diff --git a/README.ja.md b/README.ja.md
index e2f27a3e..2096735a 100644
--- a/README.ja.md
+++ b/README.ja.md
@@ -52,6 +52,7 @@ $ oj test [-c COMMAND] [TEST...]
 $ oj test-reactive [-c COMMAND] JUDGE_COMMAND
 $ oj generate-input GENERATOR_COMMAND
 $ oj generate-output [-c COMMAND] [TEST...]
+$ oj generate-reactive [-g GENERATOR_COMMAND] HACKED_COMMAND JUDGE_COMMAND
 ```
 
 詳細は `$ oj --help` を見てください。
diff --git a/README.md b/README.md
index d478632f..6c4dce5b 100644
--- a/README.md
+++ b/README.md
@@ -51,6 +51,7 @@ $ oj test [-c COMMAND] [TEST...]
 $ oj test-reactive [-c COMMAND] JUDGE_COMMAND
 $ oj generate-input GENERATOR_COMMAND
 $ oj generate-output [-c COMMAND] [TEST...]
+$ oj generate-reactive [-g GENERATOR_COMMAND] HACKED_COMMAND JUDGE_COMMAND
 ```
 
 For details, see `$ oj --help`.
diff --git a/onlinejudge_command/main.py b/onlinejudge_command/main.py
index 1bd14509..0d3b7e10 100644
--- a/onlinejudge_command/main.py
+++ b/onlinejudge_command/main.py
@@ -12,6 +12,7 @@
 import onlinejudge_command.subcommand.download as subcommand_download
 import onlinejudge_command.subcommand.generate_input as subcommand_generate_input
 import onlinejudge_command.subcommand.generate_output as subcommand_generate_output
+import onlinejudge_command.subcommand.generate_reactive as subcommand_generate_reactive
 import onlinejudge_command.subcommand.login as subcommand_login
 import onlinejudge_command.subcommand.submit as subcommand_submit
 import onlinejudge_command.subcommand.test as subcommand_test
@@ -42,6 +43,7 @@ def get_parser() -> argparse.ArgumentParser:
     subcommand_test.add_subparser(subparsers)
     subcommand_generate_output.add_subparser(subparsers)
     subcommand_generate_input.add_subparser(subparsers)
+    subcommand_generate_reactive.add_subparser(subparsers)
     subcommand_test_reactive.add_subparser(subparsers)
 
     return parser
@@ -76,6 +78,8 @@ def run_program(args: argparse.Namespace, parser: argparse.ArgumentParser) -> in
         subcommand_generate_output.run(args)
     elif args.subcommand in ['generate-input', 'g/i']:
         subcommand_generate_input.run(args)
+    elif args.subcommand in ['generate-reactive', 'g/r']:
+        subcommand_generate_reactive.run(args)
     else:
         parser.print_help(file=sys.stderr)
         return 1
diff --git a/onlinejudge_command/subcommand/generate_reactive.py b/onlinejudge_command/subcommand/generate_reactive.py
new file mode 100644
index 00000000..c5e319cf
--- /dev/null
+++ b/onlinejudge_command/subcommand/generate_reactive.py
@@ -0,0 +1,270 @@
+import argparse
+import concurrent.futures
+import contextlib
+import enum
+import hashlib
+import itertools
+import os
+import pathlib
+import subprocess
+import sys
+import tempfile
+import threading
+from logging import getLogger
+from typing import *
+
+import onlinejudge_command.format_utils as fmtutils
+from onlinejudge_command import pretty_printers, utils
+
+logger = getLogger(__name__)
+
+
+def add_subparser(subparsers: argparse.Action) -> None:
+    subparsers_add_parser: Callable[..., argparse.ArgumentParser] = subparsers.add_parser  # type: ignore
+    subparser = subparsers_add_parser('generate-reactive', aliases=['g/r'], help='generate input files for reactive problems from given generator', formatter_class=argparse.RawTextHelpFormatter, epilog='''\
+format string for --format:
+  %s                    name
+  %e                    extension: "in" or "out"
+  (both %d and %e are required.)
+
+tips:
+  For the random testing, you can read a tutorial: https://github.com/online-judge-tools/oj/blob/master/docs/getting-started.md#random-testing
+
+  There is a command to automatically generate a input generator, `oj-template` command. See https://github.com/online-judge-tools/template-generator .
+
+  This subcommand has also the feature to find a hack case.
+    e.g. for a target program `a.out`, and a reactive judge program `judge.py` which generates a random input-case by itself, run $ oj g/r ./a.out 'python3 judge.py'
+    For this case, `judge.py` should print debugging information.
