diff --git a/examples/echoserver/echoserver.c b/examples/echoserver/echoserver.c
index 435cec473..732d4e8b4 100644
--- a/examples/echoserver/echoserver.c
+++ b/examples/echoserver/echoserver.c
@@ -86,6 +86,7 @@
     #endif
 #ifndef USE_WINDOWS_API
     #include <pwd.h>
+    #include <sys/wait.h>
 #endif
     #include <signal.h>
 #if defined(__QNX__) || defined(__QNXNTO__)
@@ -139,6 +140,15 @@ static int quit = 0;
 wolfSSL_Mutex doneLock;
 #define MAX_PASSWD_RETRY 3
 static int passwdRetry = MAX_PASSWD_RETRY;
+/* ssh_worker() reads ChildRunning whether or not a shell is compiled in.
+ * With a shell, ChildSig() writes it from a SIGCHLD handler, so it has to
+ * be sig_atomic_t; without one there is no handler, and no signal.h to
+ * declare that type. Zephyr's libc has neither. */
+#ifdef WOLFSSH_SHELL
+static volatile sig_atomic_t ChildRunning = 0;
+#else
+static volatile int ChildRunning = 0;
+#endif
 
 
 #ifndef EXAMPLE_HIGHWATER_MARK
@@ -197,7 +207,7 @@ typedef struct WS_FwdCbActionCtx {
 typedef struct {
     WOLFSSH* ssh;
     WS_SOCKET_T fd;
-    word32 id;
+    word32 tid;
     int echo;
     char nonBlock;
 #if defined(WOLFSSL_PTHREADS) && defined(WOLFSSL_TEST_GLOBAL_REQ)
@@ -212,6 +222,12 @@ typedef struct {
     WS_FwdCbActionCtx fwdCbCtx;
 #endif
     WS_AppCtx shellCtx;
+#ifdef WOLFSSH_SFTP
+    int doSftp;
+#endif
+#ifdef WOLFSSH_SCP
+    int doScp;
+#endif
     byte channelBuffer[EXAMPLE_BUFFER_SZ];
     /* The EOF drain holds an unsent tail across worker passes,
      * so it cannot share channelBuffer with the read path. */
@@ -250,7 +266,7 @@ static int dump_stats(thread_ctx_t* ctx)
             "Statistics for Thread #%u:\r\n"
             "  txCount = %u\r\n  rxCount = %u\r\n"
             "  seq = %u\r\n  peerSeq = %u\r\n",
-            ctx->id, txCount, rxCount, seq, peerSeq);
+            ctx->tid, txCount, rxCount, seq, peerSeq);
     statsSz = (word32)WSTRLEN(ctx->statsBuffer);
 
     fprintf(stderr, "%s", ctx->statsBuffer);
@@ -659,8 +675,9 @@ static int wolfSSH_FwdDefaultActions(WS_FwdCbAction action, void* vCtx,
     else if (action == WOLFSSH_FWD_CHANNEL_ID) {
         appCtx->channelId = port;
     }
-    else
+    else {
         ret = WS_FWD_INVALID_ACTION;
+    }
 
     return ret;
 }
@@ -668,6 +685,277 @@ static int wolfSSH_FwdDefaultActions(WS_FwdCbAction action, void* vCtx,
 #endif /* WOLFSSH_FWD */
 
