benchmark_v3 / harness /agent_protocol.py
ymh233's picture
Add files using upload-large-folder tool
8b97eb8 verified
Raw
History Blame Contribute Delete
20.8 kB
"""Fixed tool-call interface for RealSR v3 agents.
This is the FROZEN protocol any solver shares (the baseline agent and any
plugged-in evolving / search agent alike). It owns three things:
1. The tool TAGS the model emits and how they are parsed:
<python>...code...</python> inspect data / fit
<experiment>{...}</experiment> probe a simulator (if any)
<final_formula>...module text...</final_formula> submit (ends the trial)
2. The <python> SANDBOX: validate + exec with preloaded variables, capture stdout.
3. The per-turn DISPATCH (`step`): pick the first emitted tag, run it, and return
either the submission or the feedback string to append to the conversation.
Pair this with `prompts.load_system_prompt` / `prompts.build_task_prompt` (the
matching instruction text). Your agent only has to: call your LLM -> `step(...)`
-> append feedback / stop on submit. Nothing here depends on a particular LLM
client or task object, so it is reusable across agents.
"""
from __future__ import annotations
import ast
import builtins
import io
import json
import math
import re
import signal
import traceback
from contextlib import redirect_stdout
from typing import Any, Callable, Dict, List, Optional, Tuple
import numpy as np
TOOL_TAGS = ("final_formula", "python", "experiment")
PYTHON_TIMEOUT_SECONDS = 100
PYTHON_STDOUT_MAX_CHARS = 16_000
EXPERIMENT_OUTPUT_MAX_CHARS = 16_000
BRUTE_FORCE_MAX_KNOWN_ITERATIONS = 200_000
BRUTE_FORCE_MAX_NESTED_LOOP_DEPTH = 4
# ---- tag parsers -----------------------------------------------------------
def _last_block(text: str, open_tag: str, close_tag: str) -> Optional[str]:
start = text.rfind(open_tag)
if start == -1:
return None
end = text.find(close_tag, start)
if end == -1:
return None
return text[start + len(open_tag):end].strip()
def parse_experiment(text: str) -> Optional[Dict[str, Any]]:
block = _last_block(text, "<experiment>", "</experiment>")
if block is None:
return None
try:
parsed = json.loads(block)
except Exception:
return None
return parsed if isinstance(parsed, dict) else None
def parse_final_formula(text: str) -> Tuple[bool, str]:
"""Return (ok, module_text) for the <final_formula>...</final_formula> block.
`ok` is False unless the block exists and defines `predict`."""
block = _last_block(text, "<final_formula>", "</final_formula>")
if block is None or "def predict" not in block:
return False, ""
return True, block
def extract_python(text: str) -> Optional[str]:
return _last_block(text, "<python>", "</python>")
def first_tool_tag(text: str) -> Optional[str]:
"""The tool tag that appears FIRST by source position (the model's intent
when it emits several in one turn — e.g. explore, then submit)."""
