SBCL's stdin during --load is the load file, NOT the terminal.
When uiop:run-program creates a subprocess, it inherits this
stdin, so 'stty size' reads from the load file and fails.
:input :interactive opens /dev/tty for the child's stdin,
matching the behavior of 'stty size' from an interactive shell.
stty size via subprocess is the most reliable method — it
returns the correct 59x83 from the user's terminal. Move it
before ioctl to ensure it's tried first.
uiop:run-program may redirect the child's stdin, preventing stty
from querying the terminal. 'stty size < /dev/tty' explicitly
reads from the controlling terminal regardless of stdin setup.
stty size via uiop:run-program is the most reliable method —
it works from the shell on every Unix system and bypasses
alien/ioctl quirks. Placed between stdout ioctl and /dev/tty
ioctl in the fallback chain.
uiop:run-program can inherit different terminal state than the
interactive shell. Opening /dev/tty directly and calling ioctl
on that fd is equivalent to what the shell's stty does, and
works regardless of SBCL's fd inheritance quirks.
SBCL unconditionally strips COLUMNS and LINES from the environment,
so posix-getenv always returns nil for those names. stty size is
the reliable cross-platform fallback for terminal dimensions.
- stty size returns 'rows cols'. Old code set only one dimension
when both env vars were missing; new code sets both.
- Added tput cols/lines as final env-var fallback for systems
where COLUMNS/LINES are not exported and stty is unavailable.
- Added 'export COLUMNS LINES' to the passepartout script so
SBCL can read them from the environment.
sb-posix is a built-in SBCL contrib, available via require.
Without it, sb-posix:sigwinch causes a reader error and the
eval-when for the SIGWINCH handler never executes, making
*terminal-resized-p* always nil and resize detection broken.
ioctl on stdout's fd can return 80x24 even when the terminal is
larger. Add COLUMNS/LINES from the shell as a fallback. Also adds
ioctl-based sizing to simple-backend (was hardcoded 80x24).
The try-ioctl function returns (values cols rows) only when both
are valid integers > 0. or propagates complete pairs. This avoids
the nil-in-h crash from partial ioctl results.
ignore-errors + ioctl can return (values 80 nil) when the fd exists
but isn't a terminal. or propagates partial values, causing nil in
w or h. Wrap with multiple-value-bind + when to filter.
The user's terminal reports 186x60 via stty (which uses stdin fd)
and via COLUMNS/LINES, but ioctl on stdout's fd returns 80x24.
Priority: fd 0 → backend output fd → env vars → 80x24 fallback.
The previous logic (check ioctl result, prefer env when 80x24)
added complexity and crashes. Simple or with env vars after ioctl
is safe: ioctl returns 80x24 on stdout fd mismatch, env vars
(COLUMNS/LINES from shell) provide the correct initial size.
ioctl on stdout's fd can return the default 80x24 even when the
terminal is much larger (fd mismatch). The new logic:
1. Try ioctl — if it returns >80x24, trust it (correct at runtime).
2. If ioctl returned 80x24 (suspicious default), try COLUMNS/LINES
from the shell environment instead.
3. If both fail, return whatever ioctl gave us (80x24).
This fixes initial sizing on terminals where ioctl disagrees with
the real TTY size, without breaking runtime SIGWINCH resize
(which always re-queries ioctl, and that is correct after resize).
ioctl on stdout's fd can disagree with the real terminal size when
the process is started with stdout redirected or in some terminal
multiplexer configurations. / are set by every POSIX
shell at process start and reflect the actual terminal dimensions.
Priority: ioctl → env vars → 80x24 fallback.
This covers both initial sizing and dynamic SIGWINCH-driven resize.
The CSI 18t query leaks into the threaded keyboard reader because
the response arrives on stdin after the reader thread starts. The
response bytes get queued as key events and inserted as text into
the TUI input buffer. Removing the query entirely — ioctl is
sufficient for terminal size detection on all modern terminals.
The blocking read-char in %query-terminal-size could hang if the
terminal doesn't respond to CSI 18 t. Wrapped in
sb-ext:with-timeout 0.3 to abort if no response.
%query-terminal-size uses blocking read-char on an fd 0 stream
to read the terminal's response to \033[18t. This works even when
unix-simple-poll on fd 0 returns NIL (unlike read-char-no-hang).
Added as fallback in both modern and simple backends.
The earlier fix that checked unix-simple-poll before unix-read
broke input: poll on fd 0 always returns NIL in this SBCL, so
read-raw-byte always returned nil. Reverted to original: call
unix-simple-poll (for timeout) then unix-read unconditionally.
unix-read blocks until data arrives, which is correct for a TUI.
The %query-terminal-size function sent \033[18t and tried to read
the response via read-char-no-hang on fd 0, which always returns nil
in this SBCL environment. The response leaked into user input,
displaying garbled CSI sequences. Rely on ioctl only.
- Disabled \033[?u kitty keyboard protocol in modern-backend
(converts all keys to escape sequences, breaking Ctrl+letter dispatch)
- Fixed parse-csi-sequence: use multiple-value-bind instead of let*
with destructuring-bind (lost secondary return value from read-param)
- Fixed parse-csi-params format string: pass char-code of terminator
as distinct argument for ~d, keeping the character for ~C
- Added %query-terminal-size in classes.lisp: ANSI CSI 18t fallback
for terminal size detection when ioctl fails or returns zero
The original code called unix-simple-poll then unconditionally
called unix-read, ignoring the poll result. When poll returned
nil (no data), unix-read would block indefinitely. Fixed by
checking poll result: only read if poll says data is ready.
