diff --git a/py/test/dist/dsession.py b/py/test/dist/dsession.py
index 3cf2394f0..d517993da 100644
--- a/py/test/dist/dsession.py
+++ b/py/test/dist/dsession.py
@@ -25,8 +25,8 @@ class LoopState(object):
         self.testsfailed = False
 
     def pytest_itemtestreport(self, rep):
-        if rep.colitem in self.dsession.item2nodes:
-            self.dsession.removeitem(rep.colitem, rep.node)
+        if rep.item in self.dsession.item2nodes:
+            self.dsession.removeitem(rep.item, rep.node)
         if rep.failed:
             self.testsfailed = True
 
diff --git a/py/test/dist/testing/test_dsession.py b/py/test/dist/testing/test_dsession.py
index 0c237b263..0ebce2ffa 100644
--- a/py/test/dist/testing/test_dsession.py
+++ b/py/test/dist/testing/test_dsession.py
@@ -180,10 +180,10 @@ class TestDSession:
         session.loop_once(loopstate)
 
         assert loopstate.colitems == [item2] # do not reschedule crash item
-        testrep = reprec.matchreport(names="pytest_itemtestreport")
-        assert testrep.failed
-        assert testrep.colitem == item1
-        assert str(testrep.longrepr).find("crashed") != -1
+        rep = reprec.matchreport(names="pytest_itemtestreport")
+        assert rep.failed
+        assert rep.item == item1
+        assert str(rep.longrepr).find("crashed") != -1
         #assert str(testrep.longrepr).find(node.gateway.spec) != -1
 
     def test_testnodeready_adds_to_available(self, testdir):
diff --git a/py/test/dist/testing/test_txnode.py b/py/test/dist/testing/test_txnode.py
index c96d972ba..d988a798d 100644
--- a/py/test/dist/testing/test_txnode.py
+++ b/py/test/dist/testing/test_txnode.py
@@ -115,7 +115,7 @@ class TestMasterSlaveConnection:
         rep = kwargs['rep'] 
         assert rep.passed 
         print rep
-        assert rep.colitem == item
+        assert rep.item == item
 
     def test_send_some(self, testdir, mysetup):
         items = testdir.getitems("""
diff --git a/py/test/looponfail/remote.py b/py/test/looponfail/remote.py
index 6046bc531..6ff438e55 100644
--- a/py/test/looponfail/remote.py
+++ b/py/test/looponfail/remote.py
@@ -150,4 +150,4 @@ def slave_runsession(channel, config, fullwidth, hasmarkup):
     session.config.hook.pytest_looponfailinfo(
         failreports=list(failreports), 
         rootdirs=[config.topdir])
-    channel.send([x.colitem._totrail() for x in failreports])
+    channel.send([rep.getnode()._totrail() for rep in failreports])
diff --git a/py/test/plugin/api.py b/py/test/plugin/api.py
index eceb1d7a6..6739d38f1 100644
--- a/py/test/plugin/api.py
+++ b/py/test/plugin/api.py
@@ -84,10 +84,11 @@ class PluginHooks:
     def pytest_itemtestreport(self, rep):
         """ test has been run. """
 
+    # XXX pytest_runner reports 
     def pytest_item_runtest_finished(self, item, excinfo, outerr):
         """ test has been run. """
 
-    def pytest_itemsetupreport(self, rep):
+    def pytest_itemfixturereport(self, rep):
         """ a report on running a fixture function. """ 
 
     # ------------------------------------------------------------------------------
diff --git a/py/test/plugin/pytest__pytest.py b/py/test/plugin/pytest__pytest.py
index 711d1053a..5e720eb01 100644
--- a/py/test/plugin/pytest__pytest.py
+++ b/py/test/plugin/pytest__pytest.py
@@ -91,7 +91,7 @@ class HookRecorder:
             if call._name == name:
                 del self.calls[i]
                 return call 
-        raise ValueError("could not find call %r in %r" %(name, self.calls))
+        raise ValueError("could not find call %r" %(name, ))
 
     def getcall(self, name):
         l = self.getcalls(name)
diff --git a/py/test/plugin/pytest_doctest.py b/py/test/plugin/pytest_doctest.py
index a39e53771..ddd22f1fa 100644
--- a/py/test/plugin/pytest_doctest.py
+++ b/py/test/plugin/pytest_doctest.py
@@ -141,7 +141,6 @@ class TestDoctests:
         reprec.assertoutcome(failed=1) 
 
     def test_txtfile_failing(self, testdir):
-        testdir.plugins.append("doctest")
         p = testdir.maketxtfile("""
             >>> i = 0
             >>> i + 1
diff --git a/py/test/plugin/pytest_pytester.py b/py/test/plugin/pytest_pytester.py
index 468bb2175..23406907a 100644
--- a/py/test/plugin/pytest_pytester.py
+++ b/py/test/plugin/pytest_pytester.py
@@ -305,6 +305,9 @@ class ReportRecorder(object):
     def getcall(self, name):
         return self.hookrecorder.getcall(name)
 