first_tag, first_pos = None, -1
for t in TOOL_TAGS:
p = text.find(f"<{t}>")
if p >= 0 and (first_pos < 0 or p < first_pos):
first_pos, first_tag = p, t
return first_tag
def unclosed_tags(text: str) -> List[str]:
out = []
for tag in ("python", "experiment", "final_formula"):
o, c = f"<{tag}>", f"</{tag}>"
if o in text and text.rfind(o) > text.rfind(c):
out.append(tag)
return out
# ---- python sandbox --------------------------------------------------------
_DANGEROUS_PATTERNS = [
r"import\s+os", r"import\s+sys", r"import\s+subprocess",
r"from\s+os\s+import", r"from\s+sys\s+import", r"from\s+subprocess\s+import",
r"import\s+pathlib", r"from\s+pathlib\s+import",
r"import\s+pickle", r"from\s+pickle\s+import",
r"import\s+joblib", r"from\s+joblib\s+import",
r"import\s+glob", r"from\s+glob\s+import",
r"import\s+importlib", r"from\s+importlib\s+import",
r"import\s+inspect", r"from\s+inspect\s+import",
r"import\s+shutil", r"from\s+shutil\s+import",
r"__import__", r"\beval\(", r"\bexec\(", r"\bopen\(", r"\bfile\(",
r"\binput\(", r"\braw_input\(", r"\bcompile\(",
r"\bglobals\(", r"\blocals\(", r"\bgetattr\(", r"\bsetattr\(",
r"__dict__", r"__class__", r"__mro__", r"__subclasses__",
r"\bread_text\(", r"\bread_bytes\(", r"\bread_csv\(", r"\bread_table\(",
r"\bread_excel\(", r"\bloadtxt\(", r"\bgenfromtxt\(",
r"\bGridSearchCV\b", r"\bParameterGrid\b",
]
_ALLOWED_IMPORT_ROOTS = {
"collections",
"functools",
"itertools",
"math",
"numpy",
"pandas",
"scipy",
"sklearn",
"statistics",
"warnings",
}
_BLOCKED_IMPORT_ROOTS = {
"builtins",
"glob",
"importlib",
"inspect",
"io",
"joblib",
"os",
"pathlib",
"pickle",
"shutil",
"subprocess",
"sys",
}
_BLOCKED_CALL_NAMES = {
"__import__",
"compile",
"eval",
"exec",
"file",
"getattr",
"globals",
"input",
"locals",
"open",
"raw_input",
"setattr",
}
_BLOCKED_CALL_ATTRS = {
"dump",
"dumps",
"fromfile",
"genfromtxt",
"get_handle",
"load",
"loadtxt",
"open",
"read_bytes",
"read_csv",
"read_excel",
"read_feather",
"read_hdf",
"read_json",
"read_orc",
"read_parquet",
"read_pickle",
"read_sas",
"read_stata",
"read_table",
"read_text",
"savetxt",
"tofile",
}
_SAFE_BUILTIN_NAMES = {
"ArithmeticError",
"AssertionError",
"Exception",
"FloatingPointError",
"IndexError",
"KeyError",
"LookupError",
"NameError",
"OverflowError",
"RuntimeError",
"TypeError",
"ValueError",
"ZeroDivisionError",
"abs",
"all",
"any",
"bool",
"callable",
"dict",
"enumerate",
"filter",
"float",
"hasattr",
"int",
"isinstance",
"iter",
"len",
"list",
"map",
"max",
"min",
"next",
"object",
"pow",
"print",
"range",
"repr",
"reversed",
"round",
"set",
"slice",
"sorted",
"str",
"sum",
"tuple",
"type",
"zip",
}
class _PythonExecTimeout(BaseException):
pass
def _python_timeout_handler(signum, frame): # noqa: ARG001
raise _PythonExecTimeout()
def _truncate_stdout(stdout: str) -> tuple[str, bool, int]:
n = len(stdout)
if n <= PYTHON_STDOUT_MAX_CHARS:
return stdout, False, n
omitted = n - PYTHON_STDOUT_MAX_CHARS
truncated = (
stdout[:PYTHON_STDOUT_MAX_CHARS]
+ f"\n...[python stdout truncated; omitted {omitted} characters]"
)
return truncated, True, n
def _literal_number(node: ast.AST) -> float | None:
if isinstance(node, ast.Constant) and isinstance(node.value, (int, float)):
return float(node.value)