 
+#ifdef WOLFSSH_SHELL
+static void ChildSig(int sig)
+{
+    (void)sig;
+    ChildRunning = 0;
+}
+
+
+/* Undo a forkpty() whose session request is about to be refused: the shell
+ * would run on behind a pty nothing reads, and its exit would clear
+ * ChildRunning out from under the worker loop. */
+static void ShellChildCleanup(thread_ctx_t* threadCtx, pid_t childPid)
+{
+    void (*prevSig)(int);
+
+    if (threadCtx->shellCtx.appFd >= 0) {
+        WCLOSESOCKET(threadCtx->shellCtx.appFd);
+        threadCtx->shellCtx.appFd = -1;
+    }
+
+    /* This exit is ours, not the session's. */
+    prevSig = signal(SIGCHLD, SIG_DFL);
+    kill(childPid, SIGKILL);
+    waitpid(childPid, NULL, 0);
+    signal(SIGCHLD, prevSig);
+}
+
+
+#ifdef SHELL_DEBUG
+static int termios_show(int fd)
+{
+    struct termios tios;
+    int i;
+    int rc;
+
+    WMEMSET((void *) &tios, 0, sizeof(tios));
+    rc = tcgetattr(fd, &tios);
+    printf("tcgetattr returns=%x\n", rc);
+
+    printf("iflag/oflag/cflag/lflag = %x/%x/%x/%x\n",
+            (unsigned int)tios.c_iflag, (unsigned int)tios.c_oflag,
+            (unsigned int)tios.c_cflag, (unsigned int)tios.c_lflag);
+    printf("c_ispeed/c_ospeed = %x/%x\n",
+            (unsigned int)tios.c_ispeed, (unsigned int)tios.c_ospeed);
+    for (i = 0; i < NCCS; i++) {
+        printf("c_cc[%d] = %hhx\n", i, tios.c_cc[i]);
+    }
+    return 0;
+}
+#endif
+#endif /* WOLFSSH_SHELL */
+
+
+/* One program start per connection, as RFC 4254 section 6.5 allows. A
+ * second start would fork a shell over the running one, or hand the
+ * session to sftp or scp, whose divert closes the pty. */
+static int SessionInUse(const thread_ctx_t* threadCtx)
+{
+    int inUse;
+
+    /* WS_SOCKET_T is unsigned on Windows, so the unset fd is -1 rather
+     * than anything below zero. */
+    inUse = threadCtx->shellCtx.state == APP_STATE_CONNECTED
+            || threadCtx->shellCtx.appFd != (WS_SOCKET_T)-1;
+#ifdef WOLFSSH_SFTP
+    inUse = inUse || threadCtx->doSftp;
+#endif
+#ifdef WOLFSSH_SCP
+    inUse = inUse || threadCtx->doScp;
+#endif
+
+    return inUse;
+}
+
+
+/* Registered in every build, in both modes: with no shell the echoserver
+ * still has to take the channel to mark it connected, so ssh_worker() will
+ * echo on it. Returns WS_SUCCESS to accept the request, 1 to reject it. */
+static int wsShellStartCb(WOLFSSH_CHANNEL* channel, void* ctx)
+{
+    thread_ctx_t* threadCtx = (thread_ctx_t*)ctx;
+    word32 channelId = 0;
+
+    if (threadCtx == NULL) {
+        return 1;
+    }
+
+    if (SessionInUse(threadCtx)) {
+        return 1;
+    }
+
+    /* Our own id: it is what wolfSSH_worker() reports and what the read,
+     * send, and find calls below take. */
+    if (wolfSSH_ChannelGetId(channel, &channelId, WS_CHANNEL_ID_SELF)
+            != WS_SUCCESS) {
+        return 1;
+    }
+
+#ifdef WOLFSSH_SHELL
+    /* Echo mode has no shell to start, ssh_worker() echoes the channel data
+     * back through the SSH stream. */
+    if (!threadCtx->echo) {
+        WOLFSSH* ssh;
+        const char *userName;
+        struct passwd *p_passwd;
+        struct termios tios;
+        pid_t childPid;
+        int rc;
+
+        ssh = threadCtx->ssh;
+        userName = wolfSSH_GetUsername(ssh);
+        p_passwd = getpwnam((const char *)userName);
+        if (p_passwd == NULL) {
+            /* Not actually a user on the system. */
+            #ifdef SHELL_DEBUG
+                fprintf(stderr, "user %s does not exist\n", userName);
+            #endif
+            return 1;
+        }
+
+        childPid = forkpty(&threadCtx->shellCtx.appFd, NULL, NULL, NULL);
+
+        if (childPid < 0) {
+            /* Refuse the request; the connection carries on without it. */
+            return 1;
+        }
+        else if (childPid == 0) {
+            /* Child process */
+            const char *args[] = {"-sh", NULL};
+
+            signal(SIGINT, SIG_DFL);
+
+            #ifdef SHELL_DEBUG
+                printf("userName is %s\n", userName);
+                system("env");
+            #endif
+
+            setenv("HOME", p_passwd->pw_dir, 1);
+            setenv("LOGNAME", p_passwd->pw_name, 1);
+            rc = chdir(p_passwd->pw_dir);
+            if (rc != 0) {
+                /* Never return: the child would run on inside the library
+                 * and write to the parent's socket. */
+                _exit(EXIT_FAILURE);
+            }
+
+            execv("/bin/sh", (char **)args);
+            _exit(EXIT_FAILURE);
+        }
+        #ifdef SHELL_DEBUG
+            printf("In childPid > 0; getpid=%d\n", (int)getpid());
+        #endif
+        rc = tcgetattr(threadCtx->shellCtx.appFd, &tios);
+        if (rc != 0) {
+            printf("tcgetattr failed: rc =%d,errno=%x\n", rc, errno);
+            ShellChildCleanup(threadCtx, childPid);
+            return 1;
+        }
+        rc = tcsetattr(threadCtx->shellCtx.appFd, TCSAFLUSH, &tios);
+        if (rc != 0) {
+            printf("tcsetattr failed: rc =%d,errno=%x\n", rc, errno);
+            ShellChildCleanup(threadCtx, childPid);
+            return 1;
+        }
+
+        /* Installed only now: the refusals above reap their own child. */
+        signal(SIGCHLD, ChildSig);
+
+        #ifdef SHELL_DEBUG
+            termios_show(threadCtx->shellCtx.appFd);
+        #endif
+
+        /* set initial size of terminal based on saved size */
+        #if !defined(NO_TERMIOS) && defined(WOLFSSH_TERM)
+        #if defined(HAVE_SYS_IOCTL_H)
+        wolfSSH_DoModes(ssh->modes, ssh->modesSz, threadCtx->shellCtx.appFd);
+        {
+            struct winsize s = {0};
+
+            s.ws_col = ssh->widthChar;
+            s.ws_row = ssh->heightRows;
+            s.ws_xpixel = ssh->widthPixels;
+            s.ws_ypixel = ssh->heightPixels;
+
+            ioctl(threadCtx->shellCtx.appFd, TIOCSWINSZ, &s);
+        }
+        #endif /* HAVE_SYS_IOCTL_H */
+
+        wolfSSH_SetTerminalResizeCtx(ssh, (void*)&threadCtx->shellCtx.appFd);
+        #endif /* !NO_TERMIOS && WOLFSSH_TERM */
+    }
+#endif /* WOLFSSH_SHELL */
+
+    /* Claim the channel only once it can be served. Claiming it up front
+     * would leave the worker driving a connected shell that never started. */
+    threadCtx->shellCtx.channelId = channelId;
+    threadCtx->shellCtx.state = APP_STATE_CONNECTED;
+
+    return WS_SUCCESS;
+}
+
+
+#ifdef WOLFSSH_SFTP
+static int wsSubsysStartCb(WOLFSSH_CHANNEL* channel, void* vCtx)
+{
+    int rej = 1;
+
+    if (vCtx && channel) {
+        thread_ctx_t* threadCtx;
+        const char* cmd;
+        WS_SessionType type;
+
+        threadCtx = (thread_ctx_t*)vCtx;
+
+        if (SessionInUse(threadCtx)) {
+            return 1;
+        }
+
+        cmd = wolfSSH_ChannelGetSessionCommand(channel);
+        type = wolfSSH_ChannelGetSessionType(channel);
+
+        /* A truncated subsystem string leaves the command NULL, and this
+         * runs before anything else has looked at it. The name matches
+         * whole, length and bytes, as wolfSSH_SFTP_accept() asks: granting
+         * sftp with an embedded NUL answers success on a session it then
+         * refuses. */
+        if (type == WOLFSSH_SESSION_SUBSYSTEM && cmd != NULL
+                && wolfSSH_ChannelGetSessionCommandSz(channel)
+                        == (word32)WSTRLEN("sftp")
+                && WSTRCMP(cmd, "sftp") == 0) {
+            threadCtx->doSftp = 1;
+            rej = WS_SUCCESS;
+        }
+    }
+
+    return rej;
+}
+#endif /* WOLFSSH_SFTP */
+
+
+/* An "scp ..." command starts a transfer, anything else runs as a session,
+ * the same as a shell request: the echoserver never runs the command. */
+static int wsExecStartCb(WOLFSSH_CHANNEL* channel, void* vCtx)
+{
+    int rej = 1;
+
+    if (vCtx && channel) {
+        const char* cmd = wolfSSH_ChannelGetSessionCommand(channel);
+
+        if (SessionInUse((thread_ctx_t*)vCtx)) {
+            return 1;
+        }
+
+#ifdef WOLFSSH_SCP
+        /* The prefix ChannelCommandIsScp() matches, so both modes agree. */
+        if (cmd != NULL && WSTRNCMP(cmd, "scp", 3) == 0) {
+            ((thread_ctx_t*)vCtx)->doScp = 1;
+            rej = WS_SUCCESS;
+        }
+        else
+#endif /* WOLFSSH_SCP */
+        {
+            rej = wsShellStartCb(channel, vCtx);
+        }
+        (void)cmd;
+    }
+
+    return rej;
+}
+
+
 #ifdef SHELL_DEBUG
 