+
+    e.g. for a target program `a.out`, a random input-case generator `generate.py`, and a reactive judge program `judge.py`, run $ oj g/r -g 'python3 generate.py' ./a.out 'python3 judge.py'
+
+  You can do similar things with shell
+    e.g. $ for i in `seq 100` ; do [[ -e a.pipe ]] && rm -f a.pipe ; mkfifo a.pipe && ./a.out < a.pipe | python3 judge.py > a.pipe; done
+    e.g. $ for i in `seq 100` ; do python3 generate.py > test/random-$i.in ; [[ -e a.pipe ]] && rm -f a.pipe ; mkfifo a.pipe && ./a.out < a.pipe | python3 judge.py test/random-$i.in > a.pipe; done
+''')
+    subparser.add_argument('-f', '--format', default='%s.%e', help='a format string to recognize the relationship of test cases. (default: "%%s.%%e")')
+    subparser.add_argument('-d', '--directory', type=pathlib.Path, default=pathlib.Path('test'), help='a directory name for test cases (default: test/)')
+    subparser.add_argument('-t', '--tle', type=float, help='set the time limit (in second) (default: inf)')
+    subparser.add_argument('-j', '--jobs', type=int, help='run tests in parallel')
+    subparser.add_argument('--width', type=int, default=3, help='specify the width of indices of cases. (default: 3)')
+    subparser.add_argument('--name', help='specify the base name of cases. (default: "random")')
+    subparser.add_argument('-g', '--generator', type=str, help='your program to generate test cases')
+    subparser.add_argument('hack', help='specify your wrong solution to be judged with the reactive program')
+    subparser.add_argument('judge', type=str, help='judge program using standard I/O')
+    subparser.add_argument('count', nargs='?', type=int, help='the number of cases to generate (default: 100)')
+
+
+@contextlib.contextmanager
+def BufferedExecutor(lock: Optional[threading.Lock]):
+    buf: List[Tuple[Callable, List[Any], Dict[str, Any]]] = []
+
+    def submit(f, *args, **kwargs):
+        nonlocal buf
+        if lock is None:
+            f(*args, **kwargs)
+        else:
+            buf += [(f, args, kwargs)]
+
+    result = yield submit
+
+    if lock is not None:
+        with lock:
+            for f, args, kwargs in buf:
+                f(*args, **kwargs)
+    return result
+
+
+def write_result(input_data: bytes, *, input_path: pathlib.Path, lock: Optional[threading.Lock] = None) -> None:
+    # acquire lock to print logs properly, if in parallel
+    nullcontext = contextlib.ExitStack()  # TODO: use contextlib.nullcontext after Python 3.7
+    with lock or nullcontext:
+        if not input_path.parent.is_dir():
+            os.makedirs(str(input_path.parent), exist_ok=True)
+        with input_path.open('wb') as fh:
+            fh.write(input_data)
+        logger.info(utils.SUCCESS + 'saved to: %s', input_path)
+
+
+def check_status(info: Dict[str, Any], proc: subprocess.Popen, *, submit: Callable[..., None], input_data: Optional[bytes]) -> bool:
+    submit(logger.info, 'time: %f sec', info['elapsed'])
+    if proc.returncode is None:
+        submit(logger.info, utils.FAILURE + utils.red('TLE'))
+        if input_data is not None:
+            submit(logger.info, utils.NO_HEADER + 'input:')
+            submit(logger.info, utils.NO_HEADER + '%s', pretty_printers.make_pretty_large_file_content(input_data, limit=40, head=20, tail=10))
+        submit(logger.info, 'skipped.')
+        return False
+    elif proc.returncode != 0:
+        submit(logger.info, utils.FAILURE + utils.red('RE') + ': return code %d', proc.returncode)
+        if input_data is not None:
+            submit(logger.info, utils.NO_HEADER + 'input:')
+            submit(logger.info, utils.NO_HEADER + '%s', pretty_printers.make_pretty_large_file_content(input_data, limit=40, head=20, tail=10))
+        submit(logger.info, 'skipped.')
+        return False
+    assert info['answer'] is not None
+    return True
+
+
+def check_randomness_of_generator(input_data: bytes, *, name: str, lock: Optional[threading.Lock], generated_input_hashes: Dict[bytes, str]) -> Optional[str]:
+    """check_randomness_of_generator() checks the generated inputs. This adds some overheads but is needed for foolproof. Many users forget to initialize their library and use fixed seeds.
+
+    :returns: a previous name of the input when it was already once generated. None if it's a new input.
+    """
+
+    # To prevent consuming unlimited memories, do nothing if the user's generator is properly implemented.
