SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.09447 · 2026

SWE-AGI: Benchmarking Specification-Driven Software Construction with MoonBit in the Era of Autonomous Agents

Zhirui Zhang, Yu Zhang, Heung-Yeung Shum, Ziyue Liu, Hongbo Zhang, Yubin Chen, Yixuan Sun, Haoxiang Fei, Zhiyuan Bao, Zhengyu Lei, Mingkun Xiao, Zihang Ye, and 2 more

arXiv · PDF · Open in the Atlas

Code that ran

16 of 20 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 4 violated it and 0 have no contract verdict; 12 failures of our instrument; 4 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.09447.json · badge: /api/ran/2602.09447.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
moonbitlang/SWE-AGI-Eval canonical (classifier label)
classifier basis: census
4 ran · 0 no recorded run
moonbitlang/SWE-AGI community
Papers with Code archive
12 ran · 4 no recorded run
FunctionStatusWhere it lives
_extract_test_block_from_text Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("0d26355a0593d1cd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

# Construct arguments for _extract_test_block_from_text
text = '''
/// This is a test header comment
test "my_test" {
    let x = 10;
    let y = 20;
}

test "another_test" {
    let z = 30;
}
'''

test_name = "my_test"

ARGS = {
    'text': text,
    'test_name': test_name
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_find_test_file Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("3e52f0603142fd16")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned PosixPath. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path
import tempfile
import os

# Create a temporary directory structure for testing
temp_dir = tempfile.mkdtemp()
project_dir = Path(temp_dir)

# Create some subdirectories and files
package_dir = project_dir / "mypackage"
package_dir.mkdir(exist_ok=True)

# Create a test file in the package directory
test_file = package_dir / "test_module.py"
test_file.write_text("# test file")

# Create a test file in the root project directory
root_test_file = project_dir / "root_test.py"
root_test_file.write_text("# root test")

# Create a failure dict with filename and package
failure = {
    'filename': 'test_module.py',
    'package': 'mypackage'
}

ARGS = {
    'project_dir': str(project_dir),
    'failure': failure
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_format_failure_message Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("5523b61639d585a9")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path
import json

# Create a temporary directory and test file structure
temp_dir = tempfile.mkdtemp()
project_dir = Path(temp_dir)

# Create a test file with proper structure
test_file_path = project_dir / "test_example.txt"
test_content = '''/// This is a test header comment
test "my_test_case" {
    some validation code
    assert something
}

test "another_test" {
    more code here
}
'''
test_file_path.write_text(test_content, encoding='utf-8')

# Create a package subdirectory with another test file
package_dir = project_dir / "mypackage"
package_dir.mkdir(exist_ok=True)
package_test_file = package_dir / "test_package.txt"
package_test_content = '''test "packaged_test" {
    package test code
}
'''
package_test_file.write_text(package_test_content, encoding='utf-8')

# Create various failure dictionaries to test different code paths
failure_basic = {
    'test_name': 'my_test_case',
    'message': 'Test failed',
    'kind': 'AssertionError',
    'expected': 'True',
    'actual': 'False',
    'filename': 'test_example.txt',
    'package': None
}

failure_with_dict_message = {
    'test_name': 'packaged_test',
    'message': {'error': 'validation failed', 'code': 42},
    'kind': 'ValidationError',
    'filename': 'test_package.txt',
    'package': 'mypackage'
}

failure_minimal = {
    'test_name': 'unknown_test',
    'message': 'Something went wrong',
}

failure_empty = {}

# Build ARGS to test _format_failure_message with various inputs
ARGS = {
    'failure': failure_basic,
    'project_dir': str(project_dir)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_is_selection_error Ran
ran, violated the contract we drafted
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("5f4997aeaa42de77")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

ARGS = {
    "message": "Error: not unique"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_parse_failure_jsonl Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("b5f7c09f9a07afea")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json

# The function _parse_failure_jsonl takes a single parameter: output
# which should be a string containing newline-separated JSON lines.
# Each line should be parseable as JSON and represent a dict with a 'test_name' key.

