diff --git a/sys/src/libthread/create.c b/sys/src/libthread/create.c
index 41a373c99..c58e6a837 100644
--- a/sys/src/libthread/create.c
+++ b/sys/src/libthread/create.c
@@ -97,7 +97,8 @@ procrfork(void (*f)(void *), void *arg, uint stacksize, int rforkflag)
 	int id;
 
 	p = _threadgetproc();
-	assert(p->newproc == nil);
+	while(p->newproc)
+		_sched();
 	p->newproc = _newproc(f, arg, stacksize, nil, p->thread->grp, rforkflag);
 	id = p->newproc->threads.head->id;
 	_sched();
diff --git a/sys/src/libthread/exec.c b/sys/src/libthread/exec.c
index 3bfbda3c6..b3ac6c213 100644
--- a/sys/src/libthread/exec.c
+++ b/sys/src/libthread/exec.c
@@ -45,8 +45,10 @@ procexec(Channel *pidc, char *prog, char *args[])
 	close(p->exec.fd[1]);
 	p->exec.fd[1] = n;
 
+	while(p->needexec || p->newproc)
+		_sched();
+
 	/* exec in parallel via the scheduler */
-	assert(p->needexec==0);
 	p->exec.prog = prog;
 	p->exec.args = args;
 	p->needexec = 1;
@@ -61,6 +63,8 @@ procexec(Channel *pidc, char *prog, char *args[])
 	}
 	close(p->exec.fd[0]);
 
+	if(t->ret == -1)
+		goto Bad;
 	if(pidc)
 		sendul(pidc, t->ret);
 
diff --git a/sys/src/libthread/exit.c b/sys/src/libthread/exit.c
index d36e11af4..a6ee207b2 100644
--- a/sys/src/libthread/exit.c
+++ b/sys/src/libthread/exit.c
@@ -2,7 +2,6 @@
 #include <libc.h>
 #include <thread.h>
 #include "threadimpl.h"
-#include <tos.h>
 
 char *_threadexitsallstatus;
 Channel *_threadwaitchan;
@@ -33,7 +32,7 @@ threadexitsall(char *exitstr)
 		exitstr = "";
 	_threadexitsallstatus = exitstr;
 	_threaddebug(DBGSCHED, "_threadexitsallstatus set to %p", _threadexitsallstatus);
-	mypid = _tos->pid; //getpid();
+	mypid = getpid();
 
 	/*
 	 * signal others.
@@ -47,7 +46,7 @@ threadexitsall(char *exitstr)
 		lock(&_threadpq.lock);
 		npid = 0;
 		for(p = _threadpq.head; p && npid < nelem(pid); p=p->next){
-			if(p->threadint == 0 && p->pid != mypid){
+			if(p->threadint == 0 && p->pid != mypid && p->pid != -1){
 				pid[npid++] = p->pid;
 				p->threadint = 1;
 			}
diff --git a/sys/src/libthread/id.c b/sys/src/libthread/id.c
index 289dc8698..778bdcd5c 100644
--- a/sys/src/libthread/id.c
+++ b/sys/src/libthread/id.c
@@ -2,7 +2,6 @@
 #include <libc.h>
 #include <thread.h>
 #include "threadimpl.h"
-#include <tos.h>
 
 int
 threadid(void)
diff --git a/sys/src/libthread/main.c b/sys/src/libthread/main.c
index a2a2eb0af..48a8acd52 100644
--- a/sys/src/libthread/main.c
+++ b/sys/src/libthread/main.c
@@ -28,8 +28,6 @@ main(int argc, char **argv)
 
 	rfork(RFREND);
 	mainp = &p;
-	if(setjmp(_mainjmp))
-		_schedinit(p);
 
 //_threaddebuglevel = (DBGSCHED|DBGCHAN|DBGREND)^~0;
 	_systhreadinit();
@@ -45,6 +43,7 @@ main(int argc, char **argv)
 	a->argv = argv;
 
 	p = _newproc(mainlauncher, a, mainstacksize, "threadmain", 0, 0);
+	setjmp(_mainjmp);
 	_schedinit(p);
 	abort();	/* not reached */
 }
@@ -117,18 +116,6 @@ void
 _schedexit(Proc *p)
 {
 	char ex[ERRMAX];
-	Proc **l;
-
-	lock(&_threadpq.lock);
-	for(l=&_threadpq.head; *l; l=&(*l)->next){
-		if(*l == p){
-			*l = p->next;
-			if(*l == nil)
-				_threadpq.tail = l;
-			break;
-		}
-	}
-	unlock(&_threadpq.lock);
 
 	utfecpy(ex, ex+sizeof ex, p->exitstr);
 	free(p);
diff --git a/sys/src/libthread/sched.c b/sys/src/libthread/sched.c
index 50603a3c4..888d64cc2 100644
--- a/sys/src/libthread/sched.c
+++ b/sys/src/libthread/sched.c
@@ -2,15 +2,9 @@
 #include <libc.h>
 #include <thread.h>
 #include "threadimpl.h"
-#include <tos.h>
-
-static Thread	*runthread(Proc*);
 
 static char *_psstate[] = {
-	"Moribund",
 	"Dead",
-	"Exec",
-	"Fork",
 	"Running",
 	"Ready",
 	"Rendezvous",
@@ -24,6 +18,23 @@ psstate(int s)
 	return _psstate[s];
 }
 
