Untouched open Tk bug tickets: Reproduced

Triage of 548 open Bug tickets: those with no comment or check-in from me on 2026-09-24, plus 38 added on 2026-10-04 (filed since then, Pending then, or reopened; 16 of them filed by me), plus 68 added on 2026-10-22 (all other open or Pending Bug tickets not filed by me: the 30 created before 8.5, 31 I had already worked on, and 7 filed or missed since 10-04). Reproduced on Linux/X11 under Xvfb (main 9.1 tip, 9.0, 8.7, 8.6 tip, 8.5, 8.4) and on a Windows 11 VM (9.1, 8.6, 8.5, 8.4). Rows marked patch link to a prototype patch (150 in all, in tk-triage/prototypes). Updated 2026-11-11 with what happened since: the emoji before each ticket id shows its progress. On 2026-10-24 every prototype for a still-open ticket was verified again on current main, on Linux and Windows, and given a test where practical; on 2026-10-25 seven more reproduced tickets got prototypes.

Reproduced 182

✅354fbe9fe62026Windows: -appearance set with wm attributes is lost when the wrapper is recreatedREPRODUCED (9.1 main); FIXED on branch nab-win-appearance-rebuild

Open ticket 354fbe9fe6 · 68. Win Window Operations · filed 2026-10-21

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-21 check-in 075ec1ae00 on nab-win-appearance-rebuild, in main, 9.0: Reference ticket 354fbe9fe6 in tests winWm-10.5 and 10.6.
  • 🔧 2026-11-05 check-in ca38b6462f on trunk, in main, 9.0: Fix for 354fbe9fe6: Windows: -appearance set with wm attributes is lost when the wrapper is recreated
  • 💬 2026-10-21 nab: Fix proposed on branch [nab-win-appearance-rebuild]: d89119f114, tests winWm-10.5 and 10.6 (075ec1ae00). Verified on Windows 11: both tests fail without the fix and pass with it.
  • 💬 2026-10-22 serhiy.storchaka: Confirmed on Windows 11: on main, a dark appearance forced with wm attributes -appearance dark is lost after wm resizable, wm overrideredirect, -toolwindow, -fullscreen on/off and wm transient on a mapped window. On nab-win-appearance-rebuild it is kept in all five cases, and wm.test and winWm.test pass.
  • 💬 2026-11-05 nemethi: Many thanks to Nicolas for the fix and to Serhiy for the review! Merged into main by commit ca38b646. No merge into core-9-0-branch or core-8-6-branch is planned.
Full verdict

REPRODUCED (9.1 main); FIXED on branch nab-win-appearance-rebuild

  • Platform Windows
  • Tested not on Linux (DWM-only). Confirmed that main still has the early return TCL_OK for WM_OVERRIDE_APPEARANCE in SynchronizeAppearance.
  • Cause win/tkWinWm.c:SynchronizeAppearance returns early when the appearance is forced, so the AppearanceIdleProc that UpdateWrapper schedules never applies DWMWA_USE_IMMERSIVE_DARK_MODE to the new HWND.
  • Prototype none (already on branch nab-win-appearance-rebuild: d89119f114 fix, 075ec1ae00 tests winWm-10.5/10.6, by nab, 2026-10-21). Repro: win/354fbe9fe6.tcl (checks dark/light after wm resizable, overrideredirect, -toolwindow, -fullscreen, wm transient)
  • Suggested ticket comment The fix on [nab-win-appearance-rebuild] looks right: main still returns early from SynchronizeAppearance() for a forced appearance. It can be merged once checked on Windows; win/354fbe9fe6.tcl covers the other recreation paths too.
Windows VM, run 6
  • Platform Windows
  • Verdict REPRODUCED (9.1 main); FIXED on branch nab-win-appearance-rebuild
  • Tested With wm attributes -appearance dark, winfo isdark is 0 after each of wm resizable, wm overrideredirect, -toolwindow, -fullscreen on/off and wm transient (5 of 5); light passes. Side finding: winfo isdark on a window that was never mapped (e.g. a withdrawn .) crashes Tk on Windows: TkpWindowIsDark (win/tkWinWm.c) calls Tk_GetHWND(Tk_WindowId(tkwin)) without creating the window. The repro script now probes a mapped toplevel. Branch nab-win-appearance-rebuild: all 10 checks PASS (BUG NOT PRESENT); winWm.test and wm.test pass (340 tests, 0 failed).
  • Prototype none
  • Suggested ticket comment Confirmed on main on Windows 11: the forced dark appearance is lost on all five wrapper re-creations. Separately, [winfo isdark] crashes on a window that has no HWND yet.
💬8f2fdde1d22026ttk::treeview: setting a padding on a tag makes focus ring visible againpatchREPRODUCED (main)

Open ticket 8f2fdde1d2 · 88. Themed Tk · filed 2026-10-19

💬 Since triage (updated 2026-11-11)

Status now: Open

  • 💬 2026-10-20 bohagan: A custom style isn’t needed. From the docs all you need is: ttk::style configure Row -focusthickness 0 ttk::style map Row -focusthickness {} The padding issue is a ttk issue since 9.0. See the last few comments in [forum](https://core.tcl-lang.org/tips/forumpost/caa6535943).
Full verdict

REPRODUCED (main)

