If you are developing new software or updating older software, you need to
avoid compatibility traps.  This comprehensive list of Release 2
compatibility problem areas can help you avoid and diagnose compatibility
problems  In addition, refer to the "General Amiga Development Guidelines"
listed in Chapter 1.

 General Compatibility Problem Areas 
 Release 2 Changes That Can Affect Compatibility 
 Additional Information 

The following improper Amiga programming practices are likely to fail on
new ROMs or hardware.

  * Requiring all free RAM.

  * Overwriting memory allocations.  With 32-bit addresses, a 1-byte
    overwrite of a string array can wipe out the high byte of a pointer
    or stack return address.  This bug could go unnoticed on a 24-bit
    address machine (e.g., A500) but crash the system or cause other
    problems on an A3000.

  * Improper flags or garbage in system structures.  A bit that means
    nothing under one OS may drastically change the behavior of a
    function in a newer version of the OS.  Clear structures before
    using, and use correct flags.

  * Misuse of function return values.  Use function prototypes and read
    the Autodocs for the functions you are using.  Some system functions
    return just success or failure, or nothing at all (void).  In such
    cases, the value which the function happens to return must not be
    used except as it is documented.

  * Depending on unsupported side effects or undocumented behavior.  Be
    sure to read the Autodocs, include file comments and other
    documentation.

  * Assuming current choices, configurations or initial values.  If the
    current possibilities are A, B, or C, do not assume C if it isn't A
    or B.  Check specifically for the choices currently implemented, and
    provide default behavior for unexpected values.

Amiga debugging tools such as Enforcer, Mungwall and Scratch can find many
program bugs that may affect compatibility. A program that is
Enforcer/Mungwall/Scratch clean stands a much better chance of working
well under current and future versions of the OS.

There are several areas where Relase 2 OS changes and enhancements can
cause compatibility problems for some software.

 Exec               Serial Device             Preferences 
 Expansion          Timer Device              Workbench 
 Strap              Trackdisk Device          Layers 
 DOS                CIA Timers               Graphics 
 Audio Device       Other Hardware Issues    Fonts 
 Gameport Device    Intuition                CLI/Shell 

  * Do not jump to location $FC0002 -- the start of the ROM under
    1.3 -- as part of performing a system RESET.  The 2.04 Kickstart
    ROM has a temporary compatibility hack called "Kickety-Split" which
    is a redirecting jump at $FC0002.  This hack does not appear on the
    A3000 ROM and due to space considerations will not appear on future
    machines.

  * Everything has moved.

  * The Supervisor stack is not in the same place as it was under 1.3.
    This has caused problems for some games that completely take over the
    Amiga.  If your program goes into Supervisor mode, you must either
    respect allocated memory or provide your own Supervisor stack when
    taking over the machine.

  * ExecBase is moved to expansion memory if possible.  Before, ExecBase
    would only end up in one of two fixed locations.  Now, ColdCapture
    may be called after expansion memory has been configured.

  * Exception/Interrupt vectors may move.  This means the 68010 and above
    Vector Base Register (VBR) may contain a non-zero value.  Poking
    assumed low memory vector addresses may have no effect.  You must
    read the VBR on 68010 and above to find the base.

  * No longer tolerant of wild Forbid() counts.  Under 1.3, sometimes
    this bug could go unnoticed.  Make sure that all Forbid()s are
    matched with one and only one Permit() (and vice versa).

  * When an Exec device gets an IORequest, it must validate io_Command.
    If the io_Command is 0 or out of range, the device must return
    IOERR_NOCMD and take no other action.  The filesystem now sends new
    commands and expects older devices to properly ignore them.

  * A fix to task-switching in Release 2 allows a busy task to properly
    regain the processor after an interrupt until either its quantum (4
    vertical blanks) is up or a higher priority task preempts it.  This
    can dramatically change the behavior of multitask programs where one
    task busyloops while another same-priority task Wait()s.  See
    "Task Switching" in the "Additional Information" section below.