output = """{"test_name": "test_foo", "status": "failed", "error": "assertion error"}
{"test_name": "test_bar", "status": "passed"}

{"test_name": "test_baz", "error": "timeout"}
invalid json line
{"not_a_dict": "value"}
{"test_name": "test_qux"}
"""

ARGS = {
    "output": output
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
_parse_test_summary Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("2d84a21c452eeb6f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

SUMMARY_RE = re.compile('Total tests:\\s*(\\d+),\\s*passed:\\s*(\\d+),\\s*failed:\\s*(\\d+)')

def _parse_test_summary(output):
    match = SUMMARY_RE.search(output)
    if not match:
        return None
    return {'total': int(match.group(1)), 'passed': int(match.group(2)), 'failed': int(match.group(3))}


# Test case 1: Valid output with test summary
output_valid = "Running tests...\nTotal tests: 42, passed: 38, failed: 4\nDone."

# Test case 2: Output without summary
output_invalid = "Running tests...\nNo summary found.\nDone."

ARGS = {
    'output': output_valid
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
collect_labels Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI-Eval/claude-opus-4.5/cdcl/try.py
code served (permissive licence) · get_code("48b9834c9dd08085")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def collect_labels(text: str) -> list[str]:
    labels: list[str] = []
    for line in text.splitlines():
        stripped = line.lstrip()
        if stripped.startswith('//'):
            continue
        match = re.match('^\\s*test\\s+"([^"]+)"', line)
        if match:
            labels.append(match.group(1))
    return labels


# Construct test input
test_text = """
// This is a comment
    test "label1"
  test "label2"
test "label3"
    // another comment
   test "label4"
some other line
"""

ARGS = {
    "text": test_text
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
colorize Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI-Eval/claude-opus-4.5/cdcl/try.py
code served (permissive licence) · get_code("bd08390f95c0fe75")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import sys

class Colors:
    GREEN = '\x1b[92m'
    RED = '\x1b[91m'
    YELLOW = '\x1b[93m'
    BLUE = '\x1b[94m'
    RESET = '\x1b[0m'
    BOLD = '\x1b[1m'

def supports_color() -> bool:
    if not hasattr(sys.stdout, 'isatty'):
        return False
    if not sys.stdout.isatty():
        return False
    return True

USE_COLOR = supports_color()

def colorize(text: str, color: str) -> str:
    if USE_COLOR:
        return f'{color}{text}{Colors.RESET}'
    return text

ARGS = {
    'text': 'Hello, World!',
    'color': Colors.GREEN
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
copy_project Ran
ran, violated the contract we drafted
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("d47e662df259f5cf")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import shutil
import tempfile
from pathlib import Path

# Create temporary directories for testing
src_dir = tempfile.mkdtemp(prefix='src_')
dst_dir = tempfile.mkdtemp(prefix='dst_')

# Set up source directory with some test files
src_path = Path(src_dir)
(src_path / 'file1.txt').write_text('content1')
(src_path / 'subdir').mkdir(exist_ok=True)
(src_path / 'subdir' / 'file2.txt').write_text('content2')

# Set up destination directory with some files (including ones to be cleaned)
dst_path = Path(dst_dir)
(dst_path / 'old_file.txt').write_text('old_content')
(dst_path / 'old_subdir').mkdir(exist_ok=True)
(dst_path / 'old_subdir' / 'old_file2.txt').write_text('old_content2')

# Create a _priv_test directory that should be preserved
priv_test_dir = dst_path / 'data_priv_test'
priv_test_dir.mkdir(exist_ok=True)
(priv_test_dir / 'priv_test.mbt').write_text('private_data')

ARGS = {
    'src_dir': src_dir,
    'dst_dir': dst_dir,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
generate_moonbit Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI-Eval/deepseek-v3.2/hpack/parse_huffman.py
code served (permissive licence) · get_code("7a0a58d0e71fa07f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# Analyze the function: it takes 'entries' which is iterated over
# Each entry is either None or a tuple of (bits, hexcode, length)
# bits is a string representation of bits
# hexcode is a hex string
# length is an integer