  • Platform generic (seen on X11, default theme; 9.1 only, because the Row layout gained Treeitem.focus with TIP 740)
  • Tested main/X11 with Xvfb screenshots. With ttk::style configure Row -focusthickness 0; ttk::style map Row -focusthickness {} and a tag with -padding {10 6}, the focused padded row shows a 1px black ring. An unpadded focused row shows none. 9.0 is not affected (its Row layout has no focus element).
  • Cause generic/ttk/ttkElements.c:DrawFocusRing raises thickness 0 to 1 when -focussolid is set, and all themes set -focussolid 1 for Row (the win theme uses a dotted ring, which ignores the thickness entirely). So the focus element always draws a 1px ring. Without padding, the Treeitem.row background inside it paints over the ring. A tag's -padding insets that background (Treeitem.padding) and the ring shows. FocusElementSize already reserves 0 pixels for thickness 0, so the drawing does not match the sizing.
  • Prototype 8f2fdde1d2.patch (34 lines; FocusElementDraw draws nothing when the scaled -focusthickness is < 1, and the clamp in DrawFocusRing is removed). Verified: the ring disappears for the padded row, and the default ring (focus → thickness 1) is still drawn. treeview, treetags, ttk, checkbutton, radiobutton, toggleswitch, entry, notebook and combobox tests pass. No new test (it would need pixel inspection).
  • Suggested ticket comment The focus element drew a 1px ring even with -focusthickness 0 (DrawFocusRing raised the thickness to 1 for solid rings). Normally the row background covers that ring, but tag -padding insets the background. Proposed fix: draw no ring when -focusthickness is 0, which matches the element's size computation. Separately, a tag's -padding does not grow the row height, so the padded rows' text is clipped (see the forum thread bohagan mentions). That is a different issue.
Prototype verified 2026-10-24, Linux
  • Status VERIFIED
  • Applies to main clean
  • Repro the focused row with tag -padding {10 6} and Row -focusthickness 0, checked with a PIL screen grab of the row's top edge. Before: 196 dark pixels. After: 0.
  • Test none (only pixels change; core Tk on X11 has no photo -format window)
  • Test file run treeview.test 1552/0, ttk.test 113/0, treetags.test 32/0, notebook.test 64/0 (same as main)
  • Patch changes none. Review: FocusElementDraw is the only caller of DrawFocusRing, and every theme maps Row -focusthickness to 1 when focused, so default rings are still drawn. With thickness 0, the dotted (non-solid) ring is now skipped too, which matches FocusElementSize reserving 0 px.
Prototype verified 2026-10-24, Windows VM, generic patches
✅59d6d5f6c02026Proposed fix for dancing selection in mixed RTL displaypatchREPRODUCED (main fb166f9f73, X11, bidi). Not applicable to 9.0 and 8.6: they have no bidi layout, so the text stays in left-to-right logical order …