+    def popcall(self, name):
+        return self.hookrecorder.popcall(name)
+
     def getcalls(self, names):
         """ return list of ParsedCall instances matching the given eventname. """
         return self.hookrecorder.getcalls(names)
@@ -318,7 +321,8 @@ class ReportRecorder(object):
         """ return a testreport whose dotted import path matches """
         l = []
         for rep in self.getreports(names=names):
-            if not inamepart or inamepart in rep.colitem.listnames():
+            colitem = rep.getnode()
+            if not inamepart or inamepart in colitem.listnames():
                 l.append(rep)
         if not l:
             raise ValueError("could not find test report matching %r: no test reports at all!" %
diff --git a/py/test/plugin/pytest_resultlog.py b/py/test/plugin/pytest_resultlog.py
index 2a8ab2dbe..12c9f2a1e 100644
--- a/py/test/plugin/pytest_resultlog.py
+++ b/py/test/plugin/pytest_resultlog.py
@@ -20,7 +20,7 @@ def pytest_unconfigure(config):
     resultlog = getattr(config, '_resultlog', None)
     if resultlog:
         resultlog.logfile.close()
-        del config.resultlog 
+        del config._resultlog 
         config.pluginmanager.unregister(resultlog)
 
 def generic_path(item):
@@ -55,8 +55,8 @@ class ResultLog(object):
         for line in longrepr.splitlines():
             print >>self.logfile, " %s" % line
 
-    def log_outcome(self, event, shortrepr, longrepr):
-        testpath = generic_path(event.colitem)
+    def log_outcome(self, node, shortrepr, longrepr):
+        testpath = generic_path(node)
         self.write_log_entry(testpath, shortrepr, longrepr) 
 
     def pytest_itemtestreport(self, rep):
@@ -67,7 +67,7 @@ class ResultLog(object):
             longrepr = str(rep.longrepr) 
         elif rep.skipped:
             longrepr = str(rep.longrepr.reprcrash.message)
-        self.log_outcome(rep, code, longrepr) 
+        self.log_outcome(rep.item, code, longrepr) 
 
     def pytest_collectreport(self, rep):
         if not rep.passed:
@@ -77,7 +77,7 @@ class ResultLog(object):
                 assert rep.skipped
                 code = "S"
             longrepr = str(rep.longrepr.reprcrash)
-            self.log_outcome(rep, code, longrepr)    
+            self.log_outcome(rep.collector, code, longrepr)    
 
     def pytest_internalerror(self, excrepr):
         path = excrepr.reprcrash.path 
diff --git a/py/test/plugin/pytest_runner.py b/py/test/plugin/pytest_runner.py
index 1faecf130..1fb196d5e 100644
--- a/py/test/plugin/pytest_runner.py
+++ b/py/test/plugin/pytest_runner.py
@@ -1,31 +1,39 @@
 import py
 from py.__.test.outcome import Skipped
+from py.__.test.runner import SetupState
 
-class RunnerPlugin:
-    def pytest_configure(self, config):
-        config._setupstate = SetupState()
+py.test.skip("remove me")
 
-    def pytest_unconfigure(self, config):
-        config._setupstate.teardown_all()
+def pytest_configure(config):
+    config._setupstate = SetupState()
 
-    def pytest_item_setup_and_runtest(self, item):
-        setupstate = item.config._setupstate
-        call = item.config.guardedcall(lambda: setupstate.prepare(item))
-        if call.excinfo:
-            rep = ItemSetupReport(item, call.excinfo, call.outerr)
-            item.config.hook.pytest_itemsetupreport(rep=rep)
-        else:
-            call = item.config.guardedcall(lambda: item.runtest())
-            item.config.hook.pytest_item_runtest_finished(
-                item=item, excinfo=call.excinfo, outerr=call.outerr)
-            call = item.config.guardedcall(lambda: self.teardown_exact(item))
-            if call.excinfo:
-                rep = ItemSetupReport(item, call.excinfo, call.outerr)
-                item.config.hook.pytest_itemsetupreport(rep=rep)
+def pytest_unconfigure(config):
+    config._setupstate.teardown_all()
 
-    def pytest_collector_collect(self, collector):
-        call = item.config.guardedcall(lambda x: collector._memocollect())
-        return CollectReport(collector, res, excinfo, outerr)
+def pytest_item_setup_and_runtest(item):
+    setupstate = item.config._setupstate
+
+    # setup (and implied teardown of previous items) 
+    call = item.config.guardedcall(lambda: setupstate.prepare(item))
+    if call.excinfo:
+        rep = ItemTestReport(item, call.excinfo, call.outerr)
+        item.config.hook.pytest_itemfixturereport(rep=rep)
+        return
+
+    # runtest 
+    call = item.config.guardedcall(lambda: item.runtest())
+    item.config.hook.pytest_item_runtest_finished(
+        item=item, excinfo=call.excinfo, outerr=call.outerr)
+
+    # teardown 
+    call = item.config.guardedcall(lambda: setupstate.teardown_exact(item))
+    if call.excinfo:
+        rep = ItemFixtureReport(item, call.excinfo, call.outerr)
+        item.config.hook.pytest_itemfixturereport(rep=rep)
+
+def pytest_collector_collect(collector):
+    call = item.config.guardedcall(lambda x: collector._memocollect())
+    return CollectReport(collector, res, excinfo, outerr)
 
 class BaseReport(object):
     def __repr__(self):
@@ -97,7 +105,7 @@ class CollectReport(BaseReport):
             else:
                 self.failed = True
 
-class ItemSetupReport(BaseReport):
+class ItemFixtureReport(BaseReport):
     failed = passed = skipped = False
     def __init__(self, item, excinfo=None, outerr=None):
         self.item = item 
@@ -111,33 +119,6 @@ class ItemSetupReport(BaseReport):
                 self.failed = True
             self.excrepr = item._repr_failure_py(excinfo, [])
 
