diff --git a/nonlib/Thread.C b/nonlib/Thread.C
index 2117f87..f0c6292 100644
--- a/nonlib/Thread.C
+++ b/nonlib/Thread.C
@@ -127,6 +127,6 @@ Thread::join ( void )
 void
 Thread::exit ( void *retval )
 {
-    pthread_exit( retval );
     _thread = 0;
+    pthread_exit( retval );
 }
diff --git a/timeline/src/Audio_Region.C b/timeline/src/Audio_Region.C
index d70ed37..4ce3633 100644
--- a/timeline/src/Audio_Region.C
+++ b/timeline/src/Audio_Region.C
@@ -651,7 +651,7 @@ Audio_Region::draw ( void )
         
         if ( peaks < loop_peaks_needed )
         {
-            DMESSAGE( "Peak read came up %lu peaks short", (unsigned long)loop_peaks_needed - peaks );
+//            DMESSAGE( "Peak read came up %lu peaks short", (unsigned long)loop_peaks_needed - peaks );
         }
 
         fo += loop_frames_needed;
diff --git a/timeline/src/Control_Sequence.C b/timeline/src/Control_Sequence.C
index 9368dfa..61143ef 100644
--- a/timeline/src/Control_Sequence.C
+++ b/timeline/src/Control_Sequence.C
@@ -783,11 +783,11 @@ Control_Sequence::handle ( int m )
                 test_press( FL_BUTTON1 ) )
             {
                 /* insert new control point */
-                timeline->wrlock();
+                timeline->sequence_lock.wrlock();
 
                 new Control_Point( this, timeline->xoffset + timeline->x_to_ts( Fl::event_x() - drawable_x() ), (float)(Fl::event_y() - y()) / h() );
                 
-                timeline->unlock();
+                timeline->sequence_lock.unlock();
 
                 return 1;
             }
diff --git a/timeline/src/Engine/Audio_Region.C b/timeline/src/Engine/Audio_Region.C
index f7ac319..7107b75 100644
--- a/timeline/src/Engine/Audio_Region.C
+++ b/timeline/src/Engine/Audio_Region.C
@@ -340,8 +340,12 @@ Audio_Region::write ( nframes_t nframes )
         }
     }
 
+    timeline->sequence_lock.wrlock();
+
     _range.length += nframes;
 
+    timeline->sequence_lock.unlock();
+
     return nframes;
 }
 
@@ -354,8 +358,12 @@ Audio_Region::finalize ( nframes_t frame )
 
     DMESSAGE( "finalizing capture region" );
 
+    timeline->sequence_lock.wrlock();
+
     _range.length = frame - _range.start;
 
+    timeline->sequence_lock.unlock();
+
     _clip->close();
     _clip->open();
 
diff --git a/timeline/src/Engine/Engine.C b/timeline/src/Engine/Engine.C
index c8b0a2c..3f7c84a 100644
--- a/timeline/src/Engine/Engine.C
+++ b/timeline/src/Engine/Engine.C
@@ -172,40 +172,55 @@ Engine::process ( nframes_t nframes )
 
     transport->poll();
 
-    if ( freewheeling() )
+    if ( !timeline)
+        /* handle chicken/egg problem */
+        return 0;
+    
+    nframes_t n = 0;
+   
+    n += timeline->process_input(nframes);
+    n += timeline->process_output(nframes);
+
+    if ( n != nframes * 2 )
     {
-        if ( timeline )
-        {
-            timeline->rdlock();
-
-            timeline->process( nframes );
-
-            timeline->unlock();
-        }
+        _buffers_dropped++;
+        WARNING("xrun");
     }
-    else
-    {
-        if ( !timeline)
-            /* handle chicken/egg problem */
-            return 0;
+    
+    /* if ( freewheeling() ) */
+    /* { */
+    /*     if ( timeline ) */
+    /*     { */
+    /*         timeline->rdlock(); */
 
-        if ( timeline->tryrdlock() )
-        {
-            /* the data structures we need to access here (tracks and
-             * their ports, but not track contents) may be in an
-             * inconsistent state at the moment. Just punt and drop this
-             * buffer. */
-            ++_buffers_dropped;
-            return 0;
-        }
+    /*         timeline->process( nframes ); */
 
-        /* this will initiate the process() call graph for the various
-         * number and types of tracks, which will in turn send data out
-         * the appropriate ports.  */
-        timeline->process( nframes );
+    /*         timeline->unlock(); */
+    /*     } */
+    /* } */
+    /* else */
+    /* { */
+    /*     if ( !timeline) */
+    /*         /\* handle chicken/egg problem *\/ */
+    /*         return 0; */
+ 
+    /*     if ( timeline->tryrdlock() ) */
+    /*     { */
+    /*         /\* the data structures we need to access here (tracks and */
+    /*          * their ports, but not track contents) may be in an */
+    /*          * inconsistent state at the moment. Just punt and drop this */
+    /*          * buffer. *\/ */
+    /*         ++_buffers_dropped; */
+    /*         return 0; */
+    /*     } */
 