Open ticket 59d6d5f6c0 · 18. [text] · filed 2026-10-14

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-30 check-in 0e34e33c21 on text-bidi-chunk-order, not merged: Fix f5b47c533f (see 59d6d5f6c0): a tag (e.g. the selection) splitting right-to-left text in a text widget made the line scrambled: the font back-end draws a
  • 🔧 2026-10-31 check-in a4124a7246 on text-bidi-chunk-order, not merged: Fix the insertion cursor at the start of a line of right-to-left text (see 59d6d5f6c0): it was drawn at the left end of the line. ReorderBidiChunks() now also
  • 🔧 2026-10-31 check-in 3403347da9 on trunk, in main, 9.0: Fix cursor insertion bug in RTL text and selection bug in mixed-bidi text 59d6d5f6c0; restore broken RTL text display on Windows; implement ::tk::startOfClust
  • 🔧 2026-11-01 check-in 8a2c92b32d on nab-text-mark-stretch, in main, 9.0: Fix text shimmering while a selection is dragged on macOS (see 59d6d5f6c0): since 3403347da9 every mark, the insertion cursor included, ended the base-char
  • 🔧 2026-11-02 check-in 4cb0789f36 on text-rtl-highlight, not merged: Text widget: the highlight of a part of a joined right-to-left word (e.g. one Arabic letter selected) did not match its bbox (see 59d6d5f6c0): the chunks of a
  • 🔧 2026-10-31 check-in 793f5d5a8c on dancing_bidi, in main, 9.0: Merge trunk
  • 💬 2026-10-15 kevin_walzer: We are not able to accept patches written in Python. The code should be written in C against Tk's C API (if that is the correct place to target) or in Tcl.
  • 💬 2026-10-16 anonymous: Tcl patch attached to the issue.
  • 💬 2026-10-16 serhiy.storchaka: Reproduced on X11 (main, --enable-bidi). Without tags the right-to-left text is placed correctly, but a tag which splits it, the sel tag included, scrambles the layout: the chunks are placed from left to right in logical order, while each of them is drawn from right to left. [code] So each change of the selection moves the text under the pointer, and the selection jumps while dragging. The same problem was fixed i…
  • 💬 2026-10-17 Arnault01: Thank you for the thorough testing and analysis — very helpful. You are right that my patch addresses a different (and much smaller) issue: it only keeps selection boundaries out of clusters of combining marks, which stabilizes dragging in pointed Hebrew (nikkud/teamim) but does nothing for plain RTL text. The reproducer shows the actual root cause is in the display engine: when a tag — including the "sel" tag — s…
  • 💬 2026-10-17 Arnault01: Proposed patch for point 2 below.
  • 💬 2026-10-17 Arnault01: Proposed patch for point 2 below.
  • 💬 2026-10-28 Arnault01: I have forked the github tcltk repo to work on the 2 fixes (plus one more I encountered during tests). Here is a description of the problems: # Tk text widget — "dancing selection" with RTL text: submission drafts Three independent tickets for the Tk bug tracker (core.tcl-lang.org, Tk — Text widget). Each patch is self-contained; Tickets 1 and 2 were validated in headless CI (Xvfb, Tcl 9 + Tk trunk, `--enable-bidi…
  • 💬 2026-10-28 kevin_walzer: Please provide a succinct, non-AI summary of each patch - what it does and what platform it applies to. No new Tcl commands will be considered as part of a bugfix - let’s address the “dancing” behavior. I will try to review these soon after reviewing your briefer summary.
  • 💬 2026-10-28 kevin_walzer: Thank you for the patches. I applied the three patches and tried the simple demo outlined in the ticket description - create a text widget on X11 and insert Hebrew: "שלום עליכם, זהו טקסט מימין לשמאל עם מספר מילים." I then tried to select the text, and what I saw was a serious regression: instead of a jumpy/"dancing" selection highlight, the glyphs themselves vanished and were redrawn in LTR order, and were dragged…
  • 💬 2026-10-29 kevin_walzer: I closed the "dancing_bidi" branch because I was not getting anywhere with it. Starting over with trunk.
  • 💬 2026-10-30 kevin_walzer: I re-opened the [dancing_bidi] branch and have what I believe are satisfactory results with selection and cursor tracking in RTL text. User-tested on the widget demo on X11; still need to test on macOS and Windows. Please review.
  • 💬 2026-10-30 Arnault01: Closed because three separate and dependant tickets have been created
  • 💬 2026-10-30 kevin_walzer: I am working from this ticket.
  • 💬 2026-10-30 Arnault01: Hello Kevin, Can you drive me about how to perform the code review ?
  • 💬 2026-10-30 Arnault01: Following up on the remaining mixed-bidi issues on dancing_bidi: I found the root cause, and a small patch fixes it. Cause. The maxPrefixWidth clamp in CharChunkMeasureChars() keeps the value from the layout (the width of the whole line). After layout, when CharBboxProc() measures a single character, the clamp still runs and rewrites the correct result into lineWidth + charWidth. So on mixed-direction lines [bbox]…
  • 💬 2026-10-30 Arnault01: Here is my review: Overall The branch replaces my original chunk-reordering approach with a cleaner architecture: glyphs are never moved — the font backend keeps full responsibility for drawing — while the generic layer only records where each chunk lies (for backgrounds, hit-testing, bboxes). This is the right division of labor and fixes the regression seen with my first patch. Coding guidelines Mostly good Tcl/T…
  • 💬 2026-10-30 serhiy.storchaka: I tried branch dancing_bidi (dec0422697) on X11 (Xvfb, --enable-bidi), alone and with dancing-mixed-fix.patch. A tag splitting RTL text no longer moves it, but plain left-to-right text regresses; trunk (d005489a6b, the base of the branch) passes all these tests. With the patch, textTag.test passes, but: * textDisp.test crashes after textDisp-24.20, in AdjustForTab() (tkTextDisp.c line 10333, reading ciPtr->cha…
  • 💬 2026-10-30 kevin_walzer: Serhiy or Arnault, can you set up a new branch from main and these patches apply? I was doung my own work in dancing_bidi and things may have gotten mixes up. The current status of dancing_bidi is that all visual artifacts are corrected on macOS/X11 but not yet on Windows.
  • 💬 2026-10-30 kevin_walzer: I went ahead and applied the three patches on macOS to a clean update of trunk and they do NOT solve the visual shimmering issue with pure RTL text. So I am going to continue my work on the dancing_bidi branch. The outstanding issue on Windows is visual artifacts with tags - the 'sel' tag is applied in a mirror image, so when a highlight is drawn RTL, text color changes LTR within that tag boundary.
  • 💬 2026-10-30 serhiy.storchaka: Kevin, branch text-bidi-chunk-order is based on main: it places the chunks of each stretch in visual order and measures them from their logical positions, without the maxPrefixWidth clamp. On X11 it fixes the dancing with pure and mixed RTL text, and all text tests pass. On Windows there are no regressions, but a tag still moves RTL text, because of the measuring in f3d964112f. I could not test it on macOS. Unli…
  • 💬 2026-10-30 kevin_walzer: I have applied all of Arnault's patches in the dancing_bidi, and there are two remaining bugs: 1. On all platforms, the caret/cursor is inset one glyph at the start of a paragraph of RTL text - you cannot insert it at the right margin. 2. On Windows, mix bidi text shimmers when selected - the 'sel' tag appears to be rendered in mirror relation to the cursor and text, so RTL text will draw a LTR selection bar. I ha…
  • 💬 2026-10-31 anonymous: Hello Kevin, I'm working on trying to resolve both bugs. So far I cannot reproduce bug 1. "insert the caret at the right margin". Can you provide a more detailed step list or, better, a test script ? It is possible that since the comment it has been fixed in the tip of the dancing_bidi branch (core.tcl-lang.org).
  • 💬 2026-10-31 anonymous: ltr-chunk-isrtl.patch is the proposed patch for bug 2.
  • 💬 2026-10-31 anonymous: I just added a patch for the orphan nikkud issue.
  • 💬 2026-10-31 kevin_walzer: To reproduce, run the bidi text widget demo and click the mouse at the start of the Hebrew clock on the right hand side.
  • 💬 2026-10-31 serhiy.storchaka: Reproduced on X11, also on trunk. On text-bidi-chunk-order the click gives the correct index (2.0, trunk gives 2.1), but the cursor was drawn at the left end of the line: the insertion cursor before the first chunk of a line was not moved with the reordered chunks. Fixed on text-bidi-chunk-order in a4124a7246.
  • 💬 2026-10-31 kevin_walzer: Serhiy, can you please review the dancing_bidi branch? RTL display is currently broken on trunk on Windows, and the dancing_bidi branch addresses that successfully. The function you are patching has been completely removed in my branch, and reinstating it restores the broken RTL display on Windows. We need to address the insertion cursor issue across platforms, the shimmering 'sel' issue on Windows, and retain cor…
  • 💬 2026-10-31 kevin_walzer: Just attached a screenshot that shows RTL on trunk on Windows.
  • 💬 2026-10-31 kevin_walzer: Fix for these issues committed in 3403347d; behavior is correct on X11, Windows and macOS. No new failures in textDisp.test. Thanks to all who reported these issues and assisted with patch submission.
  • 💬 2026-11-01 Arnault01: Hello Kevin, For bug 1. To reproduce, run the bidi text widget demo and click the mouse at the start of the Hebrew clock on the right hand side. Can you elaborate as to: - What is the observed behavior - What should be the correct behavior Meanwhile I start to investigate.
  • 💬 2026-11-01 Arnault01: With my AI assistant, here is what I can say about bug 1: With recent step list, caret can be placed at the right margin of the hebrew paragraph and clicks on the right position correctly the insertion at the beginning of the line (tested with X11 and Windows). So it's still "can't reproduce actually on the tip of the dancing_bidi branch". I'll attach an headless test that leads me to that conclusion.
  • 💬 2026-11-01 kevin_walzer: This is all fixed. Thank you!
  • 💬 2026-11-02 serhiy.storchaka: On trunk, the highlight of a part of a joined Arabic word does not match its bbox (X11 and Windows). In السعر (TkDefaultFont), selecting ع (bbox width 7) highlights 2 pixels, and ل (bbox width 4) highlights 10 pixels, shifted left. The chunks are measured during the layout with the string cut at the end of each chunk, so the last letter of a chunk takes its final form. Fixed on text-rtl-highlight in 4cb0789f36: …
  • 💬 2026-11-02 nab: Tested on macOS (aqua, TkDefaultFont): - 4cb0789f36(/info/4cb0789f36) fixes it on macOS too. With U+0639 of your example selected, the highlight is 3 px wide on trunk (bbox 7 px); with U+0644 it is 8 px wide and starts 5 px left of its bbox (3 px). On the branch both match their bbox. The text*.test files give the same results as trunk. textDisp-37.4 also passes on trunk on macOS. - On trunk and on the branch,…
  • 💬 2026-11-02 Arnault01: Thank you for the validation :) 2 questions: 1. Can I have a summary, detailing what is left to be fixed if any ? 2. Can I have a release date integrating this branch ? I'm eager to use it :)
Full verdict

REPRODUCED (main fb166f9f73, X11, bidi). Not applicable to 9.0 and 8.6: they have no bidi layout, so the text stays in left-to-right logical order and the selection does not move.