Same pattern as the draw-text array fix. Application code may call
backend-clear with a framebuffer array instead of a backend instance.
The array method clears all cells to default blank state.
- Replace make-alien unsigned-char buffer with make-array + vector-sap
to avoid SBCL alien type mismatch between signed-char and unsigned-char
- Convert timeout seconds to fixnum milliseconds for unix-simple-poll
(was passing float 0.1, broke on fixnum-typed sb-unix:to-msec)
- Both fixes make read-raw-byte work on SBCL 2.5.2.debian
Application code (passepartout TUI) calls draw-text with a framebuffer
(2D array) as the first argument, but draw-text only had methods for
framebuffer-backend CLOS instances. Added a method on array that sets
cells directly on the framebuffer array, matching make-framebuffer's
return type.
with-terminal macro was only in tangled .lisp (not .org). suspend-backend
and resume-backend generics + simple-backend methods + tests were also
in hand-edited .lisp only. All three added to org/backend-protocol.org
with proper prose, following the literate programming discipline.
Also added suspend/resume assertions to simple-backend-lifecycle test suite.
- Add %read-digits to read multi-digit parameters from raw terminal bytes
- Add %parse-sgr-mouse to decode ESC[<Cb;Cx;CyM/m SGR mouse sequences
into mouse-event structs with :press/:release type and :left/:middle/
:right/:scroll-up/:scroll-down/:drag button classification
- Modified parse-csi-sequence to detect the < marker (0x3C) and
delegate to %parse-sgr-mouse instead of treating it as key input
- Coordinates converted from 1-based (terminal protocol) to 0-based
(framebuffer convention)
- All 12 test suites pass at 100% (461 checks, no regressions)
- Org source (text-input.org) updated as the source of truth
Bug fixes:
- Fix OSC8 format strings (backslash escape layering) in modern-backend.org
- Test format string had single backslash instead of double, causing
unclosed CL string that cascaded through 3 subsequent test forms
- Implementation format string had leading escaped quote (not a string
opener) and triple-backslash ending (also not a string terminator)
- Fix missing closing parens in border-char-rounded and border-char-double tests
- Fix ASDF input-tests pathname (file lives in tests/, not src/components/)
New features:
- Implement suspend-backend / resume-backend protocol methods
- modern-backend: exit/enter alt screen, re-enable mouse/kitty/bracketed-paste
- simple-backend: no-ops (no terminal state to preserve)
Infrastructure:
- Update test suite to cover suspend/resume (backend + modern-backend suites)
- 454 checks, 100% pass across 14 test suites
Add assert to reject invalid mode keywords on first registration
instead of silently storing them and only crashing later in
slot-render's ecase. Valid modes: :stack, :replace, :single-winner.
The osc8-link implementation and its test both had doubled
backslashes (\\ -> \\) in their format strings, producing two
literal backslashes at runtime instead of the single backslash
needed for the OSC 8 string terminator (ST = ESC \).
Fix: change \\ to \\ in both the implementation and test format
strings. The tangled .lisp files now have correct escaped
backslashes (\) producing one backslash in the runtime string.
Additionally clean up a patch artifact that left a stray backslash
before the opening quote.
Add :mode parameter to defslot with three behaviors:
- :stack (default) — accumulate all registrations, render in order
- :replace — each registration replaces previous entries
- :single-winner — first registration wins, rest ignored
Mode is set on first defslot call and frozen for subsequent calls
to prevent conflicting mode specifications from different plugins.
Store slot data as plist (:mode <keyword> :entries <list>) instead
of bare entries list.
Add 5 new tests covering mode-specific behavior. All 9 slot tests
pass. All 13 suites pass at 100%.
- Create org/integration-tests.org (15 blocks, per-test prose)
- Add Markdown tests section to org/markdown-renderer.org (11 test blocks)
- Delete deprecated src/components/input-tests.lisp stub
- Update README.org: tree diagram, literate programming section,
development commands, remove stale test counts
All 13 test suites pass at 100%. Zero .lisp files without org origin.
The combined org file had no unique content — all prose and code were
already in scrollbox.org, tabbar.org, and container-package.org. The
old file's code blocks had the pre-bugfix render/draw-scrollbars
versions and all had :tangle no.
Also update README.org and ARCHITECTURE.org references from
scrollbox-tabbar.org to the individual org files.
Distribute the literate prose from the old combined scrollbox-tabbar.org
into three individual module org files:
- scrollbox.org: ScrollBox class, render, scrollbars, bug fixes,
plus the combined test suite (tangles scrollbox-tabbar-tests.lisp)
- tabbar.org: TabBar class, navigation, keyboard handler, render
- container-package.org: Package definition and exports
The old scrollbox-tabbar.org is retained as a documentation archive
with all code blocks set to :tangle no and a redirecting note.
Fixes the draw-scrollbars code block to use the post-bugfix version
(with layout-node origin offset ox/oy), matching the working code.
All 13 test suites pass at 100%.
The tangled handle-text-input used (key-event-text event) for character
insertion, but the test suite creates key events with :code not :text.
Restored the original handle-text-input which uses
(code-char (key-event-code event)) — matching the test expectations.
