File size: 12,609 Bytes
b1eddd6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
DebateArenaEnv — True Multi-Agent Runner.

Two genuinely independent agent objects take turns against the SAME
environment server over HTTP. Neither agent sees the other's internal
state — they only see the shared observation returned by /step.

Architecture
------------
  ProposerAgent   ──┐
                    ├──► DebateArenaEnv HTTP server (port 8000)
  ChallengerAgent ──┘

Each agent:
  - Has its own strategy (scripted here; replace with LLM calls on-site)
  - Only knows its own role's allowed tools
  - Sees only the shared observation (no access to opponent's reasoning)
  - Scores are computed independently by the Judge (rubric)

This is the file to replace scripted agents with real LLM agents:
  agent.decide(observation) → AgentAction
"""

from __future__ import annotations

import httpx
import json
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional


ENV_URL = "http://localhost:8000"


# ──────────────────────────────────────────────────────────────────────────────
# Agent base class
# ──────────────────────────────────────────────────────────────────────────────

@dataclass
class AgentMemory:
    """Each agent maintains its own private memory — not shared with opponent."""
    role: str
    topic_claim: str = ""
    verified_facts: Dict[str, bool] = field(default_factory=dict)
    my_arguments: List[str] = field(default_factory=list)
    opponent_arguments: List[str] = field(default_factory=list)
    round: int = 0


class BaseAgent:
    """
    Abstract debate agent.
    Override `decide()` to plug in a real LLM.
    """

    def __init__(self, role: str, name: str) -> None:
        self.role = role          # "proposer" | "challenger"
        self.name = name
        self.memory = AgentMemory(role=role)

    def observe(self, observation: Dict[str, Any]) -> None:
        """Update private memory from shared observation."""
        self.memory.round = observation.get("round", 0)
        topic = observation.get("topic", {})
        if topic:
            self.memory.topic_claim = topic.get("claim", "")

    def decide(self, observation: Dict[str, Any]) -> Dict[str, Any]:
        """
        Return the next tool call as {"tool": ..., "params": {...}}.
        Override this with an LLM call for real agents.
        """
        raise NotImplementedError


# ──────────────────────────────────────────────────────────────────────────────
# Scripted agents  (replace decide() with LLM on-site)
# ──────────────────────────────────────────────────────────────────────────────

class ProposerAgent(BaseAgent):
    """
    Scripted Proposer — uses the structured optimal strategy.
    On-site: replace decide() with GPT-4o / fine-tuned Llama call.
    """

    def __init__(self, topic_facts: Dict[str, bool], topic_keywords: List[str]) -> None:
        super().__init__(role="proposer", name="ProposerAgent")
        self._true_facts = [k for k, v in topic_facts.items() if v is True]
        self._keywords = topic_keywords
        self._phase = "verify"     # internal FSM: verify → argue → update → concede → close

    def decide(self, observation: Dict[str, Any]) -> Dict[str, Any]:
        self.observe(observation)
        phase = self._phase

        if phase == "verify":
            self._phase = "argue"
            fact = self._true_facts[0] if self._true_facts else "EU_AI_Act_mandates_labelling"
            self.memory.verified_facts[fact] = True
            return {
                "tool": "verify_fact",
                "params": {"fact_key": fact, "role": "proposer"},
            }

        if phase == "argue":
            self._phase = "update"
            kw = self._keywords[0] if self._keywords else "the evidence"
            arg = (
                f"The motion is correct because evidence shows {kw}. "
                f"Studies show this is supported by {self._true_facts[0]}. "
                f"Therefore the conclusion follows logically. "
                f"As a result, this policy is necessary."
            )
            self.memory.my_arguments.append(arg)
            return {
                "tool": "submit_argument",
                "params": {
                    "argument": arg,
                    "facts_cited": self._true_facts[:2],
                },
            }

        if phase == "update":
            self._phase = "concede"
            return {
                "tool": "refine_position",
                "params": {
                    "refined_claim": "A phased, audited rollout is more defensible.",
                    "reason": "The Challenger raised valid implementation concerns.",
                },
            }

        if phase == "concede":
            self._phase = "close"
            return {
                "tool": "concede_point",
                "params": {
                    "sub_point": "Enforcement mechanisms are genuinely complex.",
                    "maintain_main": "The core obligation to label remains justified.",
                },
            }

        # close
        closing = (
            "Because the evidence is clear and studies show the need, "
            "therefore this motion stands. As a result, we urge adoption."
        )
        return {
            "tool": "end_debate",
            "params": {"closing_statement": closing, "role": "proposer"},
        }


class ChallengerAgent(BaseAgent):
    """
    Scripted Challenger — rebuts with structured counter-evidence.
    On-site: replace decide() with a separate LLM instance.
    """

    def __init__(self, topic_keywords: List[str]) -> None:
        super().__init__(role="challenger", name="ChallengerAgent")
        self._keywords = topic_keywords
        self._has_rebutted = False

    def decide(self, observation: Dict[str, Any]) -> Dict[str, Any]:
        self.observe(observation)