# Build sample entries for Huffman table
# RFC 7541 Huffman codes have symbols 0-255 plus EOS symbol 256
# Each entry should be (bits_string, hex_string, length_int)

entries = [
    ('00000', '0', 5),           # symbol 0
    ('00001', '1', 5),           # symbol 1
    ('00010', '2', 5),           # symbol 2
    None,                         # symbol 3 - missing
    ('00011', '3', 5),           # symbol 4
    ('00100', '4', 5),           # symbol 5
    ('00101', '5', 5),           # symbol 6
    ('00110', '6', 5),           # symbol 7
    ('00111', '7', 5),           # symbol 8
    ('01000', '8', 5),           # symbol 9
    ('01001', '9', 5),           # symbol 10
    ('01010', 'a', 5),           # symbol 11
    ('01011', 'b', 5),           # symbol 12
    None,                         # symbol 13 - missing
    ('01100', 'c', 5),           # symbol 14
    ('01101', 'd', 5),           # symbol 15
    ('01110', 'e', 5),           # symbol 16
    ('01111', 'f', 5),           # symbol 17
    ('10000', '10', 6),          # symbol 18
    ('10001', '11', 6),          # symbol 19
    ('10010', '12', 6),          # symbol 20
]

ARGS = {
    'entries': entries
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run_cdcl_test_streaming Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("e49ed3ddb63cda58")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import sys
import io
import threading
import tempfile
from unittest.mock import Mock, MagicMock

# Create a mock wfile object that behaves like an HTTP response file
wfile = io.BytesIO()

# Create a temporary project directory
project_dir = tempfile.mkdtemp()

# Basic parameters
timeout = 10800
per_test_timeout = 10
keepalive_interval = 15
test_name = "test_example"
test_file = "test_file.py"
request_id = "req_12345"

# Create a threading.Event for the cancelled flag
cancelled = threading.Event()
cancelled.clear()  # Not cancelled

# Construct the ARGS dictionary
ARGS = {
    'wfile': wfile,
    'project_dir': project_dir,
    'timeout': timeout,
    'per_test_timeout': per_test_timeout,
    'keepalive_interval': keepalive_interval,
    'test_name': test_name,
    'test_file': test_file,
    'request_id': request_id,
    'cancelled': cancelled
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
run_moon_build Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("db5f1c7ce134271e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import subprocess
import signal
import threading
from datetime import datetime
from unittest.mock import MagicMock
import tempfile

# Create a temporary directory to use as project_dir
temp_dir = tempfile.mkdtemp()

# Create a mock cancelled event
cancelled_event = threading.Event()

# Define ARGS dict with all parameters for run_moon_build
ARGS = {
    'project_dir': temp_dir,
    'timeout': 120,
    'request_id': 'test_request_123',
    'cancelled': cancelled_event
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
send_sse_comment Ran
ran, violated the contract we drafted
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("bdd2704909cc54c3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import io

# Create a mock file-like object that supports write() and flush()
class MockWFile:
    def __init__(self):
        self.buffer = io.BytesIO()
    
    def write(self, data):
        self.buffer.write(data)
    
    def flush(self):
        pass

wfile = MockWFile()
comment = "keep-alive"

ARGS = {
    "wfile": wfile,
    "comment": comment,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
send_sse_event Ran
ran, violated the contract we drafted
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("c39d6db42bcfe568")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import io

# Create a mock file-like object that supports write() and flush()
class MockWFile:
    def __init__(self):
        self.buffer = io.BytesIO()
    
    def write(self, data):
        self.buffer.write(data)
    
    def flush(self):
        pass

# Construct valid arguments
wfile = MockWFile()
event_type = "update"
data = {
    "status": "processing",
    "progress": 42,
    "message": "Task in progress",
    "nested": {"key": "value"}
}