  • Platform generic (bidi build, 9.1)
  • Tested serhiy.storchaka's bbox script. On main, x of 1.0/1.4/1.8 goes from 949 914 877 to 684 725 953 after tag add sel 1.4 1.8, so the text moves. On 9.0.5 and 8.6.18 it stays 3 46 79.
  • Cause generic/tkTextDisp.c (classic text widget). When a tag splits an RTL display line, the chunks are placed left to right in logical order and only each chunk is drawn right to left. The fix e7c85cd941 (bd243152d2, merged as 371cf58a3e on 2026-10-14) is only on the revised_text / tip-466 branch, not on main.
  • Prototype none. Porting the chunk reordering from 371cf58a3e is about 1500 changed lines in tkTextDisp.c, tkFont.c, tkText.h and others, too big for a prototype. The submitter's Tcl patch (cluster snapping) is a separate issue, and as noted it should use tk::startOfCluster/endOfCluster.
  • Suggested ticket comment Confirmed on main. The fix is to port the chunk reordering of bd243152d2 (revised_text) to the classic tkTextDisp.c. The cluster-snapping patch should go into its own ticket and use tk::startOfCluster/tk::endOfCluster.
✅897f6672f72026Windows: wm deiconify restores a maximized window as normalREPRODUCED (own ticket, filed 2026-10-03); proposed fix in branch win-deiconify-zoomed, not merged

Open ticket 897f6672f7 · 68. Win Window Operations · filed 2026-10-03

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-07 check-in eec7513de8 on trunk, in main, 9.0: Fix for 897f6672f7: Windows: wm deiconify restores a maximized window as normal.
  • 🔧 2026-10-07 check-in 5309adeaea on core-9-0-branch, in main, 9.0, 8.6: Fix for 897f6672f7: Windows: wm deiconify restores a maximized window as normal.
  • 🔧 2026-10-07 check-in f7bee94164 on core-8-6-branch, in main, 9.0, 8.6: Fix for 897f6672f7: Windows: wm deiconify restores a maximized window as normal.
  • 💬 2026-10-07 nemethi: Serhiy, many thanks for the ticket and the fix! Merged into main, core-9-0-branch, and core-8-6-branch by commits eec7513d, 5309adea, and f7bee941.
Full verdict

REPRODUCED (own ticket, filed 2026-10-03); proposed fix in branch win-deiconify-zoomed, not merged

  • Platform see title
  • Tested filed after reproducing it; see the ticket
  • Prototype branch win-deiconify-zoomed
✅4c5184d9742026X11: wm deiconify loses the position of a window moved by the userREPRODUCED (own ticket, filed 2026-10-03); proposed fix in branch x11-withdraw-keep-position, not merged

Open ticket 4c5184d974 · 67. Unix Window Operations · filed 2026-10-03

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-03 check-in c889e90b31 on x11-withdraw-keep-position, in main, 9.0: Fix 4c5184d974: X11: wm deiconify loses the position of a window moved by the user. A withdrawn window is placed anew by the window manager unless its size hi
  • 🔧 2026-10-04 check-in 0df196cee1 on trunk, in main, 9.0: Fix 4c5184d974: X11: wm deiconify loses the position of a window moved by the user
  • 🔧 2026-10-04 check-in 71c626ee0e on core-9-0-branch, in main, 9.0, 8.6: Fix 4c5184d974: X11: wm deiconify loses the position of a window moved by the user
  • 🔧 2026-10-04 check-in b5565f3397 on core-8-6-branch, in main, 9.0, 8.6: Fix 4c5184d974: X11: wm deiconify loses the position of a window moved by the user
  • 💬 2026-10-03 serhiy.storchaka: Proposed fix in branch x11-withdraw-keep-position: when a window that has been mapped is withdrawn, Tk sets USPosition in its size hints, so the window manager keeps its position when it is mapped again. Tested with KWin, icewm, openbox and fvwm (fvwm ignores PPosition). If the screen gets smaller while the window is withdrawn, the window is now handled like a window positioned by wm geometry +x+y: icewm moves it …
  • 💬 2026-10-04 jan.nijtmans: Fixed in 0df196cee1, core-9-0-branch and core-8-6-branch Many thanks!
Full verdict

REPRODUCED (own ticket, filed 2026-10-03); proposed fix in branch x11-withdraw-keep-position, not merged

  • Platform see title
  • Tested filed after reproducing it; see the ticket
  • Prototype branch x11-withdraw-keep-position
✅3b2c972f292026X11 accessibility: a screen reader announces every widget twice when it gets the focuspatchREPRODUCED (main; 9.0 and 8.6 have no accessibility)

Open ticket 3b2c972f29 · (unused) · filed 2026-10-03

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 💬 2026-10-04 serhiy.storchaka: Confirmed with Orca 50.2. Orca reacts to the extra event because the toplevel accessible has no parent: Orca cannot find the window of the focused widget, drops its focus after the announcement, and takes the following active-descendant-changed as a new focus. Setting the parent alone also removes the doubling, and makes Orca's flat review work: [code] 01.patch is right nevertheless, so both changes are worth appl…
  • 💬 2026-10-05 kevin_walzer: This has been addressed in the alt_x11_accessibility_fixes branch - please confirm it works as expected.
  • 💬 2026-10-05 serhiy.storchaka: Confirmed in alt_x11_accessibility_fixes: each focus change now produces only one object:state-changed:focused, without active-descendant-changed. The toplevel accessible still has no parent, so the second change from my previous comment (atk_object_set_parent() for toplevels) is still needed for Orca's flat review.
  • 💬 2026-10-07 0xjnml: Confirmed with Orca 48.1 on a MATE desktop at 30d3e9a7b7: tabbing through the entry, check, radio and ttkbut demos of the widget demo, each control is announced once ("Dismiss button.", "Safety Check check box not checked."), where stock 9.1.0 announces each one twice. The script and the details are on 74d973b49d. As Serhiy noted, the window still has no parent: an AT-SPI client reads the parent of the demo's to…
  • 💬 2026-10-11 kevin_walzer: Merged fixes to trunk.
Full verdict