-class SetupState(object):
-    """ shared state for setting up/tearing down test items or collectors. """
-    def __init__(self):
-        self.stack = []
-
-    def teardown_all(self): 
-        while self.stack: 
-            col = self.stack.pop() 
-            col.teardown() 
-
-    def teardown_exact(self, item):
-        if self.stack and self.stack[-1] == item:
-            col = self.stack.pop()
-            col.teardown()
-     
-    def prepare(self, colitem): 
-        """ setup objects along the collector chain to the test-method
-            Teardown any unneccessary previously setup objects."""
-        needed_collectors = colitem.listchain() 
-        while self.stack: 
-            if self.stack == needed_collectors[:len(self.stack)]: 
-                break 
-            col = self.stack.pop() 
-            col.teardown()
-        for col in needed_collectors[len(self.stack):]: 
-            col.setup() 
-            self.stack.append(col) 
 
 # ===============================================================================
 #
@@ -206,67 +187,71 @@ class TestSetupState:
 class TestRunnerPlugin:
     def test_pytest_item_setup_and_runtest(self, testdir):
         item = testdir.getitem("""def test_func(): pass""")
-        plugin = RunnerPlugin()
-        plugin.pytest_configure(item.config)
+        pytest_configure(item.config)
         reprec = testdir.getreportrecorder(item)
-        plugin.pytest_item_setup_and_runtest(item)
+        pytest_item_setup_and_runtest(item)
         rep = reprec.getcall("pytest_itemtestreport").rep
         assert rep.passed 
-        
+       
 class TestSetupEvents:
-    disabled = True
-
-    def test_setup_runtest_ok(self, testdir):
-        reprec = testdir.inline_runsource("""
+    def pytest_funcarg__runfunc(self, request):
+        testdir = request.getfuncargvalue("testdir")
+        def runfunc(source):
+            item = testdir.getitem(source)
+            reprec = testdir.getreportrecorder(item)
+            pytest_item_setup_and_runtest(item)
+            return reprec 
+        return runfunc
+        
+    def test_setup_runtest_ok(self, runfunc):
+        reprec = runfunc("""
             def setup_module(mod):
                 pass 
             def test_func():
                 pass
         """)
-        assert not reprec.getcalls(pytest_itemsetupreport)
+        assert not reprec.getcalls("pytest_itemfixturereport")
 
-    def test_setup_fails(self, testdir):
-        reprec = testdir.inline_runsource("""
+    def test_setup_fails(self, runfunc):
+        reprec = runfunc(""" 
             def setup_module(mod):
                 print "world"
                 raise ValueError(42)
             def test_func():
                 pass
         """)
-        rep = reprec.popcall(pytest_itemsetupreport).rep
+        rep = reprec.popcall("pytest_itemfixturereport").rep
         assert rep.failed
         assert not rep.skipped
         assert rep.excrepr 
         assert "42" in str(rep.excrepr)
         assert rep.outerr[0].find("world") != -1
 
-    def test_teardown_fails(self, testdir):
-        reprec = testdir.inline_runsource("""
+    def test_teardown_fails(self, runfunc):
+        reprec = runfunc(""" 
             def test_func():
                 pass
             def teardown_function(func): 
                 print "13"
                 raise ValueError(25)
         """)
-        rep = reprec.popcall(pytest_itemsetupreport).rep 
+        rep = reprec.popcall("pytest_itemfixturereport").rep 
         assert rep.failed 
-        assert rep.item == item 
+        assert rep.item.name == "test_func" 
         assert not rep.passed
         assert "13" in rep.outerr[0]
         assert "25" in str(rep.excrepr)
 
-    def test_setup_skips(self, testdir):
-        reprec = testdir.inline_runsource("""
+    def test_setup_skips(self, runfunc):
+        reprec = runfunc(""" 
             import py
             def setup_module(mod):
                 py.test.skip("17")
             def test_func():
                 pass
         """)
-        rep = reprec.popcall(pytest_itemsetupreport)
+        rep = reprec.popcall("pytest_itemfixturereport").rep
         assert not rep.failed
         assert rep.skipped
         assert rep.excrepr 
         assert "17" in str(rep.excrepr)
-    
-
diff --git a/py/test/plugin/pytest_terminal.py b/py/test/plugin/pytest_terminal.py
index 4598024ae..b53036a7f 100644
--- a/py/test/plugin/pytest_terminal.py
+++ b/py/test/plugin/pytest_terminal.py
@@ -162,7 +162,7 @@ class TerminalReporter:
                 self.write_fspath_result(fspath, "")
 
     def pytest_itemtestreport(self, rep):
-        fspath = rep.colitem.fspath 
+        fspath = rep.item.fspath 
         cat, letter, word = self.getcategoryletterword(rep)
         if isinstance(word, tuple):
             word, markup = word
@@ -170,10 +170,10 @@ class TerminalReporter:
             markup = {}
         self.stats.setdefault(cat, []).append(rep)
         if not self.config.option.verbose:
-            fspath, lineno, msg = self._getreportinfo(rep.colitem)
+            fspath, lineno, msg = self._getreportinfo(rep.item)
             self.write_fspath_result(fspath, letter)
         else:
-            line = self._reportinfoline(rep.colitem)
+            line = self._reportinfoline(rep.item)
             if not hasattr(rep, 'node'):
                 self.write_ensure_prefix(line, word, **markup)
             else:
@@ -189,10 +189,10 @@ class TerminalReporter:
             if rep.failed:
                 self.stats.setdefault("failed", []).append(rep)
                 msg = rep.longrepr.reprcrash.message 
-                self.write_fspath_result(rep.colitem.fspath, "F")
+                self.write_fspath_result(rep.collector.fspath, "F")
             elif rep.skipped:
                 self.stats.setdefault("skipped", []).append(rep)
-                self.write_fspath_result(rep.colitem.fspath, "S")
+                self.write_fspath_result(rep.collector.fspath, "S")
 
     def pytest_testrunstart(self):
         self.write_sep("=", "test session starts", bold=True)
@@ -268,10 +268,10 @@ class TerminalReporter:
         return line % locals() + " "
         
