Spaces:
Sleeping
Sleeping
File size: 43,829 Bytes
bf9c466 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 | """play_task β LLM-driven CLI for LeanMigrate task playthrough.
Usage:
uv run python scripts/play_task.py list
uv run python scripts/play_task.py play lru_cache expression_eval
uv run python scripts/play_task.py play --all
"""
from __future__ import annotations
import datetime
import json
import os
import re
import sys
import textwrap
from pathlib import Path
from typing import Annotated, Any, Literal, Optional
from dotenv import load_dotenv
import typer
from rich.columns import Columns
from rich.console import Console
from rich.panel import Panel
from rich.rule import Rule
from rich.syntax import Syntax
from rich.table import Table
from rich.text import Text
# Load .env before anything reads os.environ (including typer envvar resolution)
load_dotenv(Path(__file__).resolve().parents[1] / ".env")
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from lean_migrate.env.grader import clamp_open_unit, oracle_result # noqa: E402
from lean_migrate.env.models import ( # noqa: E402
AnalyzeDepsAction,
InspectAction,
LeanMigrateAction,
LeanMigrateObservation,
RunTestsAction,
SubmitAction,
)
from lean_migrate.env.tasks import FunctionSpec, Task, get_task, list_tasks # noqa: E402
from lean_migrate.server.lean_migrate_environment import LeanMigrateEnvironment # noqa: E402
# ββ Pydantic schema for structured LLM output ββββββββββββββββββββββββββββββββ
from pydantic import BaseModel, Field # noqa: E402
class ActionRequest(BaseModel):
"""Structured output schema for LLM action responses."""
type: Literal["inspect", "analyze_deps", "run_tests", "submit"]
function_name: str = Field(description="Name of the function this action targets.")
candidate_code: str | None = Field(
default=None,
description="Full implementation for run_tests. Omit for other action types.",
)
lean_proof: str | None = Field(
default=None,
description="Lean proof text. Only required for proof tasks.",
)
# ββ App + console βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
app = typer.Typer(
name="play_task",
help="LLM-driven LeanMigrate task runner with rich step display.",
add_completion=False,
)
console = Console()
# ββ Rich helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _difficulty_color(d: str) -> str:
return {"easy": "green", "medium": "yellow", "hard": "red"}.get(d, "white")
def _trim(text: str, max_lines: int = 20) -> str:
lines = text.strip().splitlines()
if len(lines) <= max_lines:
return "\n".join(lines)
return (
"\n".join(lines[:max_lines])
+ f"\n[dim]β¦ ({len(lines) - max_lines} more lines)[/dim]"
)
def _split_at_semicolons(s: str) -> list[str]:
"""Depth-aware split at ';' β won't split inside brackets."""
parts: list[str] = []
depth = 0
buf: list[str] = []
for ch in s:
if ch in "([{":
depth += 1
buf.append(ch)
elif ch in ")]}":
depth -= 1
buf.append(ch)
elif ch == ";" and depth == 0:
part = "".join(buf).strip()
if part:
parts.append(part)
buf = []
else:
buf.append(ch)
last = "".join(buf).strip()
if last:
parts.append(last)
return parts
# Lean tactic keywords that start a new tactic at the top level of a `by` block.
_LEAN_TACTIC_KWS = re.compile(
r"(?<!\w)("
r"intro\b|intros\b|apply\b|refine\b|exact\b|simp\b|rfl\b|norm_num\b|"
r"decide\b|ring\b|omega\b|constructor\b|cases\b|induction\b|have\b|"
r"obtain\b|rw\s*\[|rewrite\s*\[|calc\b|show\b|suffices\b|use\b|"
r"trivial\b|tauto\b|aesop\b|linarith\b|nlinarith\b|positivity\b|"
r"field_simp\b|push_neg\b|contrapose\b|by_contra\b|by_cases\b|"
r"rcases\b|ext\b|funext\b|congr\b|subst\b|assumption\b|"
r"contradiction\b|scalar_tac\b|progress\b|next\b|all_goals\b|"
r"any_goals\b|first\b|repeat\b|done\b|skip\b"
r")"
)
def _format_lean(code: str) -> str:
"""Best-effort Lean 4 single-line β multi-line formatter."""