-        timeline->unlock();
-    }
+    /*     /\* this will initiate the process() call graph for the various */
+    /*      * number and types of tracks, which will in turn send data out */
+    /*      * the appropriate ports.  *\/ */
+    /*     timeline->process( nframes ); */
+
+    /*     timeline->unlock(); */
+    /* } */
 
     return 0;
 }
diff --git a/timeline/src/Engine/Playback_DS.C b/timeline/src/Engine/Playback_DS.C
index ce05c52..93b1b0f 100644
--- a/timeline/src/Engine/Playback_DS.C
+++ b/timeline/src/Engine/Playback_DS.C
@@ -84,7 +84,7 @@ Playback_DS::read_block ( sample_t *buf, nframes_t nframes )
     if ( !timeline )
         return;
 
-    while ( timeline->tryrdlock() )
+    while ( timeline->sequence_lock.tryrdlock() )
     {
         if ( _terminate )
             return;
@@ -100,7 +100,7 @@ Playback_DS::read_block ( sample_t *buf, nframes_t nframes )
         _frame += nframes;
     }
     
-    timeline->unlock();
+    timeline->sequence_lock.unlock();
 }
 
 void
diff --git a/timeline/src/Engine/Record_DS.C b/timeline/src/Engine/Record_DS.C
index d429f5f..4c55e67 100644
--- a/timeline/src/Engine/Record_DS.C
+++ b/timeline/src/Engine/Record_DS.C
@@ -60,10 +60,14 @@ Record_DS::write_block ( sample_t *buf, nframes_t nframes )
     if ( ! ( timeline && sequence() ) )
         return;
 
+    if ( ! _capture )
+    {
+        _capture = new Track::Capture;
 
-    if ( ! _capture->audio_file )
+    /* if ( ! _capture->audio_file ) */
         /* create the file */
         track()->record( _capture, _frame );
+    }
 
     track()->write( _capture, buf, nframes );
 
@@ -79,83 +83,162 @@ Record_DS::disk_thread ( void )
 
     const nframes_t nframes = _nframes;
 
+    _disk_io_blocks = 1;
+
     /* buffer to hold the interleaved data returned by the track reader */
     sample_t *buf = buffer_alloc( nframes * channels() * _disk_io_blocks );
     sample_t *cbuf = buffer_alloc( nframes );
 
-    const size_t block_size = nframes * sizeof( sample_t );
+//    const size_t block_size = nframes * sizeof( sample_t );
 
-    nframes_t blocks_read = 0;
+    nframes_t frames_read = 0;
 
+    bool punching_in = false;
+    bool punching_out = false;
+    bool punched_in = false;
+
+    nframes_t bS = 0;
+    nframes_t bE = 0;
+
+again:
+
+    _capture = NULL;
+
+    punched_in = false;
+    punching_out = false;
+
+    nframes_t pS = _frame;
+    nframes_t pE = _stop_frame;
+
+    if ( punching_in )
+    {
+        /* write remainder of buffer */
+        write_block( buf + ((pS - bS) * channels()),
+                         bE - pS );        
+
+        punching_in = false;
+        punched_in = true;
+    }
+    
     while ( wait_for_block() )
     {
         /* pull data from the per-channel ringbuffers and interlace it */
-        for ( int i = channels(); i--; )
+        size_t frames_to_read = nframes;
+        
+        /* we read the entire block if a partial... */
+        for ( int i = 0; i < channels(); i++ )
         {
-            while ( jack_ringbuffer_read_space( _rb[ i ] ) < block_size )
+            while ( jack_ringbuffer_read_space( _rb[ i ] ) < frames_to_read * sizeof( sample_t ) )
                 usleep( 10 * 1000 );
-
-            jack_ringbuffer_read( _rb[ i ], ((char*)cbuf), block_size );
             
-            buffer_interleave_one_channel( buf + ( blocks_read * nframes * channels() ),
+            jack_ringbuffer_read( _rb[ i ], ((char*)cbuf), frames_to_read * sizeof( sample_t ) );
+            
+            buffer_interleave_one_channel( buf,
                                            cbuf, 
                                            i,
                                            channels(),
-                                           nframes );
+                                           frames_to_read);
         }
+       
+        bS = _first_frame + frames_read;
 
-        blocks_read++;
+        frames_read += frames_to_read;
 
-        if ( blocks_read == _disk_io_blocks )
+        bE = _first_frame + frames_read;
+
+        punching_in = ! punched_in && bE > pS;
+        punching_out = punched_in && pE < bE;
+        
+        if ( punching_out )
         {
-            write_block( buf, nframes * _disk_io_blocks );
-            blocks_read = 0;
+            write_block( buf,
+                         pE - bS );
+
+            break;
+        }
+        else
+        if ( punching_in )
+        {
+            write_block( buf + ((pS - bS) * channels()),
+                         bE - pS );
+
+            punching_in = false;
+            punched_in = true;
+        }
+        else if ( punched_in )
+        {
+            write_block( buf, bE - bS );
         }
     }
 
-    DMESSAGE( "capture thread terminating" );
+//    DMESSAGE( "capture thread terminating" );
 