     def _getfailureheadline(self, rep):
-        if isinstance(rep.colitem, py.test.collect.Collector):
-            return str(rep.colitem.fspath)
+        if hasattr(rep, "collector"):
+            return str(rep.collector.fspath)
         else:
-            fspath, lineno, msg = self._getreportinfo(rep.colitem)
+            fspath, lineno, msg = self._getreportinfo(rep.item)
             return msg
 
     def _getreportinfo(self, item):
diff --git a/py/test/plugin/pytest_xfail.py b/py/test/plugin/pytest_xfail.py
index 6024eb673..8bbd29c5b 100644
--- a/py/test/plugin/pytest_xfail.py
+++ b/py/test/plugin/pytest_xfail.py
@@ -40,7 +40,7 @@ def pytest_terminal_summary(terminalreporter):
             entry = event.longrepr.reprcrash 
             key = entry.path, entry.lineno, entry.message
             reason = event.longrepr.reprcrash.message
-            modpath = event.colitem.getmodpath(includemodule=True)
+            modpath = event.item.getmodpath(includemodule=True)
             #tr._tw.line("%s %s:%d: %s" %(modpath, entry.path, entry.lineno, entry.message))
             tr._tw.line("%s %s:%d: " %(modpath, entry.path, entry.lineno))
 
@@ -48,7 +48,7 @@ def pytest_terminal_summary(terminalreporter):
     if xpassed:
         tr.write_sep("_", "UNEXPECTEDLY PASSING TESTS")
         for event in xpassed:
-            tr._tw.line("%s: xpassed" %(event.colitem,))
+            tr._tw.line("%s: xpassed" %(event.item,))
 
 # ===============================================================================
 #
diff --git a/py/test/runner.py b/py/test/runner.py
index 7b399e055..dada94c90 100644
--- a/py/test/runner.py
+++ b/py/test/runner.py
@@ -19,13 +19,13 @@ def basic_run_report(item, pdb=None):
             when = "setup"
             item.config._setupstate.prepare(item)
             try:
-                when = "execute"
+                when = "runtest"
                 if not item._deprecated_testexecution():
                     item.runtest()
             finally:
                 when = "teardown"
                 item.config._setupstate.teardown_exact(item)
-                when = "execute"
+                when = "runtest"
         finally:
             outerr = capture.reset()
     except KeyboardInterrupt:
@@ -41,18 +41,13 @@ def basic_run_report(item, pdb=None):
     return testrep
 
 def basic_collect_report(collector):
-    excinfo = res = None
-    try:
-        capture = collector.config._getcapture()
-        try:
-            res = collector._memocollect()
-        finally:
-            outerr = capture.reset()
-    except KeyboardInterrupt:
-        raise
-    except: 
-        excinfo = py.code.ExceptionInfo()
-    return CollectReport(collector, res, excinfo, outerr)
+    call = collector.config.guardedcall(
+        lambda: collector._memocollect()
+    )
+    result = None
+    if not call.excinfo:
+        result = call.result
+    return CollectReport(collector, result, call.excinfo, call.outerr)
 
 def forked_run_report(item, pdb=None):
     EXITSTATUS_TESTEXIT = 4
@@ -99,11 +94,10 @@ class BaseReport(object):
 class ItemTestReport(BaseReport):
     failed = passed = skipped = False
 
-    # XXX rename colitem to item here
-    def __init__(self, colitem, excinfo=None, when=None, outerr=None):
-        self.colitem = colitem 
-        if colitem and when != "setup":
-            self.keywords = colitem.readkeywords() 
+    def __init__(self, item, excinfo=None, when=None, outerr=None):
+        self.item = item 
+        if item and when != "setup":
+            self.keywords = item.readkeywords() 
         else:
             # if we fail during setup it might mean 
             # we are not able to access the underlying object
@@ -123,50 +117,40 @@ class ItemTestReport(BaseReport):
             elif excinfo.errisinstance(Skipped):
                 self.skipped = True 
                 shortrepr = "s"
-                longrepr = self.colitem._repr_failure_py(excinfo, outerr)
+                longrepr = self.item._repr_failure_py(excinfo, outerr)
             else:
                 self.failed = True
-                shortrepr = self.colitem.shortfailurerepr
-                if self.when == "execute":
-                    longrepr = self.colitem.repr_failure(excinfo, outerr)
+                shortrepr = self.item.shortfailurerepr
+                if self.when == "runtest":
+                    longrepr = self.item.repr_failure(excinfo, outerr)
                 else: # exception in setup or teardown 
-                    longrepr = self.colitem._repr_failure_py(excinfo, outerr)
+                    longrepr = self.item._repr_failure_py(excinfo, outerr)
                     shortrepr = shortrepr.lower()
             self.shortrepr = shortrepr 
             self.longrepr = longrepr 
-            
+
+    def getnode(self):
+        return self.item 
 
 class CollectReport(BaseReport):
     skipped = failed = passed = False 
 
-    def __init__(self, colitem, result, excinfo=None, outerr=None):
-        # XXX rename to collector 
-        self.colitem = colitem 
+    def __init__(self, collector, result, excinfo=None, outerr=None):
+        self.collector = collector 
         if not excinfo:
             self.passed = True
             self.result = result 
         else:
             self.outerr = outerr
-            self.longrepr = self.colitem._repr_failure_py(excinfo, outerr)
+            self.longrepr = self.collector._repr_failure_py(excinfo, outerr)
             if excinfo.errisinstance(Skipped):
                 self.skipped = True
                 self.reason = str(excinfo.value)
             else:
                 self.failed = True
 