# Already multi-line β just normalise indentation slightly.
if "\n" in code.strip():
lines = [ln.rstrip() for ln in code.strip().splitlines()]
return "\n".join(lines)
s = code.strip()
# Match: <decl head> := by <tactics>
m = re.match(r"^(.*?)\s*:=\s*by\s+(.+)$", s, re.DOTALL)
if m:
head, tactics_str = m.group(1).strip(), m.group(2).strip()
# First split at semicolons
tactics = _split_at_semicolons(tactics_str)
# Then split each part further at tactic keyword boundaries (depth 0 only)
expanded: list[str] = []
for part in tactics:
# Find positions of tactic-keyword starts (skip inside brackets)
splits: list[int] = [0]
depth = 0
i = 0
while i < len(part):
if part[i] in "([{":
depth += 1
elif part[i] in ")]}":
depth -= 1
elif depth == 0 and i > 0 and part[i - 1] == " ":
hit = _LEAN_TACTIC_KWS.match(part, i)
if hit:
splits.append(i)
i += 1
for j, start in enumerate(splits):
end = splits[j + 1] if j + 1 < len(splits) else len(part)
chunk = part[start:end].strip()
if chunk:
expanded.append(chunk)
if len(expanded) == 1:
return f"{head} := by\n {expanded[0]}"
return head + " := by\n" + "\n".join(f" {t}" for t in expanded)
# Match: <decl head> := <term> (no `by`)
m2 = re.match(r"^(.*?)\s*:=\s*(.+)$", s, re.DOTALL)
if m2:
head, body = m2.group(1).strip(), m2.group(2).strip()
return f"{head} :=\n {body}"
return s
def _format_code(code: str, language: str) -> str:
"""Run a language-specific formatter. Returns original on any failure."""
import subprocess
try:
if language == "lean":
return _format_lean(code)
if language == "typescript":
result = subprocess.run(
["npx", "--yes", "prettier", "--stdin-filepath", "candidate.ts"],
input=code, capture_output=True, text=True, timeout=15,
)
elif language == "python":
result = subprocess.run(
["ruff", "format", "--stdin-filename", "candidate.py", "-"],
input=code, capture_output=True, text=True, timeout=10,
)
elif language == "rust":
result = subprocess.run(
["rustfmt", "--edition", "2021"],
input=code, capture_output=True, text=True, timeout=10,
)
else:
# Unknown language β return as-is (word_wrap still applies in panel)
return code
return result.stdout if result.returncode == 0 and result.stdout.strip() else code
except Exception:
return code
def _code_panel(title: str, code: str, language: str, border: str = "blue") -> Panel:
formatted = _format_code(code, language)
try:
body = Syntax(formatted, language, theme="monokai", line_numbers=True, word_wrap=True)
except Exception:
body = Text(formatted, no_wrap=False)
return Panel(body, title=f"[bold {border}]{title}[/]", border_style=border)
def _text_panel(title: str, body: Text | str, border: str = "cyan") -> Panel:
if isinstance(body, str):
body = Text(body)
return Panel(body, title=f"[bold {border}]{title}[/]", border_style=border)
def _state_text(obs: LeanMigrateObservation) -> Text:
t = Text()
t.append("progress : ", style="bold cyan")
t.append(f"{clamp_open_unit(float(obs.progress)):.3f}\n")
t.append("verified : ", style="bold cyan")
t.append(", ".join(obs.verified) if obs.verified else "(none)")
t.append("\n")
t.append("remaining: ", style="bold cyan")
t.append(", ".join(obs.remaining) if obs.remaining else "(none)")
t.append("\n")
t.append("failing : ", style="bold cyan")
t.append(json.dumps(obs.failing) if obs.failing else "(none)")
t.append("\n")
t.append("done : ", style="bold cyan")
t.append(str(obs.done).lower())
return t
def _action_text(ad: dict[str, Any]) -> Text:
t = Text()
t.append("type : ", style="bold magenta")
t.append(ad.get("type", "?"))
t.append("\n")
t.append("function : ", style="bold magenta")
t.append(ad.get("function_name", "?"))
return t
def _obs_text(obs: LeanMigrateObservation) -> Text:
t = Text()
reward_val = clamp_open_unit(float(obs.reward or 0.0))
color = "green" if reward_val > 0 else ("red" if reward_val < 0 else "yellow")
t.append("reward : ", style=f"bold {color}")
t.append(f"{reward_val:.3f}\n", style=color)
t.append("done : ", style="bold cyan")
t.append(str(obs.done).lower())
return t
# ββ Trace helpers βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _obs_snapshot(obs: LeanMigrateObservation) -> dict[str, Any]:
"""Serialise the parts of an observation useful for offline analysis."""
return {
"progress": float(obs.progress),
"verified": list(obs.verified),
"remaining": list(obs.remaining),
"failing": dict(obs.failing),
"done": bool(obs.done),
"episode_step": int(obs.episode_step),
}
def _reward_snapshot(obs: LeanMigrateObservation) -> dict[str, Any] | None:
rd = obs.reward_details
if rd is None:
return None
return {
"score": float(rd.score),
"cumulative_score": float(rd.cumulative_score),
"tests_passed": rd.tests_passed,
"tests_total": rd.tests_total,
"proof_compiled": rd.proof_compiled,
"breakdown": dict(rd.breakdown),
"feedback": rd.feedback,
"lean_error": rd.lean_error,
}
def _step_analysis(
obs_before: LeanMigrateObservation,
obs_after: LeanMigrateObservation,
action_dict: dict[str, Any],
reward: float,
) -> dict[str, Any]:
"""Derived signals useful for reward-shaping / behaviour analysis."""