  * ExpansionBase is private - use FindConfigDev().

  * Memory from contiguous cards of the same memory type is automatically
    merged into one memory pool.

  * Romboot.library is gone.

  * Audio.device cannot be OpenDevice()ed by a boot block program.  See
    "Audio Device" below.

  * Boot from other floppies (+5,-10,-20,-30) is possible.

  * Undocumented system stack and register usage at Diag and Boot time
    have changed.

  * DOS is now written in C and assembler, not BCPL.  The BCPL compiler
    artifact which caused D0 function results to also be in D1 is gone.
    System patches in Release 2 that return some DOS function results in
    both D0 and D1 are not guaranteed to remain in the next release. Fix
    your programs!  Use Scratch to find these problems in your code.

  * DOS now has a real library base with normal LVO vectors.

  * Stack usage has all changed (variables, direction).

  * New packet and lock types.  Make sure you are not passing stack
    garbage for the second argument to Lock().

  * Process structure is bigger.  "Rolling your own" Process structure
    from a Task fails.  Use dos.library System() or CreateNewProc().

  * Unless documented otherwise, you must be a process to call DOS
    functions. DOS function dependence on special process structures can
    change with OS revisions.

  * Now not initialized until used.  This means low memory open failure
    is possible.  Check your return values from OpenDevice().  This also
    means audio.device cannot be opened during 2.0 Strap unless
    InitResident()ed first.  If OpenDevice() of audio.device fails during
    strap, you must FindResident()/InitResident() audio.device, and then
    try OpenDevice() again.  There will be a small memory loss (until
    reboot) generated by the first opener of audio.device or
    narrator.device (memory used in building of audio.device's base).

  * Initial state of hardware lines may differ.

  * Clears io_Device on CloseDevice() (since 1.3.2)

  * The most common mistake programmers make with timer.device is to send
    off a particular timerequest before the previous use of that
    timerequest has completed.  Use IO_Torture to catch this.

  * IO_QUICK requests may be deferred and be replied as documented.

  * VBLANK timer requests, as documented, now wait at least as long as
    the full number of VBlanks you asked for.  Previously, a partial
    vertical blank could count towards your requested number.  The new
    behavior is more correct and matches the docs, but it can cause
    VBlank requests to now take up to 1 VBlank longer under 2.0 as
    compared to 1.3. For example, a 1/10 second request, may take 6-7
    Vblanks instead of 5-6 VBlanks, or about 15% longer.

  * Private trackdisk structures have changed.  See trackdisk.doc for a
    compatible REMCHANGEINT.

  * Buffer is freeable, so low memory open failure is possible.

  * Do not disable interrupts (any of them), then expect trackdisk to
    function while they are disabled.

  * System use of CIA timers has changed.  Don't assume how they're used.

  * Don't depend on initial values of CIA registers.

  * Don't mess with CIABase. Use cia.resource.

  * If your code requires hardware level CIA timers, allocate the timers
    using cia.resource AddICRVector()!  This is very important. Operating
    system usage of the CIA timers has changed.  The new 2.0 timer.device
    ("Jumpy the Magic Timer Device") will try to jump to different CIAs
    so programs that properly allocate timers will have a better chance
    of getting what they want.  If possible, be flexible and design your
    code to work with whatever timer you can successfully allocate.

  * OS usage of INT6 is increasing.  Do not totally take over INT6, and
    do not terminate the server chain if an interrupt is not for you.

  * Battery-backed clock is different on A3000.  Use battclock.resource
    to access the real-time clock if battclock.resource can be opened.

  * A 68030 hardware characteristic causes longword-aligned longword
    writes to allocate a valid entry in the data cache,  even if the
    hardware area shouldn't be cached.  This can cause problems for I/O
    registers and shared memory devices.  To solve this: 1) don't do that
    2) flush the cache or 3) use Enforcer Quiet.  See the Motorola 68030
    manual under the description of the Write Allocate bit (which must be
    set for the Amiga to run with the Data Cache).