unary = isinstance(node, ast.UnaryOp) and isinstance(node.op, (ast.UAdd, ast.USub))
if unary and isinstance(node.operand, ast.Constant) and isinstance(node.operand.value, (int, float)):
v = float(node.operand.value)
return -v if isinstance(node.op, ast.USub) else v
return None
def _literal_int(node: ast.AST) -> int | None:
v = _literal_number(node)
if v is None or not float(v).is_integer():
return None
return int(v)
def _call_name(node: ast.AST) -> str:
if isinstance(node, ast.Name):
return node.id
if isinstance(node, ast.Attribute):
base = _call_name(node.value)
return f"{base}.{node.attr}" if base else node.attr
return ""
def _range_like_count(call: ast.Call) -> int | None:
name = _call_name(call.func)
args = call.args
kwargs = {kw.arg: kw.value for kw in call.keywords if kw.arg}
if name == "range":
vals = [_literal_int(a) for a in args]
if any(v is None for v in vals):
return None
if len(vals) == 1:
start, stop, step = 0, vals[0], 1
elif len(vals) == 2:
start, stop, step = vals[0], vals[1], 1
elif len(vals) == 3:
start, stop, step = vals
else:
return None
if step == 0:
return None
return max(0, math.ceil((stop - start) / step)) if step > 0 else max(0, math.ceil((start - stop) / abs(step)))
if name.endswith("linspace") or name.endswith("logspace"):
if "num" in kwargs:
return _literal_int(kwargs["num"])
if len(args) >= 3:
return _literal_int(args[2])
return 50
if name.endswith("arange"):
vals = [_literal_number(a) for a in args]
if any(v is None for v in vals):
return None
if len(vals) == 1:
start, stop, step = 0.0, vals[0], 1.0
elif len(vals) == 2:
start, stop, step = vals[0], vals[1], 1.0
elif len(vals) == 3:
start, stop, step = vals
else:
return None
if step == 0:
return None
return max(0, math.ceil((stop - start) / step)) if step > 0 else max(0, math.ceil((start - stop) / abs(step)))
return None
def _iter_count(node: ast.AST) -> int | None:
if isinstance(node, (ast.List, ast.Tuple, ast.Set)):
return len(node.elts)
if isinstance(node, ast.Call):
name = _call_name(node.func)
if name.endswith("product"):
repeat = 1
for kw in node.keywords:
if kw.arg == "repeat":
repeat = _literal_int(kw.value) or repeat
counts = [_iter_count(arg) for arg in node.args]
if not counts or any(c is None for c in counts):
return None
prod = 1
for c in counts:
prod *= c
return prod ** repeat
return _range_like_count(node)
return None
def _validate_no_large_bruteforce(tree: ast.AST) -> Tuple[bool, Optional[str]]:
class Visitor(ast.NodeVisitor):
def __init__(self):
self.loop_counts: List[int | None] = []
self.error: Optional[str] = None
def visit_For(self, node: ast.For) -> None:
if self.error:
return
count = _iter_count(node.iter)
self.loop_counts.append(count)
depth = len(self.loop_counts)
if depth >= BRUTE_FORCE_MAX_NESTED_LOOP_DEPTH:
self.error = (
f"Code appears to use high-dimensional brute-force search "
f"({depth} nested for-loops). Use vectorized fitting or a "
"small targeted search instead."
)
return
known = [c for c in self.loop_counts if c is not None]
if len(known) == depth:
prod = 1
for c in known:
prod *= c
if prod > BRUTE_FORCE_MAX_KNOWN_ITERATIONS:
self.error = (
f"Code appears to use large brute-force grid search "
f"({prod} known loop iterations > "
f"{BRUTE_FORCE_MAX_KNOWN_ITERATIONS}). Use a smaller "
"targeted search or scipy fitting."