 static void display_ascii(char *p_buf,
@@ -709,30 +997,6 @@ static void buf_dump(unsigned char *buf, int len)
     return;
 }
 
-
-#ifdef WOLFSSH_SHELL
-static int termios_show(int fd)
-{
-    struct termios tios;
-    int i;
-    int rc;
-
-    WMEMSET((void *) &tios, 0, sizeof(tios));
-    rc = tcgetattr(fd, &tios);
-    printf("tcgetattr returns=%x\n", rc);
-
-    printf("iflag/oflag/cflag/lflag = %x/%x/%x/%x\n",
-            (unsigned int)tios.c_iflag, (unsigned int)tios.c_oflag,
-            (unsigned int)tios.c_cflag, (unsigned int)tios.c_lflag);
-    printf("c_ispeed/c_ospeed = %x/%x\n",
-            (unsigned int)tios.c_ispeed, (unsigned int)tios.c_ospeed);
-    for (i = 0; i < NCCS; i++) {
-        printf("c_cc[%d] = %hhx\n", i, tios.c_cc[i]);
-    }
-    return 0;
-}
-#endif /* WOLFSSH_SHELL */
-
 #endif /* SHELL_DEBUG */
 
 
@@ -817,21 +1081,14 @@ static int termios_show(int fd)
 #endif
 
 
-int ChildRunning = 0;
-
-#ifdef WOLFSSH_SHELL
-static void ChildSig(int sig)
-{
-    (void)sig;
-    ChildRunning = 0;
-}
-#endif
-
 static int ssh_worker(thread_ctx_t* threadCtx)
 {
     WOLFSSH* ssh;
     WS_SOCKET_T sshFd;
     int rc = 0;
+    /* What the loop hands back, so a transfer taking over the session
+     * still leaves through the cleanup below it. */
+    int workerRet = 0;
     int eofAnswered = 0;
     /* Held across passes with 0 <= eofOff <= eofRead. */
     int eofRead = 0;
@@ -839,11 +1096,8 @@ static int ssh_worker(thread_ctx_t* threadCtx)
     /* Without a shell there is no child to outlive the peer's EOF, and the
      * read path echoes unconditionally. */
     int echoOnly = 1;
-#ifdef WOLFSSH_SHELL
-    const char *userName;
-    struct passwd *p_passwd;
-    WS_SOCKET_T childFd = 0;
-    pid_t childPid;
+#ifdef WOLFSSH_AGENT
+    int agentOpened = 0;
 #endif
 #if defined(WOLFSSL_PTHREADS) && defined(WOLFSSL_TEST_GLOBAL_REQ)
     pthread_t globalReq_th;
@@ -862,6 +1116,29 @@ static int ssh_worker(thread_ctx_t* threadCtx)
 
     sshFd = wolfSSH_get_fd(ssh);
 
+    if (threadCtx->shellCtx.state != APP_STATE_CONNECTED) {
+        /* The legacy path: accept() answered the session request itself,
+         * with no callback registered to claim the channel. Take it when
+         * it was granted and there is somewhere to put the data: nothing
+         * started a shell, so a shell build serves it only in echo mode.
+         * The grant is read from internal.h; the library has no public
+         * accessor for it yet. */
+        WOLFSSH_CHANNEL* sessionChannel;
+        int canServe = 1;
+
+#ifdef WOLFSSH_SHELL
+        canServe = echoOnly || threadCtx->shellCtx.appFd >= 0;
+#endif
+
+        sessionChannel = wolfSSH_ChannelNext(ssh, NULL);
+        if (canServe && sessionChannel != NULL
+                && sessionChannel->sessionGranted) {
+            threadCtx->shellCtx.state = APP_STATE_CONNECTED;
+            wolfSSH_ChannelGetId(sessionChannel,
+                    &threadCtx->shellCtx.channelId, WS_CHANNEL_ID_SELF);
+        }
+    }
+
 #if defined(WOLFSSL_PTHREADS) && defined(WOLFSSL_TEST_GLOBAL_REQ)
     /* submit Global Request for keep-alive */
     rc = pthread_create(&globalReq_th, NULL, global_req, threadCtx);
@@ -869,57 +1146,11 @@ static int ssh_worker(thread_ctx_t* threadCtx)
         printf("pthread_create() failed.\n");
 #endif
 