  * Private IntuitionBase variables have moved/changed.  Reading them is
    illegal.  Writing them is both illegal and dangerous.

  * Poking IntuitionBase MaxMouse variables is now a no-op, but stop
    poking when Intuition version is >35.

  * If you are opening on the Workbench screen, be prepared to handle
    larger screens, new modes, new fonts, and overscan.  Also see
    "Fonts" compatibility information.

  * Screen TopEdge and LeftEdge may be negative.

  * Left-Amiga-Select is used for dragging large screens. Do not use
    left-Amiga-key combinations for application command keys.  The
    left-Amiga key is reserved for system use.

  * For compatibility reasons, GetScreenData() lies if the Workbench
    screen is a mode only available after 1.3.  It will try to return the
    most sensible mode that old OpenScreen() can open.  This was
    necessary to prevent problems in applications that cloned the
    Workbench screen. To properly handle new modes, see LockPubScreen(),
    GetVPModeID(), and the SA_DisplayID tag for OpenScreenTags().

  * Using combined RAWKEY and VANILLAKEY now gives VANILLAKEY messages
    for regular keys, and RAWKEY messages for special keys (fkeys, help,
    etc.)

  * Moving a SIMPLE_REFRESH window does not necessarily cause a
    REFRESHWINDOW event because layers now preserves all the bits it can.

  * Sizing a SIMPLE_REFRESH window will not clear it.

  * MENUVERIFY/REQVERIFY/SIZEVERIFY can time out if you take too long to
    ReplyMsg().

  * Menu-key equivalents are ignored while string gadgets are active.

  * You can't type control characters into string gadgets by default.
    Use Ctrl-Amiga-char to type them in or use IControl Prefs to change
    the default behavior.

  * Width and Height parameters of AutoRequest() are ignored.

  * New default colors, new gadget images.

  * JAM1 rendering/text in border may be invisible gadgets over default
    colors.

  * The cursor for string gadgets can no longer reside outside the
    cleared container area.  If your gadget is 32 pixels wide, with
    MaxChars of 4, all 32 pixels will be cleared, instead of just 24, as
    was true in 1.3.

  * Applications and requesters that fail to specify desired fonts will
    get the fonts the user sets up in Font Preferences in Release 2.
    These could be much larger, or proportional in some cases.  Screen
    and window titlebars (and their gadgets) will be taller when
    accommodating a larger font.  Applications which open on the
    Workbench screen must adapt to variable size titlebars.  Any
    application which accepts system defaults for its screen, window,
    menu, Text or IntuiText fonts must adapt to different fonts and
    titlebar sizes.  String gadgets whose height is too small for a font
    will revert to a smaller ROM font.  There are now 2 different
    user-specifiable default system fonts which affect different
    Intuition features.  This can lead to mismatches in mixed gadgets and
    text.  See the "Intuition Screens" chapter.

  * Don't modify gadgets directly without first removing them from the
    gadget list, unless you are using a system function designed for that
    purpose, such as NewModifyProp() or SetGadgetAttrs().

  * Don't rely on NewModifyProp() to fully refresh your prop gadget after
    you've changed values in the structure.  NewModifyProp() will only
    correctly refresh changes which were passed to it as parameters.  Use
    Remove/Add/RefreshGList() for other kinds of changes.

  * Custom screens must be of type CUSTOMSCREEN or PUBLICSCREEN.  Other
    types are illegal.  One application opens its screen with
    NewScreen.Type = 0 (instead of CUSTOMSCREEN, 0x0F).  Then, when it
    opens its windows, it specifies NewWindow.Type of 0 and
    NewWindow.Screen of NULL, instead of Type = CUSTOMSCREEN and
    Screen = (their screen). That happened to work before, but no longer.