progress_delta = float(obs_after.progress) - float(obs_before.progress)
rd = obs_after.reward_details
tests_pass_rate: float | None = None
if rd and rd.tests_total:
tests_pass_rate = (rd.tests_passed or 0) / rd.tests_total
return {
"progress_delta": round(progress_delta, 6),
"tests_pass_rate": tests_pass_rate,
"lean_error_present": bool(rd and rd.lean_error),
"proof_compiled": rd.proof_compiled if rd else None,
"new_function_verified": progress_delta > 0,
"reward_sign": "positive" if reward > 0 else ("negative" if reward < 0 else "zero"),
"action_type": action_dict.get("type"),
"used_inspection": action_dict.get("type") in {"inspect", "analyze_deps"},
}
def _write_partial_trace(
out_path: Path,
task_id: str,
episode_id: str,
model: str,
api_base_url: str,
temperature: float,
task: Any,
steps: list[dict[str, Any]],
rewards: list[float],
system_prompt: str = "",
) -> None:
"""Write an in-progress snapshot so work is preserved on interruption."""
trace = {
"meta": {
"task_id": task_id,
"episode_id": episode_id,
"model": model,
"api_base_url": api_base_url,
"temperature": temperature,
"difficulty": task.difficulty,
"source_language": task.source_language,
"target_language": task.target_language,
"functions": [f.name for f in task.functions],
"proof_functions": [f.name for f in task.functions if f.is_proof_required],
"max_steps": task.max_steps,
"system_prompt": system_prompt,
"partial": True,
},
"steps": steps,
"summary": None,
}
out_path.write_text(json.dumps(trace, indent=2, ensure_ascii=False))
def _write_trace(
trace_dir: Path,
task_id: str,
episode_id: str,
model: str,
api_base_url: str,
temperature: float,
task: Any,
steps: list[dict[str, Any]],
rewards: list[float],
final_score: float,
out_path: Path | None = None,
system_prompt: str = "",
) -> Path:
"""Write the full episode trace to a JSON file and return its path."""
trace_dir.mkdir(parents=True, exist_ok=True)
ts = datetime.datetime.now(datetime.timezone.utc)
ts_str = ts.strftime("%Y%m%dT%H%M%SZ")
# ββ Summary analytics ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
action_counts: dict[str, int] = {}
lean_errors = 0
negative_rewards = 0
max_consecutive_run_tests = 0
current_run_tests_streak = 0
total_code_chars = 0
total_test_runs = 0
total_tests_passed = 0
total_tests_total = 0
for s in steps:
atype = s["action"]["type"]
action_counts[atype] = action_counts.get(atype, 0) + 1
if s["analysis"]["lean_error_present"]:
lean_errors += 1
if s["reward"] < 0:
negative_rewards += 1
if atype == "run_tests":
current_run_tests_streak += 1
max_consecutive_run_tests = max(
max_consecutive_run_tests, current_run_tests_streak
)
code = s["action"].get("candidate_code") or ""
total_code_chars += len(code)
total_test_runs += 1
rd = s["reward_details"] or {}
total_tests_passed += rd.get("tests_passed") or 0
total_tests_total += rd.get("tests_total") or 0
else:
current_run_tests_streak = 0
n_types = len(action_counts)
total_actions = len(steps)
# Shannon entropy of action types as a diversity score
import math
diversity = 0.0
if total_actions > 0 and n_types > 1:
for cnt in action_counts.values():
p = cnt / total_actions
if p > 0:
diversity -= p * math.log2(p)
diversity /= math.log2(len({"inspect", "analyze_deps", "run_tests", "submit"}))
final_obs = steps[-1]["state_after"] if steps else {}
summary = {
"total_steps": len(steps),
"final_score": round(final_score, 6),
"success": final_score >= 0.99,
"rewards": [round(r, 4) for r in rewards],
"reward_sum": round(sum(rewards), 4),
"action_type_counts": action_counts,
"action_diversity_score": round(diversity, 4),
"functions_verified": final_obs.get("verified", []),
"functions_remaining": final_obs.get("remaining", []),
"functions_failing": list((final_obs.get("failing") or {}).keys()),
"max_consecutive_run_tests": max_consecutive_run_tests,
"lean_errors_count": lean_errors,
"negative_rewards_count": negative_rewards,
"avg_tests_pass_rate": (
round(total_tests_passed / total_tests_total, 4)
if total_tests_total > 0
else None
),
"never_used_inspect": action_counts.get("inspect", 0) == 0,
"never_used_analyze_deps": action_counts.get("analyze_deps", 0) == 0,
}
trace = {
"meta": {
"task_id": task_id,
"episode_id": episode_id,
"model": model,
"api_base_url": api_base_url,
"temperature": temperature,
"timestamp": ts.isoformat(),
"difficulty": task.difficulty,
"source_language": task.source_language,
"target_language": task.target_language,
"functions": [f.name for f in task.functions],
"proof_functions": [f.name for f in task.functions if f.is_proof_required],
"max_steps": task.max_steps,
"system_prompt": system_prompt,
},
"steps": steps,
"summary": summary,
}
if out_path is None:
filename = f"{task_id}__{ts_str}__{episode_id[:8]}.json"
out_path = trace_dir / filename
out_path.write_text(json.dumps(trace, indent=2, ensure_ascii=False))
return out_path
def _print_step(
step: int,
obs_before: LeanMigrateObservation,
action_dict: dict[str, Any],
obs_after: LeanMigrateObservation,
target_language: str,
) -> None:
action_type = action_dict.get("type", "?")
