
"done": False,
"hand": [m.value for m in hand],
"drift_adapted": False,
"drift_turn": None,
}
logger.info(f"MCP start_negotiation: session={session_id}, scenario={scenario_id}, persona={persona}")
return {
"session_id": session_id,
"observation": {
"step_count": 0,
"zopa_lower": zopa[0] if zopa else 0,
"zopa_upper": zopa[1] if zopa else 0,
"nash_point": nash,
"credibility_points": CP_START,
"tension_score": 10.0,
"belief_state": initial_belief.model_dump(),
},
"scenario": {
"title": scenario.title,
"description": scenario.description,
"currency": scenario.currency,
"unit": scenario.unit,
"anchor_seller": scenario.anchor_seller,
"anchor_buyer": scenario.anchor_buyer,
},
"available_cards": [
{
"move": m,
"name": get_card(m).name,
"cp_cost": get_card(m).cp_cost,
"description": get_card(m).description,
}
for m in _sessions[session_id]["hand"]
],
"opening_message": opening_message,
}
@mcp.tool()
async def make_offer(
session_id: str,
amount: float,
message: str,
tactical_move: Optional[str] = None,
) -> dict:
"""
Make a structured offer in the negotiation.
Args:
session_id: Session ID from start_negotiation.
amount: Offer amount in the scenario's currency.
message: Natural language message accompanying the offer.
tactical_move: Optional tactical card to play. One of:
anchor_high, batna_reveal, silence
Returns:
opponent_response: AI opponent's counter-message and offer.
updated_observation: Updated ZOPA, beliefs, tension, CP.
reward: Step reward earned this turn.
done: Whether the episode has ended.
drift_event: Description of any drift event triggered (or null).
"""
if session_id not in _sessions:
return {"error": f"Session {session_id} not found. Call start_negotiation first."}
sess = _sessions[session_id]
if sess["done"]:
return {"error": "Episode is already complete. Start a new session."}
move: Optional[TacticalMove] = None
if tactical_move:
try:
move = TacticalMove(tactical_move)
except ValueError:
return {"error": f"Invalid tactical_move. Valid: {[m.value for m in TacticalMove]}"}
cost = _CP_COSTS.get(move, 0)
current_cp = sess["credibility_points"]
if current_cp < cost:
return {"error": f"Insufficient credibility points. Need {cost}, have {current_cp}."}
new_cp = min(CP_START, current_cp + CP_REGEN - cost)
# Check for drift events
turn = sess["step_count"]
scenario = get_scenario(sess["scenario_id"])
drift_event_desc: Optional[str] = None
for event in scenario.drift_events:
if event.trigger_turn == turn:
drift_event_desc = event.event
sess["drift_turn"] = turn
sess["tom"].drift_event(event.effect_on_urgency, event.effect_on_has_alternative)
logger.info(f"Drift event triggered: {event.event}")
break
# Build Gemini message history (last 10 messages for context window)
gemini_messages = []
for msg in sess["conversation"][-10:]:
role = "user" if msg["role"] == "player" else "model"
gemini_messages.append({"role": role, "parts": [msg["content"]]})
player_text = f"Player offer: {amount:,.0f}. Message: {message}"
if tactical_move:
player_text += f" [Tactical move: {tactical_move}]"
gemini_messages.append({"role": "user", "parts": [player_text]})
opponent_resp = await call_gemini(
sess["system_prompt"],
gemini_messages,
model=MODEL_ID_DEMO,
scenario_id=sess["scenario_id"],
)
opponent_utterance: str = opponent_resp.get("utterance", "I'll need to consider that.")
raw_opp = opponent_resp.get("offer_amount")
opponent_offer: Optional[float] = None
if raw_opp is not None:
try:
opponent_offer = float(raw_opp)
except (TypeError, ValueError):
opponent_offer = None
if opponent_offer is not None:
opponent_offer = validate_ai_offer_direction(
opponent_offer, float(amount), sess["scenario_id"]
)
opponent_move: Optional[str] = opponent_resp.get("tactical_move")
sess["conversation"].append({
"role": "player", "content": message,
"offer": amount, "move": tactical_move, "turn": turn + 1,
})
sess["conversation"].append({
"role": "opponent", "content": opponent_utterance,
"offer": opponent_offer, "turn": turn + 1,
})
parsed_opp_move: Optional[TacticalMove] = None
if opponent_move:
try:
parsed_opp_move = TacticalMove(opponent_move)
except ValueError:
pass
updated_belief = sess["tom"].update(
observed_offer=opponent_offer,
observed_move=parsed_opp_move,
utterance=opponent_utterance,
turn=turn,
)
# Check drift adaptation within 2 turns
if sess["drift_turn"] is not None and not sess["drift_adapted"]:
if turn <= sess["drift_turn"] + 2:
adaptation_signals = ["understand", "noted", "given that", "considering", "account"]
if any(s in message.lower() for s in adaptation_signals):
sess["drift_adapted"] = True
sess["offer_history"].append(amount)
sess["step_count"] += 1
sess["credibility_points"] = new_cp
step_reward = 5.0 * (1.0 - sess["tom"].accuracy_against(sess["hidden"]))
sess["cumulative_reward"] += step_reward
done = sess["step_count"] >= MAX_TURNS
sess["done"] = done
hidden = sess["hidden"]
zopa = compute_zopa(hidden.budget_ceiling, hidden.walk_away_price)
nash = compute_nash_bargaining_solution(hidden.budget_ceiling, hidden.walk_away_price)
tension = min(100.0, 20.0 + (sess["step_count"] / MAX_TURNS) * 80.0)
_sessions[session_id] = sess
logger.debug(f"MCP make_offer: session={session_id}, turn={sess['step_count']}, amount={amount}")
return {
"opponent_response": {
"utterance": opponent_utterance,
"offer": opponent_offer,
"tactical_move": opponent_move,
},
"updated_observation": {
"step_count": sess["step_count"],
"zopa_lower": zopa[0] if zopa else 0,
"zopa_upper": zopa[1] if zopa else 0,
"nash_point": nash,
"tension_score": tension,
"belief_state": updated_belief.model_dump(),
"credibility_points": new_cp,
},
"reward": step_reward,
"cumulative_reward": sess["cumulative_reward"],
"done": done,
"drift_event": drift_event_desc,
}
@mcp.tool()
async def get_game_state(session_id: str) -> dict:
"""
Get the full current game state for a session.