  * Referencing IntuiMessage->IAddress as a Gadget pointer on non-Gadget
    IDCMP messages, or as a Window pointer (rather than looking at the
    proper field IntuiMessage->IDCMPWindow) may now cause Enforcer hits
    or crashes.  The IAddress field always used to contain a pointer of
    some type even for IDCMP events for which no IAddress value is
    documented.  Now, for some IDCMP events, IAddress may contain a
    non-address, possibly an odd value that would crash a 68000 based
    system).

  * Using Intuition flags in the wrong structure fields (for example,
    using ACTIVEWINDOW instead of ACTIVATE).  To alleviate this problem,
    2.0 has introduced new synonyms that are less confusing than the old
    ones.  For example, IDCMP_ACTIVEWINDOW and WFLG_ACTIVATE.  This
    particular example of confusion (there are several) was the nastiest,
    since IDCMP_ACTIVEWINDOW, when stuffed into NewWindow.Flags,
    corresponds numerically to WFLG_NW_EXTENDED, which informs Intuition
    that the NewWindow structure is immediately followed by a TagItem,
    list which isn't there!  Intuition does some validation on the
    tag-list pointer, in order to partially compensate.  To make your
    compiler use the new synonyms only, add this line to your code before
    Intuition include files: #define INTUI_V36_NAMES_ONLY.

  * Do not place spaces into the StringInfo->Buffer of a GACT_LONGINT
    string gadget. Under 1.3, this worked, but the 2.0 validation routine
    that checks for illegal keystrokes looks at the contents for illegal
    (i.e., non-numeric) characters, and if any are found assumes that the
    user typed an illegal keystroke.  The user's only options may be
    shift-delete or Amiga-X. Use the correct justification instead.

  * If you specify NULL for a font in an IntuiText, don't assume you'll
    get Topaz 8.  Either explicitly supply the font you you need or be
    prepared to size accordingly.  Otherwise, your rendering will be
    wrong, and the user will have to reset his Preferences just to make
    your software work right.

  * Window borders are now drawn in the screen's DetailPen and BlockPen
    rather than the Window's pens.  For best appearance, you should pass
    an SA_Pens array to OpenScreen(). This can be done in a backwards
    compatible manner with the ExtNewScreen structure and the NS_EXTENDED
    flag.

  * The system now renders into the full width of window borders,
    although the widths themselves are unchanged.  Window borders are
    filled upon activation and inactivation.

  * Window border rendering has changed significantly for 2.0.  Note that
    the border dimensions are unchanged from 1.x (Look at
    Window->BorderLeft/Top/Right/Bottom if you don't believe us!).  If
    your gadget intersects the border area, although it may have looked
    OK under 1.3, a visual conflict may occur under 2.0.  If Intuition
    notices a gadget which is substantially in the border but not
    declared as such, it treats it as though it were (this is called
    "bordersniffing").  Never rely on Intuition to sniff these out for
    you; always declare them explicitly (see the Gadget Activation flags
    GACT_RIGHTBORDER, etc.).  See "Intuition Gadgets and Window Borders"
    in the "Additional Information" section below.

  * Some old struct Preferences fields are now ignored by SetPrefs() (for
    example FontHeight). SetPrefs() also stops listening to the pointer
    fields as soon as a new-style pointer is passed to Intuition
    (new-style pointers can be taller or deeper).

  * Preferences ViewX/YOffset only applies to the default mode. You
    cannot use these fields to move the position of all modes.

  * The Preferences LaceWB bit is not necessarily correct when Workbench
    is in a new display mode (akin to GetScreenData()).

  * The Workbench GUI now has new screen sizes, screen top/left offsets,
    depths, modes, and fonts.

  * Default Tool now searches paths.

  * New Look (boxed) icons take more space.

  * Do not use icons which have more 1bits set in PlanePick than planes
    in the ImageData (one IFF-to-Icon utility does this).  Such icons
    will appear trashed on deeper Workbenches.