fn_name = action_dict.get("function_name", "?")
console.print(
Rule(f"[bold yellow]STEP {step}[/] [white]{action_type}({fn_name})[/]")
)
# ββ State (before) + Action summary side-by-side
console.print(
Columns(
[
_text_panel("STATE (before)", _state_text(obs_before), "cyan"),
_text_panel("ACTION", _action_text(action_dict), "magenta"),
],
expand=True,
)
)
# ββ Code block (if run_tests)
code = action_dict.get("candidate_code")
if code:
lang = (
"rust"
if target_language == "rust"
else ("typescript" if target_language == "typescript" else "python")
)
console.print(_code_panel(f"CANDIDATE CODE [{lang}]", code, lang, "blue"))
# ββ Lean proof (if submit with proof)
proof = action_dict.get("lean_proof")
if proof:
console.print(_code_panel("LEAN PROOF", proof, "lean", "green"))
# ββ Observation (after)
rd = obs_after.reward_details
feedback = obs_after.last_action_feedback or ""
lean_err = (rd.lean_error if rd else None) or ""
reward_val = clamp_open_unit(float(obs_after.reward or 0.0))
border = "green" if reward_val > 0 else ("red" if reward_val < 0 else "yellow")
obs_body = Text()
obs_body.append_text(_obs_text(obs_after))
if feedback:
obs_body.append("\n\nfeedback :\n", style="bold cyan")
obs_body.append(_trim(feedback, 15))
if lean_err:
obs_body.append("\n\nlean error:\n", style="bold red")
obs_body.append(_trim(lean_err, 20))
console.print(_text_panel("OBSERVATION", obs_body, border))
# ββ LLM helpers (from inference.py) ββββββββββββββββββββββββββββββββββββββββββ
def _get_simple_sample(task: Task, fn: FunctionSpec) -> str:
cases = task.sample_inputs.get(fn.name, [])
if not cases or fn.is_proof_required:
return ""
case = cases[0]
expected = oracle_result(task.task_id, fn.name, case.args)
args_repr = ", ".join(repr(a) for a in case.args)
if len(args_repr) > 300:
args_repr = args_repr[:297] + "..."
return f"Sample: ({args_repr}) β {expected!r}"
def _get_language_notes(task: Task) -> str:
if task.target_language == "typescript":
return textwrap.dedent("""
## TypeScript notes
- Use top-level `function` declarations (not arrow functions / const assignments).
- Dependency code is injected automatically β write only the current function.
""").strip()
if task.target_language == "rust":
return textwrap.dedent("""
## Rust notes
- Use serde_json and serde (already in Cargo.toml).
- Write only the current function. Dependencies are injected as separate pub fn definitions.
- Match the inspected Rust signature exactly, including ownership and mutability. The harness passes owned JSON-decoded values into your function, so do not add `&`, `&mut`, or slice parameters unless the inspected source explicitly shows them.
- If the inspect output shows a `Source symbol`, use that exact Rust function name in your code.
- Return types serializable to serde_json::Value (Vec, Option, tuples all work).
- run_tests compiles via cargo. If run_tests returns "cargo not found", cargo is
unavailable β submit your implementation directly without run_tests.
""").strip()
return ""
def _get_proof_notes(task: Task) -> str:
if any(fn.is_proof_required for fn in task.functions):
return textwrap.dedent("""
## Proof task notes
- NEVER use `sorry` β the verifier rejects it IMMEDIATELY, gives -0.05 reward,
and does not compile Lean at all. `sorry` has zero partial credit.