Args:
session_id: Session ID from start_negotiation.
Returns:
Full game state including beliefs, offer history, ZOPA data,
CP balance, conversation history, and current leaderboard rank.
"""
if session_id not in _sessions:
return {"error": f"Session {session_id} not found."}
sess = _sessions[session_id]
hidden = sess["hidden"]
zopa = compute_zopa(hidden.budget_ceiling, hidden.walk_away_price)
nash = compute_nash_bargaining_solution(hidden.budget_ceiling, hidden.walk_away_price)
current_belief = sess["tom"].current_belief
rank = await _leaderboard.get_rank(sess["player_name"], sess["scenario_id"])
return {
"session_id": session_id,
"player_name": sess["player_name"],
"scenario_id": sess["scenario_id"],
"persona": sess["persona"],
"step_count": sess["step_count"],
"done": sess["done"],
"offer_history": sess["offer_history"],
"zopa": {"lower": zopa[0] if zopa else 0, "upper": zopa[1] if zopa else 0},
"nash_point": nash,
"belief_state": current_belief.model_dump(),
"belief_history_count": len(sess["tom"].history),
"credibility_points": sess["credibility_points"],
"cumulative_reward": sess["cumulative_reward"],
"drift_adapted": sess["drift_adapted"],
"bluffs_detected": sess["tom"].bluffs_detected,
"leaderboard_rank": rank,
"hand": sess["hand"],
}
@mcp.tool()
async def accept_deal(session_id: str) -> dict:
"""
Accept the current offer and close the negotiation.
Args:
session_id: Session ID from start_negotiation.
Returns:
final_reward: Complete reward breakdown (step + terminal).
deal_efficiency: Fraction of ZOPA captured [0, 1].
nash_comparison: How the deal compares to Nash Bargaining Solution.
episode_summary: Full grade breakdown.
"""
if session_id not in _sessions:
return {"error": f"Session {session_id} not found."}
sess = _sessions[session_id]
if sess["done"]:
return {"error": "Episode already concluded."}
if not sess["offer_history"]:
return {"error": "No offer has been made yet. Make an offer before accepting."}
final_price = sess["offer_history"][-1]
hidden = sess["hidden"]
state = ParlayState(
session_id=session_id,
scenario_id=sess["scenario_id"],
persona=PersonaType(sess["persona"]),
step_count=sess["step_count"],
cumulative_reward=sess["cumulative_reward"],
hidden_state=hidden,
belief_history=sess["tom"].history,
offer_history=sess["offer_history"],
drift_events_fired=1 if sess["drift_turn"] is not None else 0,
episode_done=True,
termination_reason="deal_accepted",
credibility_points=sess["credibility_points"],
)
grade = grade_episode(
state,
final_price=final_price,
t_close=sess["step_count"],
t_max=MAX_TURNS,
drift_adapted=sess["drift_adapted"],
bluffs_caught=sess["tom"].bluffs_detected,
)
await _leaderboard.record_result(
player_name=sess["player_name"],
scenario_id=sess["scenario_id"],
persona=sess["persona"],
total_reward=grade.total_reward,
deal_efficiency=grade.deal_efficiency,
acts_completed=1,
deal_closed=True,
)
sess["done"] = True
_sessions[session_id] = sess
nash = compute_nash_bargaining_solution(hidden.budget_ceiling, hidden.walk_away_price)
zopa = compute_zopa(hidden.budget_ceiling, hidden.walk_away_price)
zopa_width = (zopa[1] - zopa[0]) if zopa else 1.0
logger.info(
f"MCP accept_deal: session={session_id}, "
f"price={final_price:,.0f}, efficiency={grade.deal_efficiency:.3f}"
)
return {
"final_price": final_price,
"final_reward": grade.total_reward,
"deal_efficiency": grade.deal_efficiency,
"nash_comparison": {
"nash_point": nash,
"your_deal": final_price,
"vs_nash_pct": round((final_price - nash) / max(zopa_width, 1) * 100, 1),
},
"episode_summary": {
"total_reward": grade.total_reward,
"deal_efficiency": grade.deal_efficiency,
"tom_accuracy_avg": grade.tom_accuracy_avg,
"bluffs_caught": grade.bluffs_caught,
"drift_adapted": grade.drift_adapted,
},
}
@mcp.tool()
async def walk_away(session_id: str) -> dict:
"""
Walk away from the negotiation without a deal.
