diff --git a/.pylintrc b/.pylintrc
index dd792729..769ae821 100644
--- a/.pylintrc
+++ b/.pylintrc
@@ -1,5 +1,5 @@
 [MASTER]
-init-hook='sys.path.insert(0, "mocks")'
+init-hook='sys.path.append("mocks"); sys.path.append("src/lib")'
 
 [MESSAGES CONTROL]
-disable=C0111,C0103,W0603
+disable=C0111,C0103,W0603,W0703
diff --git a/src/lib/protobuf.py b/src/lib/protobuf.py
index 1fb47e95..f97bcd7f 100644
--- a/src/lib/protobuf.py
+++ b/src/lib/protobuf.py
@@ -1,150 +1,53 @@
 '''
-Streaming protobuf codec.
-
-Handles asynchronous encoding and decoding of protobuf value streams.
-
-Value format: ((field_name, field_type, field_flags), field_value)
-    field_name (str):  Field name string.
-    field_type (Type): Subclass of Type.
-    field_flags (int): Field bit flags: `FLAG_REPEATED`.
-    field_value (Any): Depends on field_type.
-                       MessageTypes have `field_value == None`.
-
-Type classes are either scalar or message-like.  `load()` generators of
-scalar types return the value, message types stream it to a target
-generator as described above.  All types can be loaded and dumped
-synchronously with `loads()` and `dumps()`.
+Extremely minimal streaming codec for a subset of protobuf.  Supports uint32,
+bytes, string, embedded message and repeated fields.
 '''
 
 from micropython import const
-from streams import StreamReader, BufferWriter
+
+_UVARINT_BUFFER = bytearray(1)
 
 
-def build_message(msg_type, callback=None, *args):
-    msg = msg_type()
-    try:
-        while True:
-            field, fvalue = yield
-            fname, ftype, fflags = field
-            if issubclass(ftype, MessageType):
-                fvalue = yield from build_message(ftype)
-            if fflags & FLAG_REPEATED:
-                prev_value = getattr(msg, fname, [])
-                prev_value.append(fvalue)
-                fvalue = prev_value
-            setattr(msg, fname, fvalue)
-    except EOFError:
-        fill_missing_fields(msg)
-        if callback is not None:
-            callback(msg, *args)
-        return msg
+async def load_uvarint(reader):
+    buffer = _UVARINT_BUFFER
+    result = 0
+    shift = 0
+    byte = 0x80
+    while byte & 0x80:
+        await reader.readinto(buffer)
+        byte = buffer[0]
+        result += (byte & 0x7F) << shift
+        shift += 7
+    return result
 
 
-def fill_missing_fields(msg):
-    for tag in msg.FIELDS:
-        field = msg.FIELDS[tag]
-        if not hasattr(msg, field[0]):
-            setattr(msg, field[0], None)
+async def dump_uvarint(writer, n):
+    buffer = _UVARINT_BUFFER
+    shifted = True
+    while shifted:
+        shifted = n >> 7
+        buffer[0] = (n & 0x7F) | (0x80 if shifted else 0x00)
+        await writer.write(buffer)
+        n = shifted
 
 
-class Type:
-
-    @classmethod
-    def loads(cls, value):
-        source = StreamReader(value, len(value))
-        loader = cls.load(source)
-        try:
-            while True:
-                loader.send(None)
-        except StopIteration as e:
-            return e.value
-
-    @classmethod
-    def dumps(cls, value):
-        target = BufferWriter()
-        dumper = cls.dump(value, target)
-        try:
-            while True:
-                dumper.send(None)
-        except StopIteration:
-            return target.buffer
-
-
-_uvarint_buffer = bytearray(1)
-
-
-class UVarintType(Type):
+class UVarintType:
     WIRE_TYPE = 0
 
-    @staticmethod
-    async def load(source):
-        value, shift, quantum = 0, 0, 0x80
-        while quantum & 0x80:
-            await source.read_into(_uvarint_buffer)
-            quantum = _uvarint_buffer[0]
-            value = value + ((quantum & 0x7F) << shift)
-            shift += 7
-        return value
 
-    @staticmethod
-    async def dump(value, target):
-        shifted = True
-        while shifted:
-            shifted = value >> 7
-            _uvarint_buffer[0] = (value & 0x7F) | (0x80 if shifted else 0x00)
-            await target.write(_uvarint_buffer)
-            value = shifted
-
-
-class BoolType(Type):
+class BoolType:
     WIRE_TYPE = 0
 
-    @staticmethod
-    async def load(source):
-        return await UVarintType.load(source) != 0
 
-    @staticmethod
-    async def dump(value, target):
-        await target.write(b'\x01' if value else b'\x00')
-
-
-class BytesType(Type):
+class BytesType:
     WIRE_TYPE = 2
 
-    @staticmethod
-    async def load(source):
-        size = await UVarintType.load(source)
-        data = bytearray(size)
-        await source.read_into(data)
-        return data
 
-    @staticmethod
-    async def dump(value, target):
-        await UVarintType.dump(len(value), target)
-        await target.write(value)
-
-
-class UnicodeType(Type):
+class UnicodeType:
     WIRE_TYPE = 2
 
-    @staticmethod
-    async def load(source):
-        size = await UVarintType.load(source)
-        data = bytearray(size)
-        await source.read_into(data)
-        return str(data, 'utf-8')
 
-    @staticmethod
-    async def dump(value, target):
-        data = bytes(value, 'utf-8')
-        await UVarintType.dump(len(data), target)
-        await target.write(data)
-
-
-FLAG_REPEATED = const(1)
-
-
-class MessageType(Type):
+class MessageType:
     WIRE_TYPE = 2
     FIELDS = {}
 
@@ -159,61 +62,141 @@ class MessageType(Type):
     def __repr__(self):
         return '<%s: %s>' % (self.__class__.__name__, self.__dict__)
 
-    @classmethod
-    async def load(cls, source=None, target=None):
-        if target is None:
-            target = build_message(cls)
-        if source is None:
-            source = StreamReader()
-        try:
-            while True:
-                fkey = await UVarintType.load(source)
-                ftag = fkey >> 3
-                wtype = fkey & 7
-                if ftag in cls.FIELDS:
-                    field = cls.FIELDS[ftag]
-                    ftype = field[1]
-                    if wtype != ftype.WIRE_TYPE:
-                        raise TypeError(
-                            'Value of tag %s has incorrect wiretype %s, %s expected.' %
-                            (ftag, wtype, ftype.WIRE_TYPE))
-                else:
-                    ftype = {0: UVarintType, 2: BytesType}[wtype]
-                    await ftype.load(source)
-                    continue
-                if issubclass(ftype, MessageType):
-                    flen = await UVarintType.load(source)
-                    slen = source.set_limit(flen)
-                    target.send((field, None))
-                    await ftype.load(source, target)
-                    source.set_limit(slen)
-                else:
-                    fvalue = await ftype.load(source)
-                    target.send((field, fvalue))
-        except EOFError as e:
-            try:
-                target.throw(e)
-            except StopIteration as e:
-                return e.value
 
-    @classmethod
-    async def dump(cls, msg, target):
-        for ftag in cls.FIELDS:
-            fname, ftype, fflags = cls.FIELDS[ftag]
-            fvalue = getattr(msg, fname, None)
-            if fvalue is None:
-                continue
-            key = (ftag << 3) | ftype.WIRE_TYPE
-            if fflags & FLAG_REPEATED:
-                for svalue in fvalue:
-                    await UVarintType.dump(key, target)
-                    if issubclass(ftype, MessageType):
-                        await BytesType.dump(ftype.dumps(svalue), target)
-                    else:
-                        await ftype.dump(svalue, target)
+class LimitedReader:
+
+    def __init__(self, reader, limit):
+        self.reader = reader
+        self.limit = limit
+
+    async def readinto(self, buf):
+        if self.limit < len(buf):
+            raise EOFError
+        else:
+            nread = await self.reader.readinto(buf)
+            self.limit -= nread
+            return nread
+
+
+class CountingWriter:
+
+    def __init__(self):
+        self.size = 0
+
+    async def write(self, buf):
+        nwritten = len(buf)
+        self.size += nwritten
+        return nwritten
+
+
+FLAG_REPEATED = const(1)
+
+
+async def load_message(reader, msg_type):
+    fields = msg_type.FIELDS
+    msg = msg_type()
+
+    while True:
+        try:
+            fkey = await load_uvarint(reader)
+        except EOFError:
+            break  # no more fields to load
+
+        ftag = fkey >> 3
+        wtype = fkey & 7
+
+        field = fields.get(ftag, None)
+
+        if field is None:  # unknown field, skip it
+            if wtype == 0:
+                await load_uvarint(reader)
+            elif wtype == 2:
+                ivalue = await load_uvarint(reader)
+                await reader.readinto(bytearray(ivalue))
             else:
-                await UVarintType.dump(key, target)
-                if issubclass(ftype, MessageType):
-                    await BytesType.dump(ftype.dumps(fvalue), target)
-                else:
-                    await ftype.dump(fvalue, target)
+                raise ValueError
+            continue
+
+        fname, ftype, fflags = field
+        if wtype != ftype.WIRE_TYPE:
+            raise TypeError  # parsed wire type differs from the schema
+
+        ivalue = await load_uvarint(reader)
+
+        if ftype is UVarintType:
+            fvalue = ivalue
+        elif ftype is BoolType:
+            fvalue = bool(ivalue)
+        elif ftype is BytesType:
+            fvalue = bytearray(ivalue)
+            await reader.readinto(fvalue)
+        elif ftype is UnicodeType:
+            fvalue = bytearray(ivalue)
+            await reader.readinto(fvalue)
+            fvalue = str(fvalue, 'utf8')
+        elif issubclass(ftype, MessageType):
+            fvalue = await load_message(LimitedReader(reader, ivalue), ftype)
+        else:
+            raise TypeError  # field type is unknown
+
+        if fflags & FLAG_REPEATED:
+            pvalue = getattr(msg, fname, [])
+            pvalue.append(fvalue)
+            fvalue = pvalue
+        setattr(msg, fname, fvalue)
+
+    # fill missing fields
+    for tag in msg.FIELDS:
+        field = msg.FIELDS[tag]
+        if not hasattr(msg, field[0]):
+            setattr(msg, field[0], None)
+
+    return msg
+
+
+async def dump_message(writer, msg):
+    repvalue = [0]
+    mtype = msg.__class__
+    fields = mtype.FIELDS
+
+    for ftag in fields:
+        field = fields[ftag]
+        fname = field[0]
+        ftype = field[1]
+        fflags = field[2]
+
+        fvalue = getattr(msg, fname, None)
+        if fvalue is None:
+            continue
+
+        fkey = (ftag << 3) | ftype.WIRE_TYPE
+
+        if not fflags & FLAG_REPEATED:
+            repvalue[0] = fvalue
+            fvalue = repvalue
+
+        for svalue in fvalue:
+            await dump_uvarint(writer, fkey)
+
+            if ftype is UVarintType:
+                await dump_uvarint(writer, svalue)
+
+            elif ftype is BoolType:
+                await dump_uvarint(writer, int(svalue))
+
+            elif ftype is BytesType:
+                await dump_uvarint(writer, len(svalue))
+                await writer.write(svalue)
+
+            elif ftype is UnicodeType:
+                await dump_uvarint(writer, len(svalue))
+                await writer.write(bytes(svalue, 'utf8'))
+
+            elif issubclass(ftype, MessageType):
+                counter = CountingWriter()
+                await dump_message(counter, svalue)
+                await dump_uvarint(writer, counter.size)
+                await dump_message(writer, svalue)
+
+            else:
+                raise TypeError
diff --git a/src/lib/streams.py b/src/lib/streams.py
deleted file mode 100644
index 09d02979..00000000
--- a/src/lib/streams.py
+++ /dev/null
@@ -1,65 +0,0 @@
-from trezor.utils import memcpy
-
-
-class StreamReader:
-
-    def __init__(self, buffer=None, limit=None):
-        if buffer is None:
-            buffer = bytearray()
-        self._buffer = buffer
-        self._limit = limit
-        self._ofs = 0
-
-    async def read_into(self, dst):
-        '''
-        Read exactly `len(dst)` bytes into writable buffer-like `dst`.
-
-        Raises `EOFError` if the internal limit was reached or the
-        backing IO strategy signalled an EOF.
-        '''
-        n = len(dst)
-
-        if self._limit is not None:
-            if self._limit < n:
-                raise EOFError()
-            self._limit -= n
-
-        buf = self._buffer
-        ofs = self._ofs
-        i = 0
-        while i < n:
-            if ofs >= len(buf):
-                buf = yield
-                ofs = 0
-            # memcpy caps on the buffer lengths, no need for exact byte count
-            nb = memcpy(dst, i, buf, ofs, n)
-            ofs += nb
-            i += nb
-        self._buffer = buf
-        self._ofs = ofs
-
-    def set_limit(self, n):
-        '''
-        Makes this reader to signal EOF after reading `n` bytes.
-
-        Returns the number of bytes that the reader can read after
-        raising EOF (intended to be restored with another call to
-        `set_limit`).
-        '''
-        if self._limit is not None and n is not None:
-            rem = self._limit - n
-        else:
-            rem = None
-        self._limit = n
-        return rem
-
-
-class BufferWriter:
-
-    def __init__(self, buffer=None):
-        if buffer is None:
-            buffer = bytearray()
-        self.buffer = buffer
-
-    async def write(self, b):
-        self.buffer.extend(b)
diff --git a/src/main.py b/src/main.py
index 1b6077e1..f0c43214 100644
--- a/src/main.py
+++ b/src/main.py
@@ -5,6 +5,8 @@ from trezor import config
 from trezor import msg
 from trezor import ui
 from trezor import wire
+from trezor import loop
+from trezor.wire import codec_v2
 