- If you cannot write the full proof yet, DO NOT submit. Call `inspect` on the
proof function to read the exact theorem and full Lean spec, then reason step by step.
- Useful first-try tactics: `by decide`, `by native_decide`, `by simp [...]`,
`by omega`, `by rfl`, `by constructor <;> simp`.
- Use qualified ADT constructors from the spec (e.g. `Spec.MyType.mk`).
""").strip()
return ""
def _system_prompt(task: Task) -> str:
fn_sections = []
for fn in task.functions:
proof_tag = " [PROOF REQUIRED β submit with lean_proof field]" if fn.is_proof_required else ""
fn_sections.append(
f"### {fn.name}{proof_tag}\n"
f"Description: {fn.description}\n"
f"Depends on: {fn.depends_on or ['(none)']}\n"
)
language_notes = _get_language_notes(task)
proof_notes = _get_proof_notes(task)
extra = "\n\n".join(s for s in [language_notes, proof_notes] if s)
return textwrap.dedent(
f"""
You are an expert software engineer migrating {task.source_language} code to
{task.target_language} with Lean 4 verification.
Task: {task.display_name}
## Workflow
1. Call analyze_deps on any function to see the full dependency graph and migration order.
2. Call inspect on each function to read its source code and Lean spec before implementing.
3. Call run_tests with your implementation and iterate until all cases pass.
4. Call submit when tests pass β the system uses the last run_tests code automatically.
5. Write only the current function's code; verified dependencies are injected automatically.
6. For Rust tasks, treat the inspected function signature as authoritative. If a compile error mentions a type mismatch, fix the signature to match the harness before changing the body.
7. If inspect output names a Rust `Source symbol`, prefer that exact symbol over the Lean label when naming the function.
Example sequence for a 2-function task where B depends on A:
analyze_deps(B) β see order: A first, then B
inspect(A) β read source + Lean spec
run_tests(A, <code>) β iterate until passing
submit(A)
inspect(B) β read source + Lean spec
run_tests(B, <code>) β iterate until passing
submit(B)
## Actions (return ONLY JSON)
- inspect: {{"type": "inspect", "function_name": "foo"}}
- analyze_deps: {{"type": "analyze_deps", "function_name": "foo"}}
- run_tests: {{"type": "run_tests", "function_name": "foo", "candidate_code": "..."}}
- submit: {{"type": "submit", "function_name": "foo"}}
Proof only: {{"type": "submit", "function_name": "foo", "lean_proof": "..."}}
{extra}
## Functions to migrate
{"".join(fn_sections)}
"""
).strip()
def _llm_action(
client: Any,
model: str,
temperature: float,
task: Task,
obs: LeanMigrateObservation,
history: list[dict],
) -> tuple[dict[str, Any], str, str]:
"""Return (action_dict, user_prompt, raw_model_output)."""
current_step = len(history) + 1
# ββ Dynamic warnings ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
warnings: list[str] = []
recent_sorry = [
h for h in history[-8:]
if "sorry" in (h.get("feedback") or "").lower() and h.get("reward", 0) < 0
]
if len(recent_sorry) >= 2:
warnings.append(
f"β You have submitted `sorry` {len(recent_sorry)} times. "
"The verifier ALWAYS rejects sorry (-0.05 each time, no exceptions). "
"Write a complete proof: try `by decide`, `by native_decide`, `by simp [...]`, `by omega`."
)
if task.target_language == "rust":
recent_runtime_errors = sum(
1 for h in history[-4:]
if any(k in (h.get("feedback") or "").lower()
for k in ("cargo not found", "cannot find a runtime", "install cargo"))
)
if recent_runtime_errors >= 2:
warnings.append(
"β cargo is unavailable in this environment. "
"Do NOT use run_tests β submit your implementation directly."