        # Only rebut once; then wait for proposer to close
        if not self._has_rebutted:
            self._has_rebutted = True
            kw = self._keywords[0] if self._keywords else "the primary issue"
            rebuttal = (
                f"The Proposer makes a valid point on {kw}, "
                f"however this means that implementation remains complex. "
                f"Consequently, a phased approach would be better. "
                f"Because evidence shows that rapid mandates often fail."
            )
            self.memory.opponent_arguments.append(rebuttal)
            return {
                "tool": "submit_rebuttal",
                "params": {
                    "rebuttal": rebuttal,
                    "facts_cited": [],
                    "expose_fallacy": None,
                },
            }

        # Challenger done — let Proposer close
        return {}   # no action this turn


# ──────────────────────────────────────────────────────────────────────────────
# Multi-agent runner
# ──────────────────────────────────────────────────────────────────────────────

def call_env(tool: str, params: Dict[str, Any]) -> Dict[str, Any]:
    r = httpx.post(f"{ENV_URL}/step",
                   json={"tool": tool, "params": params},
                   timeout=10)
    r.raise_for_status()
    return r.json()


def run_multiagent_episode(topic_id: str) -> Dict[str, Any]:
    """
    Run one full episode with two independent agents taking turns.
    Returns episode summary with reward and rubric breakdown.
    """
    # ── Reset shared environment ──────────────────────────────────────────────
    r = httpx.post(f"{ENV_URL}/reset", json={"topic_id": topic_id}, timeout=10)
    r.raise_for_status()
    obs = r.json()["observation"]

    topic = obs["topic"]
    max_rounds = topic["max_rounds"]

    print(f"\n{'═'*70}")
    print(f"  DEBATE: {topic_id.upper()} — {topic['domain']}")
    print(f"  Motion: {topic['claim'][:65]}…")
    print(f"{'═'*70}")

    # Import topic data for agent initialisation
    from server.tasks import TOPIC_BANK
    topic_obj = next(t for t in TOPIC_BANK if t.topic_id == topic_id)

    # ── Instantiate two INDEPENDENT agents ────────────────────────────────────
    proposer   = ProposerAgent(topic_obj.known_facts, topic_obj.evidence_keywords)
    challenger = ChallengerAgent(topic_obj.evidence_keywords)

    print(f"\n  Agents:  [{proposer.name}]  vs  [{challenger.name}]")
    print(f"  Roles:    proposer              challenger\n")

    turn = 0
    last_result: Dict[str, Any] = {}

    # ── Turn-based loop — agents alternate ────────────────────────────────────
    # Order: Proposer verifies → Proposer argues → Challenger rebuts →
    #        Proposer updates → Proposer concedes → Proposer closes
    turn_order = [proposer, proposer, challenger, proposer, proposer, proposer]

    for agent in turn_order:
        action = agent.decide(obs)
        if not action:
            continue    # agent passed this turn

        tool   = action["tool"]
        params = action["params"]

        result = call_env(tool, params)
        obs    = result.get("observation", obs)
        reward = result.get("reward", 0)
        done   = result.get("done", False)

        tag = f"[{agent.name:<16}] {tool:<20}"
        print(f"  Turn {turn+1:02d}  {tag}  reward={reward:.3f}")
        if result.get("result"):
            print(f"           └─ {result['result'][:80]}")
        turn += 1

        if done:
            last_result = result
            break

    # ── Episode summary ───────────────────────────────────────────────────────
    final_obs = last_result.get("observation", obs)
    winner     = final_obs.get("winner", "?")
    final_reward = last_result.get("reward", 0)

    print(f"\n{'─'*70}")
    print(f"  Winner: {winner.upper()}   |   Final reward: {final_reward:.3f}")
    print(f"\n  Rubric breakdown:")
    for line in last_result.get("breakdown", "").split("\n"):
        if line.strip():
            print(f"    {line}")

    return {
        "topic_id": topic_id,
        "winner": winner,
        "reward": final_reward,
        "proposer": proposer.name,
        "challenger": challenger.name,
        "turns": turn,
    }


# ──────────────────────────────────────────────────────────────────────────────
# Entry point
# ──────────────────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    print("DebateArenaEnv — Multi-Agent Runner")
    print("Two independent agents, one shared environment server")
    print(f"Environment: {ENV_URL}\n")

    summaries = []
    for topic_id in ["easy", "medium", "hard"]:
        summary = run_multiagent_episode(topic_id)
        summaries.append(summary)

    print(f"\n{'═'*70}")
    print("  MULTI-AGENT EPISODE SUMMARY")
    print(f"{'═'*70}")
    print(f"  {'Topic':<10} {'Winner':<15} {'Reward':>8}  {'Turns':>6}")
    print(f"  {'─'*50}")
    for s in summaries:
        print(f"  {s['topic_id']:<10} {s['winner']:<15} {s['reward']:>8.3f}  {s['turns']:>6}")
    print(f"{'═'*70}\n")