ARGS = {
    "wfile": wfile,
    "event_type": event_type,
    "data": data
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
status_symbol Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI-Eval/claude-opus-4.5/cdcl/try.py
code served (permissive licence) · get_code("6fed0709f2c2de4c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import sys

class Colors:
    GREEN = '\x1b[92m'
    RED = '\x1b[91m'
    YELLOW = '\x1b[93m'
    BLUE = '\x1b[94m'
    RESET = '\x1b[0m'
    BOLD = '\x1b[1m'

def supports_color() -> bool:
    if not hasattr(sys.stdout, 'isatty'):
        return False
    if not sys.stdout.isatty():
        return False
    return True

USE_COLOR = supports_color()

def colorize(text: str, color: str) -> str:
    if USE_COLOR:
        return f'{color}{text}{Colors.RESET}'
    return text

def status_symbol(status: str) -> str:
    if status == 'pass':
        return colorize('PASS', Colors.GREEN)
    if status == 'fail':
        return colorize('FAIL', Colors.RED)
    if status == 'timeout':
        return colorize('TIMEOUT', Colors.YELLOW)
    if status == 'oom':
        return colorize('OOM', Colors.YELLOW)
    return colorize(status.upper(), Colors.YELLOW)

ARGS = {
    'status': 'pass'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
terminate_process_gracefully Ran
ran; our contract draft was wrong, not the code
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("719368dd9fb84c7b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import subprocess
import os
import signal

# Create a simple subprocess that we can terminate
# We'll create a process that sleeps for a long time
proc = subprocess.Popen(
    ['sleep', '100'],
    stdout=subprocess.PIPE,
    stderr=subprocess.PIPE,
    preexec_fn=os.setsid  # Create a new process group so we can kill it
)

ARGS = {
    'proc': proc,
    'grace_period': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
handle_proj_submit_streaming Not yet run
unverified (harvested, no recorded run)
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("52c4ad71a7b7b591")
How it ran

no invocation recorded.

log Not yet run
unverified (harvested, no recorded run)
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("6c8362c015ee20dd")
How it ran

no invocation recorded.

register_process Not yet run
unverified (harvested, no recorded run)
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("786d0d4d4541702b")
How it ran

no invocation recorded.

run_moon_test Not yet run
unverified (harvested, no recorded run)
moonbitlang/SWE-AGI/docker/server.py
code served (permissive licence) · get_code("f0c203f3c5165c98")
How it ran

no invocation recorded.

Also present in other papers' code

Of 20 samples lifted from this paper's linked repositories, 20 carry a function identity and 0 do not. Of the 20 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 20 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Although large language models (LLMs) have demonstrated impressive coding capabilities, their ability to autonomously build production-scale software from explicit specifications remains an open question. In this paper, we introduce SWE-AGI, the first open-source benchmark for evaluating end-to-end, specification-driven construction of software systems written in MoonBit. SWE-AGI tasks require LLM-based agents to implement a range of software systems, including parsers, interpreters, binary decoders, and SAT solvers, strictly from authoritative standards and RFCs under a fixed API scaffold. Each task involves implementing 10 3 -10 4 lines of core logic, corresponding to weeks or months of engineering effort for an experienced human developer. By leveraging the nascent MoonBit ecosystem, SWE-AGI minimizes data leakage, forcing agents to rely on long-horizon architectural reasoning rather than code retrieval. Across frontier models, gpt-5.3-codex achieves the best overall performance (solving 19/22 tasks, 86.4%), outperforming claude-opus-4.6 (15/22, 68.2%), and kimi-2.5 exhibits the strongest performance among open-source models. However, performance degrades sharply with increasing task difficulty, particularly on hard, specification-intensive systems. Behavioral analysis further reveals that as codebases scale, code reading, rather than writing, becomes the dominant bottleneck in AI-assisted development. Overall, while specification-driven autonomous software engineering is increasingly viable, substantial challenges remain before it can reliably support production-scale development.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2602.09447")
get_code_for_paper("2602.09447")
have("2602.09447")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2602.09447.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2602.09447.svg)](https://syntology.ai/paper/2602.09447)

Connect an agent — have() is free.