)
# ββ Build prompt sections βββββββββββββββββββββββββββββββββββββββββββββββββ
state_section = (
f"<STATE>\n"
f"Step: {current_step} / {task.max_steps}\n"
f"Progress: {obs.progress:.2f}\n"
f"Verified: {', '.join(obs.verified) if obs.verified else 'None'}\n"
f"Remaining: {', '.join(obs.remaining)}\n"
f"Failing: {json.dumps(obs.failing)}\n"
f"</STATE>"
)
# Full last observation β no truncation; inspect output can be large
last_obs_section = ""
if obs.last_action_feedback:
last_obs_section = (
f"<LAST_OBSERVATION>\n"
f"{obs.last_action_feedback}\n"
f"</LAST_OBSERVATION>"
)
# Only the single most recent step (not a history window)
last_step_section = ""
if history:
h = history[-1]
last_step_section = (
f"<LAST_STEP>\n"
f"Step {h['step']}: type={h['action'].get('type')} "
f"fn={h['action'].get('function_name')} "
f"reward={h['reward']:.3f}\n"
f"</LAST_STEP>"
)
warnings_section = (
("<WARNINGS>\n" + "\n".join(warnings) + "\n</WARNINGS>")
if warnings else ""
)
prompt = "\n\n".join(
s for s in [
state_section,
last_obs_section,
last_step_section,
warnings_section,
"Decide the next action. Return ONLY valid JSON.",
]
if s
)
try:
response = client.beta.chat.completions.parse(
model=model,
messages=[
{"role": "system", "content": _system_prompt(task)},
{"role": "user", "content": prompt},
],
temperature=temperature,
response_format=ActionRequest,
)
raw_output = response.choices[0].message.content or ""
parsed = response.choices[0].message.parsed
if parsed is None:
raise ValueError("No structured output returned")
return parsed.model_dump(), prompt, raw_output
except Exception as exc:
console.print(f"[red]LLM error:[/red] {exc}")
return (
{
"type": "run_tests",
"function_name": obs.remaining[0] if obs.remaining else "",
"candidate_code": "# llm error",
},
prompt,
f"<error: {exc}>",
)
def _parse_action(ad: dict[str, Any]) -> LeanMigrateAction:
t = ad.get("type", "run_tests")
fn = ad.get("function_name", "")
if t == "inspect":
return InspectAction(type="inspect", function_name=fn)
if t == "analyze_deps":
return AnalyzeDepsAction(type="analyze_deps", function_name=fn)
if t == "run_tests":
return RunTestsAction(
type="run_tests",
function_name=fn,
candidate_code=ad.get("candidate_code", ""),
)
if t == "submit":
return SubmitAction(
type="submit",
function_name=fn,
target_code=ad.get("target_code"),
lean_proof=ad.get("lean_proof"),
)
return RunTestsAction(type="run_tests", function_name="", candidate_code="")
# ββ Commands ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@app.command("list")
def cmd_list() -> None:
"""List all available LeanMigrate tasks."""
tasks_info = list_tasks()
from rich import box as rich_box
table = Table(
title="[bold]LeanMigrate Tasks[/bold]",
show_header=True,
header_style="bold cyan",
box=rich_box.SIMPLE_HEAD,
padding=(0, 1),
expand=True,
)
table.add_column("task_id", style="bold yellow", no_wrap=True)
table.add_column("display_name")
table.add_column("diff", no_wrap=True)
table.add_column("migration", no_wrap=True)
table.add_column("functions")
table.add_column("steps", justify="right", no_wrap=True)
for info in tasks_info:
task = get_task(info["task_id"])
diff = info.get("difficulty", "?")
color = _difficulty_color(diff)
non_proof = [f.name for f in task.functions if not f.is_proof_required]
proof = [f.name for f in task.functions if f.is_proof_required]
fn_str = ", ".join(non_proof)
if proof:
fn_str += f" [dim](+{', '.join(proof)})[/dim]"
table.add_row(
info["task_id"],
task.display_name,
f"[{color}]{diff}[/{color}]",
f"{task.source_language} β {task.target_language}",
fn_str,
str(task.max_steps),
)
console.print()
console.print(table)
console.print()
console.print(
"[dim]Play a task:[/dim] uv run python scripts/play_task.py play <task_id>"
)
console.print(
"[dim]Play all: [/dim] uv run python scripts/play_task.py play --all"
)
console.print()
@app.command("play")
def cmd_play(
task_ids: Annotated[
Optional[list[str]],
typer.Argument(help="Task IDs to play (omit to use --all)."),
] = None,
all_tasks: Annotated[bool, typer.Option("--all", help="Play every task.")] = False,
model: Annotated[
Optional[str],
typer.Option(
"--model",
"-m",
envvar="MODEL_NAME",
help="LLM model name. [env: MODEL_NAME]",
),
] = None,
api_base_url: Annotated[
Optional[str],
typer.Option(
"--api-base-url",
envvar="API_BASE_URL",
help="OpenAI-compatible API base URL. [env: API_BASE_URL]",
),
] = None,
api_key: Annotated[
Optional[str],
typer.Option(
"--api-key",
envvar="HF_TOKEN",
help="API key. [env: HF_TOKEN or GEMINI_API_KEY]",
),
] = None,
max_steps: Annotated[
int,
typer.Option("--max-steps", envvar="MAX_STEPS", help="Max LLM steps per task."),
] = 50,
temperature: Annotated[
float,
typer.Option(
"--temperature", envvar="TEMPERATURE", help="Sampling temperature."