)
return
self.generic_visit(node)
self.loop_counts.pop()
v = Visitor()
v.visit(tree)
return (v.error is None, v.error)
def _validate_no_file_or_unsafe_imports(tree: ast.AST) -> Tuple[bool, Optional[str]]:
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for alias in node.names:
root = alias.name.split(".", 1)[0]
if root in _BLOCKED_IMPORT_ROOTS or root not in _ALLOWED_IMPORT_ROOTS:
return False, f"Import of module {alias.name!r} is not allowed in the sandbox"
elif isinstance(node, ast.ImportFrom):
if node.module is None:
return False, "Relative imports are not allowed in the sandbox"
root = node.module.split(".", 1)[0]
if root in _BLOCKED_IMPORT_ROOTS or root not in _ALLOWED_IMPORT_ROOTS:
return False, f"Import from module {node.module!r} is not allowed in the sandbox"
elif isinstance(node, ast.Call):
name = _call_name(node.func)
attr = name.rsplit(".", 1)[-1]
if name in _BLOCKED_CALL_NAMES or attr in _BLOCKED_CALL_ATTRS:
return False, f"Call to {name!r} is not allowed in the sandbox"
elif isinstance(node, ast.Attribute):
if node.attr.startswith("__") and node.attr.endswith("__"):
return False, f"Access to dunder attribute {node.attr!r} is not allowed in the sandbox"
elif isinstance(node, ast.Name):
if node.id.startswith("__") and node.id.endswith("__"):
return False, f"Access to dunder name {node.id!r} is not allowed in the sandbox"
return True, None
def validate_python(code: str) -> Tuple[bool, Optional[str]]:
try:
tree = ast.parse(code)
except SyntaxError as e:
return False, f"Syntax error: {e}"
for pat in _DANGEROUS_PATTERNS:
if re.search(pat, code, re.IGNORECASE):
return False, f"Code contains blocked pattern: {pat}"
ok, err = _validate_no_file_or_unsafe_imports(tree)
if not ok:
return False, err
ok, err = _validate_no_large_bruteforce(tree)
if not ok:
return False, err
return True, None
def _safe_import(name, globals=None, locals=None, fromlist=(), level=0): # noqa: A002, ARG001
if level != 0:
raise ImportError("relative imports are not allowed in the sandbox")
root = str(name).split(".", 1)[0]
if root in _BLOCKED_IMPORT_ROOTS or root not in _ALLOWED_IMPORT_ROOTS:
raise ImportError(f"import of {name!r} is not allowed in the sandbox")
return builtins.__import__(name, globals, locals, fromlist, level)
def _safe_builtins() -> Dict[str, Any]:
out = {name: getattr(builtins, name) for name in _SAFE_BUILTIN_NAMES}
out["__import__"] = _safe_import
return out
def build_sandbox(train_df=None, X_train=None, y_train=None, group_ids=None,
input_cols=None, target_col=None) -> Dict[str, Any]:
"""Preloaded variables exposed inside <python>. Pass whatever you have;
`np`/`scipy`/`pd` are added automatically by `run_python`."""
sb: Dict[str, Any] = {}
if train_df is not None:
sb["train_df"] = train_df
if X_train is not None:
sb["X_train"] = X_train
if y_train is not None:
sb["y_train"] = y_train
if group_ids is not None:
sb["group_ids_train"] = group_ids
if input_cols is not None:
sb["input_cols"] = list(input_cols)
if target_col is not None:
sb["target_col"] = target_col
return sb
def run_python(code: str, sandbox: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
"""Validate + exec `code` with `sandbox` preloaded; capture stdout. Only
print() output is returned to the model."""
ok, err = validate_python(code)
if not ok:
return {"success": False, "error_type": "ValidationError",
"error_message": err, "code": code}
ns: Dict[str, Any] = {"np": np, "__builtins__": _safe_builtins()}
try:
import scipy # noqa: F401
ns["scipy"] = scipy
except ImportError:
pass
try:
import pandas as pd # noqa: F401
ns["pd"] = pd
except ImportError:
pass
ns.update(sandbox or {})
buf = io.StringIO()
old_handler = signal.getsignal(signal.SIGALRM)
signal.signal(signal.SIGALRM, _python_timeout_handler)
old_timer = signal.setitimer(signal.ITIMER_REAL, PYTHON_TIMEOUT_SECONDS)
try:
with redirect_stdout(buf):
exec(code, ns)
stdout, truncated, original_len = _truncate_stdout(buf.getvalue().strip())
return {
"success": True,
"stdout": stdout,
"stdout_truncated": truncated,
"stdout_original_chars": original_len,