-class ItemSetupReport(BaseReport):
-    failed = passed = skipped = False
-    def __init__(self, item, excinfo=None, outerr=None):
-        self.item = item 
-        self.outerr = outerr 
-        if not excinfo:
-            self.passed = True
-        else:
-            if excinfo.errisinstance(Skipped):
-                self.skipped = True 
-            else:
-                self.failed = True
-            self.excrepr = item._repr_failure_py(excinfo, [])
+    def getnode(self):
+        return self.collector 
 
 class SetupState(object):
     """ shared state for setting up/tearing down test items or collectors. """
diff --git a/py/test/testing/test_collect.py b/py/test/testing/test_collect.py
index 8b32349d2..d7c1b0e65 100644
--- a/py/test/testing/test_collect.py
+++ b/py/test/testing/test_collect.py
@@ -176,7 +176,7 @@ class TestCollectPluginHooks:
         assert "world" in wascalled
         # make sure the directories do not get double-appended 
         colreports = reprec.getreports("pytest_collectreport")
-        names = [rep.colitem.name for rep in colreports]
+        names = [rep.collector.name for rep in colreports]
         assert names.count("hello") == 1
 
 class TestCustomConftests:
@@ -214,8 +214,8 @@ class TestCustomConftests:
         """)
         testdir.mkdir("hello")
         reprec = testdir.inline_run(testdir.tmpdir)
-        names = [rep.colitem.name for rep in reprec.getreports("pytest_collectreport")]
+        names = [rep.collector.name for rep in reprec.getreports("pytest_collectreport")]
         assert 'hello' not in names 
         reprec = testdir.inline_run(testdir.tmpdir, "--XX")
-        names = [rep.colitem.name for rep in reprec.getreports("pytest_collectreport")]
+        names = [rep.collector.name for rep in reprec.getreports("pytest_collectreport")]
         assert 'hello' in names 
diff --git a/py/test/testing/test_genitems.py b/py/test/testing/test_genitems.py
index 2ebfbe302..f917e5386 100644
--- a/py/test/testing/test_genitems.py
+++ b/py/test/testing/test_genitems.py
@@ -73,7 +73,7 @@ class TestKeywordSelection:
             reprec = testdir.inline_run("-s", "-k", keyword, file_test)
             passed, skipped, failed = reprec.listoutcomes()
             assert len(failed) == 1
-            assert failed[0].colitem.name == name 
+            assert failed[0].item.name == name 
             assert len(reprec.getcalls('pytest_deselected')) == 1
 
         for keyword in ['test_one', 'est_on']:
@@ -101,7 +101,7 @@ class TestKeywordSelection:
             print "keyword", repr(keyword)
             passed, skipped, failed = reprec.listoutcomes()
             assert len(passed) == 1
-            assert passed[0].colitem.name == "test_2"
+            assert passed[0].item.name == "test_2"
             dlist = reprec.getcalls("pytest_deselected")
             assert len(dlist) == 1
             assert dlist[0].items[0].name == 'test_1'
diff --git a/py/test/testing/test_runner.py b/py/test/testing/test_runner.py
index 9bc8638c1..41bdc1433 100644
--- a/py/test/testing/test_runner.py
+++ b/py/test/testing/test_runner.py
@@ -1,9 +1,10 @@
-
-from py.__.test.config import SetupState
+import py
+from py.__.test import runner 
+from py.__.code.excinfo import ReprExceptionInfo
 
 class TestSetupState:
     def test_setup(self, testdir):
-        ss = SetupState()
+        ss = runner.SetupState()
         item = testdir.getitem("def test_func(): pass")
         l = [1]
         ss.prepare(item)
@@ -14,7 +15,7 @@ class TestSetupState:
 
     def test_setup_scope_None(self, testdir):
         item = testdir.getitem("def test_func(): pass")
-        ss = SetupState()
+        ss = runner.SetupState()
         l = [1]
         ss.prepare(item)
         ss.addfinalizer(l.pop, colitem=None)
@@ -26,93 +27,263 @@ class TestSetupState:
         ss.teardown_all()
         assert not l 
 
-class TestSetupStateFunctional:
-    disabled = True
-    def test_setup_ok(self, testdir):
-        item = testdir.getitem("""
-            def setup_module(mod):
-                pass 
+
+class BaseFunctionalTests:
+    def test_funcattr(self, testdir):
+        rep = testdir.runitem("""
+            import py
+            @py.test.mark(xfail="needs refactoring")
+            def test_func():
+                raise Exit()
+        """)
+        assert rep.keywords['xfail'] == "needs refactoring" 
+
+    def test_passfunction(self, testdir):
+        rep = testdir.runitem("""
             def test_func():
                 pass
         """)
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        res = setup.do_setup(item)
-        assert res
+        assert rep.passed 
+        assert not rep.failed
+        assert rep.shortrepr == "."
+        assert not hasattr(rep, 'longrepr')
+                
+    def test_failfunction(self, testdir):
+        rep = testdir.runitem("""
+            def test_func():
+                assert 0
+        """)
+        assert not rep.passed 
+        assert not rep.skipped 
+        assert rep.failed 
+        assert rep.when == "runtest"
+        assert isinstance(rep.longrepr, ReprExceptionInfo)
+        assert str(rep.shortrepr) == "F"
 