  * New Look colors have black and white swapped (as compared to 1.3).

  * The Workbench screen may not be open at startup-sequence time until
    some output occurs to the initial Shell window.  This can break
    startup-sequence-started games that think they can steal WB's screen
    bitplanes.  Do not steal the Workbench screen's bitplanes.  (For
    compatibility, booting off pre-2.0 disks forces the initial screen
    open. This is not guaranteed to remain in the system.)  Use startup
    code that can detach when RUN (such as cback.o) and use
    CloseWorkbench() to regain the screen's memory.  In addition, see
    "Workbench and Startup" in the "Additional Information" section
    below.

  * Use NewLayerInfo() to create, not FattenLayerInfo(), ThinLayerInfo(),
    InitLayers().

  * Simple-refresh preserves all of the pixels it can.  Sizing a
    SIMPLE_REFRESH window no longer clears the whole window.

  * Speed of layer operations is different.  Don't depend on layer
    operations to finish before or after other asynchronous actions.

  * Do not rely on the order of Copper list instructions.  The Release 2
    MrgCop() function builds different Copper lists to that of 1.3, by
    including new registers in the list (e.g., MOVE xxxx,DIWHIGH).  This
    changes the positions of  the  other  instructions.  We  know  of one
    game that 'assumes' the BPLxPTRs  would be at a certain offset in the
    Copper list, and that is now broken on machines running 2.0 with the
    new Denise chip.

  * Graphics and layers functions which use the blitter generally return
    after starting the final blit.  If you are mixing graphics rendering
    calls and processor access of the same memory, you must WaitBlit()
    before touching (or deallocating) the source or destination memory
    with the processor.  For example, the Text() function is faster in
    Release 2, causing some programs to trash partial lines of text.

  * ColorMap structure is bigger.  Programs must use GetColorMap() to
    create one.

  * Blitter rtns decide ascend/descend on 1st plane only.

  * Changing the display mode of an existing screen or viewport while
    open is still not a supported operation.

  * GfxBase DisplayFlags and row/cols may not match Workbench screen.

  * Do not hardcode modulo values - use BitMap->BytesPerRow.

  * If the graphics Autodocs say that you need a TmpRas of a certain size
    for some functions, then you must make that the minimum size.  In
    some cases, before 2.0, you may have gotten away with using a smaller
    TmpRas with some functions (for example Flood()).  To be more robust,
    graphics now checks the TmpRas size and will fail the function call
    if the TmpRas is too small.

  * ECS chips under 2.0 generate displays differently. The display window
    registers now control DMA.

  * LoadRGB4() used to poke colors into the active copperlist with no
    protection against deallocation of that copperlist while it was being
    poked.  Under 2.0, semaphore protection of the copperlist was added
    to LoadRGB4() which makes it totally incorrect and extremely
    dangerous to call LoadRGB4() during an interrupt.  The general
    symptom of this problem is that a system deadlock can be caused by
    dragging one screen up and down while another is cycling.  Color
    cycling should be performed from within a task, not an interrupt.  In
    general, the only functions which may be safely called from within an
    interrupt are the small list of Exec functions documented in the
    "Exec Interrupts" chapter.

  * Some font format changes (old format supported).

  * Private format of .font files has changed (use FixFonts to create).

  * Default fonts may be larger, proportional.

  * Topaz is now sans-serif.

  * Any size font will be created via scaling as long as TextAttr.Flags
    FPF_DESIGNED bit is not set.  If you were asking for an extreme size,
    like size 1 to get smallest available, or 999 to get largest
    available, you will get a big (or very very small) surprise now.

  * Do not use -1 for TextAttr.Flags or Styles, nor as the flags for
    AvailFonts (one high bit now causes AvailFonts to return different
    structures).  Only set what you know you want.  A kludge has been
    added to the OS to protect applications which currently pass -1 for
    AvailFonts flags.