 config.init()
 
diff --git a/src/trezor/loop.py b/src/trezor/loop.py
index 8ac38921..d634159a 100644
--- a/src/trezor/loop.py
+++ b/src/trezor/loop.py
@@ -103,11 +103,11 @@ def run_forever():
             log_delay_rb[log_delay_pos] = delay
             log_delay_pos = (log_delay_pos + 1) % log_delay_rb_len
 
-        if trezormsg.poll(_paused_tasks, msg_entry, delay):
+        if io.poll(_paused_tasks, msg_entry, delay):
             # message received, run tasks paused on the interface
             msg_tasks = _paused_tasks.pop(msg_entry[0], ())
             for task in msg_tasks:
-                _step_task(task, msg_entry[1])
+                    _step_task(task, msg_entry[1])
         else:
             # timeout occurred, run the first scheduled task
             if _scheduled_tasks:
@@ -292,6 +292,72 @@ class Wait(Syscall):
             raise
 
 
+class Put(Syscall):
+
+    def __init__(self, chan, value=None):
+        self.chan = chan
+        self.value = value
+
+    def __call__(self, value):
+        self.value = value
+        return self
+
+    def handle(self, task):
+        self.chan.schedule_put(schedule_task, task, self.value)
+
+
+class Take(Syscall):
+
+    def __init__(self, chan):
+        self.chan = chan
+
+    def __call__(self):
+        return self
+
+    def handle(self, task):
+        if self.chan.schedule_take(schedule_task, task) and self.chan.id is not None:
+            _pause_task(self.chan, self.chan.id)
+
+
+class Chan:
+
+    def __init__(self, id=None):
+        self.id = id
+        self.putters = []
+        self.takers = []
+        self.put = Put(self)
+        self.take = Take(self)
+
+    def schedule_publish(self, schedule, value):
+        if self.takers:
+            for taker in self.takers:
+                schedule(taker, value)
+            self.takers.clear()
+            return True
+        else:
+            return False
+
+    def schedule_put(self, schedule, putter, value):
+        if self.takers:
+            taker = self.takers.pop(0)
+            schedule(taker, value)
+            schedule(putter, value)
+            return True
+        else:
+            self.putters.append((putter, value))
+            return False
+
+    def schedule_take(self, schedule, taker):
+        if self.putters:
+            putter, value = self.putters.pop(0)
+            schedule(taker, value)
+            schedule(putter, value)
+            return True
+        else:
+            self.takers.append(taker)
+            return False
+
+
 select = Select
 sleep = Sleep
 wait = Wait
diff --git a/src/trezor/main.py b/src/trezor/main.py
index 2dfa6304..dfafbcfd 100644
--- a/src/trezor/main.py
+++ b/src/trezor/main.py
@@ -1,42 +1,31 @@
-import sys
-sys.path.append('lib')
-
 import gc
+import micropython
+import sys
+
+sys.path.append('lib')
 
 from trezor import loop
 from trezor import workflow
 from trezor import log
 
 log.level = log.DEBUG
-# log.level = log.INFO
-
-
-def perf_info_debug():
-    while True:
-        queue_len = len(loop._scheduled_tasks)
-
-        delay_avg = sum(loop.log_delay_rb) / loop.log_delay_rb_len
-        delay_last = loop.log_delay_rb[loop.log_delay_pos]
-
-        mem_alloc = gc.mem_alloc()
-        gc.collect()
-        log.debug(__name__, "mem_alloc: %s/%s, delay_avg: %d, delay_last: %d, queue_len: %d",
-                  mem_alloc, gc.mem_alloc(), delay_avg, delay_last, queue_len)
-
-        yield loop.Sleep(1000000)
 
 
 def perf_info():
+    prev = 0
+    peak = 0
+    sleep = loop.sleep(100000)
     while True:
         gc.collect()
-        log.info(__name__, "mem_alloc: %d", gc.mem_alloc())
-        yield loop.Sleep(1000000)
+        used = gc.mem_alloc()
+        if used != prev:
+            prev = used
+            peak = max(peak, used)
+            print('peak %d, used %d' % (peak, used))
+        yield sleep
 
 
 def run(default_workflow):
-    # if __debug__:
-    #     loop.schedule_task(perf_info_debug())
-    # else:
-    #     loop.schedule_task(perf_info())
+    # loop.schedule_task(perf_info())
     workflow.start_default(default_workflow)
     loop.run_forever()
diff --git a/src/trezor/messages/__init__.py b/src/trezor/messages/__init__.py
index e626cc27..dac057b0 100644
--- a/src/trezor/messages/__init__.py
+++ b/src/trezor/messages/__init__.py
@@ -1,13 +1,13 @@
 from . import wire_types
 
 
-def get_protobuf_type_name(wire_type):
+def get_type_name(wire_type):
     for name in dir(wire_types):
         if getattr(wire_types, name) == wire_type:
             return name
 
 
-def get_protobuf_type(wire_type):
-    name = get_protobuf_type_name(wire_type)
+def get_type(wire_type):
+    name = get_type_name(wire_type)
     module = __import__('trezor.messages.%s' % name, None, None, (name, ), 0)
     return getattr(module, name)
diff --git a/src/trezor/wire/__init__.py b/src/trezor/wire/__init__.py
index 2917abe5..93a660f4 100644
--- a/src/trezor/wire/__init__.py
+++ b/src/trezor/wire/__init__.py
@@ -1,186 +1,134 @@
-import ubinascii
 import protobuf
 
 from trezor import log
 from trezor import loop
 from trezor import messages
-from trezor import msg
 from trezor import workflow
 
 from . import codec_v1
 from . import codec_v2
-from . import sessions
 
-_interface = None
-
-_workflow_callbacks = {}  # wire type -> function returning workflow
-_workflow_args = {}  # wire type -> args
+workflows = {}
 
 
-def register(wire_type, callback, *args):
-    if wire_type in _workflow_callbacks:
-        raise KeyError('Message %d already registered' % wire_type)
-    _workflow_callbacks[wire_type] = callback
-    _workflow_args[wire_type] = args
+def register(wire_type, handler, *args):
+    if wire_type in workflows:
+        raise KeyError
+    workflows[wire_type] = (handler, args)
 
 
-def setup(iface):
-    global _interface
-
-    # setup wire interface for reading and writing
-    _interface = iface
-
-    # implicitly register v1 codec on its session.  v2 sessions are
-    # opened/closed explicitely through session control messages.
-    _session_open(codec_v1.SESSION)
-
-    # run session dispatcher
-    loop.schedule_task(_dispatch_reports())
+def setup(interface):
+    session_supervisor = codec_v2.SesssionSupervisor(interface,
+                                                     session_handler)
+    session_supervisor.open(codec_v1.SESSION_ID)
+    loop.schedule_task(session_supervisor.listen())
 
 
-async def read(session_id, *wire_types):
-    log.info(__name__, 'session %x: read(%s)', session_id, wire_types)
-    signal = loop.Signal()
-    sessions.listen(session_id, _handle_response, wire_types, signal)
-    return await signal
+class Context:
+    def __init__(self, interface, session_id):
+        self.interface = interface
+        self.session_id = session_id
+
+    def get_reader(self):
+        if self.session_id == codec_v1.SESSION_ID:
+            return codec_v1.Reader(self.interface)
+        else:
+            return codec_v2.Reader(self.interface, self.session_id)
+
+    def get_writer(self, mtype, msize):
+        if self.session_id == codec_v1.SESSION_ID:
+            return codec_v1.Writer(self.interface, mtype, msize)
+        else:
+            return codec_v2.Writer(self.interface, self.session_id, mtype, msize)
+
+    async def read(self, types):
+        reader = self.get_reader()
+        await reader.open()
+        if reader.type not in types:
+            raise UnexpectedMessageError(reader)
+        return await protobuf.load_message(reader,
+                                           messages.get_type(reader.type))
+
+    async def write(self, msg):
+        counter = protobuf.CountingWriter()
+        await protobuf.dump_message(counter, msg)
+        writer = self.get_writer(msg.MESSAGE_WIRE_TYPE, counter.size)
+        await protobuf.dump_message(writer, msg)
+        await writer.close()
+
+    async def call(self, msg, types):
+        await self.write(msg)
+        return await self.read(types)
 
