Spaces:
Sleeping
Sleeping
File size: 13,935 Bytes
dfbd16e | 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 | # server/multi_agent.py
"""
Multi-Agent Comparison Engine.
Runs multiple agent configurations against the SAME task variant
and produces a side-by-side comparison report.
Agent configurations:
- Deterministic (rule-based, no LLM) β baseline
- Test-first (forces reading tests before anything)
- Search-first (forces search_code before reads)
- LLM-based (if HF_TOKEN provided)
This is the key feature that answers: "Which agent strategy wins?"
"""
import time
import copy
from typing import List, Dict, Any, Optional, Callable
from dataclasses import dataclass, field
@dataclass
class AgentRunResult:
"""Result of one agent configuration running one episode."""
agent_name: str
task: str
variant_id: str
final_score: float
total_steps: int
cumulative_reward: float
duration_seconds: float
action_sequence: List[str]
files_read: List[str]
files_written: List[str]
strategy: str # Detected strategy label
strategy_score: float
failure_type: str
reliability_index: float
step_timeline: List[dict]
def to_dict(self) -> dict:
return {
"agent_name": self.agent_name,
"task": self.task,
"variant_id": self.variant_id,
"final_score": round(self.final_score, 3),
"total_steps": self.total_steps,
"cumulative_reward": round(self.cumulative_reward, 3),
"duration_seconds": round(self.duration_seconds, 2),
"action_sequence": self.action_sequence,
"files_read": self.files_read,
"files_written": self.files_written,
"strategy": self.strategy,
"strategy_score": round(self.strategy_score, 3),
"failure_type": self.failure_type,
"reliability_index": round(self.reliability_index, 3),
"step_timeline": self.step_timeline,
}
@dataclass
class ComparisonReport:
"""Side-by-side comparison of multiple agent configurations."""
task: str
variant_id: str
runs: List[AgentRunResult] = field(default_factory=list)
def to_dict(self) -> dict:
if not self.runs:
return {"error": "No runs to compare"}
# Rank by score then steps
ranked = sorted(self.runs, key=lambda r: (-r.final_score, r.total_steps))
winner = ranked[0]
return {
"task": self.task,
"variant_id": self.variant_id,
"winner": winner.agent_name,
"winner_score": winner.final_score,
"summary_table": [
{
"rank": i + 1,
"agent": r.agent_name,
"score": round(r.final_score, 3),
"steps": r.total_steps,
"reward": round(r.cumulative_reward, 3),
"strategy": r.strategy,
"failure": r.failure_type,
"reliability": round(r.reliability_index, 3),
}
for i, r in enumerate(ranked)
],
"detailed_runs": [r.to_dict() for r in self.runs],
"insights": self._generate_insights(ranked),
}
def _generate_insights(self, ranked: List[AgentRunResult]) -> List[str]:
insights = []
if len(ranked) < 2:
return insights
best = ranked[0]
worst = ranked[-1]
if best.final_score > worst.final_score + 0.2:
insights.append(
f"'{best.agent_name}' significantly outperformed '{worst.agent_name}' "
f"({best.final_score:.2f} vs {worst.final_score:.2f})"
)
step_diffs = [(r.agent_name, r.total_steps) for r in ranked]
most_efficient = min(ranked, key=lambda r: r.total_steps if r.final_score >= 0.5 else float('inf'))
if most_efficient.final_score >= 0.5:
insights.append(
f"Most step-efficient successful agent: '{most_efficient.agent_name}' "
f"({most_efficient.total_steps} steps)"
)
strategies = [r.strategy for r in ranked]
if len(set(strategies)) > 1:
insights.append(
f"Strategy variance observed: {set(strategies)} β "
f"'{best.agent_name}' used {best.strategy} which proved most effective."
)
return insights
class MultiAgentComparison:
"""
Runs multiple deterministic agent strategies against the same environment.
Usage (in-process, no LLM required):
from server.environment import CodebaseNavEnvironment
from server.models import RepoAction
env = CodebaseNavEnvironment()
engine = MultiAgentComparison()
report = engine.compare(env, task="task1")
"""
# ββ Built-in agent strategies βββββββββββββββββββββββββββββββββββββββββββββ
@staticmethod
def _agent_test_first(obs: dict, step: int, context: dict) -> dict:
"""Strategy: Read tests before any source file."""
tree = obs.get("repo_tree", [])
files_read = set(obs.get("files_read", []))
test_files = sorted([f for f in tree if f.startswith("tests/")])
src_files = sorted([f for f in tree if f.startswith("src/") and f.endswith(".py")])
spec_files = sorted([f for f in tree if f.endswith(".md")])
# Phase 1: Tests first
for tf in test_files:
if tf not in files_read:
return {"action_type": "read_file", "path": tf}
# Phase 2: Source files
for sf in src_files:
if sf not in files_read:
return {"action_type": "read_file", "path": sf}
# Phase 3: Run tests
if test_files and context.get("tests_run", 0) == 0:
context["tests_run"] = 1
return {"action_type": "run_tests", "path": test_files[0]}
return {"action_type": "submit"}
@staticmethod
def _agent_search_first(obs: dict, step: int, context: dict) -> dict:
"""Strategy: Use search_code to locate the bug before reading."""