  * Many more commands are now built-in (no longer in C:).  This can
    break installation scripts that copy C:commandname, and programs that
    try to Lock() or Open() C:commandname to check for the command's
    existence.

  * The limit of 20 CLI processes is gone and the DOSBase CLI table has
    changed to accommodate this.  Under V36 and higher, you should use
    new 2.0 functions rather than accessing the CLI table directly.

  * Shell windows now have close gadgets.  The EOF character is passed
    for the close gadget of a Shell.  This is -1L with CON: getchar(),
    and the Close Gadget raw event ESC seq with RAW:.

  * Shells now use the simple-refresh character-mapped console.

  * By default, CON: now opens SIMPLE_REFRESH windows using the V36/V37
    console character mapped mode.  Because of some differences between
    character mapped consoles, and SMART_REFRESH non-mapped consoles,
    this may cause incompatibilities with some applications.  For
    example, the Amiga private sequences to set left/top offset, and set
    line/page length behave differently in character mapped console
    windows.  The only known work-around is to recompile asking for a
    CON: (or RAW:) window using the SMART flag.

  * Simple refresh/character mapped console windows now support
    highlighting and copying text with the mouse.  This feature, as well
    as pasting text should be transparent to programs which use CON: for
    console input, and output.  Pasted text will appear in your input
    stream as if the user had typed it.

  * While CONCLIP (see s:startup-sequence) is running, programs may
    receive "0 v" in their input stream indicating the user wants to
    paste text from the clipboard.  This shouldn't cause any problems for
    programs which parse correctly (however we know that it does; the
    most common problems are outputting the sequence, or confusing it
    with another sequence like that for FKEY 1 which is "0~").

  * The console.device now renders a ghosted cursor in inactive console
    windows (both SMART_REFRESH, and SIMPLE_REFRESH with character maps).
    Therefore, rendering over the console's cursor with graphics.library
    calls can trash the cursor; if you must do this, first turn off the
    cursor.

  * Some degree of unofficial support has been put in for programs which
    use SMART_REFRESH console windows, and use graphics.library calls
    mixed with console.device sequences to scroll, draw text, clear, etc.
    This is not supported in SIMPLE_REFRESH windows with character maps,
    and is strongly discouraged in all cases.

  * Closing an Intuition window before closing the attached
    console.device will now crash or hang the machine.

  * Under 1.2 and 1.3, vacated portions of a console window (e.g., areas
    vacated because of a clear, or a scroll) were filled in with the
    character cell color.  As of V36 this is no longer true, vacated
    areas are filled in with the global background color which can be set
    using the SGR sequence "[>##m" where ## is a value between 0-7.
    In order to set the background color under Release 2, send the SGR to
    set background color, and a form feed to clear the screen.

  * Note that SIMPLE_REFRESH character mapped consoles are immediately
    redrawn with the global background color when changed--this is not
    possible with SMART_REFRESH windows.

 Task Switching 
 Intuition Gadgets and Window Borders 
 Workbench and Startup 

The 1.3 Kickstart contained two task switching bugs.  After an interrupt,
a task could lose the CPU to another equal priority task, even if the
first task's time was not up.  The second bug allowed a task whose time
was up to hold on to the CPU either forever, or until a higher priority
task was scheduled.  Two busy-waiting tasks at high priority would never
share the CPU.  Because the input.device runs at priority 20, usually the
effect of these bugs was masked out for low priority tasks.  The
ExecBase->Quantum field had little effect because of the bugs.

For 2.0, a task runs until either its Quantum is up, or a higher priority
task preempts it.  When the Quantum time is up, the task will now lose the
CPU.  The Quantum was set to 16/60 second for 1.3, and 4/60 second for 2.0.