REPRODUCED (main; 9.0 and 8.6 have no accessibility)

  • Platform X11
  • Tested Orca 50.2 (not the reporter's 48.1) run under dbus-run-session on Xvfb :101 with a private HOME and XDG_RUNTIME_DIR. The speechd Python module was replaced by a stub, so the run was silent and everything Orca would speak was read from its --debug-file. The app was repro.tcl with a non-blocking focus tour (focus -force every 3 s). On main every widget is spoken twice: 'Button' 'button.' 'Save.' x2, 'Entry' 'entry.' x2, ... An AT-SPI event dump (python3-gi Atspi) shows the reported active-descendant-changed after every focus event. Both 01.patch and my one-line prototype give a single announcement each.
  • Cause the real trigger is in unix/tkUnixAccessibility.c:RegisterToplevelWindow(). The toplevel accessible never gets a parent: atk_object_set_parent() is only called for non-toplevels, and the root lists the toplevels only through tk_get_n_children/tk_ref_child. Orca's log shows "[window] claims to have no parent" and "None is top-level object for [button]", so Orca sets its active window to None. Right after presenting the focus it clears its locus of focus ("Focus [button] is not in None"). The active-descendant-changed then emitted on the parent (tk_grab_focus, RegisterWidgetRecursive, UpdateAtkFocusChain, TkAtkAccessible_FocusHandler) sets the locus again, and Orca presents the widget a second time. With a parent on the toplevel, Orca ignores the active-descendant event as "existing locus of focus". 01.patch is still sensible, since the signal is meant for item containers, but it only hides the missing parent. The same missing parent also leaves Orca's flat review with no window, and the script falls back to "default".
  • Prototype 3b2c972f29.patch (17 lines, 1 code line: atk_object_set_parent(accessible, tk_root_accessible) for toplevels; no test, Tk has no AT-SPI test harness). With it Orca 50.2 announces each focus change once and uses the "Tk Application" script.
  • Suggested ticket comment With Orca 50.2 the second announcement comes from the toplevel accessible having no parent. Orca cannot find the window of the focus, drops its locus of focus after announcing, and the following active-descendant-changed sets it again. Setting the toplevel's parent to the application object in RegisterToplevelWindow() alone removes the doubling (01.patch also does). Proposed fix in branch <name>.
✅324059e3d02026X11 accessibility: checkbuttons have role "unknown" (the role table has "Checkbox", accessibility.tcl sets "Checkbutton")patchREPRODUCED (main; 9.0 and 8.6 have no accessibility)

Open ticket 324059e3d0 · (unused) · filed 2026-10-03

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 💬 2026-10-17 kevin_walzer: Fixed and merged.
Full verdict

REPRODUCED (main; 9.0 and 8.6 have no accessibility)

  • Platform X11
  • Tested AT-SPI tree of repro.tcl via python3-gi Atspi (private bus, Xvfb :101). Main shows [unknown] name='Checkbutton', and with 03.patch [check box]. On X11 with a reader running, ttk::checkbutton is aliased to tk::checkbutton (accessibility.tcl), and the class fallback in GetAtkRoleForWidget finds no "Checkbutton" entry either. Also checked:
    • After the checkbutton is turned off, ATK still reports CHECKED, because the value becomes "not selected".
    • Two radiobuttons sharing one variable are both reported CHECKED on main.
    • tk_dialog with the Dialog mapping added shows [dialog].
  • Cause unix/tkUnixAccessibility.c roleMap lacks "Checkbutton" and "Dialog", the roles set by library/accessibility.tcl and documented in doc/accessible.n. Two related defects:
    • tk_ref_state_set() treats any value whose first character is not '0' as checked, and _announce_button_state/_updateselection store "selected"/"not selected".
    • The Radiobutton/TRadiobutton <Map> bindings store [%W cget -variable], the variable name, as the value, so every radiobutton starts as checked.
  • Prototype 324059e3d0.patch (30 lines: 03.patch plus {"Dialog", ATK_ROLE_DIALOG} and "not selected" treated as unchecked; no test). The tree shows [check box], CHECKED is cleared after unchecking, and the dialog is [dialog]. The radiobutton value is not fixed there.
  • Suggested ticket comment Confirmed; 03.patch is right, and "Dialog" needs an entry too (ATK_ROLE_DIALOG). Once the role is fixed, the checked state is wrong after unchecking ("not selected" counts as checked in tk_ref_state_set), and all radiobuttons start checked because their <Map> binding stores the variable name as the value.
✅87bd2b899a2026accessibility.tcl: the class defaults overwrite accessible names set by the application on every map (_init’s "only once" guard never fires)REPRODUCED (main; 9.0 and 8.6 have no accessibility)

Open ticket 87bd2b899a · (unused) · filed 2026-10-03

✅ Since triage (updated 2026-11-11)

Status now: Closed (Wont Fix)

    Full verdict

    REPRODUCED (main; 9.0 and 8.6 have no accessibility)

    • Platform generic (library/accessibility.tcl); tested on X11
    • Tested "repro.tcl -names" on build-trunk with a pgrep stand-in reporting "orca" (dbus-run-session, Xvfb :101).
      • Caveat: with the repro's fixed wm geometry . 360x240 the extra "Open" button is not mapped under Xvfb's default fonts (winfo ismapped 0), so the output wrongly shows the name preserved. With 360x400 it gives exactly the reported "after map: name=Button / description=Open".
      • With 04.patch: name "Open file" kept after map and re-map, notebook value still follows the tab (General -> Advanced).
      • The <Destroy> cleanup in 04.patch handles paths with glob characters (.b*, .c[x]).
    • Cause library/accessibility.tcl:_init returns only if get_acc_role fails and its error message equals the role, which never happens. It then unconditionally calls set_acc_role/name/description/value/state/action, and runs again on every <Map> (and on <<NotebookTabChanged>>).
    • Prototype none (04.patch works as described). Review notes:
      • It depends on bind all <Destroy> {+...}. An application's bind all <Destroy> without "+" silently removes the cleanup. generic/tkAccessibility.c already has a C WindowDestroyHandler that frees the attribute table, and this bookkeeping could live there.
      • The -takefocus 1 in _init stays (see 595d72d0a6).
    • Suggested ticket comment Confirmed on main once the extra button fits in the window (with 360x240 it is not mapped here, so the repro's "after map" line shows the old name). 04.patch fixes it and keeps the notebook refresh. The bookkeeping could be freed by the existing C destroy handler instead of bind all <Destroy>.
    ✅74d973b49d2026X11 accessibility: widgets are named by their class ("Button, button, Save"); X11 reads the attributes differently from Windows and macOSREPRODUCED (main; 9.0 and 8.6 have no accessibility)