-#ifdef WOLFSSH_SHELL
-    if (!threadCtx->echo) {
-
-        userName = wolfSSH_GetUsername(ssh);
-        p_passwd = getpwnam((const char *)userName);
-        if (p_passwd == NULL) {
-            /* Not actually a user on the system. */
-            #ifdef SHELL_DEBUG
-                fprintf(stderr, "user %s does not exist\n", userName);
-            #endif
-            return WS_FATAL_ERROR;
-        }
-
-        ChildRunning = 1;
-        childPid = forkpty(&childFd, NULL, NULL, NULL);
-
-        if (childPid < 0) {
-            /* forkpty failed, so return */
-            ChildRunning = 0;
-            return WS_FATAL_ERROR;
-        }
-        else if (childPid == 0) {
-            /* Child process */
-            const char *args[] = {"-sh", NULL};
-
-            signal(SIGINT, SIG_DFL);
-
-            #ifdef SHELL_DEBUG
-                printf("userName is %s\n", userName);
-                system("env");
-            #endif
-
-            setenv("HOME", p_passwd->pw_dir, 1);
-            setenv("LOGNAME", p_passwd->pw_name, 1);
-            rc = chdir(p_passwd->pw_dir);
-            if (rc != 0) {
-                return WS_FATAL_ERROR;
-            }
-
-            execv("/bin/sh", (char **)args);
-        }
-    }
-#endif
     {
         /* Parent process */
-#ifdef WOLFSSH_SHELL
-        struct termios tios;
-#endif
+        int wantWrite = 0;
 #ifdef WOLFSSH_AGENT
         WS_SOCKET_T agentFd = -1;
-        WS_SOCKET_T agentListenFd = threadCtx->agentCtx.listenFd;
         word32 agentChannelId = -1;
 #endif
 #ifdef WOLFSSH_FWD
@@ -927,55 +1158,12 @@ static int ssh_worker(thread_ctx_t* threadCtx)
         word32 fwdBufferIdx = 0;
 #endif
 
-#ifdef WOLFSSH_SHELL
-        if (!threadCtx->echo) {
-            #ifdef SHELL_DEBUG
-                printf("In childPid > 0; getpid=%d\n", (int)getpid());
-            #endif
-            signal(SIGCHLD, ChildSig);
-
-            rc = tcgetattr(childFd, &tios);
-            if (rc != 0) {
-                printf("tcgetattr failed: rc =%d,errno=%x\n", rc, errno);
-                return WS_FATAL_ERROR;
-            }
-            rc = tcsetattr(childFd, TCSAFLUSH, &tios);
-            if (rc != 0) {
-                printf("tcsetattr failed: rc =%d,errno=%x\n", rc, errno);
-                return WS_FATAL_ERROR;
-            }
-
-            #ifdef SHELL_DEBUG
-                termios_show(childFd);
-            #endif
-        }
-        else
-            ChildRunning = 1;
-#else
         ChildRunning = 1;
-#endif
-
-#if !defined(NO_TERMIOS) && defined(WOLFSSH_TERM) && defined(WOLFSSH_SHELL)
-#if defined(HAVE_SYS_IOCTL_H)
-    /* if not echoing, set initial size of terminal based on saved size */
-    if (!threadCtx->echo) {
-        struct winsize s = {0,0,0,0};
-
-        wolfSSH_DoModes(ssh->modes, ssh->modesSz, childFd);
-        s.ws_col = ssh->widthChar;
-        s.ws_row = ssh->heightRows;
-        s.ws_xpixel = ssh->widthPixels;
-        s.ws_ypixel = ssh->heightPixels;
-
-        ioctl(childFd, TIOCSWINSZ, &s);
-
-        wolfSSH_SetTerminalResizeCtx(ssh, (void*)&childFd);
-    }
-#endif /* HAVE_SYS_IOCTL_H */
-#endif /* !NO_TERMIOS && WOLFSSH_TERM && WOLFSSH_SHELL */
 
         while (ChildRunning) {
             fd_set readFds;
+            fd_set writeFds;
+            int writable;
             WS_SOCKET_T maxFd;
             int cnt_r;
             int cnt_w;
@@ -984,18 +1172,37 @@ static int ssh_worker(thread_ctx_t* threadCtx)
             FD_SET(sshFd, &readFds);
             maxFd = sshFd;
 