),
] = 0.2,
quiet: Annotated[
bool,
typer.Option("--quiet", "-q", help="Suppress per-step rich output."),
] = False,
trace_dir: Annotated[
Optional[Path],
typer.Option(
"--trace-dir",
help="Directory to write per-episode JSON traces. Defaults to <project>/traces/.",
),
] = None,
) -> None:
"""Play LeanMigrate tasks with an LLM, showing rich step-by-step output."""
from openai import OpenAI
all_ids = [t["task_id"] for t in list_tasks()]
if all_tasks:
selected = all_ids
elif task_ids:
bad = [t for t in task_ids if t not in all_ids]
if bad:
console.print(f"[red]Unknown task(s):[/red] {', '.join(bad)}")
console.print(f"Available: {', '.join(all_ids)}")
raise typer.Exit(code=1)
selected = list(task_ids)
else:
console.print("[red]Specify task IDs or --all.[/red]")
console.print(f"Available: {', '.join(all_ids)}")
raise typer.Exit(code=1)
# Resolve LLM config: CLI flag β .env / os.environ β hardcoded fallback
resolved_model = model or os.getenv("MODEL_NAME", "gemini-2.5-flash-preview-04-17")
resolved_url = api_base_url or os.getenv(
"API_BASE_URL", "https://generativelanguage.googleapis.com/v1beta/openai/"
)
resolved_key = api_key or os.getenv("HF_TOKEN") or os.getenv("GEMINI_API_KEY")
if not resolved_key:
console.print("[red]Error:[/red] API key required.")
console.print(
"Set [bold]HF_TOKEN[/bold] or [bold]GEMINI_API_KEY[/bold] in your "
".env file or pass [bold]--api-key[/bold]."
)
raise typer.Exit(code=1)
console.print(
Panel(
Text.assemble(
("model : ", "bold cyan"),
resolved_model,
"\n",
("base_url : ", "bold cyan"),
resolved_url,
"\n",
("tasks : ", "bold cyan"),
", ".join(selected),
),
title="[bold white]LLM CONFIG[/]",
border_style="white",
)
)
resolved_trace_dir = trace_dir or (ROOT / "traces")
client = OpenAI(base_url=resolved_url, api_key=resolved_key)
scores: dict[str, float] = {}
for task_id in selected:
scores[task_id] = _play_one_task(
client=client,
task_id=task_id,
model=resolved_model,
api_base_url=resolved_url,
max_steps=max_steps,
temperature=temperature,
quiet=quiet,
trace_dir=resolved_trace_dir,
)
# ββ Final summary ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
console.print(Rule("[bold white]PLAYTHROUGH SUMMARY[/]"))
summary_table = Table(
show_header=True, header_style="bold cyan", box=None, padding=(0, 2)
)
summary_table.add_column("task_id", style="bold yellow")
summary_table.add_column("score", justify="right")
summary_table.add_column("result")
total = 0.0
for tid, score in scores.items():
color = "green" if score >= 0.99 else ("yellow" if score > 0 else "red")
result = (
"β complete" if score >= 0.99 else ("partial" if score > 0 else "failed")
)
summary_table.add_row(tid, f"{score:.3f}", f"[{color}]{result}[/{color}]")
total += score
if len(scores) > 1:
avg = total / len(scores)
summary_table.add_section()
summary_table.add_row("[bold]overall[/bold]", f"[bold]{avg:.3f}[/bold]", "")
console.print(summary_table)
def _play_one_task(
client: Any,
task_id: str,
model: str,
api_base_url: str,
max_steps: int,
temperature: float,
quiet: bool,
trace_dir: Path,
) -> float:
task = get_task(task_id)
env = LeanMigrateEnvironment()
obs = env.reset(task_id=task_id)
# ββ Task header ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
diff = task.difficulty
color = _difficulty_color(diff)
header = Text()
header.append("task : ", style="bold cyan")
header.append(task_id + "\n")
header.append("name : ", style="bold cyan")
header.append(task.display_name + "\n")
header.append("difficulty: ", style="bold cyan")
header.append(diff, style=f"bold {color}")
header.append("\n")
header.append("migration : ", style="bold cyan")
header.append(f"{task.source_language} β {task.target_language}\n")
header.append("episode : ", style="bold cyan")
header.append(obs.episode_id + "\n")
header.append("model : ", style="bold cyan")
header.append(model + "\n")
header.append("max_steps: ", style="bold cyan")
header.append(str(max_steps))
console.print()
console.print(Rule(f"[bold white]{task.display_name}[/] [dim]({task_id})[/dim]"))
console.print(_text_panel("TASK", header, "white"))
rewards: list[float] = []
history: list[dict] = []
trace_steps: list[dict[str, Any]] = []
code_cache: dict[str, str] = {}
steps_taken = 0
episode_system_prompt = _system_prompt(task)