    Open ticket 74d973b49d · (unused) · filed 2026-10-03

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-10-03 kevin_walzer: You refer to a "atk-util-get-root branch" - I do not see any indication of that. Are you able to incorporate your suggested changes in your multiple tickets into a single branch that I can review?
    • 💬 2026-10-04 0xjnml: Thanks for looking at these. I can't create branches in the Fossil repository, so the attached x11-accessibility-combined.patch has all of the changes in one patch (patch -p1 from the Tk source root). It applies to trunk, which has the same two files as 9.1.0, and also applies unchanged on top of the atk-util-get-root branch. That branch is Serhiy's fix for ticket b9d85fc100, two check-ins of 2026-09-25: https://c…
    • 💬 2026-10-04 kevin_walzer: I don't quite understand why, but this large patch winds up causing significant regressions with ATK - to the point where the only narration in Tk's widget demo comes from the manual speech from spd-say. So I don't think I can use this in its current form. To move the review forward, I have committed a separate branch, [alt_x11_accessibility_fixes], that does the following: 1. Tk now narrates toplevel windows. 2. …
    • 💬 2026-10-07 0xjnml: You are right about the regression, and it was my mistake. I reproduced it with Orca 48.1 on a MATE desktop (Debian 13): wish built from the 9.1.0 tarball plus x11-accessibility-combined.patch, running the widget demo through the attached demo-tour.tcl, which opens the entry1, check, radio and ttkbut demos and moves the focus through them with tk_focusNext (23 focus changes). Orca said nothing at all. Its debug lo…
    • 💬 2026-10-07 kevin_walzer: I thought I merged the branch with the toplevel parent patch. I can’t tell from your report if that is working or not? In my testing toplevel window titles are narrated once. I did not accept every suggestion you made. I am leaving the widget naming as is. Have you tried deferring your custom names until after the widgets are mapped? That should override the default values.
    • 💬 2026-10-08 kevin_walzer: I found Serhiy's patch - it was just one line - and applied it, and now the toplevel title is narrated only once. To me, narrating the toplevel name is sufficient - I chose "Tk Application" because "tk appname" most often defaults to "wish.$majorversion.$minorversion," and I was not fond of that. So my inclination is to leave that as is.
    • 💬 2026-10-08 kevin_walzer: Can this branch be merged?
    • 💬 2026-10-11 kevin_walzer: As OP has not responded, merging updates to trunk and calling this ticket fixed.
    Full verdict

    REPRODUCED (main; 9.0 and 8.6 have no accessibility)

    • Platform X11 (the attribute semantics are generic)
    • Tested AT-SPI tree of repro.tcl on main, identical to the ticket's ([button] name='Button' description='Save', [entry] name='Entry' description='Entry'). Orca 50.2 (silent speechd stub) speaks 'Button' 'button.' 'Save.' and 'Entry' 'entry.'. With 02.patch the tree is [button] name='Save', [entry] name='', application 'hold.tcl', and Orca says 'Save' 'button.' and 'entry.'. Code check: win/tkWinAccessibility.c get_accName -> TkAccDescription and accHelp -> "help"; macosx accessibilityLabel -> "description"; unix GetAtkNameForWidget -> "name".
    • Cause unix/tkUnixAccessibility.c:GetAtkNameForWidget()/tk_get_description() use "name"/"description", while the class bindings in library/accessibility.tcl put the role word into "name" and the text into "description". The Windows and macOS bridges read "description" as the label. Note that doc/accessible.n documents the opposite: name is the "Short label identifying the widget", description a "Longer string".
    • Prototype none (design decision for Kevin; 02.patch exists). Review of 02.patch:
      • The windowingsystem ne "x11" guard covers only 10 of the 15 help bindings. TScrollbar, Menubutton, TMenubutton, TNotebook and Text still set help on X11, which with item 2 becomes the ATK description that Orca reads on focus.
      • It aligns X11 with the Win/mac code but against doc/accessible.n. The alternative is to put the text into "name" in the class bindings and have Win/mac prefer "name".
      • TNotebook stays named "Notebook" and Treeview is named by its column list.
      • The [tk appname] root name touches the same code as atk-util-get-root df6cd07ca4.
      • tk_get_name() leaks the g_utf8_make_valid() string on every call (as before the patch).
    • Suggested ticket comment (reply to kevin_walzer): The branch is atk-util-get-root, check-in df6cd07ca4, the fix for b9d85fc100 (the root object is registered through AtkUtil instead of overriding atk_get_root()). I can collect the X11 parts of these tickets (role table, toplevel parent, speech) in one branch on top of it. Which attribute is the label (name or description) needs deciding first, since doc/accessible.n says name while the Windows and macOS bridges use description.
    ✅48b39019bd2026X11 accessibility: Tk speaks typed keys and state changes itself, synchronously, on top of the screen readerpatch speechd-guardpatchREPRODUCED (main; 9.0 and 8.6 have no accessibility)