     /* flush what remains in the buffer out to disk */
 
+    /* { */
+    /*     while ( blocks_read-- > 0 || ( ! sem_trywait( &_blocks ) && errno != EAGAIN ) ) */
+    /*     { */
+    /*         for ( int i = channels(); i--; ) */
+    /*         { */
+    /*             jack_ringbuffer_read( _rb[ i ], (char*)cbuf, block_size ); */
+
+    /*             buffer_interleave_one_channel( buf, cbuf, i, channels(), nframes ); */
+    /*         } */
+
+    /*         const nframes_t frames_remaining = (_stop_frame - _frame ) - _frames_written; */
+
+    /*         if ( frames_remaining < nframes ) */
+    /*         { */
+    /*             /\* this is the last block, might be partial  *\/ */
+    /*             write_block( buf, frames_remaining ); */
+    /*             break; */
+    /*         } */
+    /*         else */
+    /*             write_block( buf, nframes ); */
+    /*     } */
+    /* } */
+
+  
+   
+    if ( _capture )
     {
-        while ( blocks_read-- > 0 || ( ! sem_trywait( &_blocks ) && errno != EAGAIN ) )
+        DMESSAGE( "finalzing capture" );
+        Track::Capture *c = _capture;
+        
+        _capture = NULL;
+        
+        /* now finalize the recording */
+        
+//        if ( c->audio_file )
+        track()->finalize( c, _stop_frame );
+
+        delete c;
+    }
+
+    if ( ! _terminate ) 
+    {
+        nframes_t in, out;
+        
+        if ( timeline->next_punch( _stop_frame, &in, &out ) )
         {
-            for ( int i = channels(); i--; )
-            {
-                jack_ringbuffer_read( _rb[ i ], (char*)cbuf, block_size );
+            _frame = in;
+            _stop_frame = out;
+            _frames_written = 0;
 
-                buffer_interleave_one_channel( buf, cbuf, i, channels(), nframes );
-            }
+            punched_in = false;
+            punching_out = false;
+            
+            punching_in = bE > in;
 
-            const nframes_t frames_remaining = (_stop_frame - _frame ) - _frames_written;
+            DMESSAGE( "Next punch: %lu:%lu", (unsigned long)in,(unsigned long)out );
 
-            if ( frames_remaining < nframes )
-            {
-                /* this is the last block, might be partial  */
-                write_block( buf, frames_remaining );
-                break;
-            }
-            else
-                write_block( buf, nframes );
+            goto again;
         }
     }
 
     free(buf);
     free(cbuf);
 
-    DMESSAGE( "finalzing capture" );
-
-    Track::Capture *c = _capture;
-
-    _capture = NULL;
-
-    /* now finalize the recording */
-
-    if ( c->audio_file )
-        track()->finalize( c, _stop_frame );
-
-    delete c;
-
     flush();
 
     _terminate = false;
@@ -168,7 +251,7 @@ Record_DS::disk_thread ( void )
 
 /** begin recording */
 void
-Record_DS::start ( nframes_t frame )
+Record_DS::start ( nframes_t frame, nframes_t start_frame, nframes_t stop_frame )
 {
     THREAD_ASSERT( UI );
 
@@ -183,9 +266,10 @@ Record_DS::start ( nframes_t frame )
 
     DMESSAGE( "recording started at frame %lu", (unsigned long)frame);
 
-    _frame = frame;
+    _frame = start_frame;
+    _stop_frame = stop_frame;
 
-    _capture = new Track::Capture;
+    _first_frame = frame;
 
     run();
 
@@ -227,40 +311,36 @@ Record_DS::process ( nframes_t nframes )
     if ( ! _recording )
         return 0;
 
-     if ( transport->frame < _frame  )
-         return 0;
+     /* if ( transport->frame < _frame  ) */
+     /*     return 0; */
 
 /*    DMESSAGE( "recording actually happening at %lu (start frame %lu)", (unsigned long)transport->frame, (unsigned long)_frame); */
 
     nframes_t offset = 0;
 
-    if ( _frame > transport->frame &&
-         _frame < transport->frame + nframes )
-    {
-        /* The record start frame falls somewhere within the current
-           buffer.  We must discard the unneeded portion and only
-           stuff the part requested into the ringbuffer. */
+/*     if ( _frame > transport->frame && */
+/*          _frame < transport->frame + nframes ) */
+/*     { */
+/*         /\* The record start frame falls somewhere within the current */
+/*            buffer.  We must discard the unneeded portion and only */
+/*            stuff the part requested into the ringbuffer. *\/ */
 
-        offset = _frame - transport->frame;
+/*         offset = _frame - transport->frame; */
 
-/*         DMESSAGE( "offset = %lu", (unsigned long)offset ); */
-    }
+/* /\*         DMESSAGE( "offset = %lu", (unsigned long)offset ); *\/ */
+/*     } */
 
     const size_t offset_size = offset * sizeof( sample_t );
     const size_t block_size = ( nframes * sizeof( sample_t ) ) - offset_size;
 