In general, the 2.0 change makes the system more efficient by eliminating
unnecessary task switches on interrupt-busy systems (for example, during
serial input).  However, the change has caused problems for some programs
that use two tasks of equal priority, one busy-waiting and one Wait()ing
on events such as serial input.  Previously, each incoming serial
character interrupt would cause task context switch allowing the
event-handling task to run immediately.  Under 2.0 the two tasks share the
processor fairly.

If 2.0 Intuition finds a gadget whose hit area (Gadget LeftEdge/TopEdge/
Width/Height) is substantially inside the border, it will be treated as
though it was declared in the border.  This is called "bordersniffing".
Gadgets declared as being in the border or detected by Intuition as being
in the border are refreshed each time after the border is refreshed, and
thus aren't clobbered.

Noteworthy special cases:

 1) A gadget that has several pixels not in the border is not
    bordersniffed. An example would be an 18-pixel high gadget in the
    bottom border of a SIZEBBOTTOM window.  About half the gadget will be
    clobbered by the border rendering.

 2) A gadget that is not substantially in the border but has imagery that
    extends into the border cannot be sniffed out by Intuition.

 3) A gadget that is substantially in the border but has imagery that
    extends into the main part of the window will be sniffed out as a
    border gadget, and this could change the refreshing results.  A
    common trick to put imagery in a window is to put a 1x1 or 0x0 dummy
    gadget at window location (0,0) and attach the window imagery to it.
    To support this, Intuition will never bordersniff gadgets of size 1x1
    or smaller.

    All these cases can be fixed by setting the appropriate
    GACT_xxxBORDER gadget Activation flag.

 4) In rare cases, buttons rendered with Border structures and JAM1 text
    may appear invisible under Release 2.

The height of the window's title bar is affected by the current font
settings.  See the discussion of "Screen Attributes" in the "Intuition
Screens" chapter. To predict your window's titlebar height before you call
OpenWindow():

    topborder = screen->WBorTop + screen->Font->ta_YSize + 1

The screen's font may not legally be changed after a screen is opened.

Be sure the screen cannot go away on you.  This is true if:

    1) You opened the screen yourself.

    2) You currently have a window open on the screen.

    3) You currently hold a lock on this screen (see LockPubScreen()).

IntuiText rendered into a window (either through PrintIText() or as a
gadget's GadgetText) defaults to the Window RastPort font, but can be
overridden using its ITextFont field.  Text rendered with the Text()
function appears in the Window RastPort font.

The Window's RPort's font shown above is the initial font that Intuition
sets for you in your window's RastPort. It is legal to change that
subsequently with SetFont().

Under 1.3, the Workbench Screen and initial Shell (CLI) opened before the
first line in s:startup-sequence.  Some naughty programmers, in an attempt
to recover memory, would search for the bitplane pointers and appropriate
the memory for their own use.  This behavior is unsafe.

By default 2.0 opens the initial CLI on the first _output_ from the
s:startup-sequence.  This allows screen modes and other parameters to be
set before the user sees the screen.  However, this broke so many programs
that we put in the "silent-startup" hack.  A disk installed with 1.3
install opens the screen as before.  A disk installed under 2.0 opens
silently. Never steal the Workbench bitplanes.  You don't know where they
are, how big they are, what format they may be in, or even if they are
allocated.  Recovering the memory is a bit tricky.

Under 2.0, simply avoid any output from your s:startup-sequence.  If your
program opens a screen it will be the first screen the user ever sees.
Note that if ENDCLI is ever hit, the screen will pop open.

Under 1.3, after ENDCLI, use CloseWorkbench() to close the screen.  This
also works under 2.0.  Loop on CloseWorkbench() with a delay between
loops.  Continue looping until CloseWorkbench() succeeds or too much time
has passed.  Note that a new program called EndRun is available for
starting non-returning programs from the startup-sequence.  EndRun will
reduce memory fragmentation and will close Workbench if it is open.
EndRun.lzh will be available in Commodore's Amiga listings area on BIX.