+    limit = 1000
+    if len(generated_input_hashes) >= limit:
+        return None
+
+    input_digest = hashlib.sha1(input_data).digest()
+    nullcontext = contextlib.ExitStack()  # TODO: use contextlib.nullcontext after Python 3.7
+    with lock or nullcontext:
+        if len(generated_input_hashes) < limit:
+            if input_digest in generated_input_hashes:
+                return generated_input_hashes[input_digest]
+            else:
+                generated_input_hashes[input_digest] = name
+                if len(generated_input_hashes) == limit:
+                    logger.info('Conflict checking of generated inputs is disabled now because it seems the given input generator has enough randomness.')  # This prints a log line but it's safe because here is in a lock.
+    return None
+
+
+class JudgeStatus(enum.Enum):
+    AC = 'AC'
+    WA = 'WA'
+    RE = 'RE'
+
+
+@contextlib.contextmanager
+def fifo() -> Generator[Tuple[Any, Any], None, None]:
+    fdr, fdw = os.pipe()
+    fhr = os.fdopen(fdr, 'r')
+    fhw = os.fdopen(fdw, 'w')
+    yield fhr, fhw
+    fhw.close()
+    fhr.close()
+    # os.close(fdw), os.close(fdr) are unnecessary
+
+
+def run_reactive(hack: str, judge: str, generated_file: Optional[pathlib.Path]) -> JudgeStatus:
+    with fifo() as (fhr1, fhw1):
+        with fifo() as (fhr2, fhw2):
+            with subprocess.Popen(hack, shell=True, stdin=fhr2, stdout=fhw1, stderr=sys.stderr) as proc1:
+                if generated_file is not None:
+                    judge_command = ' '.join([judge, str(generated_file.resolve())])
+                else:
+                    judge_command = judge
+
+                with subprocess.Popen(judge_command, shell=True, stdin=fhr1, stdout=fhw2, stderr=sys.stderr) as proc2:
+                    proc1.communicate()
+                    proc2.communicate()
+
+                    if proc1.returncode != 0:
+                        return JudgeStatus.RE
+                    elif proc2.returncode == 0:
+                        return JudgeStatus.AC
+                    else:
+                        return JudgeStatus.WA
+
+
+def try_hack_once(generator: Optional[str], hack: str, judge: str, *, tle: Optional[float], attempt: int, lock: Optional[threading.Lock] = None, generated_input_hashes: Dict[bytes, str]) -> Tuple[bool, Optional[bytes]]:
+    with BufferedExecutor(lock) as submit:
+
+        # print the header
+        submit(logger.info, '')
+        submit(logger.info, '%d-th attempt', attempt)
+
+        # generate input if generator is given
+        input_data: Optional[bytes] = None
+        if generator is not None:
+            submit(logger.info, 'generate input...')
+            info, proc = utils.exec_command(generator, stdin=None, timeout=tle)
+            input_data = info['answer']
+            if not check_status(info, proc, submit=submit, input_data=input_data):
+                return (False, None)
+            assert input_data is not None
+
+        # check the randomness of generator
+        name = '{}-th attempt'.format(attempt)
+        if input_data is not None:
+            conflicted_name = check_randomness_of_generator(input_data, name=name, lock=lock, generated_input_hashes=generated_input_hashes)
+            if conflicted_name is not None:
+                submit(logger.warning, 'The same input is already generated at %s. Please use a random input generator.', conflicted_name)
+                submit(logger.info, utils.NO_HEADER + 'input:')
+                submit(logger.info, utils.NO_HEADER + '%s', pretty_printers.make_pretty_large_file_content(input_data, limit=40, head=20, tail=10))
+
+        # hack
+        submit(logger.info, 'hack...')
+        if input_data is not None:
+            with tempfile.NamedTemporaryFile(delete=True) as fh:
+                with open(fh.name, 'wb') as fh1:
+                    fh1.write(input_data)
+                status = run_reactive(hack, judge, pathlib.Path(fh.name))
+        else:
+            status = run_reactive(hack, judge, None)
+
+        if status != JudgeStatus.AC and input_data is not None:
+            logger.info(utils.FAILURE + '' + utils.red(status.value))
+            logger.info(utils.NO_HEADER + 'input:\n%s', pretty_printers.make_pretty_large_file_content(input_data, limit=40, head=20, tail=10))
+
+        # return the result
+        if status == JudgeStatus.AC:
+            return (False, None)
+        elif generator is None:
+            return (True, None)
+        else:
+            return (True, input_data)
+
+
+def run(args: argparse.Namespace) -> None:
+    if not args.generator:
+        logger.info('--generator is not given. It will assume the reactive judge generates a random input-case each time.')