-    for ( int i = channels(); i--;  )
+    for ( int i = 0; i < channels(); i++ )
     {
         /* read the entire input buffer */
         void *buf = track()->input[ i ].buffer( nframes );
 
-        /* FIXME: this results in a ringbuffer size that is no longer
-         necessarily a multiple of nframes...  how will the other side
-         handle that? */
-
         if ( engine->freewheeling() )
         {
-            while ( jack_ringbuffer_write_space( _rb[i] ) < block_size )
+            while ( transport->recording && jack_ringbuffer_write_space( _rb[i] ) < block_size )
                 usleep( 10 * 1000 );
 
             jack_ringbuffer_write( _rb[ i ], ((char*)buf) + offset_size, block_size );
@@ -271,12 +351,14 @@ Record_DS::process ( nframes_t nframes )
             {
                 memset( buf, 0, block_size );
                 /* FIXME: we need to resync somehow */
+                WARNING( "xrun" );
                 ++_xruns;
             }
-            else
-            {
-                jack_ringbuffer_write( _rb[ i ], ((char*)buf) + offset_size, block_size );
-            }
+
+            jack_ringbuffer_write( _rb[ i ], ((char*)buf) + offset_size, block_size );
+
+//            DMESSAGE( "wrote %lu", (unsigned long) nframes );
+
         }
     }
 
diff --git a/timeline/src/Engine/Record_DS.H b/timeline/src/Engine/Record_DS.H
index 776a7e0..fb6741c 100644
--- a/timeline/src/Engine/Record_DS.H
+++ b/timeline/src/Engine/Record_DS.H
@@ -36,7 +36,8 @@ class Record_DS : public Disk_Stream
 
     nframes_t _frames_written;
     volatile nframes_t _stop_frame;
-
+    volatile nframes_t _first_frame;
+    
     volatile bool _recording;
 
     Audio_File_SF *_af;                             /* capture file */
@@ -56,8 +57,9 @@ public:
 
             _capture = NULL;
             _recording = false;
-            _stop_frame = -1;
+            _stop_frame = 0;
             _frames_written = 0;
+            _first_frame = 0;
         }
 
     virtual ~Record_DS ( ) { shutdown(); }
@@ -67,7 +69,7 @@ public:
     const Audio_Region * capture_region ( void ) const;
     Track::Capture * capture ( void );
 
-    void start ( nframes_t frame );
+    void start ( nframes_t frame, nframes_t start_frame, nframes_t stop_frame = 0 );
     void stop ( nframes_t frame );
     nframes_t process ( nframes_t nframes );
 
diff --git a/timeline/src/Engine/Timeline.C b/timeline/src/Engine/Timeline.C
index e6f6537..9bcfa0f 100644
--- a/timeline/src/Engine/Timeline.C
+++ b/timeline/src/Engine/Timeline.C
@@ -28,6 +28,8 @@
 #include "Thread.H"
 #include "../Cursor_Sequence.H"
 
+#include "Engine.H"
+
 #include <unistd.h>
 
 /** Initiate recording for all armed tracks */
@@ -48,14 +50,15 @@ Timeline::record ( void )
         _created_new_takes = true;
     }
 
-    transport->recording = true;
-
     deactivate();
 
     Loggable::block_start();
 
     nframes_t frame = transport->frame;
 
+    nframes_t _punch_out_frame = 0;
+//    nframes_t _punch_in_frame = 0;
+
     if ( transport->punch_enabled() )
     {
         DMESSAGE( "Finding next punch region following frame %lu...", (unsigned long)frame);
@@ -65,7 +68,7 @@ Timeline::record ( void )
         
         if (p || n )
         {
-            if ( p && frame > p->start() && frame < p->start() + p->length() )
+            if ( p && frame >= p->start() && frame < p->start() + p->length() )
             {
                 /* recording started in the middle of a punch
                  * cursor... Just start recording and punch out at the
@@ -86,33 +89,23 @@ Timeline::record ( void )
         }
     }
 
-    _punch_in_frame = frame;
+//    _punch_in_frame = frame;
 
-    punch_in( frame );
-
-    return true;
-}
-
-void
-Timeline::punch_in ( nframes_t frame )
-{
-    if ( _punched_in )
-    {
-        WARNING( "Programming error. Attempt to punch in twice" );
-        return;
-    }
-
-    DMESSAGE( "Going to record starting at frame %lu", (unsigned long)frame );
 
     for ( int i = tracks->children(); i-- ; )
     {
         Track *t = (Track*)tracks->child( i );
 
         if ( t->armed() && t->record_ds )
-            t->record_ds->start( frame );
+        {
+            t->record_ds->start( transport->frame, frame, _punch_out_frame );
+        }
     }
 
-    _punched_in = true;
+    transport->recording = true;
+
+
+    return true;
 }
 
 void
@@ -142,10 +135,6 @@ Timeline::punch_out ( nframes_t frame )
     }
     
     DMESSAGE( "All record threads stopped." );
-
-    _punched_in = false;
-    _punch_in_frame = 0;
-    _punch_out_frame = 0;
 }
 