    Open ticket 48b39019bd · (unused) · filed 2026-10-03

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-10-03 kevin_walzer: In developing the accessibility module, I never saw any direct narration of the keypress by Orca in Tk's widgets; that is why we are manually calling spd-say. Is there a setting in Orca that can allow it to narrate the keys as they are pressed in, say, Tk's text widget as opposed as into the console/terminal? On macOS and Windows, these keypresses and words are narrated by the custom announcement/event change API'…
    • 💬 2026-10-04 0xjnml: Orca has two kinds of echo, both on its Echo preferences page: - Key echo ("Enable key echo", on by default) speaks each key as it is pressed. Orca gets the key events from AT-SPI itself, not from the widget, so it works in Tk windows as in any other. In our session with Orca 48.1 and its default settings, it said "tab" on every Tab press in a Tk 9.1.0 window, before announcing the newly focused control. - Echo by…
    • 💬 2026-10-04 kevin_walzer: Implementing a full text interface was considered but rejected. The accessibility API is tightly connected to the Tk_Window design, and non-window data like the text widget contents or listbox rows adds too much complexity at the C level because there is no natural mapping. Using Tk’s normal event process to dispatch this data to custom announcements is cleaner. Perhaps we can look at moving these calls to libspee…
    • 💬 2026-10-05 kevin_walzer: The update committed to the alt_x11_accessibility_fixes branch replaces the shelling out to spd-say with directing invocation through libspeechd, which runs in the same process and should improve performance. This will also mean that the call to [tk accessible check_screenreader] only runs on Tk startup, as it is designed to do.
    • 💬 2026-10-05 serhiy.storchaka: In alt_x11_accessibility_fixes, configure on X11 now stops with an error if the development files of ATK/Atk-bridge or of Speech Dispatcher are not installed. Both were optional before. This breaks building Tk on FreeBSD, NetBSD, OpenBSD and DragonFly, where they are not installed by default (configure stops at the ATK check). Please keep both optional: build without accessibility if ATK is not found, and without …
    • 💬 2026-10-05 kevin_walzer: Accessibility is critical and has been omitted from Tk for too long. We wouldn’t make Xlib an optional dependency. Why should accessibility be optional?
    • 💬 2026-10-05 serhiy.storchaka: Most of Tk's dependencies are optional, including Xft and bidi. Accessibility is important, but not critical: Tk was without it for decades. I am afraid that making it required will harm portability and slow down adoption of Tk 9.1.
    • 💬 2026-10-05 kevin_walzer: Bidi must be explicitly disabled - it is the default. I am willing to do the same with Atk and libspeechd. Would that work?
    • 💬 2026-10-05 serhiy.storchaka: Bidi can be disabled, and since 1dfc71ae4a it is disabled if not detected. I think the same should be for ATK and libspeechd.
    • 💬 2026-10-06 kevin_walzer: I have made accessibility a configure switch via --enable-accessibility, which will be the default. Please review.
    • 💬 2026-10-06 serhiy.storchaka: Thanks, configure works now. But if ATK is found and libspeechd is not, the build fails: tkUnixAccessibility.c includes libspeechd.h and calls spd_*() without checking HAVE_SPEECHD. The attached patch adds the missing #ifdef HAVE_SPEECHD; without it Tk just does not speak. With it, the branch builds with ATK and without libspeechd on Linux, FreeBSD, NetBSD, OpenBSD and DragonFly, and still builds with libspeechd.
    • 💬 2026-10-06 kevin_walzer: What systems would have ATK installed and not libspeechd? Without speech, the screen reader experience is seriously degraded, and Tk would erroneously report that accessibility was enabled, leading to all kinds of bug reports. What is the use case for this patch?
    • 💬 2026-10-06 serhiy.storchaka: ATK comes with GTK, so it is installed almost everywhere. The libspeechd development files are usually not, unless installed on purpose. Without libspeechd, accessibility still works: Orca gets roles, names, states and focus through AT-SPI and speaks them itself. Only Tk's own announcements are lost. A missing optional library should not make the build fail. If that is a concern, configure can report that speech i…
    • 💬 2026-10-06 kevin_walzer: But Tk’s own speech is a significant part of the design. It is not an optional feature. Why is silent failure of this feature at runtime preferable to noisy failure at build time, with a message to install the required components?
    • 💬 2026-10-06 kevin_walzer: The more fundamental point is that Unix users are accustomed to installing a complex chain of dependencies, their ports/packaging systems will help, and the entire accessibility stack needs to be installed for accessibility to work correctly. I am unwilling to make libspeechd optional. How do we address this? I am inclined to go back to making it all a hard dependency again, but I am open to suggestions.
    • 💬 2026-10-06 kevin_walzer: Solution: libspeechd will be an entire requirement at build time for —enable-accessibility to work, and will fall back to —disable-accessibility if not found.
    • 💬 2026-10-06 serhiy.storchaka: That is fine with me: an error if --enable-accessibility is explicitly specified and ATK or libspeechd is not found, and falling back to --disable-accessibility if it is not specified and they are not found. This solves the build failure, and the attached patch is then not needed.
    • 💬 2026-10-07 kevin_walzer: I have updated configure.ac with these changes. Please test the build. If it works for you, I plan to merge this branch to trunk.
    • 💬 2026-10-07 serhiy.storchaka: Thanks. I tested 30d3e9a7b7 on Linux in six configurations: without options, with --enable-accessibility and with --disable-accessibility, each with or without ATK or libspeechd. All of them configure and build, and wish starts. When ATK and libspeechd are found, Tk is linked with both; otherwise accessibility is disabled with a warning. But --enable-accessibility given explicitly also falls back with a warning …
    • 💬 2026-10-07 kevin_walzer: Since this appears to work correctly in the various scenarios you tested, I will merge this to trunk in its present form soon. Thank you for testing.
    • 💬 2026-10-11 kevin_walzer: Merging fixes to trunk.
    Full verdict

    REPRODUCED (main; 9.0 and 8.6 have no accessibility)