+            FD_ZERO(&writeFds);
+            if (wantWrite)
+                FD_SET(sshFd, &writeFds);
+
+            #ifdef WOLFSSH_AGENT
+            /* The peer's auth-agent-req lands after wolfSSH_accept() has
+             * already returned in application-driven mode, so the channel
+             * answering it is opened here rather than inside accept(). The
+             * call reports WS_BAD_ARGUMENT until the request arrives. */
+            if (!agentOpened
+                    && wolfSSH_AGENT_ChannelOpen(ssh) == WS_SUCCESS) {
+                agentOpened = 1;
+            }
+            #endif
+
             #ifdef WOLFSSH_SHELL
-            if (!threadCtx->echo) {
-                FD_SET(childFd, &readFds);
-                if (childFd > maxFd)
-                    maxFd = childFd;
+            if (threadCtx->shellCtx.state == APP_STATE_CONNECTED
+                    && threadCtx->shellCtx.appFd >= 0) {
+                FD_SET(threadCtx->shellCtx.appFd, &readFds);
+                if (threadCtx->shellCtx.appFd > maxFd)
+                    maxFd = threadCtx->shellCtx.appFd;
             }
             #endif /* WOLFSSH_SHELL */
             #ifdef WOLFSSH_AGENT
-            if (threadCtx->agentCtx.state == APP_STATE_LISTEN) {
-                FD_SET(agentListenFd, &readFds);
-                if (agentListenFd > maxFd)
-                    maxFd = agentListenFd;
+            /* The poll above creates this listener mid-loop; re-read it
+             * each pass rather than caching it. */
+            if (threadCtx->agentCtx.state == APP_STATE_LISTEN
+                    && threadCtx->agentCtx.listenFd >= 0) {
+                FD_SET(threadCtx->agentCtx.listenFd, &readFds);
+                if (threadCtx->agentCtx.listenFd > maxFd)
+                    maxFd = threadCtx->agentCtx.listenFd;
             }
             if (agentFd >= 0
                     && threadCtx->agentCtx.state == APP_STATE_CONNECTED) {
@@ -1020,12 +1227,16 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                     maxFd = fwdFd;
             }
             #endif /* WOLFSSH_FWD */
-            rc = select((int)maxFd + 1, &readFds, NULL, NULL, NULL);
+
+            rc = select((int)maxFd + 1, &readFds,
+                    wantWrite ? &writeFds : NULL, NULL, NULL);
             if (rc == -1) {
                 break;
             }
+            writable = wantWrite && FD_ISSET(sshFd, &writeFds);
+            wantWrite = 0;
 
-            if (FD_ISSET(sshFd, &readFds)) {
+            if (FD_ISSET(sshFd, &readFds) || writable) {
                 word32 lastChannel = 0;
 
                 /* The following tries to read from the first channel inside
@@ -1035,6 +1246,18 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                    channel. The additional channel is only used with the
                    agent. */
                 cnt_r = wolfSSH_worker(ssh, &lastChannel);
+                #ifdef WOLFSSH_SFTP
+                if (threadCtx->doSftp) {
+                    workerRet = WS_SFTP_COMPLETE;
+                    break;
+                }
+                #endif
+                #ifdef WOLFSSH_SCP
+                if (threadCtx->doScp) {
+                    workerRet = WS_SCP_INIT;
+                    break;
+                }
+                #endif
                 /* Take the worker's status before the drain below: its
                  * reads and sends latch their own into ssh->error. */
                 rc = wolfSSH_get_error(ssh);
@@ -1045,8 +1268,11 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                  * us. Off the channel's own state, not the WS_EOF status: the
                  * flush inside wolfSSH_worker() can supersede that, and it is
                  * raised once. Echo mode only; a shell child on a pty is
-                 * still producing, so its EOF waits for the child to exit. */
-                if (!eofAnswered && echoOnly) {
+                 * still producing, so its EOF waits for the child to exit.
+                 * A claimed session only: unclaimed, shellCtx.channelId is
+                 * still 0, which is the first channel the peer is given. */
+                if (!eofAnswered && echoOnly
+                        && threadCtx->shellCtx.state == APP_STATE_CONNECTED) {
                     WOLFSSH_CHANNEL* eofChannel;
 
                     eofChannel = wolfSSH_ChannelFind(ssh,
@@ -1113,7 +1339,8 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                      * wolfSSH_ChannelIdRead() has no isKeying gate; the window
                      * credit it owes is parked until the rekey finishes. */
                     if (rc == WS_CHAN_RXD || rc == WS_REKEYING) {
-                        if (lastChannel == threadCtx->shellCtx.channelId) {
+                        if (threadCtx->shellCtx.state == APP_STATE_CONNECTED &&
+                                lastChannel == threadCtx->shellCtx.channelId) {
                             cnt_r = wolfSSH_ChannelIdRead(ssh,
                                     threadCtx->shellCtx.channelId,
                                     threadCtx->channelBuffer,
@@ -1130,7 +1357,8 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                             #endif
                             #ifdef WOLFSSH_SHELL
                                 if (!threadCtx->echo) {
-                                    cnt_w = (int)write(childFd,
+                                    cnt_w = (int)write(
+                                            threadCtx->shellCtx.appFd,
                                             threadCtx->channelBuffer, cnt_r);
                                 }
                                 else {
@@ -1243,6 +1471,14 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                          * above, which has already run this pass. */
                         continue;
                     }
+                    else if (rc == WS_WANT_WRITE) {
+                        /* The send is owed, not lost: wait for the socket to
+                         * take it. Application-driven mode answers session
+                         * requests here, so a blocked reply would otherwise
+                         * end a session accept() used to carry through. */
+                        wantWrite = 1;
+                        continue;
+                    }
                     else if (rc != WS_WANT_READ) {
                         #ifdef SHELL_DEBUG
                             printf("Break:read sshFd returns %d: errno =%x\n",
@@ -1252,11 +1488,11 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                     }
                 }
             }
-
             #ifdef WOLFSSH_SHELL
-            if (!threadCtx->echo) {
-                if (FD_ISSET(childFd, &readFds)) {
-                    cnt_r = (int)read(childFd,
+            if (threadCtx->shellCtx.state == APP_STATE_CONNECTED
+                    && threadCtx->shellCtx.appFd >= 0) {
+                if (FD_ISSET(threadCtx->shellCtx.appFd, &readFds)) {
+                    cnt_r = (int)read(threadCtx->shellCtx.appFd,
                             threadCtx->shellCtx.buffer,
                             sizeof threadCtx->shellCtx.buffer);
                     /* This read will return 0 on EOF */
@@ -1335,12 +1571,13 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                     }
                 }
             }
-            if (threadCtx->agentCtx.state == APP_STATE_LISTEN) {
-                if (FD_ISSET(agentListenFd, &readFds)) {
+            if (threadCtx->agentCtx.state == APP_STATE_LISTEN
+                    && threadCtx->agentCtx.listenFd >= 0) {
+                if (FD_ISSET(threadCtx->agentCtx.listenFd, &readFds)) {
                     #ifdef SHELL_DEBUG
                         printf("accepting agent connection\n");
                     #endif
-                    agentFd = accept(agentListenFd, NULL, NULL);
+                    agentFd = accept(threadCtx->agentCtx.listenFd, NULL, NULL);
                     if (agentFd == -1) {
                         rc = errno;
                         if (rc != SOCKET_EWOULDBLOCK) {
@@ -1371,8 +1608,7 @@ static int ssh_worker(thread_ctx_t* threadCtx)
                         fwdFd = -1;
                         threadCtx->fwdCtx.appFd = -1;
                         if (threadCtx->fwdCbCtx.hostName != NULL) {
-                            WFREE(threadCtx->fwdCbCtx.hostName,
-                                    NULL, 0);
+                            WFREE(threadCtx->fwdCbCtx.hostName, NULL, 0);
                             threadCtx->fwdCbCtx.hostName = NULL;
                         }
                         threadCtx->fwdCtx.state = APP_STATE_LISTEN;
@@ -1497,8 +1733,10 @@ static int ssh_worker(thread_ctx_t* threadCtx)
             #endif /* WOLFSSH_FWD */
         }
 #ifdef WOLFSSH_SHELL
-        if (!threadCtx->echo)
-            WCLOSESOCKET(childFd);
+        if (threadCtx->shellCtx.appFd >= 0) {
+            WCLOSESOCKET(threadCtx->shellCtx.appFd);
+            threadCtx->shellCtx.appFd = -1;
+        }
 #endif
     }
 
