| """Have Beam diagnose a failed learnability check from the environment's source and write a fix. |
| |
| dotenvx run -- python scripts/fix.py connect4_env --evidence runs/fixes/connect4_env.evidence.md |
| |
| Beam sees the environment's server code and models plus the evidence (what the ladder measured). It answers |
| with a diagnosis and the full new contents of each file it changes. The patched environment is written to |
| runs/fixes/envs/<env>/ (a copy; the vendored repo is untouched), with Beam's reply and a diff, so the |
| ladder can re-run against it with --envs-dir runs/fixes/envs. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import difflib |
| import json |
| import re |
| import shutil |
| import sys |
| from pathlib import Path |
|
|
| sys.path.insert(0, str(Path(__file__).resolve().parent.parent)) |
| from shortlist import beam |
| from shortlist.serve import OPENENV, ROOT |
|
|
| SYSTEM = """You maintain reinforcement-learning environments for training language models. |
| You are given an environment's source code and evidence from an automated learnability check. |
| Find the root cause in the code, then fix it with the smallest change that makes the reward reflect the agent's skill. |
| Keep the action and observation schemas, class names and the server API unchanged, so existing clients keep working. |
| |
| Answer in exactly this shape: |
| DIAGNOSIS: |
| <a short plain-English explanation of the root cause, naming the file and lines> |
| |
| FIX: |
| <one or two sentences on what you changed> |
| |
| Then, for every file you change, its complete new contents: |
| ### FILE: <path relative to the environment folder> |
| ```python |
| <complete file> |
| ```""" |
|
|
|
|
| def gather(env_dir: Path) -> dict[str, str]: |
| files = {} |
| for path in sorted(env_dir.rglob("*.py")): |
| rel = path.relative_to(env_dir).as_posix() |
| if "test" in rel or rel.startswith("."): |
| continue |
| files[rel] = path.read_text() |
| return files |
|
|
|
|
| def main() -> None: |
| parser = argparse.ArgumentParser() |
| parser.add_argument("env") |
| parser.add_argument("--evidence", required=True) |
| parser.add_argument("--effort", default="high") |
| args = parser.parse_args() |
|
|
| src = OPENENV / "envs" / args.env |
| files = gather(src) |
| evidence = Path(args.evidence).read_text() |
| code = "\n\n".join(f"### FILE: {rel}\n```python\n{text}\n```" for rel, text in files.items()) |
| messages = [ |
| {"role": "system", "content": SYSTEM}, |
| {"role": "user", "content": f"Environment: {args.env}\n\nEvidence from the learnability check:\n{evidence}\n\nSource:\n{code}"}, |
| ] |
| |
| response = beam.client().chat.completions.create( |
| model=beam.MODEL, messages=messages, reasoning_effort=args.effort, max_completion_tokens=beam.MAX_TOKENS[args.effort], |
| ) |
| beam.USAGE.add(response.usage) |
| message = response.choices[0].message |
| text = message.content or "" |
| out = ROOT / "runs" / "fixes" |
| out.mkdir(parents=True, exist_ok=True) |
| (out / f"{args.env}.beam-reply.md").write_text(text) |
| (out / f"{args.env}.beam-reasoning.md").write_text(getattr(message, "reasoning_content", "") or "") |
|
|
| changed = {m.group(1).strip(): m.group(2) for m in re.finditer(r"### FILE: (.+?)\n```(?:python)?\n(.*?)\n```", text, re.S)} |
| if not changed: |
| print("Beam returned no file changes. Reply saved."); print(text[:2000]); return |
| dest = out / "envs" / args.env |
| if dest.exists(): |
| shutil.rmtree(dest) |
| shutil.copytree(src, dest, ignore=shutil.ignore_patterns("__pycache__", "*.egg-info")) |
| diff = [] |
| root = dest.resolve() |
| for rel, new in changed.items(): |
| |
| target = (dest / rel).resolve() |
| if Path(rel).is_absolute() or not target.is_relative_to(root) or target.suffix != ".py": |
| print(f"Refusing to write {rel!r}: outside the environment folder or not a .py file.") |
| continue |
| old = target.read_text() if target.exists() else "" |
| target.write_text(new.rstrip() + "\n") |
| diff += difflib.unified_diff(old.splitlines(True), (new.rstrip() + "\n").splitlines(True), f"a/envs/{args.env}/{rel}", f"b/envs/{args.env}/{rel}") |
| (out / f"{args.env}.patch").write_text("".join(diff)) |
| diagnosis = text.split("### FILE:")[0].strip() |
| print(diagnosis) |
| print(f"\nChanged: {list(changed)}\nPatched copy: {dest}\nUsage: {json.dumps(beam.USAGE.summary())}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