 /** stop recording for all armed tracks. Does not affect transport. */
@@ -156,13 +145,13 @@ Timeline::stop ( void )
 
     nframes_t frame = transport->frame;
 
-    if ( transport->punch_enabled() )
-    {
-        const Sequence_Widget *w = punch_cursor_track->prev( frame );
+    /* if ( transport->punch_enabled() ) */
+    /* { */
+    /*     const Sequence_Widget *w = punch_cursor_track->prev( frame ); */
 
-        if ( w && w->start() + w->length() < frame )
-            frame = w->start() + w->length();
-    }
+    /*     if ( w && w->start() + w->length() < frame ) */
+    /*         frame = w->start() + w->length(); */
+    /* } */
     
     punch_out( frame );
    
@@ -178,20 +167,26 @@ Timeline::stop ( void )
 /* Engine */
 /**********/
 
-/** call process() on each track header */
 nframes_t
-Timeline::process ( nframes_t nframes )
+Timeline::process_input ( nframes_t nframes )
 {
-    /* there is no need to acquire a readlock here because track *
-       addition/removal locks process() and track process() calls deal with
-       ringbuffers instead of reading the sequence data directly.  */
 
-    for ( int i = tracks->children(); i-- ; )
-    {
-        Track *t = (Track*)tracks->child( i );
+    /* if ( engine->freewheeling() ) */
+    /* { */
+    /*     rdlock(); */
+    /* } */
+    /* else */
+    /* { */
+    /*     if ( tryrdlock() ) */
+    /*         return 0; */
+    /* } */
 
-        t->process_output( nframes );
-    }
+    if ( ! transport->recording )
+        return nframes;
+
+    /* we don't lock here in order to avoid deadlocks. we already know
+     * that only the record diskthread will be altering timeline
+     * structures during recording since the GUI is disabled. */
 
     for ( int i = tracks->children(); i-- ; )
     {
@@ -200,10 +195,46 @@ Timeline::process ( nframes_t nframes )
         t->process_input( nframes );
     }
 
-    /* FIXME: BOGUS */
+    /* unlock(); */
+
     return nframes;
 }
 
+nframes_t
+Timeline::process_output ( nframes_t nframes )
+{
+
+    /* if ( engine->freewheeling() ) */
+    /* { */
+    /*    rdlock(); */
+    /* } */
+    /* else */
+    /* { */
+    /*     if ( tryrdlock() ) */
+    /*         return 0; */
+    /* } */
+
+    bool r = transport->recording || engine->freewheeling();
+
+    if ( ! r )
+        if ( track_lock.tryrdlock() )
+            return 0;
+//        rdlock();
+
+    for ( int i = tracks->children(); i-- ; )
+    {
+        Track *t = (Track*)tracks->child( i );
+
+        t->process_output( nframes );
+    }
+
+    if ( ! r )
+        track_lock.unlock();
+
+    return nframes;
+
+}
+
 void
 Timeline::seek ( nframes_t frame )
 {
diff --git a/timeline/src/Engine/Track.C b/timeline/src/Engine/Track.C
index f6352ef..4bfdb5f 100644
--- a/timeline/src/Engine/Track.C
+++ b/timeline/src/Engine/Track.C
@@ -261,13 +261,15 @@ Track::resize_buffers ( nframes_t nframes )
 
 #include <time.h>
 
+static unsigned long uuid_counter = 0;
+
 /** very cheap UUID generator... */
 unsigned long long
 uuid ( void )
 {
     time_t t = time( NULL );
 
-    return (unsigned long long) t;
+    return (unsigned long long) t + uuid_counter++;
 }
 
 /** create capture region and prepare to record */
@@ -294,11 +296,12 @@ Track::record ( Capture *c, nframes_t frame )
 
     /* must acquire a write lock because the Audio_Region constructor
      * will add the region to the specified sequence, which might affect playback */
-    timeline->wrlock();
+
+    timeline->sequence_lock.wrlock();
 
     c->region = new Audio_Region( c->audio_file, sequence(), frame );
 
-    timeline->unlock();
+    timeline->sequence_lock.unlock();
 
 //    Fl::unlock();
 
@@ -313,6 +316,10 @@ Track::record ( Capture *c, nframes_t frame )
         /* in freewheeling mode, assume we're bouncing and only
          * compensate for capture latency */
         _capture_offset = max;
+
+        /* FIXME: hack to compensate for jack 1 period delay when looped back */
+//   _capture_offset += engine->playback_latency() / 2;
+        _capture_offset += engine->nframes();
     }
     else
     {
@@ -353,17 +360,18 @@ Track::finalize ( Capture *c, nframes_t frame )
     DMESSAGE( "finalizing audio file" );
     c->audio_file->finalize();
 
-    timeline->wrlock();
-
     DMESSAGE( "Adjusting capture by %lu frames.", (unsigned long)_capture_offset );
 
+    timeline->sequence_lock.wrlock();
+
     c->region->offset( _capture_offset );
+
+    timeline->sequence_lock.unlock();
+
     _capture_offset = 0;
 