@@ -1506,10 +1744,13 @@ static int ssh_worker(thread_ctx_t* threadCtx)
     pthread_join(globalReq_th, NULL);
 #endif
 
-    return 0;
+    return workerRet;
 }
 
 
+/* Seconds to wait on the socket between sftp and scp accept attempts. */
+#define ES_ACCEPT_TIMEOUT 1
+
 #ifdef WOLFSSH_SFTP
 
 #define TEST_SFTP_TIMEOUT_SHORT 0
@@ -1756,8 +1997,10 @@ static THREAD_RETURN WOLFSSH_THREAD server_worker(void* vArgs)
     else {
         ret = NonBlockSSH_accept(threadCtx->ssh);
     }
+
 #ifdef WOLFSSH_SCP
-    /* finish off SCP operation */
+    /* The legacy path: accept() reports the scp command and does the
+     * transfer on re-entry. */
     if (ret == WS_SCP_INIT) {
         if (!threadCtx->nonBlock)
             ret = wolfSSH_accept(threadCtx->ssh);
@@ -1773,6 +2016,8 @@ static THREAD_RETURN WOLFSSH_THREAD server_worker(void* vArgs)
             break;
 
         #ifdef WOLFSSH_SFTP
+        /* The legacy path: wolfSSH_accept() ran the subsystem request
+         * itself and handed back a session ready to serve. */
         case WS_SFTP_COMPLETE:
             ret = sftp_worker(threadCtx);
             break;
@@ -1780,6 +2025,48 @@ static THREAD_RETURN WOLFSSH_THREAD server_worker(void* vArgs)
 
         case WS_SUCCESS:
             ret = ssh_worker(threadCtx);
+            #ifdef WOLFSSH_SCP
+            if (ret == WS_SCP_INIT) {
+                /* On a non-blocking socket the transfer comes back part
+                 * done; resume it rather than tearing the session down
+                 * mid-file. */
+                do {
+                    ret = wolfSSH_SCP_accept(threadCtx->ssh);
+                    error = wolfSSH_get_error(threadCtx->ssh);
+                    if (ret != WS_SCP_COMPLETE
+                            && (error == WS_WANT_READ
+                                || error == WS_WANT_WRITE)) {
+                        tcp_select(wolfSSH_get_fd(threadCtx->ssh),
+                                ES_ACCEPT_TIMEOUT);
+                    }
+                } while (ret != WS_SCP_COMPLETE
+                        && (error == WS_WANT_READ || error == WS_WANT_WRITE));
+                if (ret == WS_SCP_COMPLETE) {
+                    printf("scp file transfer completed\n");
+                    ret = 0;
+                }
+            }
+            #endif
+            #ifdef WOLFSSH_SFTP
+            if (ret == WS_SFTP_COMPLETE) {
+                do {
+                    ret = wolfSSH_SFTP_accept(threadCtx->ssh);
+                    error = wolfSSH_get_error(threadCtx->ssh);
+                    /* Wait on the socket between attempts; without this the
+                     * gap before the client's SFTP INIT is a busy spin. */
+                    if (ret != WS_SFTP_COMPLETE
+                            && (error == WS_WANT_READ
+                                || error == WS_WANT_WRITE)) {
+                        tcp_select(wolfSSH_get_fd(threadCtx->ssh),
+                                ES_ACCEPT_TIMEOUT);
+                    }
+                } while (ret != WS_SFTP_COMPLETE
+                        && (error == WS_WANT_READ || error == WS_WANT_WRITE));
+            }
+            if (ret == WS_SFTP_COMPLETE) {
+                ret = sftp_worker(threadCtx);
+            }
+            #endif
             break;
     }
 
@@ -3139,6 +3426,7 @@ static void ShowUsage(void)
 #ifdef WOLFSSH_SHELL
     printf(" -f            echo input\n");
 #endif
+    printf(" -A            drive channels from the application callbacks\n");
     printf(" -p <num>      port to connect on, default %d\n", wolfSshPort);
     printf(" -N            use non-blocking sockets\n");
 #ifdef WOLFSSH_SFTP
@@ -3192,7 +3480,7 @@ static void ShowUsage(void)
 }
 
 
-#define ECHOSERVER_OPTLIST "?1a:d:DefEp:R:Ni:j:i:I:J:K:P:k:b:x:m:c:s:G:HW:"
+#define ECHOSERVER_OPTLIST "?1a:Ad:DefEp:R:Ni:j:i:I:J:K:P:k:b:x:m:c:s:G:HW:"
 