tree = obs.get("repo_tree", [])
files_read = set(obs.get("files_read", []))
failing = obs.get("failing_tests", [])
# Step 1: search for the failing test function name
if step == 1 and failing:
fn_name = failing[0].split(".")[-1] if failing else "bug"
context["searched"] = True
return {"action_type": "search_code", "query": fn_name}
# Step 2: Read files based on search
test_files = sorted([f for f in tree if f.startswith("tests/")])
src_files = sorted([f for f in tree if f.startswith("src/") and f.endswith(".py")])
for tf in test_files:
if tf not in files_read:
return {"action_type": "read_file", "path": tf}
for sf in src_files:
if sf not in files_read:
return {"action_type": "read_file", "path": sf}
if test_files and context.get("tests_run", 0) == 0:
context["tests_run"] = 1
return {"action_type": "run_tests", "path": test_files[0]}
return {"action_type": "submit"}
@staticmethod
def _agent_minimal(obs: dict, step: int, context: dict) -> dict:
"""Strategy: Minimal effort β read one file, submit immediately."""
tree = obs.get("repo_tree", [])
files_read = set(obs.get("files_read", []))
src_files = [f for f in tree if f.startswith("src/") and f.endswith(".py")]
if src_files and not files_read:
return {"action_type": "read_file", "path": src_files[0]}
return {"action_type": "submit"}
@staticmethod
def _agent_exhaustive(obs: dict, step: int, context: dict) -> dict:
"""Strategy: Read everything, run tests, then submit."""
tree = obs.get("repo_tree", [])
files_read = set(obs.get("files_read", []))
all_readable = [f for f in tree if f.endswith(".py") or f.endswith(".md")]
for f in all_readable:
if f not in files_read:
return {"action_type": "read_file", "path": f}
test_files = [f for f in tree if f.startswith("tests/")]
if test_files and context.get("tests_run", 0) == 0:
context["tests_run"] = 1
return {"action_type": "run_tests", "path": test_files[0]}
if test_files and context.get("tests_run2", 0) == 0:
context["tests_run2"] = 1
return {"action_type": "run_tests"}
return {"action_type": "submit"}
AGENT_CONFIGS = {
"test-first": _agent_test_first.__func__,
"search-first": _agent_search_first.__func__,
"minimal": _agent_minimal.__func__,
"exhaustive": _agent_exhaustive.__func__,
}
def compare(
self,
env, # CodebaseNavEnvironment instance
task: str = "task1",
agents: Optional[List[str]] = None,
shared_variant: Optional[str] = None,
) -> ComparisonReport:
"""
Run all (or selected) agents against the same task and compare.
The environment is reset to the same variant for each agent.
"""
from server.models import RepoAction
from server.strategy_detector import StrategyDetector
from server.failure_classifier import FailureClassifier
from server.advanced_metrics import AdvancedMetricsEngine
agent_names = agents or list(self.AGENT_CONFIGS.keys())
strategy_detector = StrategyDetector()
failure_classifier = FailureClassifier()
metrics_engine = AdvancedMetricsEngine()
runs: List[AgentRunResult] = []
variant_id = None
for agent_name in agent_names:
agent_fn = self.AGENT_CONFIGS.get(agent_name)
if not agent_fn:
continue
# Reset environment
reset_result = env.reset(task=task)
obs = reset_result.observation
variant_id = reset_result.info.get("variant_id", "unknown")
context = {}
start = time.time()
max_steps = 15
files_read = []
files_written = []
cumulative_reward = 0.0
action_sequence = []
step_timeline = []
obs_dict = obs.model_dump()
for step_num in range(1, max_steps + 1):
if env.done:
break
action_dict = agent_fn(obs_dict, step_num, context)
action = RepoAction(
action_type=action_dict.get("action_type", "submit"),
path=action_dict.get("path"),
query=action_dict.get("query"),
content=action_dict.get("content"),
)
result = env.step(action)
obs = result.observation
obs_dict = obs.model_dump()
cumulative_reward += result.reward
action_sequence.append(action.action_type)
if action.path and action.action_type == "read_file":
files_read.append(action.path)
if action.path and action.action_type == "write_file":
files_written.append(action.path)
step_timeline.append({
"step": step_num,
"action": action.action_type,
"path": action.path,
"reward": round(result.reward, 3),
})
if result.done:
break
# Force submit if not done
if not env.done:
result = env.step(RepoAction(action_type="submit"))
cumulative_reward += result.reward
action_sequence.append("submit")
duration = time.time() - start
final_score = env.final_score
# Get trajectory for analysis
trajectory = env.get_trajectory()
traj_steps = trajectory.get("steps", []) if trajectory else []
variant_meta = {}
if env.variant:
variant_meta = env.variant.meta
# Detect strategy
strategy_report = strategy_detector.detect(
traj_steps, task, variant_meta, files_read, final_score
)
# Classify failure
failure_report = failure_classifier.classify(
episode_id=trajectory.get("episode_id", "") if trajectory else "",
task=task,
trajectory_steps=traj_steps,
variant_meta=variant_meta,
files_read=files_read,
files_written=files_written,
final_score=final_score,
)
# Advanced metrics
adv_metrics = metrics_engine.compute(
traj_steps, variant_meta, final_score, files_read, files_written
)
runs.append(AgentRunResult(
agent_name=agent_name,
task=task,
variant_id=variant_id or "unknown",
final_score=final_score,
total_steps=len(action_sequence),
cumulative_reward=cumulative_reward,
duration_seconds=duration,
action_sequence=action_sequence,
files_read=files_read,
files_written=files_written,
strategy=strategy_report.strategy,
strategy_score=strategy_report.score,
failure_type=failure_report.primary_failure,
reliability_index=adv_metrics.reliability_index,
step_timeline=step_timeline,
))
return ComparisonReport(
task=task,
variant_id=variant_id or "unknown",
runs=runs,
)
|