-    def test_setup_fails(self, testdir):
-        item = testdir.getitem("""
-            def setup_module(mod):
-                print "world"
+    def test_skipfunction(self, testdir):
+        rep = testdir.runitem("""
+            import py
+            def test_func():
+                py.test.skip("hello")
+        """)
+        assert not rep.failed 
+        assert not rep.passed 
+        assert rep.skipped 
+        #assert rep.skipped.when == "runtest"
+        #assert rep.skipped.when == "runtest"
+        #assert rep.skipped == "%sreason == "hello"
+        #assert rep.skipped.location.lineno == 3
+        #assert rep.skipped.location.path
+        #assert not rep.skipped.failurerepr 
+
+    def test_skip_in_setup_function(self, testdir):
+        rep = testdir.runitem("""
+            import py
+            def setup_function(func):
+                py.test.skip("hello")
+            def test_func():
+                pass
+        """)
+        print rep
+        assert not rep.failed 
+        assert not rep.passed 
+        assert rep.skipped 
+        #assert rep.skipped.reason == "hello"
+        #assert rep.skipped.location.lineno == 3
+        #assert rep.skipped.location.lineno == 3
+
+    def test_failure_in_setup_function(self, testdir):
+        rep = testdir.runitem("""
+            import py
+            def setup_function(func):
                 raise ValueError(42)
             def test_func():
                 pass
         """)
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        res = setup.do_setup(item)
-        assert not res
-        rep = reprec.popcall(pytest_itemsetupreport).rep
-        assert rep.failed
-        assert not rep.skipped
-        assert rep.excrepr 
-        assert "42" in str(rep.excrepr)
-        assert rep.outerr[0].find("world") != -1
-
-    def test_teardown_fails(self, testdir):
-        item = testdir.getitem("""
-            def test_func():
-                pass
-            def teardown_function(func): 
-                print "13"
-                raise ValueError(25)
-        """)
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        res = setup.do_setup(item)
-        assert res 
-        rep = reprec.popcall(pytest_itemsetupreport).rep
-        assert rep.passed
-        setup.do_teardown(item)
-        rep = reprec.popcall(pytest_itemsetupreport).rep
-        assert rep.item == item 
+        print rep
+        assert not rep.skipped 
+        assert not rep.passed 
         assert rep.failed 
-        assert not rep.passed
-        assert "13" in rep.outerr[0]
-        assert "25" in str(rep.excrepr)
+        assert rep.when == "setup"
 
-    def test_setupitem_skips(self, testdir):
-        item = testdir.getitem("""
+    def test_failure_in_teardown_function(self, testdir):
+        rep = testdir.runitem("""
             import py
-            def setup_module(mod):
-                py.test.skip("17")
+            def teardown_function(func):
+                raise ValueError(42)
             def test_func():
                 pass
         """)
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        setup.do_setup(item)
-        rep = reprec.popcall(pytest_itemsetupreport).rep
-        assert not rep.failed
-        assert rep.skipped
-        assert rep.excrepr 
-        assert "17" in str(rep.excrepr)
+        print rep
+        assert not rep.skipped 
+        assert not rep.passed 
+        assert rep.failed 
+        assert rep.when == "teardown" 
+        assert rep.longrepr.reprcrash.lineno == 3
+        assert rep.longrepr.reprtraceback.reprentries 
 
-    def test_runtest_ok(self, testdir):
-        item = testdir.getitem("def test_func(): pass")
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        setup.do_fixture_and_runtest(item)
-        rep = reprec.popcall(pytest_itemtestreport).rep 
-        assert rep.passed 
+    def test_custom_failure_repr(self, testdir):
+        testdir.makepyfile(conftest="""
+            import py
+            class Function(py.test.collect.Function):
+                def repr_failure(self, excinfo, outerr):
+                    return "hello" 
+        """)
+        rep = testdir.runitem("""
+            import py
+            def test_func():
+                assert 0
+        """)
+        assert not rep.skipped 
+        assert not rep.passed 
+        assert rep.failed 
+        #assert rep.outcome.when == "runtest"
+        #assert rep.failed.where.lineno == 3
+        #assert rep.failed.where.path.basename == "test_func.py" 
+        #assert rep.failed.failurerepr == "hello"
 
-    def test_runtest_fails(self, testdir):
-        item = testdir.getitem("def test_func(): assert 0")
-        reprec = testdir.getreportrecorder(item.config)
-        setup = SetupState()
-        setup.do_fixture_and_runtest(item)
-        event = reprec.popcall(pytest_item_runtest_finished)
-        assert event.excinfo 
+    def test_failure_in_setup_function_ignores_custom_failure_repr(self, testdir):
+        testdir.makepyfile(conftest="""
+            import py
+            class Function(py.test.collect.Function):
+                def repr_failure(self, excinfo):
+                    assert 0
+        """)
+        rep = testdir.runitem("""
+            import py
+            def setup_function(func):
+                raise ValueError(42)
+            def test_func():
+                pass
+        """)
+        print rep
+        assert not rep.skipped 
+        assert not rep.passed 
+        assert rep.failed 
+        #assert rep.outcome.when == "setup"
+        #assert rep.outcome.where.lineno == 3
+        #assert rep.outcome.where.path.basename == "test_func.py" 
+        #assert instanace(rep.failed.failurerepr, PythonFailureRepr)
+
+    def test_capture_in_func(self, testdir):
+        rep = testdir.runitem("""
+            import py
+            def setup_function(func):
+                print >>py.std.sys.stderr, "in setup"
+            def test_func():
+                print "in function"
+                assert 0
+            def teardown_func(func):
+                print "in teardown"
+        """)
+        assert rep.failed 
+        # out, err = rep.failed.outerr
+        # assert out == ['in function\nin teardown\n']
+        # assert err == ['in setup\n']
         