 #ifdef WOLFSSH_WINDOWS_CERT_STORE
 /* Detects whether argv or the environment requests a host key from the
@@ -3323,6 +3611,7 @@ THREAD_RETURN WOLFSSH_THREAD echoserver_test(void* args)
     int userEcc = 0;
     int peerEcc = 0;
     int echo = 0;
+    int appChannels = 0;
     int ch;
     word16 port = wolfSshPort;
     char* readyFile = NULL;
@@ -3384,6 +3673,10 @@ THREAD_RETURN WOLFSSH_THREAD echoserver_test(void* args)
                     #endif
                     break;
 
+                case 'A':
+                    appChannels = 1;
+                    break;
+
                 case 'p':
                     if (myoptarg == NULL) {
                         ES_ERROR("NULL port value\n");
@@ -3625,6 +3918,25 @@ THREAD_RETURN WOLFSSH_THREAD echoserver_test(void* args)
 #ifdef WOLFSSH_FWD
     wolfSSH_CTX_SetFwdCb(ctx, wolfSSH_FwdDefaultActions, NULL);
 #endif
+    /* With -A the echoserver drives its own channels: accept() stops at
+     * userauth and these callbacks start the shell, subsystem or transfer.
+     * Off by default, so the path this example has always taken keeps an
+     * in-tree demo. The two are exclusive: the callbacks answer the session
+     * requests the accept state machine would otherwise answer itself. */
+    /* The shell callback is the only place the pty is forked, and an exec
+     * request that is not a transfer runs as a session, so both are
+     * registered in both modes. accept() honours a registered callback with
+     * application-driven channels off, so the legacy path keeps the shell it
+     * has always started for either request. The subsystem callback is not
+     * registered there: accept() serves sftp itself. */
+    wolfSSH_CTX_SetChannelReqShellCb(ctx, wsShellStartCb);
+    wolfSSH_CTX_SetChannelReqExecCb(ctx, wsExecStartCb);
+    if (appChannels) {
+        wolfSSH_CTX_SetAppChannels(ctx, 1);
+#ifdef WOLFSSH_SFTP
+        wolfSSH_CTX_SetChannelReqSubsysCb(ctx, wsSubsysStartCb);
+#endif
+    }
 
 #ifndef NO_FILESYSTEM
     if (sshPubKeyList) {
@@ -4044,6 +4356,7 @@ THREAD_RETURN WOLFSSH_THREAD echoserver_test(void* args)
     #endif
         wolfSSH_SetUserAuthCtx(ssh, &pwMapList);
         wolfSSH_SetKeyingCompletionCbCtx(ssh, (void*)ssh);
+        wolfSSH_SetChannelReqCtx(ssh, (void*)threadCtx);
 
         /* Use the session object for its own highwater callback ctx */
         if (defaultHighwater > 0) {
@@ -4103,13 +4416,13 @@ THREAD_RETURN WOLFSSH_THREAD echoserver_test(void* args)
             tcp_set_nonblocking(&clientFd);
 
         wolfSSH_set_fd(ssh, (int)clientFd);
+        threadCtx->fd = clientFd;
 
 #if defined(WOLFSSL_PTHREADS) && defined(WOLFSSL_TEST_GLOBAL_REQ)
         threadCtx->ctx = ctx;
 #endif
         threadCtx->ssh = ssh;
-        threadCtx->fd = clientFd;
-        threadCtx->id = threadCount++;
+        threadCtx->tid = threadCount++;
         threadCtx->nonBlock = nonBlock;
         threadCtx->echo = echo;
         threadCtx->shellCtx.privateData = NULL;
diff --git a/scripts/scp.test b/scripts/scp.test
index 20f9737eb..cfa69739c 100755
--- a/scripts/scp.test
+++ b/scripts/scp.test
@@ -137,6 +137,67 @@ else
     exit 1
 fi
 
+# With -A the echoserver binds the scp command itself and runs the transfer
+# through wolfSSH_SCP_accept(); the cases above take the accept() re-entry
+# instead. -N reaches that call's want-read/want-write retry path, which a
+# blocking server completes in one call. The client stays blocking, which the
+# server does not care about.
+echo "Test basic copy from server to local, app-driven server"
+./examples/echoserver/echoserver -A -N -1 -R $ready_file &
+server_pid=$!
+create_port
+$run_client ./examples/scpclient/wolfscp -u jill -P upthehill -p $port -S $PWD/scripts/scp.test:$PWD/scp.test
+RESULT=$?
+remove_ready_file
+stop_server
+check_timeout $RESULT "basic copy from server to local, app-driven server"
+
+if test -e $PWD/scp.test; then
+    rm $PWD/scp.test
+else
+    echo -e "\n\nfailed to get file from app-driven server"
+    do_cleanup
+    exit 1
+fi
+
+echo "Test basic copy from local to server, app-driven server"
+./examples/echoserver/echoserver -A -N -1 -R $ready_file &
+server_pid=$!
+create_port
+$run_client ./examples/scpclient/wolfscp -u jill -P upthehill -p $port -L $PWD/scripts/scp.test:$PWD/scp.test
+RESULT=$?
+remove_ready_file
+stop_server
+check_timeout $RESULT "basic copy from local to server, app-driven server"
+
+if test -e $PWD/scp.test; then
+    rm $PWD/scp.test
+else
+    echo -e "\n\nfailed to send file to app-driven server"
+    do_cleanup
+    exit 1
+fi
+
+# The same handoff on a blocking server, where wolfSSH_SCP_accept() runs to
+# completion in one call rather than through the retry loop above.
+echo "Test basic copy from server to local, blocking app-driven server"
+./examples/echoserver/echoserver -A -1 -R $ready_file &
+server_pid=$!
+create_port
+$run_client ./examples/scpclient/wolfscp -u jill -P upthehill -p $port -S $PWD/scripts/scp.test:$PWD/scp.test
+RESULT=$?
+remove_ready_file
+stop_server
+check_timeout $RESULT "basic copy from server to local, blocking app-driven server"
+
+if test -e $PWD/scp.test; then
+    rm $PWD/scp.test
+else
+    echo -e "\n\nfailed to get file from blocking app-driven server"
+    do_cleanup
+    exit 1
+fi
+
 echo "Test of getting empty file"
 touch $PWD/scripts/empty
 ./examples/echoserver/echoserver -1 -R $ready_file &
diff --git a/scripts/sftp.test b/scripts/sftp.test
index 3d5304cfe..9e551df15 100755
--- a/scripts/sftp.test
+++ b/scripts/sftp.test
@@ -112,6 +112,41 @@ if [ $RESULT -ne 0 ]; then
     exit 1
 fi
 