 
-async def write(session_id, pbuf_msg):
-    log.info(__name__, 'session %x: write(%s)', session_id, pbuf_msg)
-    pbuf_type = pbuf_msg.__class__
-    msg_data = pbuf_type.dumps(pbuf_msg)
-    msg_type = pbuf_type.MESSAGE_WIRE_TYPE
-    sessions.get_codec(session_id).encode(
-        session_id, msg_type, msg_data, _write_report)
-
-
-async def call(session_id, pbuf_msg, *response_types):
-    await write(session_id, pbuf_msg)
-    return await read(session_id, *response_types)
+class UnexpectedMessageError(Exception):
+    def __init__(self, reader):
+        super().__init__()
+        self.reader = reader
 
 
 class FailureError(Exception):
-
-    def to_protobuf(self):
-        from trezor.messages.Failure import Failure
-        code, message = self.args
-        return Failure(code=code, message=message)
+    def __init__(self, code, message):
+        super().__init__()
+        self.code = code
+        self.message = message
 
 
-class CloseWorkflow(Exception):
-    pass
+class Workflow:
+    def __init__(self, default):
+        self.handlers = {}
+        self.default = default
 
-
-def protobuf_workflow(session_id, msg_type, data_len, callback, *args):
-    return _build_protobuf(msg_type, _start_protobuf_workflow, session_id, callback, args)
-
-
-def _start_protobuf_workflow(pbuf_msg, session_id, callback, args):
-    wf = callback(session_id, pbuf_msg, *args)
-    wf = _wrap_protobuf_workflow(wf, session_id)
-    workflow.start(wf)
-
-
-async def _wrap_protobuf_workflow(wf, session_id):
-    try:
-        result = await wf
-
-    except CloseWorkflow:
-        return
-
-    except FailureError as e:
-        await write(session_id, e.to_protobuf())
-        raise
-
-    except Exception as e:
-        from trezor.messages.Failure import Failure
-        from trezor.messages.FailureType import FirmwareError
-        await write(session_id, Failure(
-            code=FirmwareError, message='Firmware Error'))
-        raise
-
-    else:
-        if result is not None:
-            await write(session_id, result)
-        return result
-
-    finally:
-        if session_id in sessions.opened:
-            sessions.listen(session_id, _handle_workflow)
-
-
-def _build_protobuf(msg_type, callback, *args):
-    pbuf_type = messages.get_protobuf_type(msg_type)
-    builder = protobuf.build_message(pbuf_type, callback, *args)
-    builder.send(None)
-    return pbuf_type.load(target=builder)
-
-
-def _handle_response(session_id, msg_type, data_len, response_types, signal):
-    if msg_type in response_types:
-        return _build_protobuf(msg_type, signal.send)
-    else:
-        signal.send(CloseWorkflow())
-        return _handle_workflow(session_id, msg_type, data_len)
-
-
-def _handle_workflow(session_id, msg_type, data_len):
-    if msg_type in _workflow_callbacks:
-        callback = _workflow_callbacks[msg_type]
-        args = _workflow_args[msg_type]
-        return callback(session_id, msg_type, data_len, *args)
-    else:
-        return _handle_unexpected(session_id, msg_type, data_len)
-
-
-def _handle_unexpected(session_id, msg_type, data_len):
-    log.warning(
-        __name__, 'session %x: skip type %d, len %d', session_id, msg_type, data_len)
-
-    # read the message in full
-    try:
+    async def __call__(self, interface, session_id):
+        ctx = Context(interface, session_id)
         while True:
-            yield
-    except EOFError:
-        pass
+            try:
+                reader = ctx.get_reader()
+                await reader.open()
+                try:
+                    handler = self.handlers[reader.type]
+                except KeyError:
+                    handler = self.default
+                try:
+                    await handler(ctx, reader)
+                except UnexpectedMessageError as unexp_msg:
+                    reader = unexp_msg.reader
+            except Exception as e:
+                log.exception(__name__, e)
 
+
+async def protobuf_workflow(ctx, reader, handler, *args):
+    msg = await protobuf.load_message(reader, messages.get_type(reader.type))
+    try:
+        res = await handler(reader.sid, msg, *args)
+    except Exception as exc:
+        if not isinstance(exc, UnexpectedMessageError):
+            await ctx.write(make_failure_msg(exc))
+        raise
+    else:
+        if res:
+            await ctx.write(res)
+
+
+async def handle_unexp_msg(ctx, reader):
+    # receive the message and throw it away
+    while reader.size > 0:
+        buf = bytearray(reader.size)
+        await reader.readinto(buf)
     # respond with an unknown message error
     from trezor.messages.Failure import Failure
     from trezor.messages.FailureType import UnexpectedMessage
-    failure = Failure(code=UnexpectedMessage, message='Unexpected message')
-    failure = Failure.dumps(failure)
-    sessions.get_codec(session_id).encode(
-        session_id, Failure.MESSAGE_WIRE_TYPE, failure, _write_report)
+    await ctx.write(
+        Failure(code=UnexpectedMessage, message='Unexpected message'))
 
 
-def _write_report(report):
-    # if __debug__:
-    #     log.debug(__name__, 'write report %s', ubinascii.hexlify(report))
-    msg.send(_interface, report)
-
-
-def _dispatch_reports():
-    read = loop.select(_interface)
-    while True:
-        report = yield read
-        # if __debug__:
-        #     log.debug(__name__, 'read report %s', ubinascii.hexlify(report))
-        sessions.dispatch(
-            memoryview(report), _session_open, _session_close, _session_unknown)
-
-
-def _session_open(session_id=None):
-    session_id = sessions.open(session_id)
-    sessions.listen(session_id, _handle_workflow)
-    sessions.get_codec(session_id).encode_session_open(
-        session_id, _write_report)
-
-
-def _session_close(session_id):
-    sessions.close(session_id)
-    sessions.get_codec(session_id).encode_session_close(
-        session_id, _write_report)
-
-
-def _session_unknown(session_id, report_data):
-    log.warning(__name__, 'report on unknown session %x', session_id)
+def make_failure_msg(exc):
+    from trezor.messages.Failure import Failure
+    from trezor.messages.FailureType import FirmwareError
+    if isinstance(exc, FailureError):
+        code = exc.code
+        message = exc.message
+    else:
+        code = FirmwareError
+        message = 'Firmware Error'
+    return Failure(code=code, message=message)
diff --git a/src/trezor/wire/codec_v1.py b/src/trezor/wire/codec_v1.py
index 6657f60a..9a3dfc87 100644
--- a/src/trezor/wire/codec_v1.py
+++ b/src/trezor/wire/codec_v1.py
@@ -1,114 +1,145 @@
 from micropython import const
-
 import ustruct
 
-SESSION = const(0)
-REP_MARKER = const(63)  # ord('?')
-REP_MARKER_LEN = const(1)  # len('?')
+from trezor import io
+from trezor import loop
+from trezor import utils
 
 _REP_LEN = const(64)
-_MSG_HEADER_MAGIC = const(35)  # org('#')
-_MSG_HEADER = '>BBHL'  # magic, magic, wire type, data length
-_MSG_HEADER_LEN = ustruct.calcsize(_MSG_HEADER)
+
+_REP_MARKER = const(63)  # ord('?')
+_REP_MAGIC = const(35)  # org('#')
+_REP_INIT = '>BBBHL'  # marker, magic, magic, wire type, data length
+_REP_INIT_DATA = const(9)  # offset of data in the initial report
+_REP_CONT_DATA = const(1)  # offset of data in the continuation report
+
+SESSION_ID = const(0)
 
 
-def detect(data):
-    return data[0] == REP_MARKER
+class Reader:
+    '''
+    Decoder for legacy codec over the HID layer.  Provides readable
+    async-file-like interface.
+    '''
+
+    def __init__(self, iface):
+        self.iface = iface
+        self.type = None
+        self.size = None
+        self.data = None
+        self.ofs = 0
+
+    def __repr__(self):
+        return '<ReaderV1: type=%d size=%dB>' % (self.type, self.size)
+
+    async def open(self):
+        '''
+        Begin the message transmission by waiting for initial V2 message report
+        on this session.  `self.type` and `self.size` are initialized and
+        available after `open()` returns.
+        '''
+        read = loop.select(self.iface | loop.READ)
+        while True:
+            # wait for initial report
+            report = await read
+            marker = report[0]
+            if marker == _REP_MARKER:
+                _, m1, m2, mtype, msize = ustruct.unpack(_REP_INIT, report)
+                if m1 != _REP_MAGIC or m2 != _REP_MAGIC:
+                    raise ValueError
+                break
+
+        # load received message header
+        self.type = mtype
+        self.size = msize
+        self.data = report[_REP_INIT_DATA:_REP_INIT_DATA + msize]
+        self.ofs = 0
+
+    async def readinto(self, buf):
+        '''
+        Read exactly `len(buf)` bytes into `buf`, waiting for additional
+        reports, if needed.  Raises `EOFError` if end-of-message is encountered
+        before the full read can be completed.
+        '''
+        if self.size < len(buf):
+            raise EOFError
+
+        read = loop.select(self.iface | loop.READ)
+        nread = 0
+        while nread < len(buf):
+            if self.ofs == len(self.data):
+                # we are at the end of received data
+                # wait for continuation report
+                while True:
+                    report = await read
+                    marker = report[0]
+                    if marker == _REP_MARKER:
+                        break
+                self.data = report[_REP_CONT_DATA:_REP_CONT_DATA + self.size]
+                self.ofs = 0
+
+            # copy as much as possible to target buffer
+            nbytes = utils.memcpy(buf, nread, self.data, self.ofs, len(buf))
+            nread += nbytes
+            self.ofs += nbytes
+            self.size -= nbytes
+
+        return nread
 
 
-def parse_report(data):
-    if len(data) != _REP_LEN:
-        raise ValueError('Invalid buffer size')
-    return None, SESSION, data[1:]
+class Writer:
+    '''
+    Encoder for legacy codec over the HID layer.  Provides writable
+    async-file-like interface.
+    '''
 
+    def __init__(self, iface, mtype, msize):
+        self.iface = iface
+        self.type = mtype
+        self.size = msize
+        self.data = bytearray(_REP_LEN)
+        self.ofs = _REP_INIT_DATA
 
-def parse_message(data):
-    magic1, magic2, msg_type, data_len = ustruct.unpack(_MSG_HEADER, data)
-    if magic1 != _MSG_HEADER_MAGIC or magic2 != _MSG_HEADER_MAGIC:
-        raise ValueError('Corrupted magic bytes')
-    return msg_type, data_len, data[_MSG_HEADER_LEN:]
+        # load the report with initial header
+        ustruct.pack_into(_REP_INIT, self.data, 0, _REP_MARKER, _REP_MAGIC, _REP_MAGIC, mtype, msize)
 
+    def __repr__(self):
+        return '<WriterV2: type=%d size=%dB>' % (self.type, self.size)
 