+static void
+unlinkproc(Proc *p)
+{
+	Proc **l;
+
+	lock(&_threadpq.lock);
+	for(l=&_threadpq.head; *l; l=&(*l)->next){
+		if(*l == p){
+			*l = p->next;
+			if(*l == nil)
+				_threadpq.tail = l;
+			break;
+		}
+	}
+	unlock(&_threadpq.lock);
+}
+
 void
 _schedinit(void *arg)
 {
@@ -31,8 +42,8 @@ _schedinit(void *arg)
 	Thread *t, **l;
 
 	p = arg;
+	p->pid = getpid();
 	_threadsetproc(p);
-	p->pid = _tos->pid; //getpid();
 	while(setjmp(p->sched))
 		;
 	_threaddebug(DBGSCHED, "top of schedinit, _threadexitsallstatus=%p", _threadexitsallstatus);
@@ -67,8 +78,15 @@ _schedinit(void *arg)
 			p->needexec = 0;
 		}
 		if(p->newproc){
-			t->ret = _schedfork(p->newproc);
-			p->newproc = nil;
+			Thread *x = p->newproc->threads.head;
+			if(x->moribund){
+				x->proc = p;
+				_threadready(x);
+				unlinkproc(p->newproc);
+				free(p->newproc);
+				p->newproc = nil;
+			} else if(_schedfork(p->newproc) != -1)
+				p->newproc = nil;
 		}
 		t->state = t->nextstate;
 		if(t->state == Ready)
@@ -90,44 +108,12 @@ needstack(int n)
 	
 	if((uchar*)&x - n < (uchar*)t->stk){
 		fprint(2, "%s %lud: &x=%p n=%d t->stk=%p\n",
-			argv0, _tos->pid, &x, n, t->stk);
-		fprint(2, "%s %lud: stack overflow\n", argv0, _tos->pid);
+			argv0, (ulong)p->pid, &x, n, t->stk);
+		fprint(2, "%s %lud: stack overflow\n", argv0, (ulong)p->pid);
 		abort();
 	}
 }
 
-void
-_sched(void)
-{
-	Proc *p;
-	Thread *t;
-
-Resched:
-	p = _threadgetproc();
-	if((t = p->thread) != nil){
-		needstack(128);
-		_threaddebug(DBGSCHED, "pausing, state=%s", psstate(t->state));
-		if(setjmp(t->sched)==0)
-			longjmp(p->sched, 1);
-		return;
-	}else{
-		t = runthread(p);
-		if(t == nil){
-			_threaddebug(DBGSCHED, "all threads gone; exiting");
-			_schedexit(p);
-		}
-		_threaddebug(DBGSCHED, "running %d.%d", t->proc->pid, t->id);
-		p->thread = t;
-		if(t->moribund){
-			_threaddebug(DBGSCHED, "%d.%d marked to die");
-			goto Resched;
-		}
-		t->state = Running;
-		t->nextstate = Ready;
-		longjmp(t->sched, 1);
-	}
-}
-
 static Thread*
 runthread(Proc *p)
 {
@@ -154,6 +140,39 @@ runthread(Proc *p)
 	return t;
 }
 
+void
+_sched(void)
+{
+	Proc *p;
+	Thread *t;
+
+Resched:
+	p = _threadgetproc();
+	if((t = p->thread) != nil){
+		needstack(128);
+		_threaddebug(DBGSCHED, "pausing, state=%s", psstate(t->state));
+		if(setjmp(t->sched)==0)
+			longjmp(p->sched, 1);
+		return;
+	}else{
+		t = runthread(p);
+		if(t == nil){
+			_threaddebug(DBGSCHED, "all threads gone; exiting");
+			unlinkproc(p);
+			_schedexit(p);	/* frees proc */
+		}
+		_threaddebug(DBGSCHED, "running %d.%d", t->proc->pid, t->id);
+		p->thread = t;
+		if(t->moribund){
+			_threaddebug(DBGSCHED, "%d.%d marked to die");
+			goto Resched;
+		}
+		t->state = Running;
+		t->nextstate = Ready;
+		longjmp(t->sched, 1);
+	}
+}
+
 void
 _threadready(Thread *t)
 {