-    
+    def test_systemexit_does_not_bail_out(self, testdir):
+        try:
+            rep = testdir.runitem("""
+                def test_func():
+                    raise SystemExit(42)
+            """)
+        except SystemExit:
+            py.test.fail("runner did not catch SystemExit")
+        assert rep.failed
+        assert rep.when == "runtest"
+
+    def test_exit_propagates(self, testdir):
+        from py.__.test.outcome import Exit
+        try:
+            testdir.runitem("""
+                from py.__.test.outcome import Exit
+                def test_func():
+                    raise Exit()
+            """)
+        except Exit:
+            pass
+        else: 
+            py.test.fail("did not raise")
+
+
+class TestExecutionNonForked(BaseFunctionalTests):
+    def getrunner(self):
+        return runner.basic_run_report 
+
+    def test_keyboardinterrupt_propagates(self, testdir):
+        from py.__.test.outcome import Exit
+        try:
+            testdir.runitem("""
+                def test_func():
+                    raise KeyboardInterrupt("fake")
+            """)
+        except KeyboardInterrupt, e:
+            pass
+        else: 
+            py.test.fail("did not raise")
+
+    def test_pdb_on_fail(self, testdir):
+        l = []
+        rep = testdir.runitem("""
+            def test_func():
+                assert 0
+        """, pdb=l.append)
+        assert rep.failed
+        assert rep.when == "runtest"
+        assert len(l) == 1
+
+    def test_pdb_on_skip(self, testdir):
+        l = []
+        rep = testdir.runitem("""
+            import py
+            def test_func():
+                py.test.skip("hello")
+        """, pdb=l.append)
+        assert len(l) == 0
+        assert rep.skipped 
+
+class TestExecutionForked(BaseFunctionalTests): 
+    def getrunner(self):
+        if not hasattr(py.std.os, 'fork'):
+            py.test.skip("no os.fork available")
+        return runner.forked_run_report
+
+    def test_suicide(self, testdir):
+        rep = testdir.runitem("""
+            def test_func():
+                import os
+                os.kill(os.getpid(), 15)
+        """)
+        assert rep.failed
+        assert rep.when == "???"
+
+class TestCollectionReports:
+    def test_collect_result(self, testdir):
+        col = testdir.getmodulecol("""
+            def test_func1():
+                pass
+            class TestClass:
+                pass
+        """)
+        rep = runner.basic_collect_report(col)
+        assert not rep.failed
+        assert not rep.skipped
+        assert rep.passed 
+        res = rep.result 
+        assert len(res) == 2
+        assert res[0].name == "test_func1" 
+        assert res[1].name == "TestClass" 
+
+    def test_skip_at_module_scope(self, testdir):
+        col = testdir.getmodulecol("""
+            import py
+            py.test.skip("hello")
+            def test_func():
+                pass
+        """)
+        rep = runner.basic_collect_report(col)
+        assert not rep.failed 
+        assert not rep.passed 
+        assert rep.skipped 
+
 