+# With -A the echoserver answers the subsystem request in its own callback
+# and serves the session through wolfSSH_SFTP_accept(); the cases above take
+# the accept() re-entry instead.
+if [ "$nonblockingOnly" = 0 ]; then
+    echo "Test connection to an app-driven server"
+    ./examples/echoserver/echoserver -A -1 -R "$ready_file" &
+    server_pid=$!
+    create_port
+    echo "exit" | ./examples/sftpclient/wolfsftp -u jill -P upthehill -p "$port"
+    RESULT=$?
+    remove_ready_file
+    if [ $RESULT -ne 0 ]; then
+        echo
+        echo "failed to connect to app-driven server"
+        do_cleanup
+        exit 1
+    fi
+fi
+
+# The same handoff on a non-blocking server, which reaches the accept call's
+# want-read and want-write retries.
+echo "Test non blocking connection to an app-driven server"
+./examples/echoserver/echoserver -A -N -1 -R "$ready_file" &
+server_pid=$!
+create_port
+echo "exit" | ./examples/sftpclient/wolfsftp -N -u jill -P upthehill -p "$port"
+RESULT=$?
+remove_ready_file
+if [ $RESULT -ne 0 ]; then
+    echo
+    echo "failed to connect to non blocking app-driven server"
+    do_cleanup
+    exit 1
+fi
+
 # Test want write return from highwater callback
 if [ "$nonblockingOnly" = 0 ]; then
     echo "Test want write return from highwater callback"
diff --git a/scripts/sshclient.test b/scripts/sshclient.test
index af145f968..762f457fe 100755
--- a/scripts/sshclient.test
+++ b/scripts/sshclient.test
@@ -83,6 +83,7 @@ trap do_trap INT TERM
 # -f keeps the server in echo mode. Without it a build with shell support
 # tries to fork a login shell for the user, which fails since jill isn't a
 # real account, and the session ends before anything crosses the channel.
+# Any arguments are passed on to the echoserver.
 start_server() {
     # The -1 server exits after its connection, but a client run that failed
     # before connecting leaves one listening. Reap it, server_pid is about
@@ -95,7 +96,7 @@ start_server() {
     fi
 
     rm -f "$ready_file"
-    ./examples/echoserver/echoserver -1 -f -R "$ready_file" \
+    ./examples/echoserver/echoserver -1 -f "$@" -R "$ready_file" \
         > "$work_dir/server.log" 2>&1 &
     server_pid=$!
 
@@ -282,6 +283,25 @@ if [ -s "$work_dir/terminal.log" ]; then
         && fail "the client sent a command it wasn't given"
 fi
 
+# -A hands the session requests to the echoserver's own callbacks: the shell
+# callback claims the channel the worker echoes on, and an exec request that
+# is not a transfer runs as a session through the same callback.
+echo "Test terminal session, app-driven server"
+start_server -A
+run_client "" -E "$work_dir/appterm.log" -p $port jill@127.0.0.1
+[ $? -ne 0 ] && fail "failed to open the terminal session on an app-driven server"
+
+grep -q "hello" "$client_out" \
+    || fail "the app-driven echoserver's reply didn't make it back"
+
+echo "Test a session given a command, app-driven server"
+start_server -A
+run_client "" -E "$work_dir/appcommand.log" -p $port jill@127.0.0.1 "echo three"
+[ $? -ne 0 ] && fail "failed to open the session on an app-driven server"
+
+grep -q "hello" "$client_out" \
+    || fail "the app-driven echoserver's reply didn't make it back"
+
 echo "Test the usage message"
 ./apps/wolfssh/wolfssh -Z 2>&1 | grep -q "usage:" \
     || fail "no usage message for a bad option"
diff --git a/tests/api.c b/tests/api.c
index 55009434c..1d309a8f3 100644
--- a/tests/api.c
+++ b/tests/api.c
@@ -2611,6 +2611,8 @@ static void test_wolfSSH_SCP_CB(void)
     AssertIntEQ(wolfSSH_SetScpErrorMsg(NULL, err), WS_BAD_ARGUMENT);
     AssertIntEQ(wolfSSH_SetScpErrorMsg(ssh, NULL), WS_BAD_ARGUMENT);
 
+    AssertIntEQ(wolfSSH_SCP_accept(NULL), WS_BAD_ARGUMENT);
+
     wolfSSH_free(ssh);
     wolfSSH_CTX_free(ctx);
 }
@@ -8004,6 +8006,9 @@ static void test_wolfSSH_KeyboardInteractive(void)
     argsCount = 0;
     args[argsCount++] = ".";
     args[argsCount++] = "-1";
+    /* Echo mode: "test" is not an account on the host, so the echoserver's
+     * shell callback would refuse the shell request this client sends. */
+    args[argsCount++] = "-f";
     args[argsCount++] = "-i";
     args[argsCount++] = "test:test";
     args[argsCount++] = "-p";