-def serialize_message_header(data, msg_type, msg_len):
-    if len(data) < REP_MARKER_LEN + _MSG_HEADER_LEN:
-        raise ValueError('Invalid buffer size')
-    if msg_type < 0 or msg_type > 65535:
-        raise ValueError('Value is out of range')
-    ustruct.pack_into(
-        _MSG_HEADER, data, REP_MARKER_LEN,
-        _MSG_HEADER_MAGIC, _MSG_HEADER_MAGIC, msg_type, msg_len)
+    async def write(self, buf):
+        '''
+        Encode and write every byte from `buf`.  Does not need to be called in
+        case message has zero length.  Raises `EOFError` if the length of `buf`
+        exceeds the remaining message length.
+        '''
+        if self.size < len(buf):
+            raise EOFError
 
+        write = loop.select(self.iface | loop.WRITE)
+        nwritten = 0
+        while nwritten < len(buf):
+            # copy as much as possible to report buffer
+            nbytes = utils.memcpy(self.data, self.ofs, buf, nwritten, len(buf))
+            nwritten += nbytes
+            self.ofs += nbytes
+            self.size -= nbytes
 
-def decode_stream(session_id, callback, *args):
-    '''Decode a v1 wire message from the report data and stream it to target.
+            if self.ofs == _REP_LEN:
+                # we are at the end of the report, flush it
+                await write
+                io.send(self.iface, self.data)
+                self.ofs = _REP_CONT_DATA
 
-Receives report payloads.  After first report, creates target by calling
-`callback(session_id, msg_type, data_len, *args)` and sends chunks of message
-data.  Throws `EOFError` to target after last data chunk.
+        return nwritten
 
-Pass report payloads as `memoryview` for cheaper slicing.
-'''
+    async def close(self):
+        '''Flush and close the message transmission.'''
+        if self.ofs != _REP_CONT_DATA:
+            # we didn't write anything or last write() wasn't report-aligned,
+            # pad the final report and flush it
+            while self.ofs < _REP_LEN:
+                self.data[self.ofs] = 0x00
+                self.ofs += 1
 
-    message = yield  # read first report
-    msg_type, data_len, data = parse_message(message)
-
-    target = callback(session_id, msg_type, data_len, *args)
-    target.send(None)
-
-    while data_len > 0:
-
-        data_chunk = data[:data_len]  # slice off the garbage at the end
-        data = data[len(data_chunk):]  # slice off what we have read
-        data_len -= len(data_chunk)
-        target.send(data_chunk)
-
-        if data_len > 0:
-            data = yield  # read next report
-
-    target.throw(EOFError())
-
-
-def encode(session_id, msg_type, msg_data, callback):
-    '''Encode a full v1 wire message directly to reports and stream it to callback.
-
-Callback receives `memoryview`s of HID reports which are valid until the
-callback returns.
-'''
-    report = memoryview(bytearray(_REP_LEN))
-    report[0] = REP_MARKER
-    serialize_message_header(report, msg_type, len(msg_data))
-
-    source_data = memoryview(msg_data)
-    target_data = report[REP_MARKER_LEN + _MSG_HEADER_LEN:]
-
-    while True:
-        # move as much as possible from source to target
-        n = min(len(target_data), len(source_data))
-        target_data[:n] = source_data[:n]
-        source_data = source_data[n:]
-        target_data = target_data[n:]
-
-        # fill the rest of the report with 0x00
-        x = 0
-        to_fill = len(target_data)
-        while x < to_fill:
-            target_data[x] = 0
-            x += 1
-
-        callback(report)
-
-        if not source_data:
-            break
-
-        # reset to skip the magic, not the whole header anymore
-        target_data = report[REP_MARKER_LEN:]
-
-
-def encode_session_open(session_id, callback):
-    # v1 codec does not have explicit session support
-    pass
-
-
-def encode_session_close(session_id, callback):
-    # v1 codec does not have explicit session support
-    pass
+            await loop.select(self.iface | loop.WRITE)
+            io.send(self.iface, self.data)
diff --git a/src/trezor/wire/codec_v2.py b/src/trezor/wire/codec_v2.py
index 63f7fd90..808b4fa4 100644
--- a/src/trezor/wire/codec_v2.py
+++ b/src/trezor/wire/codec_v2.py
@@ -1,190 +1,232 @@
 from micropython import const
 import ustruct
-import ubinascii
 
-# trezor wire protocol #2:
-#
-# # hid report (64B)
-# - report marker (1B)
-# - session id (4B, BE)
-# - payload (59B)
-#
-# # message
-# - streamed as payloads of hid reports
-# - message type (4B, BE)
-# - data length (4B, BE)
-# - data (var-length)
-# - data crc32 checksum (4B, BE)
-#
-# # sessions
-# - reports are interleaved, need to be dispatched by session id
+from trezor import io
+from trezor import loop
+from trezor import utils
+from trezor.crypto import random
 
-REP_MARKER_HEADER = const(72)  # ord('H')
-REP_MARKER_DATA = const(68)  # ord('D')
-REP_MARKER_OPEN = const(79)  # ord('O')
-REP_MARKER_CLOSE = const(67)  # ord('C')
-
-_REP_HEADER = '>BL'  # marker, session id
-_MSG_HEADER = '>LL'  # msg type, data length
-_MSG_FOOTER = '>L'  # data checksum
+# TREZOR wire protocol #2:
+#
+# # Initial message report
+# uint8_t  marker;        // REP_MARKER_INIT
+# uint32_t session_id;    // Big-endian
+# uint32_t message_type;  // Big-endian
+# uint32_t message_size;  // Big-endian
+# uint8_t  data[];
+#
+# # Continuation message report
+# uint8_t  marker;      // REP_MARKER_CONT
+# uint32_t session_id;  // Big-endian
+# uint32_t sequence;    // Big-endian, 0 for 1st continuation report
+# uint8_t  data[];
 
 _REP_LEN = const(64)
-_REP_HEADER_LEN = ustruct.calcsize(_REP_HEADER)
-_MSG_HEADER_LEN = ustruct.calcsize(_MSG_HEADER)
-_MSG_FOOTER_LEN = ustruct.calcsize(_MSG_FOOTER)
+
+_REP_MARKER_INIT = const(0x01)
+_REP_MARKER_CONT = const(0x02)
+_REP_MARKER_OPEN = const(0x03)
+_REP_MARKER_CLOSE = const(0x04)
+
+_REP = '>BL'         # marker, session_id
+_REP_INIT = '>BLLL'  # marker, session_id, message_type, message_size
+_REP_CONT = '>BLL'   # marker, session_id, sequence
+_REP_INIT_DATA = const(13)  # offset of data in init report
+_REP_CONT_DATA = const(9)   # offset of data in cont report
 
 
-def parse_report(data):
-    if len(data) != _REP_LEN:
-        raise ValueError('Invalid buffer size')
-    marker, session_id = ustruct.unpack(_REP_HEADER, data)
-    return marker, session_id, data[_REP_HEADER_LEN:]
-
-
-def parse_message(data):
-    if len(data) != _REP_LEN - _REP_HEADER_LEN:
-        raise ValueError('Invalid buffer size')
-    msg_type, data_len = ustruct.unpack(_MSG_HEADER, data)
-    return msg_type, data_len, data[_MSG_HEADER_LEN:]
-
-
-def parse_message_footer(data):
-    if len(data) != _MSG_FOOTER_LEN:
-        raise ValueError('Invalid buffer size')
-    data_checksum, = ustruct.unpack(_MSG_FOOTER, data)
-    return data_checksum,
-
-
-def serialize_report_header(data, marker, session_id):
-    if len(data) < _REP_HEADER_LEN:
-        raise ValueError('Invalid buffer size')
-    ustruct.pack_into(_REP_HEADER, data, 0, marker, session_id)
-
-
-def serialize_message_header(data, msg_type, msg_len):
-    if len(data) < _REP_HEADER_LEN + _MSG_HEADER_LEN:
-        raise ValueError('Invalid buffer size')
-    ustruct.pack_into(_MSG_HEADER, data, _REP_HEADER_LEN, msg_type, msg_len)
-
-
-def serialize_message_footer(data, checksum):
-    if len(data) < _MSG_FOOTER_LEN:
-        raise ValueError('Invalid buffer size')
-    ustruct.pack_into(_MSG_FOOTER, data, 0, checksum)
-
-
-def serialize_opened_session(data, session_id):
-    serialize_report_header(data, REP_MARKER_OPEN, session_id)
-
-
-class MessageChecksumError(Exception):
-    pass
-
-
-def decode_stream(session_id, callback, *args):
-    '''Decode a wire message from the report data and stream it to target.
-
-Receives report payloads.  After first report, creates target by calling
-`callback(session_id, msg_type, data_len, *args)` and sends chunks of message
-data.
-Throws `EOFError` to target after last data chunk, in case of valid checksum.
-Throws `MessageChecksumError` to target if data doesn't match the checksum.
-
-Pass report payloads as `memoryview` for cheaper slicing.
-'''
-    message = yield  # read first report
-    msg_type, data_len, data_tail = parse_message(message)
-
-    target = callback(session_id, msg_type, data_len, *args)
-    target.send(None)
-
-    checksum = 0  # crc32
-
-    while data_len > 0:
-
-        data_chunk = data_tail[:data_len]  # slice off the garbage at the end
-        data_tail = data_tail[len(data_chunk):]  # slice off what we have read
-        data_len -= len(data_chunk)
-        target.send(data_chunk)
-
-        checksum = ubinascii.crc32(data_chunk, checksum)
-
-        if data_len > 0:
-            data_tail = yield  # read next report
-
-    msg_footer = data_tail[:_MSG_FOOTER_LEN]
-    if len(msg_footer) < _MSG_FOOTER_LEN:
-        data_tail = yield  # read report with the rest of checksum
-        footer_tail = data_tail[:_MSG_FOOTER_LEN - len(msg_footer)]
-        msg_footer = bytearray(msg_footer)
-        msg_footer.extend(footer_tail)
-
-    data_checksum, = parse_message_footer(msg_footer)
-    if data_checksum != checksum:
-        target.throw(MessageChecksumError((checksum, data_checksum)))
-    else:
-        target.throw(EOFError())
-
-
-def encode(session_id, msg_type, msg_data, callback):
-    '''Encode a full wire message directly to reports and stream it to callback.
-
-Callback receives `memoryview`s of HID reports which are valid until the
-callback returns.
+class Reader:
+    '''
+    Decoder for v2 codec over the HID layer.  Provides readable async-file-like
+    interface.
     '''
-    report = memoryview(bytearray(_REP_LEN))
-    serialize_report_header(report, REP_MARKER_HEADER, session_id)
-    serialize_message_header(report, msg_type, len(msg_data))
 
-    source_data = memoryview(msg_data)
-    target_data = report[_REP_HEADER_LEN + _MSG_HEADER_LEN:]
+    def __init__(self, iface, sid):
+        self.iface = iface
+        self.sid = sid
+        self.type = None
+        self.size = None
+        self.data = None
+        self.ofs = 0
+        self.seq = 0
 
-    checksum = ubinascii.crc32(msg_data)
+    def __repr__(self):
+        return '<Reader: sid=%x type=%d size=%dB>' % (self.sid, self.type, self.size)
 