     /* have to do this last, as it logs the create */
     c->region->finalize( frame );
-
-    timeline->unlock();
 }
 
 void
diff --git a/timeline/src/Sequence.C b/timeline/src/Sequence.C
index df1850d..7d8ed2c 100644
--- a/timeline/src/Sequence.C
+++ b/timeline/src/Sequence.C
@@ -327,7 +327,7 @@ Sequence::handle ( int m )
         case FL_SHORTCUT:
             if ( Fl::test_shortcut( FL_CTRL + FL_Right ) )
             {
-                const Sequence_Widget *w = next( transport->frame );
+                const Sequence_Widget *w = next( transport->frame + 1 );
                 
                 if ( w )
                     transport->locate( w->start() );
@@ -399,9 +399,9 @@ Sequence::handle ( int m )
                     {
                         /* accept objects dragged from other sequences of this type */
                         
-                        timeline->wrlock();
+                        timeline->sequence_lock.wrlock();
                         add( Sequence_Widget::pushed() );
-                        timeline->unlock();
+                        timeline->sequence_lock.unlock();
                         
                         damage( FL_DAMAGE_USER1 );
                         
@@ -503,9 +503,9 @@ Sequence::handle ( int m )
                             Sequence_Widget::belowmouse( NULL );
                         }
 
-                        timeline->wrlock();
+                        timeline->sequence_lock.wrlock();
                         delete t;
-                        timeline->unlock();
+                        timeline->sequence_lock.unlock();
                 }
 
                 Loggable::block_end();
@@ -557,8 +557,8 @@ const Sequence_Widget *
         Sequence::next ( nframes_t from ) const
     {
         for ( list <Sequence_Widget*>::const_iterator i = _widgets.begin(); i != _widgets.end(); i++ )
-//            if ( (*i)->start() >= from )
-            if ( (*i)->start() > from )
+            if ( (*i)->start() >= from )
+//            if ( (*i)->start() > from )
                 return *i;
 
         if ( _widgets.size() )
diff --git a/timeline/src/Sequence_Widget.C b/timeline/src/Sequence_Widget.C
index 70ba673..350c49c 100644
--- a/timeline/src/Sequence_Widget.C
+++ b/timeline/src/Sequence_Widget.C
@@ -172,7 +172,7 @@ Sequence_Widget::begin_drag ( const Drag &d )
 {
     _drag = new Drag( d );
 
-    timeline->wrlock();
+    timeline->sequence_lock.wrlock();
 
     /* copy current values */
     
@@ -181,7 +181,7 @@ Sequence_Widget::begin_drag ( const Drag &d )
     /* tell display to use temporary */
     _r = &_dragging_range;
 
-    timeline->unlock();
+    timeline->sequence_lock.unlock();
 }
 
 void
@@ -190,7 +190,7 @@ Sequence_Widget::end_drag ( void )
     /* swap in the new value */
     /* go back to playback and display using same values */
     
-    timeline->wrlock();
+    timeline->sequence_lock.wrlock();
 
     _range = _dragging_range;
     _r = &_range;
@@ -201,7 +201,7 @@ Sequence_Widget::end_drag ( void )
     /* this will result in a sort */
     sequence()->handle_widget_change( _r->start, _r->length );
 
-    timeline->unlock();
+    timeline->sequence_lock.unlock();
 }
 
 /** set position of widget on the timeline. */
@@ -449,9 +449,9 @@ Sequence_Widget::handle ( int m )
             if ( test_press( FL_BUTTON1 + FL_CTRL ) && ! _drag->state )
             {
                 /* duplication */
-                timeline->wrlock();
+                timeline->sequence_lock.wrlock();
                 sequence()->add( this->clone() );
-                timeline->unlock();
+                timeline->sequence_lock.unlock();
 
                 _drag->state = 1;
                 return 1;
diff --git a/timeline/src/Tempo_Sequence.C b/timeline/src/Tempo_Sequence.C
index c07f91b..50c16b3 100644
--- a/timeline/src/Tempo_Sequence.C
+++ b/timeline/src/Tempo_Sequence.C
@@ -62,11 +62,11 @@ Tempo_Sequence::handle ( int m )
 
                 if ( Tempo_Point::edit( &t ) )
                 {
-                    timeline->wrlock();
+                    timeline->sequence_lock.wrlock();
                     
                     new Tempo_Point( timeline->x_to_offset( Fl::event_x() ), t );
 
-                    timeline->unlock();
+                    timeline->sequence_lock.unlock();
 
                     timeline->redraw();
                 }
diff --git a/timeline/src/Time_Sequence.C b/timeline/src/Time_Sequence.C
index 972bfd9..f877b8a 100644
--- a/timeline/src/Time_Sequence.C
+++ b/timeline/src/Time_Sequence.C
@@ -65,11 +65,11 @@ Time_Sequence::handle ( int m )
 
                 if ( Time_Point::edit( &t ) )
                 {
-                    timeline->wrlock();
+                    timeline->sequence_lock.wrlock();
 
                     new Time_Point( timeline->x_to_offset( Fl::event_x() ), t.beats_per_bar, t.beat_type );
 