# Pre-determine trace path so partial writes and final write share the same file.
trace_dir.mkdir(parents=True, exist_ok=True)
episode_start_ts = datetime.datetime.now(datetime.timezone.utc)
_partial_trace_path = (
trace_dir
/ f"{task_id}__{episode_start_ts.strftime('%Y%m%dT%H%M%SZ')}__{obs.episode_id[:8]}.json"
)
for step in range(1, max_steps + 1):
if obs.done:
break
obs_before = obs
try:
action_dict, llm_user_prompt, llm_raw_output = _llm_action(
client, model, temperature, task, obs, history
)
# Maintain code cache: record tested code, inject on submit
fn_name = action_dict.get("function_name", "")
if action_dict.get("type") == "run_tests":
code_cache[fn_name] = action_dict.get("candidate_code", "")
elif action_dict.get("type") == "submit":
if fn_name in code_cache:
action_dict["target_code"] = code_cache[fn_name]
elif action_dict.get("candidate_code"):
# Agent skipped run_tests (e.g. Rust without cargo) and attached code directly
action_dict["target_code"] = action_dict["candidate_code"]
code_cache[fn_name] = action_dict["candidate_code"]
action = _parse_action(action_dict)
obs = env.step(action)
raw_reward = float(obs.reward or 0.0)
display_reward = clamp_open_unit(raw_reward)
rewards.append(raw_reward) # raw for accurate analysis
steps_taken = step
feedback = obs.last_action_feedback or ""
rd = obs.reward_details
lean_err = (rd.lean_error if rd else None) or ""
if lean_err:
feedback = (feedback + " | Lean: " + lean_err).strip()
history.append(
{
"step": step,
"action": action_dict,
"reward": raw_reward, # raw so LLM sees genuine negative signal
"feedback": feedback,
}
)
# ββ Trace record βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
trace_steps.append(
{
"step": step,
"timestamp": datetime.datetime.now(datetime.timezone.utc).isoformat(),
"state_before": _obs_snapshot(obs_before),
"action": {
"type": action_dict.get("type"),
"function_name": action_dict.get("function_name"),
"candidate_code": action_dict.get("candidate_code"),
"lean_proof": action_dict.get("lean_proof"),
"target_code": action_dict.get("target_code"),
},
"llm": {
"user_prompt": llm_user_prompt,
"raw_output": llm_raw_output,
},
"reward": round(raw_reward, 6),
"reward_details": _reward_snapshot(obs),
"state_after": _obs_snapshot(obs),
"feedback": feedback,
"analysis": _step_analysis(obs_before, obs, action_dict, raw_reward),
}
)
# Persist partial trace after every step so interruption loses nothing.
_write_partial_trace(
out_path=_partial_trace_path,
task_id=task_id,
episode_id=obs.episode_id,
model=model,
api_base_url=api_base_url,
temperature=temperature,
task=task,
steps=trace_steps,
rewards=rewards,
system_prompt=episode_system_prompt,
)
if not quiet:
_print_step(step, obs_before, action_dict, obs, task.target_language)
if obs.done:
break
except KeyboardInterrupt:
console.print(
f"\n[yellow]Interrupted at step {step}. "
f"Partial trace saved β {_partial_trace_path}[/yellow]"
)
raise
# ββ Task footer ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
final_score = clamp_open_unit(float(obs.progress))
success = final_score >= 0.99
color = "green" if success else ("yellow" if final_score > 0 else "red")
result_label = (
"COMPLETE" if success else ("PARTIAL" if final_score > 0 else "FAILED")
)
footer = Text()
footer.append("result : ", style="bold cyan")
footer.append(result_label, style=f"bold {color}")
footer.append("\n")
footer.append("score : ", style="bold cyan")
footer.append(f"{final_score:.3f}\n", style=color)
footer.append("steps : ", style="bold cyan")
footer.append(f"{steps_taken}\n")
footer.append("verified : ", style="bold cyan")
footer.append(", ".join(obs.verified) if obs.verified else "(none)")
# ββ Write trace (overwrites partial with full analytics) βββββββββββββββββββ
trace_path = _write_trace(
trace_dir=trace_dir,
task_id=task_id,
episode_id=obs.episode_id,
model=model,
api_base_url=api_base_url,
temperature=temperature,
task=task,
steps=trace_steps,
rewards=rewards,
final_score=final_score,
out_path=_partial_trace_path,
system_prompt=episode_system_prompt,
)
footer.append("\ntrace : ", style="bold cyan")
footer.append(str(trace_path), style="dim")
console.print(Rule())
console.print(_text_panel(f"RESULT {task_id}", footer, color))
print(
f"[END] task={task_id} success={str(success).lower()} "
f"steps={steps_taken} score={final_score:.2f} "
f"rewards={','.join(f'{clamp_open_unit(r):.2f}' for r in rewards)} "
f"trace={trace_path}",
flush=True,
)
return final_score
if __name__ == "__main__":
app()
|