-    msg_footer = bytearray(_MSG_FOOTER_LEN)
-    serialize_message_footer(msg_footer, checksum)
+    async def open(self):
+        '''
+        Begin the message transmission by waiting for initial V2 message report
+        on this session. `self.type` and `self.size` are initialized and
+        available after `open()` returns.
+        '''
+        read = loop.select(self.iface | loop.READ)
+        while True:
+            # wait for initial report
+            report = await read
+            marker, sid, mtype, msize = ustruct.unpack(_REP_INIT, report)
+            if sid == self.sid and marker == _REP_MARKER_INIT:
+                break
 
-    first = True
+        # load received message header
+        self.type = mtype
+        self.size = msize
+        self.data = report[_REP_INIT_DATA:_REP_INIT_DATA + msize]
+        self.ofs = 0
+        self.seq = 0
 
-    while True:
-        # move as much as possible from source to target
-        n = min(len(target_data), len(source_data))
-        target_data[:n] = source_data[:n]
-        source_data = source_data[n:]
-        target_data = target_data[n:]
+    async def readinto(self, buf):
+        '''
+        Read exactly `len(buf)` bytes into `buf`, waiting for additional
+        reports, if needed.  Raises `EOFError` if end-of-message is encountered
+        before the full read can be completed.
+        '''
+        if self.size < len(buf):
+            raise EOFError
 
-        # continue with the footer if source is empty and we have space
-        if not source_data and target_data and msg_footer:
-            source_data = msg_footer
-            msg_footer = None
-            continue
+        read = loop.select(self.iface | loop.READ)
+        nread = 0
+        while nread < len(buf):
+            if self.ofs == len(self.data):
+                # we are at the end of received data
+                # wait for continuation report
+                while True:
+                    report = await read
+                    marker, sid, seq = ustruct.unpack(_REP_CONT, report)
+                    if sid == self.sid and marker == _REP_MARKER_CONT:
+                        if seq != self.seq:
+                            raise ValueError
+                        break
+                self.data = report[_REP_CONT_DATA:_REP_CONT_DATA + self.size]
+                self.seq += 1
+                self.ofs = 0
 
-        # fill the rest of the report with 0x00
-        x = 0
-        to_fill = len(target_data)
-        while x < to_fill:
-            target_data[x] = 0
-            x += 1
+            # copy as much as possible to target buffer
+            nbytes = utils.memcpy(buf, nread, self.data, self.ofs, len(buf))
+            nread += nbytes
+            self.ofs += nbytes
+            self.size -= nbytes
 
-        callback(report)
-
-        if not source_data and not msg_footer:
-            break
-
-        # reset to skip the magic and session ID
-        if first:
-            serialize_report_header(report, REP_MARKER_DATA, session_id)
-            first = False
-        target_data = report[_REP_HEADER_LEN:]
+        return nread
 
 
-def encode_session_open(session_id, callback):
-    report = bytearray(_REP_LEN)
-    serialize_report_header(report, REP_MARKER_OPEN, session_id)
-    callback(report)
+class Writer:
+    '''
+    Encoder for v2 codec over the HID layer.  Provides writable async-file-like
+    interface.
+    '''
+
+    def __init__(self, iface, sid, mtype, msize):
+        self.iface = iface
+        self.sid = sid
+        self.type = mtype
+        self.size = msize
+        self.data = bytearray(_REP_LEN)
+        self.ofs = _REP_INIT_DATA
+        self.seq = 0
+
+        # load the report with initial header
+        ustruct.pack_into(_REP_INIT, self.data, 0,
+                          _REP_MARKER_INIT, sid, mtype, msize)
+
+    async def write(self, buf):
+        '''
+        Encode and write every byte from `buf`.  Does not need to be called in
+        case message has zero length.  Raises `EOFError` if the length of `buf`
+        exceeds the remaining message length.
+        '''
+        if self.size < len(buf):
+            raise EOFError
+
+        write = loop.select(self.iface | loop.WRITE)
+        nwritten = 0
+        while nwritten < len(buf):
+            # copy as much as possible to report buffer
+            nbytes = utils.memcpy(self.data, self.ofs, buf, nwritten, len(buf))
+            nwritten += nbytes
+            self.ofs += nbytes
+            self.size -= nbytes
+
+            if self.ofs == _REP_LEN:
+                # we are at the end of the report, flush it, and prepare header
+                await write
+                io.send(self.iface, self.data)
+                ustruct.pack_into(_REP_CONT, self.data, 0,
+                                  _REP_MARKER_CONT, self.sid, self.seq)
+                self.ofs = _REP_CONT_DATA
+                self.seq += 1
+
+        return nwritten
+
+    async def close(self):
+        '''Flush and close the message transmission.'''
+        if self.ofs != _REP_CONT_DATA:
+            # we didn't write anything or last write() wasn't report-aligned,
+            # pad the final report and flush it
+            while self.ofs < _REP_LEN:
+                self.data[self.ofs] = 0x00
+                self.ofs += 1
+
+            await loop.select(self.iface | loop.WRITE)
+            io.send(self.iface, self.data)
 
 
-def encode_session_close(session_id, callback):
-    report = bytearray(_REP_LEN)
-    serialize_report_header(report, REP_MARKER_CLOSE, session_id)
-    callback(report)
+class SesssionSupervisor:
+    '''Handles session open/close requests on v2 protocol layer.'''
+
+    def __init__(self, iface, handler):
+        self.iface = iface
+        self.handler = handler
+        self.handling_tasks = {}
+        self.session_report = bytearray(_REP_LEN)
+
+    async def listen(self):
+        '''
+        Listen for open/close requests on configured interface.  After open
+        request, session is started and a new task is scheduled to handle it.
+        After close request, the handling task is closed and session terminated.
+        Both requests receive responses confirming the operation.
+        '''
+        read = loop.select(self.iface | loop.READ)
+        write = loop.select(self.iface | loop.WRITE)
+        while True:
+            report = await read
+            repmarker, repsid = ustruct.unpack(_REP, report)
+            # because tasks paused on I/O have a priority over time-scheduled
+            # tasks, we need to `yield` explicitly before sending a response to
+            # open/close request.  Otherwise the handler would have no chance to
+            # run and schedule communication.
+            if repmarker == _REP_MARKER_OPEN:
+                newsid = self.newsid()
+                self.open(newsid)
+                yield
+                await write
+                self.sendopen(newsid)
+            elif repmarker == _REP_MARKER_CLOSE:
+                self.close(repsid)
+                yield
+                await write
+                self.sendclose(repsid)
+
+    def open(self, sid):
+        if sid not in self.handling_tasks:
+            task = self.handling_tasks[sid] = self.handler(self.iface, sid)
+            loop.schedule_task(task)
+
+    def close(self, sid):
+        if sid in self.handling_tasks:
+            task = self.handling_tasks.pop(sid)
+            task.close()
+
+    def newsid(self):
+        while True:
+            sid = random.uniform(0xffffffff) + 1
+            if sid not in self.handling_tasks:
+                return sid
+
+    def sendopen(self, sid):
+        ustruct.pack_into(_REP, self.session_report, 0, _REP_MARKER_OPEN, sid)
+        io.send(self.iface, self.session_report)
+
+    def sendclose(self, sid):
+        ustruct.pack_into(_REP, self.session_report, 0, _REP_MARKER_CLOSE, sid)
+        io.send(self.iface, self.session_report)
diff --git a/src/trezor/wire/sessions.py b/src/trezor/wire/sessions.py
deleted file mode 100644
index 9fc01608..00000000
--- a/src/trezor/wire/sessions.py
+++ /dev/null
@@ -1,82 +0,0 @@
-from trezor import log
-from trezor.crypto import random
-
-from . import codec_v1
-from . import codec_v2
-
-opened = set()  # opened session ids
-readers = {}  # session id -> generator
-
-
-def generate():
-    while True:
-        session_id = random.uniform(0xffffffff) + 1
-        if session_id not in opened:
-            return session_id
-
-
-def open(session_id=None):
-    if session_id is None:
-        session_id = generate()
-    log.info(__name__, 'session %x: open', session_id)
-    opened.add(session_id)
-    return session_id
-
-
-def close(session_id):
-    log.info(__name__, 'session %x: close', session_id)
-    opened.discard(session_id)
-    readers.pop(session_id, None)
-
-
-def get_codec(session_id):
-    if session_id == codec_v1.SESSION:
-        return codec_v1
-    else:
-        return codec_v2
-
-
-def listen(session_id, handler, *args):
-    if session_id not in opened:
-        raise KeyError('Session %x is unknown' % session_id)
-    if session_id in readers:
-        raise KeyError('Session %x is already being listened on' % session_id)
-    log.info(__name__, 'session %x: listening', session_id)
-    decoder = get_codec(session_id).decode_stream(session_id, handler, *args)
-    decoder.send(None)
-    readers[session_id] = decoder
-
-
-def dispatch(report, open_callback, close_callback, unknown_callback):
-    '''
-    Dispatches payloads of reports adhering to one of the wire codecs.
-    '''
-
-    if codec_v1.detect(report):
-        marker, session_id, report_data = codec_v1.parse_report(report)
-    else:
-        marker, session_id, report_data = codec_v2.parse_report(report)
-
-        if marker == codec_v2.REP_MARKER_OPEN:
-            log.debug(__name__, 'request for new session')
-            open_callback()
-            return
-        elif marker == codec_v2.REP_MARKER_CLOSE:
-            log.debug(__name__, 'request for closing session %x', session_id)
-            close_callback(session_id)
-            return
-
-    if session_id not in readers:
-        log.warning(__name__, 'report on unknown session %x', session_id)
-        unknown_callback(session_id, report_data)
-        return
-
-    log.debug(__name__, 'report on session %x', session_id)
-    reader = readers[session_id]
-
-    try:
-        reader.send(report_data)
-    except StopIteration:
-        readers.pop(session_id)
-    except Exception as e:
-        log.exception(__name__, e)
diff --git a/src/trezor/workflow.py b/src/trezor/workflow.py
index 19d208d7..6b881dde 100644
--- a/src/trezor/workflow.py
+++ b/src/trezor/workflow.py
@@ -17,14 +17,14 @@ def start_default(genfunc):
 
 def close_default():
     global _default
-    log.info(__name__, 'close default %s', _default)
-    _default.close()
-    _default = None
+    if _default is not None:
+        log.info(__name__, 'close default %s', _default)
+        _default.close()
+        _default = None
 
 
 def start(workflow):
-    if _default is not None:
-        close_default()
+    close_default()
     _started.append(workflow)
     log.info(__name__, 'start %s', workflow)
     loop.schedule_task(_watch(workflow))
diff --git a/tests/test_trezor.wire.codec_v1.py b/tests/test_trezor.wire.codec_v1.py
index 4638f144..915eb0f2 100644
--- a/tests/test_trezor.wire.codec_v1.py
+++ b/tests/test_trezor.wire.codec_v1.py
@@ -1,178 +1,164 @@
-from common import *
+import sys
 
-import ustruct
+sys.path.append('../src')
+sys.path.append('../src/lib')
 