"code": code,
}
except _PythonExecTimeout:
stdout, truncated, original_len = _truncate_stdout(buf.getvalue().strip())
return {
"success": False,
"error_type": "TimeoutError",
"error_message": f"Python execution exceeded {PYTHON_TIMEOUT_SECONDS}s",
"stdout": stdout,
"stdout_truncated": truncated,
"stdout_original_chars": original_len,
"code": code,
}
except Exception as e:
stdout, truncated, original_len = _truncate_stdout(buf.getvalue().strip())
return {"success": False, "error_type": type(e).__name__,
"error_message": str(e), "traceback": traceback.format_exc(),
"stdout": stdout, "stdout_truncated": truncated,
"stdout_original_chars": original_len, "code": code}
finally:
signal.setitimer(signal.ITIMER_REAL, 0)
signal.signal(signal.SIGALRM, old_handler)
if old_timer[0] > 0:
signal.setitimer(signal.ITIMER_REAL, old_timer[0], old_timer[1])
# ---- feedback strings ------------------------------------------------------
def format_python_feedback(result: Dict[str, Any]) -> str:
if not result["success"]:
stdout = result.get("stdout") or ""
stdout_block = f"\nPartial stdout before failure:\n{stdout}" if stdout else ""
return (f"<python_output>\nPython execution failed: "
f"{result['error_type']}: {result['error_message']}"
f"{stdout_block}\n</python_output>")
out = result["stdout"] if result["stdout"] else "(no stdout)"
return f"<python_output>\n{out}\n</python_output>"
def format_experiment_feedback(result: Dict[str, Any]) -> str:
text = json.dumps(result, default=str)
if len(text) > EXPERIMENT_OUTPUT_MAX_CHARS:
omitted = len(text) - EXPERIMENT_OUTPUT_MAX_CHARS
text = (
text[:EXPERIMENT_OUTPUT_MAX_CHARS]
+ f"\n...[experiment output truncated; omitted {omitted} characters]"
)
return f"<experiment_output>\n{text}\n</experiment_output>"
INVALID_RESPONSE_MSG = (
"Invalid response. Output exactly one XML block and no prose:\n"
"<python>...code...</python>\n"
"<experiment>{\"<input>\": [vals], \"n_samples\": 3}</experiment> "
"(simulator tasks only; JSON literals only, no Python expressions)\n"
"<final_formula>...complete Python module...</final_formula>"
)
def unclosed_msg(tag: str) -> str:
return (f"Your `<{tag}>` block was not closed (no `</{tag}>` after the open "
f"tag). The reply was likely truncated by the token budget — try a "
f"shorter response or split the work across turns. Re-emit a complete "
f"primitive.")
# ---- one-call dispatch -----------------------------------------------------
def step(response_text: str, sandbox: Optional[Dict[str, Any]] = None,
run_experiment: Optional[Callable[..., dict]] = None) -> Dict[str, Any]:
"""Apply the fixed protocol to ONE model turn.
Picks the first-emitted tool tag and acts on it. Returns one of:
{"action": "submit", "submission": <module text>}
{"action": "python", "feedback": <str>, "ok": <bool>}
{"action": "experiment", "feedback": <str>, "ok": <bool>}
{"action": "invalid", "feedback": <str>}
The caller appends `feedback` to the conversation and continues, or stops
when action == "submit". `run_experiment(**payload)` is only called for
`<experiment>` tags (simulator tasks); omit it for fix-data tasks.
"""
tag = first_tool_tag(response_text)
if tag == "final_formula":
ok, submitted = parse_final_formula(response_text)
if ok:
return {"action": "submit", "submission": submitted}
if tag == "python":
code = extract_python(response_text)
if code is not None:
res = run_python(code, sandbox)
return {"action": "python", "ok": res["success"],
"feedback": format_python_feedback(res)}
if tag == "experiment":
exp = parse_experiment(response_text)
if exp is not None:
if run_experiment is None:
fb = ('<experiment_output>\n{"error": "this task does not support '
'`<experiment>` (no simulator backing)."}\n</experiment_output>')
return {"action": "experiment", "ok": False, "feedback": fb}
result = run_experiment(**exp)
return {"action": "experiment", "ok": "error" not in result,
"feedback": format_experiment_feedback(result)}
unc = unclosed_tags(response_text)
return {"action": "invalid",
"feedback": unclosed_msg(unc[0]) if unc else INVALID_RESPONSE_MSG}