diff --git a/py/test/testing/test_runner_functional.py b/py/test/testing/test_runner_functional.py
deleted file mode 100644
index 478e4f651..000000000
--- a/py/test/testing/test_runner_functional.py
+++ /dev/null
@@ -1,263 +0,0 @@
-import py
-from py.__.test.runner import basic_run_report, forked_run_report, basic_collect_report
-from py.__.code.excinfo import ReprExceptionInfo
-
-class BaseTests:
-    def test_funcattr(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            @py.test.mark(xfail="needs refactoring")
-            def test_func():
-                raise Exit()
-        """)
-        assert ev.keywords['xfail'] == "needs refactoring" 
-
-    def test_passfunction(self, testdir):
-        ev = testdir.runitem("""
-            def test_func():
-                pass
-        """)
-        assert ev.passed 
-        assert not ev.failed
-        assert ev.shortrepr == "."
-        assert not hasattr(ev, 'longrepr')
-                
-    def test_failfunction(self, testdir):
-        ev = testdir.runitem("""
-            def test_func():
-                assert 0
-        """)
-        assert not ev.passed 
-        assert not ev.skipped 
-        assert ev.failed 
-        assert ev.when == "execute"
-        assert isinstance(ev.longrepr, ReprExceptionInfo)
-        assert str(ev.shortrepr) == "F"
-
-    def test_skipfunction(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            def test_func():
-                py.test.skip("hello")
-        """)
-        assert not ev.failed 
-        assert not ev.passed 
-        assert ev.skipped 
-        #assert ev.skipped.when == "execute"
-        #assert ev.skipped.when == "execute"
-        #assert ev.skipped == "%sreason == "hello"
-        #assert ev.skipped.location.lineno == 3
-        #assert ev.skipped.location.path
-        #assert not ev.skipped.failurerepr 
-
-    def test_skip_in_setup_function(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            def setup_function(func):
-                py.test.skip("hello")
-            def test_func():
-                pass
-        """)
-        print ev
-        assert not ev.failed 
-        assert not ev.passed 
-        assert ev.skipped 
-        #assert ev.skipped.reason == "hello"
-        #assert ev.skipped.location.lineno == 3
-        #assert ev.skipped.location.lineno == 3
-
-    def test_failure_in_setup_function(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            def setup_function(func):
-                raise ValueError(42)
-            def test_func():
-                pass
-        """)
-        print ev
-        assert not ev.skipped 
-        assert not ev.passed 
-        assert ev.failed 
-        assert ev.when == "setup"
-
-    def test_failure_in_teardown_function(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            def teardown_function(func):
-                raise ValueError(42)
-            def test_func():
-                pass
-        """)
-        print ev
-        assert not ev.skipped 
-        assert not ev.passed 
-        assert ev.failed 
-        assert ev.when == "teardown" 
-        assert ev.longrepr.reprcrash.lineno == 3
-        assert ev.longrepr.reprtraceback.reprentries 
-
-    def test_custom_failure_repr(self, testdir):
-        testdir.makepyfile(conftest="""
-            import py
-            class Function(py.test.collect.Function):
-                def repr_failure(self, excinfo, outerr):
-                    return "hello" 
-        """)
-        ev = testdir.runitem("""
-            import py
-            def test_func():
-                assert 0
-        """)
-        assert not ev.skipped 
-        assert not ev.passed 
-        assert ev.failed 
-        #assert ev.outcome.when == "execute"
-        #assert ev.failed.where.lineno == 3
-        #assert ev.failed.where.path.basename == "test_func.py" 
-        #assert ev.failed.failurerepr == "hello"
-
-    def test_failure_in_setup_function_ignores_custom_failure_repr(self, testdir):
-        testdir.makepyfile(conftest="""
-            import py
-            class Function(py.test.collect.Function):
-                def repr_failure(self, excinfo):
-                    assert 0
-        """)
-        ev = testdir.runitem("""
-            import py
-            def setup_function(func):
-                raise ValueError(42)
-            def test_func():
-                pass
-        """)
-        print ev
-        assert not ev.skipped 
-        assert not ev.passed 
-        assert ev.failed 
-        #assert ev.outcome.when == "setup"
-        #assert ev.outcome.where.lineno == 3
-        #assert ev.outcome.where.path.basename == "test_func.py" 
-        #assert instanace(ev.failed.failurerepr, PythonFailureRepr)
-
-    def test_capture_in_func(self, testdir):
-        ev = testdir.runitem("""
-            import py
-            def setup_function(func):
-                print >>py.std.sys.stderr, "in setup"
-            def test_func():
-                print "in function"
-                assert 0
-            def teardown_func(func):
-                print "in teardown"
-        """)
-        assert ev.failed 
-        # out, err = ev.failed.outerr
-        # assert out == ['in function\nin teardown\n']
-        # assert err == ['in setup\n']
-        
-    def test_systemexit_does_not_bail_out(self, testdir):
-        try:
-            ev = testdir.runitem("""
-                def test_func():
-                    raise SystemExit(42)
-            """)
-        except SystemExit:
-            py.test.fail("runner did not catch SystemExit")
-        assert ev.failed
-        assert ev.when == "execute"
-
-    def test_exit_propagates(self, testdir):
-        from py.__.test.outcome import Exit
-        try:
-            testdir.runitem("""
-                from py.__.test.outcome import Exit
-                def test_func():
-                    raise Exit()
-            """)
-        except Exit:
-            pass
-        else: 
-            py.test.fail("did not raise")
-
-
-class TestExecutionNonForked(BaseTests):
-    def getrunner(self):
-        return basic_run_report 
-
-    def test_keyboardinterrupt_propagates(self, testdir):
-        from py.__.test.outcome import Exit
-        try:
-            testdir.runitem("""
-                def test_func():
-                    raise KeyboardInterrupt("fake")
-            """)
-        except KeyboardInterrupt, e:
-            pass
-        else: 
-            py.test.fail("did not raise")
-
-    def test_pdb_on_fail(self, testdir):
-        l = []
-        ev = testdir.runitem("""
-            def test_func():
-                assert 0
-        """, pdb=l.append)
-        assert ev.failed
-        assert ev.when == "execute"
-        assert len(l) == 1
-
-    def test_pdb_on_skip(self, testdir):
-        l = []
-        ev = testdir.runitem("""
-            import py
-            def test_func():
-                py.test.skip("hello")
-        """, pdb=l.append)
-        assert len(l) == 0
-        assert ev.skipped 
-
-class TestExecutionForked(BaseTests): 
-    def getrunner(self):
-        if not hasattr(py.std.os, 'fork'):
-            py.test.skip("no os.fork available")
-        return forked_run_report
-
-    def test_suicide(self, testdir):
-        ev = testdir.runitem("""
-            def test_func():
-                import os
-                os.kill(os.getpid(), 15)
-        """)
-        assert ev.failed
-        assert ev.when == "???"
-
-class TestCollectionEvent:
-    def test_collect_result(self, testdir):
-        col = testdir.getmodulecol("""
-            def test_func1():
-                pass
-            class TestClass:
-                pass
-        """)
-        ev = basic_collect_report(col)
-        assert not ev.failed
-        assert not ev.skipped
-        assert ev.passed 
-        res = ev.result 
-        assert len(res) == 2
-        assert res[0].name == "test_func1" 
-        assert res[1].name == "TestClass" 
-
-    def test_skip_at_module_scope(self, testdir):
-        col = testdir.getmodulecol("""
-            import py
-            py.test.skip("hello")
-            def test_func():
-                pass
-        """)
-        ev = basic_collect_report(col)
-        assert not ev.failed 
-        assert not ev.passed 
-        assert ev.skipped 
-
-
diff --git a/py/test/testing/test_session.py b/py/test/testing/test_session.py
index 30446b3d8..042c62976 100644
--- a/py/test/testing/test_session.py
+++ b/py/test/testing/test_session.py
@@ -22,9 +22,9 @@ class SessionTests:
         assert len(skipped) == 0
         assert len(passed) == 1
         assert len(failed) == 3  
-        assert failed[0].colitem.name == "test_one_one"
-        assert failed[1].colitem.name == "test_other"
-        assert failed[2].colitem.name == "test_two"
+        assert failed[0].item.name == "test_one_one"
+        assert failed[1].item.name == "test_other"
+        assert failed[2].item.name == "test_two"
         itemstarted = reprec.getcalls("pytest_itemstart")
         assert len(itemstarted) == 4
         colstarted = reprec.getcalls("pytest_collectstart")