+from utest import *
+from ustruct import pack, unpack
+from ubinascii import hexlify, unhexlify
+
+from trezor import msg
+from trezor.loop import Select, Syscall, READ, WRITE
 from trezor.crypto import random
 from trezor.utils import chunks
-
 from trezor.wire import codec_v1
 
-class TestWireCodecV1(unittest.TestCase):
-    # pylint: disable=C0301
 
-    def test_detect(self):
-        for i in range(0, 256):
-            if i == ord(b'?'):
-                self.assertTrue(codec_v1.detect(bytes([i]) + b'\x00' * 63))
-            else:
-                self.assertFalse(codec_v1.detect(bytes([i]) + b'\x00' * 63))
+def test_reader():
+    rep_len = 64
+    interface = 0xdeadbeef
+    message_type = 0x4321
+    message_len = 250
+    reader = codec_v1.Reader(interface, codec_v1.SESSION_ID)
 
-    def test_parse(self):
-        d = bytes(range(0, 55))
-        m = b'##\x00\x00\x00\x00\x00\x37' + d
-        r = b'?' + m
+    message = bytearray(range(message_len))
+    report_header = bytearray(unhexlify('3f23234321000000fa'))
 
-        rm, rs, rd = codec_v1.parse_report(r)
-        self.assertEqual(rm, None)
-        self.assertEqual(rs, 0)
-        self.assertEqual(rd, m)
+    # open, expected one read
+    first_report = report_header + message[:rep_len - len(report_header)]
+    assert_async(reader.open(), [(None, Select(READ | interface)), (first_report, StopIteration()),])
+    assert_eq(reader.type, message_type)
+    assert_eq(reader.size, message_len)
 
-        mt, ml, md = codec_v1.parse_message(m)
-        self.assertEqual(mt, 0)
-        self.assertEqual(ml, len(d))
-        self.assertEqual(md, d)
+    # empty read
+    empty_buffer = bytearray()
+    assert_async(reader.readinto(empty_buffer), [(None, StopIteration()),])
+    assert_eq(len(empty_buffer), 0)
+    assert_eq(reader.size, message_len)
 
-        for i in range(0, 1024):
-            if i != 64:
-                with self.assertRaises(ValueError):
-                    codec_v1.parse_report(bytes(range(0, i)))
+    # short read, expected no read
+    short_buffer = bytearray(32)
+    assert_async(reader.readinto(short_buffer), [(None, StopIteration()),])
+    assert_eq(len(short_buffer), 32)
+    assert_eq(short_buffer, message[:len(short_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer))
 
-        for hx in range(0, 256):
-            for hy in range(0, 256):
-                if hx != ord(b'#') and hy != ord(b'#'):
-                    with self.assertRaises(ValueError):
-                        codec_v1.parse_message(bytes([hx, hy]) + m[2:])
+    # aligned read, expected no read
+    aligned_buffer = bytearray(rep_len - len(report_header) - len(short_buffer))
+    assert_async(reader.readinto(aligned_buffer), [(None, StopIteration()),])
+    assert_eq(aligned_buffer, message[len(short_buffer):][:len(aligned_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer) - len(aligned_buffer))
 
-    def test_serialize(self):
-        data = bytearray(range(0, 10))
-        codec_v1.serialize_message_header(data, 0x1234, 0x56789abc)
-        self.assertEqual(data, b'\x00##\x12\x34\x56\x78\x9a\xbc\x09')
+    # one byte read, expected one read
+    next_report_header = bytearray(unhexlify('3f'))
+    next_report = next_report_header + message[rep_len - len(report_header):][:rep_len - len(next_report_header)]
+    onebyte_buffer = bytearray(1)
+    assert_async(reader.readinto(onebyte_buffer), [(None, Select(READ | interface)), (next_report, StopIteration()),])
+    assert_eq(onebyte_buffer, message[len(short_buffer):][len(aligned_buffer):][:len(onebyte_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer) - len(aligned_buffer) - len(onebyte_buffer))
 
-        data = bytearray(9)
-        with self.assertRaises(ValueError):
-            codec_v1.serialize_message_header(data, 65536, 0)
+    # too long read, raises eof
+    assert_async(reader.readinto(bytearray(reader.size + 1)), [(None, EOFError()),])
 
-        for i in range(0, 8):
-            data = bytearray(i)
-            with self.assertRaises(ValueError):
-                codec_v1.serialize_message_header(data, 0x1234, 0x56789abc)
+    # long read, expect multiple reads
+    start_size = reader.size
+    long_buffer = bytearray(start_size)
+    report_payload = message[rep_len - len(report_header) + rep_len - len(next_report_header):]
+    report_payload_head = report_payload[:rep_len - len(next_report_header) - len(onebyte_buffer)]
+    report_payload_rest = report_payload[len(report_payload_head):]
+    report_payload_rest = list(chunks(report_payload_rest, rep_len - len(next_report_header)))
+    report_payloads = [report_payload_head] + report_payload_rest
+    next_reports = [next_report_header + r for r in report_payloads]
+    expected_syscalls = []
+    for i, _ in enumerate(next_reports):
+        prev_report = next_reports[i - 1] if i > 0 else None
+        expected_syscalls.append((prev_report, Select(READ | interface)))
+    expected_syscalls.append((next_reports[-1], StopIteration()))
+    assert_async(reader.readinto(long_buffer), expected_syscalls)
+    assert_eq(long_buffer, message[-start_size:])
+    assert_eq(reader.size, 0)
 
-    def test_decode_empty(self):
-        message = b'##' + b'\xab\xcd' + b'\x00\x00\x00\x00' + b'\x00' * 55
+    # one byte read, raises eof
+    assert_async(reader.readinto(onebyte_buffer), [(None, EOFError()),])
 
-        record = []
-        genfunc = self._record(record, 0xdeadbeef, 0xabcd, 0, 'dummy')
-        decoder = codec_v1.decode_stream(0xdeadbeef, genfunc, 'dummy')
-        decoder.send(None)
 
-        try:
-            decoder.send(message)
-        except StopIteration as e:
-            res = e.value
-        self.assertEqual(res, None)
-        self.assertEqual(len(record), 1)
-        self.assertIsInstance(record[0], EOFError)
+def test_writer():
+    rep_len = 64
+    interface = 0xdeadbeef
+    message_type = 0x87654321
+    message_len = 1024
+    writer = codec_v1.Writer(interface, codec_v1.SESSION_ID, message_type, message_len)
 
-    def test_decode_one_report_aligned(self):
-        data = bytes(range(0, 55))
-        message = b'##' + b'\xab\xcd' + b'\x00\x00\x00\x37' + data
+    # init header corresponding to the data above
+    report_header = bytearray(unhexlify('3f2323432100000400'))
 
-        record = []
-        genfunc = self._record(record, 0xdeadbeef, 0xabcd, 55, 'dummy')
-        decoder = codec_v1.decode_stream(0xdeadbeef, genfunc, 'dummy')
-        decoder.send(None)
+    assert_eq(writer.data, report_header + bytearray(rep_len - len(report_header)))
 
-        try:
-            decoder.send(message)
-        except StopIteration as e:
-            res = e.value
-        self.assertEqual(res, None)
-        self.assertEqual(len(record), 2)
-        self.assertEqual(record[0], data)
-        self.assertIsInstance(record[1], EOFError)
+    # empty write
+    start_size = writer.size
+    assert_async(writer.write(bytearray()), [(None, StopIteration()),])
+    assert_eq(writer.data, report_header + bytearray(rep_len - len(report_header)))
+    assert_eq(writer.size, start_size)
 
-    def test_decode_generated_range(self):
-        for data_len in range(1, 512):
-            data = random.bytes(data_len)
-            data_chunks = [data[:55]] + list(chunks(data[55:], 63))
+    # short write, expected no report
+    start_size = writer.size
+    short_payload = bytearray(range(4))
+    assert_async(writer.write(short_payload), [(None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(short_payload))
+    assert_eq(writer.data,
+              report_header
+              + short_payload
+              + bytearray(rep_len - len(report_header) - len(short_payload)))
 
-            msg_type = 0xabcd
-            header = b'##' + ustruct.pack('>H', msg_type) + ustruct.pack('>L', data_len)
+    # aligned write, expected one report
+    start_size = writer.size
+    aligned_payload = bytearray(range(rep_len - len(report_header) - len(short_payload)))
+    msg.send = mock_call(msg.send, [
+        (interface, report_header
+         + short_payload
+         + aligned_payload
+         + bytearray(rep_len - len(report_header) - len(short_payload) - len(aligned_payload))), ])
+    assert_async(writer.write(aligned_payload), [(None, Select(WRITE | interface)), (None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(aligned_payload))
+    msg.send.assert_called_n_times(1)
+    msg.send = msg.send.original
 
-            message = header + data
-            message_chunks = [c + '\x00' * (63 - len(c)) for c in list(chunks(message, 63))]
+    # short write, expected no report, but data starts with correct seq and cont marker
+    report_header = bytearray(unhexlify('3f'))
+    start_size = writer.size
+    assert_async(writer.write(short_payload), [(None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(short_payload))
+    assert_eq(writer.data[:len(report_header) + len(short_payload)],
+              report_header + short_payload)
 