+
+    if args.name is None:
+        args.name = 'hack'
+
+    if args.count is None:
+        args.count = 1
+
+    def iterate_path():
+        for i in itertools.count():
+            name = '{}-{}'.format(args.name, str(i).zfill(args.width))
+            input_path = fmtutils.path_from_format(args.directory, args.format, name=name, ext='in')
+            if not input_path.exists():
+                yield (name, input_path)
+
+    # generate cases
+    generated_input_hashes: Dict[bytes, str] = {}
+    if args.jobs is None:
+        for _, input_path in itertools.islice(iterate_path(), args.count):
+            # hack serially
+            for attempt in itertools.count(1):
+                (hacked, input_data) = try_hack_once(args.generator, hack=args.hack, judge=args.judge, tle=args.tle, attempt=attempt, generated_input_hashes=generated_input_hashes)
+                if hacked:
+                    if input_data is not None:
+                        write_result(input_data, input_path=input_path)
+                    break
+    else:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=args.jobs) as executor:
+            lock = threading.Lock()
+            futures: List[concurrent.futures.Future] = []
+
+            # hack concurrently
+            attempt = 0
+            for _ in range(args.jobs):
+                attempt += 1
+                futures += [executor.submit(try_hack_once, args.generator, hack=args.hack, judge=args.judge, tle=args.tle, attempt=attempt, lock=lock, generated_input_hashes=generated_input_hashes)]
+            for _, input_path, in itertools.islice(iterate_path(), args.count):
+                hacked = False
+                while not hacked:
+                    concurrent.futures.wait(futures, return_when=concurrent.futures.FIRST_COMPLETED)
+                    for i in range(len(futures)):
+                        if not futures[i].done():
+                            continue
+                        [hacked, input_data] = futures[i].result()
+                        attempt += 1
+                        futures[i] = executor.submit(try_hack_once, args.generator, hack=args.hack, judge=args.judge, tle=args.tle, attempt=attempt, lock=lock, generated_input_hashes=generated_input_hashes)
+                        if hacked:
+                            break
+                if input_data is not None:
+                    write_result(input_data, input_path=input_path, lock=lock)
diff --git a/onlinejudge_command/subcommand/test_reactive.py b/onlinejudge_command/subcommand/test_reactive.py
index c31a9a64..bd78b17c 100644
--- a/onlinejudge_command/subcommand/test_reactive.py
+++ b/onlinejudge_command/subcommand/test_reactive.py
@@ -1,12 +1,17 @@
 import argparse
+import concurrent.futures
 import contextlib
+import enum
 import os
+import pathlib
 import subprocess
 import sys
+import threading
 from logging import getLogger
 from typing import *
 
-import onlinejudge_command.utils as utils
+import onlinejudge_command.format_utils as fmtutils
+from onlinejudge_command import pretty_printers, utils
 
 logger = getLogger(__name__)
 
@@ -21,7 +26,14 @@ def add_subparser(subparsers: argparse.Action) -> None:
     e.g. $ while oj t/i 'python3 judge.py' ; do : ; done
 ''')
     subparser.add_argument('-c', '--command', default=utils.get_default_command(), help='your solution to be tested. (default: "{}")'.format(utils.get_default_command()))
+    subparser.add_argument('-f', '--format', default='%s.%e', help='a format string to recognize the relationship of test cases. (default: "%%s.%%e")')
+    subparser.add_argument('-d', '--directory', type=pathlib.Path, default=pathlib.Path('test'), help='a directory name for test cases (default: test/)')
+    subparser.add_argument('-j', '--jobs', metavar='N', type=int, help='specifies the number of jobs to run simultaneously  (default: no parallelization)')
+    subparser.add_argument('-i', '--from-file', action='store_true', help='tests each case (specified by -d option or test argument), whose filename is passed to the judge as sys.argv[1]')
+    subparser.add_argument('--no-ignore-backup', action='store_false', dest='ignore_backup')
+    subparser.add_argument('--ignore-backup', action='store_true', help='ignore backup files and hidden files (i.e. files like "*~", "\\#*\\#" and ".*") (default)')
     subparser.add_argument('judge', help='judge program using standard I/O')
+    subparser.add_argument('test', nargs='*', type=pathlib.Path, help='paths of test cases. (if empty: globbed from --format)')
 
 
 @contextlib.contextmanager
@@ -36,6 +48,13 @@ def fifo() -> Generator[Tuple[Any, Any], None, None]:
 
 
