Repository navigation
fix(classic): Windows freeze, glass, caption, and silent death - #86
Draft
StephenSHorton wants to merge 23 commits into
Draft
StephenSHorton wants to merge 23 commits into
StephenSHorton wants to merge 23 commits into
Conversation
Cap CSI 2026 syncBuf, OSC 8 capture, and OSC 99 accumulation so agent floods cannot grow without bound. Drain inBuf in 256KiB UI slices, retry Mac job-queue drops instead of wedging bytesMsg, and treat an in-flight ConPTY Write as hot so ResizePseudoConsole cannot overlap it. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Real-machine logs on PR #86 showed a freeze with flat ~41 MB RSS and a live WM_CLOSE path, so memory was not the hang. The Win32 present path used a device bitmap plus a 25 Hz blink/rain flood; on mixed NVIDIA+AMD adapters that can stall DWM while the process still handles close. Switch the backbuffer to a system-memory DIB, coalesce blinks, cap rain to ~14 fps, auto-disable ambient after slow paints, and add a UI-thread watchdog (plus SUZURI_SAFE_MODE / frame-timing) so the next freeze dumps goroutine stacks. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
The watchdog dump on 7b0eb25 showed the UI thread spinning in WM_MOUSELEAVE → trackCaptionHover → TrackMouseEvent. The leave handler cleared captionLeaveTrk and then called trackCaptionHover(-1,-1), which re-armed TME_LEAVE. With the cursor already outside the client — or over HTCAPTION on the custom title bar — Windows posts another WM_MOUSELEAVE immediately. That loop starves paint and input while WM_CLOSE still works. This is the same title-bar hover code that shipped in 0.9.160 / 0.9.161. Leave now only clears hover; TrackMouseEvent is armed from client moves and refused for negative coordinates. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
…otices The custom title strip still sat under a DWM caption because the first WM_NCCALCSIZE ran before the HWND was registered, glass-off cleared the frame extend, and caption buttons returned HTCLIENT. Re-apply SWP_FRAMECHANGED after register, keep a 1px DWM extend for the shadow, round Win11 corners, and return HTMAXBUTTON/HTMINBUTTON/HTCLOSE so Snap Layouts and system caption clicks work. Leave handlers still never re-arm TrackMouseEvent. Rain was capped at ~14 fps from the dual-GPU guess. Tie it back to the 25 fps blink clock and only auto-disable ambient on genuinely slow presents (150ms×3 or 400ms). Notice cards were a 2× bitmap shown too large on Stephen's desktop. Size the layered window from monitor DPI, host font, and a 36% window cap, then StretchBlt 1:1 into that dest. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
Default-view rain was stuck on the input-only paint path after the first frames (skipAmbientFrame turned rain ticks into bar-only blits, then markShellDirty only every 240ms). Settings is a dim modal so it always full-painted the matrix. Drive rain from a display-paced frame loop capped at 60 fps, and never take the input-only shortcut while ambient or Settings is up. DWM has no blur radius, so the slider only flipped Mica vs Acrylic and looked dead. Map 0 / 1–39 / 40–80 to Mica / Acrylic / Tabbed and push an acrylic frost alpha on every step so empty chroma-key cells actually show a changing backdrop through the custom caption. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
9f5b3a2 stacked ACCENT_ENABLE_ACRYLICBLURBEHIND on Mica/Tabbed with AccentFlags=0, so DWM ignored the packed tint and washed the holes with the Windows accent color. Win11 now uses SYSTEMBACKDROP only; Win10 uses Flags=2 and a theme-void AABBGGRR tint. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
HostBackdrop Mica/Acrylic/Tabbed jumped at 40 (Tabbed is much darker) and went solid — often light gray — when inactive. Use one Flags=2 accent-acrylic for every blur step, frost alpha 20–196, re-apply on WM_ACTIVATE/WM_NCACTIVATE, and pin caption/border to the theme void. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Win11 rounded corners came from DWMWCP_ROUND in applyWindowChromeFrame. Force DWMWCP_DONOTROUND on create, glass toggle, and activate so DWM cannot put the rounding back. Shadow and custom caption stay. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Min/max/close were a thin GDI stamp plus a half-step/theme shadow copy. Draw each icon once with the X's stroke and a soft coverage fringe, DPI-scale the 46dip buttons, and skip NCUAH caption redraw so DWM cannot stamp a second glyph. Hover and press still lighten/darken. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
The doubled min/max/close were DWM caption sprites (HTMIN/MAX/CLOSE plus TITLEBARINFOEX). Those hit codes now map to HTCLIENT and DWM's button rects are marked invisible. Bell/coffee/min/max/close all rasterize from one Lucide-style 24px stroke set with the X's weight and AA. Strip cells no longer emit 🔔/☕. One paint per slot per frame. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
de25fe7 died ~31s after launch with no panic, no WER, no crash file. WM_NCACTIVATE/WM_ACTIVATE were calling DwmExtend/SetAccent on the same stack; those APIs re-enter WndProc and can overflow the native callback. Unfocused full paints then jumped 20-40x (rain + grok-fork + SetPixel icons + DWM) before the process vanished. Never touch DWM composition from NCACTIVATE; post a coalesced glass refresh; abort nested WndProc past depth 12; blit caption ink as one DIB; do not toast from WM_PAINT; log GDI/USER counts; write every exit reason to the crash trail (os.Exit, VEH, console-ctrl, WM_QUIT). Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
cursor
Bot
changed the title