-            record = []
-            genfunc = self._record(record, 0xdeadbeef, msg_type, data_len, 'dummy')
-            decoder = codec_v1.decode_stream(0xdeadbeef, genfunc, 'dummy')
-            decoder.send(None)
-
-            res = 1
-            try:
-                for c in message_chunks:
-                    decoder.send(c)
-            except StopIteration as e:
-                res = e.value
-            self.assertEqual(res, None)
-            self.assertEqual(len(record), len(data_chunks) + 1)
-            for i in range(0, len(data_chunks)):
-                self.assertEqual(record[i], data_chunks[i])
-            self.assertIsInstance(record[-1], EOFError)
-
-    def test_encode_empty(self):
-        record = []
-        target = self._record(record)()
-        target.send(None)
-
-        codec_v1.encode(codec_v1.SESSION, 0xabcd, b'', target.send)
-        self.assertEqual(len(record), 1)
-        self.assertEqual(record[0], b'?##\xab\xcd\x00\x00\x00\x00' + '\0' * 55)
-
-    def test_encode_one_report_aligned(self):
-        data = bytes(range(0, 55))
-
-        record = []
-        target = self._record(record)()
-        target.send(None)
-
-        codec_v1.encode(codec_v1.SESSION, 0xabcd, data, target.send)
-        self.assertEqual(record, [b'?##\xab\xcd\x00\x00\x00\x37' + data])
-
-    def test_encode_generated_range(self):
-        for data_len in range(1, 1024):
-            data = random.bytes(data_len)
-
-            msg_type = 0xabcd
-            header = b'##' + ustruct.pack('>H', msg_type) + ustruct.pack('>L', data_len)
-
-            message = header + data
-            reports = [b'?' + c for c in chunks(message, 63)]
-            reports[-1] = reports[-1] + b'\x00' * (64 - len(reports[-1]))
-
-            received = 0
-            def genfunc():
-                nonlocal received
-                while True:
-                    self.assertEqual((yield), reports[received])
-                    received += 1
-            target = genfunc()
-            target.send(None)
-
-            codec_v1.encode(codec_v1.SESSION, msg_type, data, target.send)
-            self.assertEqual(received, len(reports))
-
-    def _record(self, record, *_args):
-        def genfunc(*args):
-            self.assertEqual(args, _args)
-            while True:
-                try:
-                    v = yield
-                except Exception as e:
-                    record.append(e)
-                else:
-                    record.append(v)
-        return genfunc
+    # long write, expected multiple reports
+    start_size = writer.size
+    long_payload_head = bytearray(range(rep_len - len(report_header) - len(short_payload)))
+    long_payload_rest = bytearray(range(start_size - len(long_payload_head)))
+    long_payload = long_payload_head + long_payload_rest
+    expected_payloads = [short_payload + long_payload_head] + list(chunks(long_payload_rest, rep_len - len(report_header)))
+    expected_reports = [report_header + r for r in expected_payloads]
+    expected_reports[-1] += bytearray(bytes(1) * (rep_len - len(expected_reports[-1])))
+    # test write
+    expected_write_reports = expected_reports[:-1]
+    msg.send = mock_call(msg.send, [(interface, rep) for rep in expected_write_reports])
+    assert_async(writer.write(long_payload), len(expected_write_reports) * [(None, Select(WRITE | interface))] + [(None, StopIteration())])
+    assert_eq(writer.size, start_size - len(long_payload))
+    msg.send.assert_called_n_times(len(expected_write_reports))
+    msg.send = msg.send.original
+    # test write raises eof
+    msg.send = mock_call(msg.send, [])
+    assert_async(writer.write(bytearray(1)), [(None, EOFError())])
+    msg.send.assert_called_n_times(0)
+    msg.send = msg.send.original
+    # test close
+    expected_close_reports = expected_reports[-1:]
+    msg.send = mock_call(msg.send, [(interface, rep) for rep in expected_close_reports])
+    assert_async(writer.close(), len(expected_close_reports) * [(None, Select(WRITE | interface))] + [(None, StopIteration())])
+    assert_eq(writer.size, 0)
+    msg.send.assert_called_n_times(len(expected_close_reports))
+    msg.send = msg.send.original
 
 
 if __name__ == '__main__':
-    unittest.main()
+    run_tests()
diff --git a/tests/test_trezor.wire.codec_v2.py b/tests/test_trezor.wire.codec_v2.py
index 60226b4e..56fc6ab0 100644
--- a/tests/test_trezor.wire.codec_v2.py
+++ b/tests/test_trezor.wire.codec_v2.py
@@ -1,219 +1,167 @@
-from common import *
+import sys
 
-import ustruct
-import ubinascii
+sys.path.append('../src')
+sys.path.append('../src/lib')
 
-from trezor.crypto import random
+from utest import *
+from ustruct import pack, unpack
+from ubinascii import hexlify, unhexlify
+
+from trezor import msg
+from trezor.loop import Select, Syscall, READ, WRITE
 from trezor.utils import chunks
-
 from trezor.wire import codec_v2
 
-class TestWireCodec(unittest.TestCase):
-    # pylint: disable=C0301
 
-    def test_parse(self):
-        d = b'O' + b'\x01\x23\x45\x67' + bytes(range(0, 59))
+def test_reader():
+    rep_len = 64
+    interface = 0xdeadbeef
+    session_id = 0x12345678
+    message_type = 0x87654321
+    message_len = 250
+    reader = codec_v2.Reader(interface, session_id)
 
-        m, s, d = codec_v2.parse_report(d)
-        self.assertEqual(m, b'O'[0])
-        self.assertEqual(s, 0x01234567)
-        self.assertEqual(d, bytes(range(0, 59)))
+    message = bytearray(range(message_len))
+    report_header = bytearray(unhexlify('011234567887654321000000fa'))
 
-        t, l, d = codec_v2.parse_message(d)
-        self.assertEqual(t, 0x00010203)
-        self.assertEqual(l, 0x04050607)
-        self.assertEqual(d, bytes(range(8, 59)))
+    # open, expected one read
+    first_report = report_header + message[:rep_len - len(report_header)]
+    assert_async(reader.open(), [(None, Select(READ | interface)), (first_report, StopIteration()),])
+    assert_eq(reader.type, message_type)
+    assert_eq(reader.size, message_len)
 
-        f, = codec_v2.parse_message_footer(d[0:4])
-        self.assertEqual(f, 0x08090a0b)
+    # empty read
+    empty_buffer = bytearray()
+    assert_async(reader.readinto(empty_buffer), [(None, StopIteration()),])
+    assert_eq(len(empty_buffer), 0)
+    assert_eq(reader.size, message_len)
 
-        for i in range(0, 1024):
-            if i != 64:
-                with self.assertRaises(ValueError):
-                    codec_v2.parse_report(bytes(range(0, i)))
-            if i != 59:
-                with self.assertRaises(ValueError):
-                    codec_v2.parse_message(bytes(range(0, i)))
-            if i != 4:
-                with self.assertRaises(ValueError):
-                    codec_v2.parse_message_footer(bytes(range(0, i)))
+    # short read, expected no read
+    short_buffer = bytearray(32)
+    assert_async(reader.readinto(short_buffer), [(None, StopIteration()),])
+    assert_eq(len(short_buffer), 32)
+    assert_eq(short_buffer, message[:len(short_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer))
 
-    def test_serialize(self):
-        data = bytearray(range(0, 6))
-        codec_v2.serialize_report_header(data, 0x12, 0x3456789a)
-        self.assertEqual(data, b'\x12\x34\x56\x78\x9a\x05')
+    # aligned read, expected no read
+    aligned_buffer = bytearray(rep_len - len(report_header) - len(short_buffer))
+    assert_async(reader.readinto(aligned_buffer), [(None, StopIteration()),])
+    assert_eq(aligned_buffer, message[len(short_buffer):][:len(aligned_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer) - len(aligned_buffer))
 
-        data = bytearray(range(0, 6))
-        codec_v2.serialize_opened_session(data, 0x3456789a)
-        self.assertEqual(data, bytes([codec_v2.REP_MARKER_OPEN]) + b'\x34\x56\x78\x9a\x05')
+    # one byte read, expected one read
+    next_report_header = bytearray(unhexlify('021234567800000000'))
+    next_report = next_report_header + message[rep_len - len(report_header):][:rep_len - len(next_report_header)]
+    onebyte_buffer = bytearray(1)
+    assert_async(reader.readinto(onebyte_buffer), [(None, Select(READ | interface)), (next_report, StopIteration()),])
+    assert_eq(onebyte_buffer, message[len(short_buffer):][len(aligned_buffer):][:len(onebyte_buffer)])
+    assert_eq(reader.size, message_len - len(short_buffer) - len(aligned_buffer) - len(onebyte_buffer))
 
-        data = bytearray(range(0, 14))
-        codec_v2.serialize_message_header(data, 0x01234567, 0x89abcdef)
-        self.assertEqual(data, b'\x00\x01\x02\x03\x04\x01\x23\x45\x67\x89\xab\xcd\xef\x0d')
+    # too long read, raises eof
+    assert_async(reader.readinto(bytearray(reader.size + 1)), [(None, EOFError()),])
 
-        data = bytearray(range(0, 5))
-        codec_v2.serialize_message_footer(data, 0x89abcdef)
-        self.assertEqual(data, b'\x89\xab\xcd\xef\x04')
+    # long read, expect multiple reads
+    start_size = reader.size
+    long_buffer = bytearray(start_size)
+    report_payload = message[rep_len - len(report_header) + rep_len - len(next_report_header):]
+    report_payload_head = report_payload[:rep_len - len(next_report_header) - len(onebyte_buffer)]
+    report_payload_rest = report_payload[len(report_payload_head):]
+    report_payload_rest = list(chunks(report_payload_rest, rep_len - len(next_report_header)))
+    report_payloads = [report_payload_head] + report_payload_rest
+    next_reports = [bytearray(unhexlify('0212345678') + pack('>L', i + 1)) + r for i, r in enumerate(report_payloads)]
+    expected_syscalls = []
+    for i, _ in enumerate(next_reports):
+        prev_report = next_reports[i - 1] if i > 0 else None
+        expected_syscalls.append((prev_report, Select(READ | interface)))
+    expected_syscalls.append((next_reports[-1], StopIteration()))
+    assert_async(reader.readinto(long_buffer), expected_syscalls)
+    assert_eq(long_buffer, message[-start_size:])
+    assert_eq(reader.size, 0)
 
-        for i in range(0, 13):
-            data = bytearray(i)
-            if i < 4:
-                with self.assertRaises(ValueError):
-                    codec_v2.serialize_message_footer(data, 0x00)
-            if i < 5:
-                with self.assertRaises(ValueError):
-                    codec_v2.serialize_report_header(data, 0x00, 0x00)
-                with self.assertRaises(ValueError):
-                    codec_v2.serialize_opened_session(data, 0x00)
-            with self.assertRaises(ValueError):
-                codec_v2.serialize_message_header(data, 0x00, 0x00)
+    # one byte read, raises eof
+    assert_async(reader.readinto(onebyte_buffer), [(None, EOFError()),])
 
-    def test_decode_empty(self):
-        message = b'\xab\xcd\xef\x12' + b'\x00\x00\x00\x00' + b'\x00' * 51
 
-        record = []
-        genfunc = self._record(record, 0xdeadbeef, 0xabcdef12, 0, 'dummy')
-        decoder = codec_v2.decode_stream(0xdeadbeef, genfunc, 'dummy')
-        decoder.send(None)
+def test_writer():
+    rep_len = 64
+    interface = 0xdeadbeef
+    session_id = 0x12345678
+    message_type = 0x87654321
+    message_len = 1024
+    writer = codec_v2.Writer(interface, session_id, message_type, message_len)
 
-        try:
-            decoder.send(message)
-        except StopIteration as e:
-            res = e.value
-        self.assertEqual(res, None)
-        self.assertEqual(len(record), 1)
-        self.assertIsInstance(record[0], EOFError)
+    # init header corresponding to the data above
+    report_header = bytearray(unhexlify('01123456788765432100000400'))
 
-    def test_decode_one_report_aligned_correct(self):
-        data = bytes(range(0, 47))
-        footer = b'\x2f\x1c\x12\xce'
-        message = b'\xab\xcd\xef\x12' + b'\x00\x00\x00\x2f' + data + footer
+    assert_eq(writer.data, report_header + bytearray(64 - len(report_header)))
 
-        record = []
-        genfunc = self._record(record, 0xdeadbeef, 0xabcdef12, 47, 'dummy')
-        decoder = codec_v2.decode_stream(0xdeadbeef, genfunc, 'dummy')
-        decoder.send(None)
+    # empty write
+    start_size = writer.size
+    assert_async(writer.write(bytearray()), [(None, StopIteration()),])
+    assert_eq(writer.data, report_header + bytearray(rep_len - len(report_header)))
+    assert_eq(writer.size, start_size)
 