Args:
session_id: Session ID from start_negotiation.
Returns:
episode_summary: Final metrics (no deal recorded on leaderboard).
counterfactual: What the optimal deal would have been.
reward: Partial reward earned (penalty applied for no deal).
"""
if session_id not in _sessions:
return {"error": f"Session {session_id} not found."}
sess = _sessions[session_id]
if sess["done"]:
return {"error": "Episode already concluded."}
hidden = sess["hidden"]
state = ParlayState(
session_id=session_id,
scenario_id=sess["scenario_id"],
persona=PersonaType(sess["persona"]),
step_count=sess["step_count"],
cumulative_reward=sess["cumulative_reward"],
hidden_state=hidden,
belief_history=sess["tom"].history,
offer_history=sess["offer_history"],
drift_events_fired=1 if sess["drift_turn"] is not None else 0,
episode_done=True,
termination_reason="walk_away",
credibility_points=sess["credibility_points"],
)
grade = grade_episode(state, final_price=None, t_max=MAX_TURNS)
nash = compute_nash_bargaining_solution(hidden.budget_ceiling, hidden.walk_away_price)
zopa = compute_zopa(hidden.budget_ceiling, hidden.walk_away_price)
sess["done"] = True
_sessions[session_id] = sess
logger.info(f"MCP walk_away: session={session_id}, partial_reward={grade.total_reward:.2f}")
return {
"result": "walk_away",
"reward": grade.total_reward,
"episode_summary": {
"total_reward": grade.total_reward,
"deal_efficiency": 0.0,
},
"counterfactual": {
"optimal_deal": nash,
"zopa": {"lower": zopa[0] if zopa else 0, "upper": zopa[1] if zopa else 0},
"message": (
f"The Nash Bargaining Solution was {nash:,.0f}. "
f"Walking away left value on the table."
),
},
}
@mcp.tool()
async def get_leaderboard(
scenario_id: Optional[str] = None,
limit: int = 10,
) -> dict:
"""
Get the global or per-scenario leaderboard.
Args:
scenario_id: Optional. Filter to a specific scenario. Leave null for global.
limit: Number of entries to return (default: 10, max: 50).
Returns:
entries: Top leaderboard entries with player name, score, efficiency, and persona.
total_entries: Total number of entries in this leaderboard.
"""
limit = min(max(1, limit), 50)
valid_scenarios = list(SCENARIOS.keys())
if scenario_id and scenario_id not in valid_scenarios:
return {"error": f"Invalid scenario_id. Valid: {valid_scenarios}"}
entries = await _leaderboard.get_top(scenario_id=scenario_id, limit=limit)
return {
"scenario_id": scenario_id or "global",
"entries": entries,
"total_entries": len(entries),
}
@mcp.tool()
async def list_scenarios() -> dict:
"""
List all available B2B negotiation scenarios.
Returns:
scenarios: List of all scenarios with id, title, description,
ZOPA range, difficulty rating, and available drift events.
"""
return {
"scenarios": [
{
"id": s.id,
"title": s.title,
"description": s.description,
"currency": s.currency,
"unit": s.unit,
"zopa": {"lower": s.zopa[0], "upper": s.zopa[1]},
"anchor_seller": s.anchor_seller,
"anchor_buyer": s.anchor_buyer,
"difficulty": s.difficulty,
"drift_events": [
{"turn": e.trigger_turn, "event": e.event}
for e in s.drift_events
],
}
for s in SCENARIOS.values()
]
}
@mcp.tool()
async def list_personas() -> dict:
"""
List all available AI negotiator personas.
Returns:
personas: List of all personas with name, symbol, Big Five scores,
aggression, patience, bluff_rate, and tactical style summary.
"""
return {
"personas": [
{
"id": persona_type.value,
"name": cfg.name,
"symbol": cfg.symbol,
"emoji": cfg.emoji,
"big_five": cfg.big_five,
"aggression": cfg.aggression,
"patience": cfg.patience,
"bluff_rate": cfg.bluff_rate,
"tom_depth": cfg.tom_depth,
"style_summary": cfg.style[:120] + "...",
"drift_trigger": cfg.drift_trigger,
"color_var": cfg.color_var,
}
for persona_type, cfg in PERSONAS.items()
]
}
|