A grok-fork build flood while alt-tabbed ingested 256KiB chunks and InvalidateRect'd every one. Those full GDI presents (rain + alt-screen + glass) jumped to 150-376ms; the child/ConPTY exit did not take the host down. Unfocused ingest still drains so the pipe does not block, but presents are capped at 10fps and rain drops while PTY is hot. Crash trail: VEH+UEF write code/addr/module (+ Go stack except on overflow), WER LocalDumps under HKCU, 5s heartbeat (heap/GDI/handles/ paint ms), unclean-previous-session detect, log rotate at 2MiB, and a reason on WM_CLOSE/ENDSESSION/QUIT/DESTROY/os.Exit. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
VEH/UEF now record AV read/write, a native backtrace, and the Go stack. WER LocalDumps is registered for both suzuri.exe and the actual image name. SetCrashOutput is labeled as throw/fatal (not only panic). Warn/error log lines fsync. Recovered panics and WM_QUIT write the crash trail. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
SetDIBitsToDevice of a 32-bit caption stamp onto the sysmem DIB backbuffer marked it alpha-aware. DWM color-key holes then became opaque black and every later GDI call on that DC jumped from ~10ms to 150–3692ms. Restore SetPixel for caption ink, blit images as 24-bit, and keep default-bg / theme-void cells as glass holes. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
DwmExtend top=-1 gave DWM a caption band, so native min/max/close drew on top of ours even after remapping those hits to HTCLIENT. Match WT: cyTopHeight=0 (sides/bottom still sheet-of-glass), NCCALCSIZE eats the caption, never DwmDefWindowProc. HTMAXBUTTON stays for Snap Layouts; min/close stay HTCLIENT. Caption ink is a 24-bit DIB BitBlt. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
d2c7da8 set DwmExtend top=0 so DWM would not stamp caption sprites. Accent-acrylic color-key needs -1 on every side; BLACK_BRUSH holes then went solid black even at an idle prompt. Extend is a full sheet again. WS_SYSMENU/MINIMIZEBOX/MAXIMIZEBOX come off so DWM has no caption buttons; we still draw ours and handle SC_MINIMIZE/MAXIMIZE/RESTORE/CLOSE. WS_CAPTION|THICKFRAME stay. The paint-stall rain guard no longer assigns ShellAmbient=none, so persistWindowPlacement cannot write that into config.json. Toast: "rain paused for this session". Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
…icons The input panel used the shell's 4px left inset, so glass showed as a strip beside the dark bar. Flush the panel (and a leftover bottom hairline) to the client edges; text still has its inner pad. Caption/bell/coffee glyphs were a 2/5 strip box (~10px at 96 DPI) — too small for the thick stroke. Same stroke and AA, box +25% so all five match Windows Terminal's optical size. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Two stacked causes: (1) SWP_FRAMECHANGED after DwmExtend dropped the sheet-of-glass, so BLACK_BRUSH in the client was opaque black and only a frame sliver showed the desktop. (2) frost mapped blur 36 → alpha 99 on a black HC void tint — accent acrylic reads as a solid dark slab. Extend after FRAMECHANGED and log S_OK. Remap frost so blur 0 is clear, blur 36 stays under 64, and the slider frosts smoothly. Lift #000 void in GradientColor RGB so the color key stays the holes, not the tint. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
63a9568 logged extend_ok and frost=50 but the desktop never showed: HostBackdrop was forced to None and accent acrylic does not compose after WS_SYSMENU is stripped. Restore the pre-PR86 Mica/Acrylic backdrop, keep accent off on Win11 when veil=0, and present a 24-bit backbuffer so RGB(0,0,0) stays the color key. NCCALCSIZE no longer calls DefWindowProc (that mixed caption insets with a 7px padded frame and made height alternate 1399/1392). Top stays put; only L/R/bottom take the resize border. SUZURI_GLASS_DEBUG=1 logs a sample void pixel plus accent/extend return codes. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Title-bar drag sends WA_CLICKACTIVE then ENTERSIZEMOVE. The posted wmSuzuriGlassRefresh was dispatched inside the modal drag loop, pushGlassBackdrop stalled ~2s, and the second drag AV'd in ntdll. Do not PostMessage or call DwmExtend/SetAccent/HostBackdrop while inSizeMove. WA_CLICKACTIVE and NCACTIVATE during a drag only set a deferred flag; flush after EXITSIZEMOVE (posted, not on that stack). Keep the WndProc depth abort and crash trail. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Win11 HostBackdrop (veil=0) ignores Blur — Desktop Acrylic has no point radius. Show the row as idle with 'Windows controls frost' and ignore left/right. Rim is macOS-only on GDI, so it shows 'macOS only'. Blur stays live on macOS, Windows 10, and Win11 when Veil > 0 (accent acrylic, frost follows the slider). Drag-crash deferral is already on this branch (eea6262). Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
Grok leaves a braille spinner in the OSC title after a turn. busy() treated that leftover as in-progress work, so the pane/tab glyph never stopped. ConPTY resize already ignored titleBusy for the same reason. The strip spinner now follows an explicit agent report, a running Warp-bar/foreground command, or recent PTY output — not a stale title frame. Idle Grok waiting for input stays quiet. Co-authored-by: Stephen Horton <StephenSHorton@users.noreply.github.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Head is
8a211fe.Idle Grok spinner
The pane/tab busy glyph spun for every Grok chat, including idle ones waiting for input.
Cause: Grok leaves a braille spinner in the OSC window title after a turn.
busy()treated that leftovertitleBusyflag as in-progress work, so the strip never stopped. ConPTY resize already ignoredtitleBusy(“Grok spinners never clear”).Fix: The spinner follows an explicit agent report (
working), a running Warp-bar/foreground command, or recent PTY output. A stale title frame is not work.Also on this PR
bac1d53: disable no-op Win11 Blur / Rim sliders in Settings.eea6262: never run DWM glass mid title-bar drag.c67fd3f: HostBackdrop glass + 24-bit color-key + 7px resize.Leave PR #85 as-is.