-        try:
-            decoder.send(message)
-        except StopIteration as e:
-            res = e.value
-        self.assertEqual(res, None)
-        self.assertEqual(len(record), 2)
-        self.assertEqual(record[0], data)
-        self.assertIsInstance(record[1], EOFError)
+    # short write, expected no report
+    start_size = writer.size
+    short_payload = bytearray(range(4))
+    assert_async(writer.write(short_payload), [(None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(short_payload))
+    assert_eq(writer.data,
+              report_header
+              + short_payload
+              + bytearray(rep_len - len(report_header) - len(short_payload)))
 
-    def test_decode_one_report_aligned_incorrect(self):
-        data = bytes(range(0, 47))
-        footer = bytes(4)  # wrong checksum
-        message = b'\xab\xcd\xef\x12' + b'\x00\x00\x00\x2f' + data + footer
+    # aligned write, expected one report
+    start_size = writer.size
+    aligned_payload = bytearray(range(rep_len - len(report_header) - len(short_payload)))
+    msg.send = mock_call(msg.send, [
+        (interface, report_header
+         + short_payload
+         + aligned_payload
+         + bytearray(rep_len - len(report_header) - len(short_payload) - len(aligned_payload))), ])
+    assert_async(writer.write(aligned_payload), [(None, Select(WRITE | interface)), (None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(aligned_payload))
+    msg.send.assert_called_n_times(1)
+    msg.send = msg.send.original
 
-        record = []
-        genfunc = self._record(record, 0xdeadbeef, 0xabcdef12, 47, 'dummy')
-        decoder = codec_v2.decode_stream(0xdeadbeef, genfunc, 'dummy')
-        decoder.send(None)
+    # short write, expected no report, but data starts with correct seq and cont marker
+    report_header = bytearray(unhexlify('021234567800000000'))
+    start_size = writer.size
+    assert_async(writer.write(short_payload), [(None, StopIteration()),])
+    assert_eq(writer.size, start_size - len(short_payload))
+    assert_eq(writer.data[:len(report_header) + len(short_payload)],
+              report_header + short_payload)
 
-        try:
-            decoder.send(message)
-        except StopIteration as e:
-            res = e.value
-        self.assertEqual(res, None)
-        self.assertEqual(len(record), 2)
-        self.assertEqual(record[0], data)
-        self.assertIsInstance(record[1], codec_v2.MessageChecksumError)
-
-    def test_decode_generated_range(self):
-        for data_len in range(1, 512):
-            data = random.bytes(data_len)
-            data_chunks = [data[:51]] + list(chunks(data[51:], 59))
-
-            msg_type = 0xabcdef12
-            data_csum = ubinascii.crc32(data)
-            header = ustruct.pack('>L', msg_type) + ustruct.pack('>L', data_len)
-            footer = ustruct.pack('>L', data_csum)
-
-            message = header + data + footer
-            message_chunks = [c + '\x00' * (59 - len(c)) for c in list(chunks(message, 59))]
-
-            record = []
-            genfunc = self._record(record, 0xdeadbeef, msg_type, data_len, 'dummy')
-            decoder = codec_v2.decode_stream(0xdeadbeef, genfunc, 'dummy')
-            decoder.send(None)
-
-            res = 1
-            try:
-                for c in message_chunks:
-                    decoder.send(c)
-            except StopIteration as e:
-                res = e.value
-            self.assertEqual(res, None)
-            self.assertEqual(len(record), len(data_chunks) + 1)
-            for i in range(0, len(data_chunks)):
-                self.assertEqual(record[i], data_chunks[i])
-            self.assertIsInstance(record[-1], EOFError)
-
-    def test_encode_empty(self):
-        record = []
-        target = self._record(record)()
-        target.send(None)
-
-        codec_v2.encode(0xdeadbeef, 0xabcdef12, b'', target.send)
-        self.assertEqual(len(record), 1)
-        self.assertEqual(record[0], b'H\xde\xad\xbe\xef\xab\xcd\xef\x12\x00\x00\x00\x00' + '\0' * 51)
-
-    def test_encode_one_report_aligned(self):
-        data = bytes(range(0, 47))
-        footer = b'\x2f\x1c\x12\xce'
-
-        record = []
-        target = self._record(record)()
-        target.send(None)
-
-        codec_v2.encode(0xdeadbeef, 0xabcdef12, data, target.send)
-        self.assertEqual(record, [b'H\xde\xad\xbe\xef\xab\xcd\xef\x12\x00\x00\x00\x2f' + data + footer])
-
-    def test_encode_generated_range(self):
-        for data_len in range(1, 1024):
-            data = random.bytes(data_len)
-
-            msg_type = 0xabcdef12
-            session_id = 0xdeadbeef
-
-            data_csum = ubinascii.crc32(data)
-            header = ustruct.pack('>L', msg_type) + ustruct.pack('>L', data_len)
-            footer = ustruct.pack('>L', data_csum)
-            session_header = ustruct.pack('>L', session_id)
-
-            message = header + data + footer
-            report0 = b'H' + session_header + message[:59]
-            reports = [b'D' + session_header + c for c in chunks(message[59:], 59)]
-            reports.insert(0, report0)
-            reports[-1] = reports[-1] + b'\x00' * (64 - len(reports[-1]))
-
-            received = 0
-            def genfunc():
-                nonlocal received
-                while True:
-                    self.assertEqual((yield), reports[received])
-                    received += 1
-            target = genfunc()
-            target.send(None)
-
-            codec_v2.encode(session_id, msg_type, data, target.send)
-            self.assertEqual(received, len(reports))
-
-    def _record(self, record, *_args):
-        def genfunc(*args):
-            self.assertEqual(args, _args)
-            while True:
-                try:
-                    v = yield
-                except Exception as e:
-                    record.append(e)
-                else:
-                    record.append(v)
-        return genfunc
+    # long write, expected multiple reports
+    start_size = writer.size
+    long_payload_head = bytearray(range(rep_len - len(report_header) - len(short_payload)))
+    long_payload_rest = bytearray(range(start_size - len(long_payload_head)))
+    long_payload = long_payload_head + long_payload_rest
+    expected_payloads = [short_payload + long_payload_head] + list(chunks(long_payload_rest, rep_len - len(report_header)))
+    expected_reports = [
+        bytearray(unhexlify('0212345678') + pack('>L', seq)) + rep
+        for seq, rep in enumerate(expected_payloads)]
+    expected_reports[-1] += bytearray(bytes(1) * (rep_len - len(expected_reports[-1])))
+    # test write
+    expected_write_reports = expected_reports[:-1]
+    msg.send = mock_call(msg.send, [(interface, rep) for rep in expected_write_reports])
+    assert_async(writer.write(long_payload), len(expected_write_reports) * [(None, Select(WRITE | interface))] + [(None, StopIteration())])
+    assert_eq(writer.size, start_size - len(long_payload))
+    msg.send.assert_called_n_times(len(expected_write_reports))
+    msg.send = msg.send.original
+    # test write raises eof
+    msg.send = mock_call(msg.send, [])
+    assert_async(writer.write(bytearray(1)), [(None, EOFError())])
+    msg.send.assert_called_n_times(0)
+    msg.send = msg.send.original
+    # test close
+    expected_close_reports = expected_reports[-1:]
+    msg.send = mock_call(msg.send, [(interface, rep) for rep in expected_close_reports])
+    assert_async(writer.close(), len(expected_close_reports) * [(None, Select(WRITE | interface))] + [(None, StopIteration())])
+    assert_eq(writer.size, 0)
+    msg.send.assert_called_n_times(len(expected_close_reports))
+    msg.send = msg.send.original
 
 
 if __name__ == '__main__':
-    unittest.main()
+    run_tests()
diff --git a/tests/utest.py b/tests/utest.py
new file mode 100644
index 00000000..34cc52e3
--- /dev/null
+++ b/tests/utest.py
@@ -0,0 +1,142 @@
+import sys
+import uio
+import ure
+
+__all__ = [
+    'run_tests',
+    'run_test',
+    'assert_eq',
+    'assert_not_eq',
+    'assert_is_instance',
+    'mock_call',
+]
+
+
+# Running
+
+
+def run_tests(mod_name='__main__'):
+    ntotal = 0
+    nok = 0
+    nfailed = 0
+
+    for name, test in get_tests(mod_name):
+        result = run_test(test)
+        report_test(name, test, result)
+        ntotal += 1
+        if result:
+            nok += 1
+        else:
+            nfailed += 1
+            break
+    report_total(ntotal, nok, nfailed)
+
+    if nfailed > 0:
+        sys.exit(1)
+
+
+def get_tests(mod_name):
+    module = __import__(mod_name)
+    for name in dir(module):
+        if name.startswith('test_'):
+            yield name, getattr(module, name)
+
+
+def run_test(test):
+    try:
+        test()
+    except Exception as e:
+        report_exception(e)
+        return False
+    else:
+        return True
+
+
+# Reporting
+
+
+def report_test(name, test, result):
+    if result:
+        print('OK', name)
+    else:
+        print('ERR', name)
+
+
+def report_exception(exc):
+    sio = uio.StringIO()
+    sys.print_exception(exc, sio)
+    print(sio.getvalue())
+
+
+def report_total(total, ok, failed):
+    print('Total:', total, 'OK:', ok, 'Failed:', failed)
+
+
+# Assertions
+
+
+def assert_eq(a, b, msg=None):
+    assert a == b, msg or format_eq(a, b)
+
+
+def assert_not_eq(a, b, msg=None):
+    assert a != b, msg or format_not_eq(a, b)
+
+
+def assert_is_instance(obj, cls, msg=None):
+    assert isinstance(obj, cls), msg or format_is_instance(obj, cls)
+
+
+def assert_eq_obj(a, b, msg=None):
+    assert_is_instance(a, b.__class__, msg)
+    assert_eq(a.__dict__, b.__dict__, msg)
+
+
+def format_eq(a, b):
+    return '\n%r\nvs (expected)\n%r' % (a, b)
+
+
+def format_not_eq(a, b):
+    return '%r not expected to be equal %r' % (a, b)
+
+
+def format_is_instance(obj, cls):
+    return '%r expected to be instance of %r' % (obj, cls)
+
+
+def assert_async(task, syscalls):
+    for prev_result, expected in syscalls:
+        if isinstance(expected, Exception):
+            with assert_raises(expected.__class__):
+                task.send(prev_result)
+        else:
+            syscall = task.send(prev_result)
+            assert_eq_obj(syscall, expected)
+
+
+class assert_raises:
+
+    def __init__(self, exc_type):
+        self.exc_type = exc_type
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        assert exc_type is not None, '%r not raised' % self.exc_type
+        return issubclass(exc_type, self.exc_type)
+
+
+class mock_call:
+
+    def __init__(self, original, expected):
+        self.original = original
+        self.expected = expected
+        self.record = []
+
+    def __call__(self, *args):
+        self.record.append(args)
+        assert_eq(args, self.expected.pop(0))
+
+    def assert_called_n_times(self, n, msg=None):
+        assert_eq(len(self.record), n, msg)
\ No newline at end of file