-                    timeline->unlock();
+                    timeline->sequence_lock.unlock();
 
                     timeline->redraw();
                 }
diff --git a/timeline/src/Timeline.C b/timeline/src/Timeline.C
index c280ecd..42100ce 100644
--- a/timeline/src/Timeline.C
+++ b/timeline/src/Timeline.C
@@ -360,7 +360,7 @@ Timeline::add_take_for_armed_tracks ( void )
 {
     THREAD_ASSERT( UI );
 
-    wrlock();
+    track_lock.wrlock();
 
     for ( int i = tracks->children(); i-- ; )
     {
@@ -370,7 +370,7 @@ Timeline::add_take_for_armed_tracks ( void )
             t->sequence( new Audio_Sequence( t ) );
     }
 
-    unlock();
+    track_lock.unlock();
 }
 
 void
@@ -598,9 +598,6 @@ Timeline::Timeline ( int X, int Y, int W, int H, const char* L ) : BASE( X, Y, W
     play_cursor_track = NULL;
 
     _created_new_takes = 0;
-    _punched_in = 0;
-    _punch_in_frame = 0;
-    _punch_out_frame = 0;
     osc_thread = 0;
     _sample_rate = 44100;
 
@@ -1271,6 +1268,8 @@ Timeline::draw ( void )
      * another thread must use Fl::lock()/unlock()! */ 
     THREAD_ASSERT( UI );
 
+//    rdlock();
+
     int X, Y, W, H;
 
     int bdx = 0;
@@ -1368,6 +1367,8 @@ Timeline::draw ( void )
 
 done:
 
+//    unlock();
+
     /* panzoomer->redraw(); */
 //    update_child( *panzoomer );
 
@@ -1416,6 +1417,26 @@ Timeline::draw_cursor ( nframes_t frame, Fl_Color color, void (*symbol)(Fl_Color
     fl_pop_clip();
 }
 
+/** set /in/ and /out/ to start and end of next punch region from /frame/  */
+bool
+Timeline::next_punch ( nframes_t frame, nframes_t *in, nframes_t *out ) const
+{
+    if ( !transport->punch_enabled() )
+        return false;
+
+    const Sequence_Widget *w = punch_cursor_track->next( frame );
+    
+    if ( w && w->start() >= frame )
+    {
+        *in = w->start();
+        *out = w->start() + w->length();
+
+        return true;
+    }
+
+    return false;
+}
+
 void
 Timeline::draw_playhead ( void )
 {
@@ -1429,36 +1450,6 @@ Timeline::redraw_playhead ( void )
 //    static nframes_t last_playhead = -1;
     static int last_playhead_x = -1;
 
-    /* FIXME: kind of a hackish way to invoke punch / looping stuff from the UI thread... */
-
-    if ( transport->rolling &&
-         transport->rec_enabled() &&
-         transport->punch_enabled() )
-    {
-        if ( _punched_in &&
-             transport->frame > _punch_in_frame && 
-             transport->frame > _punch_out_frame )
-        {
-            punch_out( _punch_out_frame );
-        }
-        else if ( ! _punched_in )
-        {
-            /* we've passed one or more punch regions... punch in for the next, if available. */
-            const Sequence_Widget *w = punch_cursor_track->next( transport->frame );
-            
-            DMESSAGE( "Delayed punch in" );
-            if ( w && 
-                 w->start() > transport->frame )
-            {
-                _punch_in_frame = w->start();
-                _punch_out_frame = w->start() + w->length();
-
-                punch_in( w->start() );
-            }
-        }            
-    }
-    
-
     if ( transport->rolling )
     {
         if ( play_cursor_track->active_cursor() )
@@ -1987,36 +1978,36 @@ Timeline::remove_track ( Track *track )
 void
 Timeline::command_move_track_up ( Track *track )
 {
-    wrlock();
+    track_lock.wrlock();
     insert_track( track, find_track( track ) - 1 );
-    unlock();
+    track_lock.unlock();
 }
 
 void
 Timeline::command_move_track_down ( Track *track )
 {
-    wrlock();
+    track_lock.wrlock();
     insert_track( track, find_track( track ) + 2 );
-    unlock();
+    track_lock.unlock();
 }
 
 void
 Timeline::command_remove_track ( Track *track )
 {
-    wrlock();
+    track_lock.wrlock();
     remove_track(track);
-    unlock();
+    track_lock.unlock();
 }
 
 void
 Timeline::command_quit ( void )
 {
-    timeline->wrlock();
+    track_lock.wrlock();
 
     Project::close();
 
-    timeline->unlock();
-  
+    track_lock.unlock();
+
     command_save();
   
     while ( Fl::first_window() ) Fl::first_window()->hide();
@@ -2025,9 +2016,10 @@ Timeline::command_quit ( void )
 void
 Timeline::command_undo ( void )
 {
-    wrlock();
+    /* FIXME: sequence lock too? */
+    track_lock.wrlock();
     Project::undo();
-    unlock();
+    track_lock.unlock();
 }
 
 bool
@@ -2036,9 +2028,9 @@ Timeline::command_load ( const char *name, const char *display_name )
     if ( ! name )
         return false;
   