 def run(args: argparse.Namespace) -> bool:
+    if not args.from_file:
+        return run_single(args)
+    else:
+        return run_from_file(args)
+
+
+def run_single(args: argparse.Namespace) -> bool:
     with fifo() as (fhr1, fhw1):
         with fifo() as (fhr2, fhw2):
             with subprocess.Popen(args.command, shell=True, stdin=fhr2, stdout=fhw1, stderr=sys.stderr) as proc1:
@@ -53,3 +72,98 @@ def run(args: argparse.Namespace) -> bool:
                         logger.info(utils.FAILURE + 'WA: judge returns %d', proc2.returncode)
                         result = False
                     return result
+
+
+class JudgeStatus(enum.Enum):
+    AC = 'AC'
+    WA = 'WA'
+    RE = 'RE'
+
+
+def run_from_file(args: argparse.Namespace) -> bool:
+    # list tests
+    if not args.test:
+        args.test = fmtutils.glob_with_format(args.directory, args.format)  # by default
+    if args.ignore_backup:
+        args.test = fmtutils.drop_backup_or_hidden_files(args.test)
+    tests = fmtutils.construct_relationship_of_files(args.test, args.directory, args.format)
+
+    # run tests
+    history: List[Dict[str, Any]] = []
+    if args.jobs is None:
+        for name, paths in sorted(tests.items()):
+            history += [test_single_case(name, paths['in'], args=args)]
+    else:
+        if os.name == 'nt':
+            logger.warning("-j/--jobs opiton is unstable on Windows environment")
+        with concurrent.futures.ThreadPoolExecutor(max_workers=args.jobs) as executor:
+            lock = threading.Lock()
+            futures: List[concurrent.futures.Future] = []
+            for name, paths in sorted(tests.items()):
+                futures += [executor.submit(test_single_case, name, paths['in'], lock=lock, args=args)]
+            for future in futures:
+                history += [future.result()]
+
+    # summarize
+    ac_count = 0
+    for result in history:
+        if result['status'] == 'AC':
+            ac_count += 1
+
+    # print the summary
+    logger.info('')
+    if ac_count == len(tests):
+        logger.info(utils.SUCCESS + 'test ' + utils.green('success') + ': %d cases', len(tests))
+    else:
+        logger.info(utils.FAILURE + 'test ' + utils.red('failed') + ': %d AC / %d cases', ac_count, len(tests))
+
+    # return the result
+    return ac_count == len(tests)
+
+
+def test_single_case(test_name: str, test_input_path: pathlib.Path, *, lock: Optional[threading.Lock] = None, args: argparse.Namespace) -> Dict[str, Any]:
+    # print the header earlier if not in parallel
+    if lock is None:
+        logger.info('')
+        logger.info('%s', test_name)
+
+    with fifo() as (fhr1, fhw1):
+        with fifo() as (fhr2, fhw2):
+            with subprocess.Popen(args.command, shell=True, stdin=fhr2, stdout=fhw1, stderr=sys.stderr) as proc1:
+                judge_command = ' '.join([
+                    args.judge,
+                    str(test_input_path.resolve()),
+                ])
+                with subprocess.Popen(judge_command, shell=True, stdin=fhr1, stdout=fhw2, stderr=sys.stderr) as proc2:
+                    proc1.communicate()
+                    proc2.communicate()
+
+                    # lock is require to avoid mixing logs if in parallel
+                    nullcontext = contextlib.ExitStack()  # TODO: use contextlib.nullcontext() after updating Python to 3.7
+                    with lock or nullcontext:
+                        if lock is not None:
+                            logger.info('')
+                            logger.info('%s', test_name)
+
+                    status = JudgeStatus.AC
+                    if proc1.returncode != 0:
+                        logger.info(utils.FAILURE + 'RE: solution returns %d', proc1.returncode)
+                        status = JudgeStatus.RE
+                    if proc2.returncode == 0:
+                        logger.info(utils.SUCCESS + 'AC')
+                    else:
+                        logger.info(utils.FAILURE + 'WA: judge returns %d', proc2.returncode)
+                        status = JudgeStatus.WA
+
+                    if status != JudgeStatus.AC:
+                        with test_input_path.open('rb') as inf:
+                            logger.info(utils.NO_HEADER + 'input:\n%s', pretty_printers.make_pretty_large_file_content(inf.read(), limit=40, head=20, tail=10))
+
+                    testcase = {
+                        'name': test_name,
+                        'input': str(test_input_path.resolve()),
+                    }
+                    return {
+                        'status': status.value,
+                        'testcase': testcase,
+                    }