-    wrlock();
+    track_lock.wrlock();
     int r = Project::open( name );
-    unlock();
+    track_lock.unlock();
 
     if ( r < 0 )
     {
diff --git a/timeline/src/Timeline.H b/timeline/src/Timeline.H
index d3da3c8..022c3af 100644
--- a/timeline/src/Timeline.H
+++ b/timeline/src/Timeline.H
@@ -77,9 +77,8 @@ struct Rectangle
     Rectangle ( int X, int Y, int W, int H ) : x( X ), y( Y ), w( W ), h( H ) {}
 };
 
-class Timeline : public Fl_Group, public RWLock
+class Timeline : public Fl_Group
 {
-
     class Timeline_Panzoomer;
 
     static void draw_clip_rulers ( void * v, int X, int Y, int W, int H );
@@ -125,6 +124,9 @@ class Timeline : public Fl_Group, public RWLock
 
 public:
 
+    RWLock track_lock;                                          /* tracks/sequences */
+    RWLock sequence_lock;                                       /* sequence contents */
+
     void redraw_overlay ( void );
 
     void insert_track ( Track *track, Track *before );
@@ -166,9 +168,6 @@ public:
 
     Fl_Menu_Button *menu;
 
-    int _punched_in;
-    nframes_t _punch_out_frame;
-    nframes_t _punch_in_frame;
     bool _created_new_takes;
 
     nframes_t xoffset;
@@ -257,6 +256,7 @@ public:
     void zoom_out ( void );
     void zoom_fit ( void );
 
+    bool next_punch ( nframes_t frame, nframes_t *in, nframes_t *out ) const;
 
     /* Engine */
     int  total_input_buffer_percent ( void );
@@ -267,7 +267,6 @@ public:
 
     bool record ( void );
     void stop ( void );
-    void punch_in ( nframes_t frame );
     void punch_out ( nframes_t frame );
 
     void wait_for_buffers ( void );
@@ -315,6 +314,7 @@ private:
 
     /* Engine */
     void resize_buffers ( nframes_t nframes );
-    nframes_t process ( nframes_t nframes );
+    nframes_t process_input ( nframes_t nframes );
+    nframes_t process_output ( nframes_t nframes );
     void seek ( nframes_t frame );
 };
diff --git a/timeline/src/Track.C b/timeline/src/Track.C
index bad2943..b416407 100644
--- a/timeline/src/Track.C
+++ b/timeline/src/Track.C
@@ -737,10 +737,10 @@ void
 Track::command_configure_channels ( int n )
 {
     /* due to locking this should only be invoked by direct user action */
-    timeline->wrlock();
+    timeline->track_lock.wrlock();
     configure_inputs( n );
     configure_outputs( n );
-    timeline->unlock();
+    timeline->track_lock.unlock();
 }
 
 void
@@ -787,9 +787,9 @@ Track::menu_cb ( const Fl_Menu_ *m )
         /* add audio track */
         char *name = get_unique_control_name( "Control" );
 
-        timeline->wrlock();
+        timeline->track_lock.wrlock();
         new Control_Sequence( this, name );
-        timeline->unlock();
+        timeline->track_lock.unlock();
     }
     else if ( ! strcmp( picked, "/Overlay controls" ) )
     {
@@ -868,9 +868,9 @@ Track::menu_cb ( const Fl_Menu_ *m )
     }
     else if ( !strcmp( picked, "Takes/New" ) )
     {
-        timeline->wrlock();
+        timeline->track_lock.wrlock();
         sequence( (Audio_Sequence*)sequence()->clone_empty() );
-        timeline->unlock();
+        timeline->track_lock.unlock();
     }
     else if ( !strcmp( picked, "Takes/Remove" ) )
     {
@@ -878,7 +878,7 @@ Track::menu_cb ( const Fl_Menu_ *m )
             {
                 Loggable::block_start();
 
-                timeline->wrlock();
+                timeline->track_lock.wrlock();
 
                 Audio_Sequence *s = sequence();
 
@@ -886,7 +886,7 @@ Track::menu_cb ( const Fl_Menu_ *m )
 
                 delete s;
 
-                timeline->unlock();
+                timeline->track_lock.unlock();
 
                 Loggable::block_end();
             }
@@ -906,9 +906,9 @@ Track::menu_cb ( const Fl_Menu_ *m )
     {
         Audio_Sequence* s = (Audio_Sequence*)m->mvalue()->user_data();
 
-        timeline->wrlock();
+        timeline->track_lock.wrlock();
         sequence( s );
-        timeline->unlock();
+        timeline->track_lock.unlock();
     }
 }
 
diff --git a/timeline/src/Waveform.C b/timeline/src/Waveform.C
index a5a7fb9..a6bb8b9 100644
--- a/timeline/src/Waveform.C
+++ b/timeline/src/Waveform.C
@@ -27,6 +27,7 @@
 #include <algorithm>
 using std::min;
 using std::max;
+#include <dsp.h>