    • Platform X11
    • Tested:
      • spd-say/espeak stand-ins that log and sleep 0.5 s: focusing the entry 508 ms, two keys 1022 ms, toggle 525 ms. Sent: "entry blank", "a", "b", "Enable selected".
      • Answer to Kevin's question: Orca 50.2 with default settings (org.gnome.Orca key-echo = true; character-echo and word-echo false) on an X11 session, with keys typed into the ttk::entry by XTEST on Xvfb :101, speaks 'a' and 'b' itself (Orca debug log). Tk sent the same 'a' and 'b' to spd-say, so the user hears every key twice.
      • Orca's key echo comes from the keyboard events, not from the widget, so it works in any toolkit. Character and word echo need text-changed events and a real AtkText, and Tk's AtkText is a stub (get_text NULL, character count 0).
      • With the prototype, Orca still echoes each key once, Tk no longer sends keys, and the same steps take 2/1/6 ms.
    • Cause library/accessibility.tcl:
      • ::tk::accessible::speak runs exec spd-say synchronously from the event loop. That also blocks the AT-SPI bus, which Tk serves from the same loop: during the blocking tour, Orca's queries timed out and it processed the events seconds late.
      • It calls check_screenreader (popen "pgrep -x orca") every time.
      • _handle_keypress speaks every printable key on X11.
    • Prototype 48b39019bd.patch (38 lines; no test). speak runs exec spd-say -- $text ... & with the espeak fallback (dropping the useless quote escaping; "--" protects texts starting with "-"), and _handle_keypress no longer speaks on X11. Word, focus and state speech and the per-call check are left as they are.
    • Suggested ticket comment (reply to kevin_walzer): Orca's "key echo" is on by default and works from the keyboard events, independent of the widget. With Orca 50.2 on X11, keys typed into a ttk::entry are echoed by Orca and then again by Tk's spd-say. Character and word echo would need a real AtkText, which is currently a stub. Running spd-say in the background (exec ... &) at least stops the event loop, and with it the AT-SPI replies, from blocking.
    ✅595d72d0a62026accessibility.tcl: with a screen reader running, labels and other static widgets become Tab stopspatchREPRODUCED (main; 9.0 and 8.6 have no accessibility)

    Open ticket 595d72d0a6 · (unused) · filed 2026-10-03

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Wont Fix)

    • 💬 2026-10-03 kevin_walzer: This is by design. In standard Tk, labels do not take focus and it is not clear how the screen reader would capture their text without the -takefocus 1 flag set on X11. Atk's focus is tightly linked with Tk's focus.Are you saying that Orca can capture this text without explicit focus being set by Tk?
    • 💬 2026-10-04 0xjnml: Yes. Orca reads text that does not take focus with its flat review: it lays out the on-screen objects of the active window and reads them line by line (desktop layout: KP minus enters flat review, KP 8 reads the current line, KP 7 and KP 9 move up and down). That is how a label is read in GTK applications too, where labels are not Tab stops either. In Tk 9.1 on X11 this did not work at first. Orca read only "blank…
    • 💬 2026-10-04 kevin_walzer: How do you suggest we address the focus ring that is drawn when the widget has accessible focus? I want Orca to have feature parity with VoiceOver and NVDA, which draw these elements themselves.
    • 💬 2026-10-05 kevin_walzer: I disagree with this proposed change. It would make the behavior on X11 different from macOS and Windows. It is very important for Tk's keyboard navigation functionality to remain consistent.
    Full verdict

    REPRODUCED (main; 9.0 and 8.6 have no accessibility)

    • Platform generic in accessibility.tcl (not aqua); tested on X11
    • Tested:
      • Tab order of repro.tcl: with the reader stand-in .f.save .f.name .f.enable .f.status; without a reader .f.save .f.name .f.enable. The label's -takefocus becomes 1 after map.
      • Answer to Kevin's question: Orca 50.2 flat review, driven through Orca's D-Bus remote controller (ToggleFlatReviewMode, GoEnd, PresentLine, SayAll) while the focus stays on the Save button. With the toplevel parent fix, Orca enumerates the unfocused label ("4 zones", including [label 'Ready.']) but reads "empty", because Tk's AtkText reports 0 characters for every widget. With the prototype the label's text is exposed through AtkText, and GoEnd/PresentLine and SayAll speak "Ready." without the label ever taking focus.
      • Without the parent fix, Orca has no window and flat review is empty anyway.
    • Cause library/accessibility.tcl:_init runs $w configure -takefocus 1 for every widget it initialises, including Label/TLabel. unix/tkUnixAccessibility.c:tk_text_get_text()/tk_text_get_character_count() are stubs, so a reader's review commands get no text for static widgets. (The ATK state set already omits FOCUSABLE for labels, which matches leaving them out of the Tab order.)
    • Prototype 595d72d0a6.patch (142 lines; no test). _init no longer sets -takefocus for the Label role, and for labels AtkText returns the label text (get_text, character count, get_string_at_offset, character and range extents = the widget box). Tab order with a reader matches the order without one, and flat review reads the label (together with 3b2c972f29.patch). The other static classes (TProgressbar, Scrollbar, Canvas) are left as they are.
    • Suggested ticket comment (reply to kevin_walzer): Yes. Orca reads static text with flat review and "read window" by walking the accessible tree, not through the focus. With Orca 50.2 the unfocused label is found, but read as "empty" because Tk's AtkText returns no text. Exposing the label text through AtkText makes flat review say "Ready." with the focus left on the Save button, so -takefocus 1 is not needed for labels.
    ✅3f1ad1ff502026Windows: pointer warp sometimes lands one pixel shortREPRODUCED (own ticket, filed 2026-10-02); proposed fix in branch win-warp-rounding, not merged

    Open ticket 3f1ad1ff50 · 68. Win Window Operations · filed 2026-10-02

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-19 check-in 36a857169e on trunk, in main, 9.0: Fix 3f1ad1ff50: Windows: pointer warp could land one pixel short of the requested position
    • 🔧 2026-10-19 check-in 5ded332853 on core-9-0-branch, in main, 9.0, 8.6: Fix 3f1ad1ff50: Windows: pointer warp could land one pixel short of the requested position. Don't bother reserving a stub entry for TkpChangeFocus2
    • 🔧 2026-10-19 check-in 846fd91bf5 on core-8-6-branch, in main, 9.0, 8.6: Fix 3f1ad1ff50: Windows: pointer warp could land one pixel short of the requested position
    • 💬 2026-10-19 jan.nijtmans: Now merged to 36a85716, core-9-0-branch and core-8-6-branch
    Full verdict

    REPRODUCED (own ticket, filed 2026-10-02); proposed fix in branch win-warp-rounding, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch win-warp-rounding
    🟢deca13e8332026Tests fail at high display scaleREPRODUCED (own ticket, filed 2026-10-02); proposed fix in branch tests-high-dpi, not merged

    Open ticket deca13e833 · 86. Test Tools · filed 2026-10-02

    🟢 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-05 check-in d19371a616 on trunk, in main, 9.0: Selected fix from deca13e833: correct state left by tests: pack-20.6 did not restore "pack propagate ."; event.test leaves the pointer over the larger windows
    • 🔧 2026-10-05 check-in d83b2974b0 on core-9-0-branch, in main, 9.0: Selected fix from deca13e833: correct state left by tests: pack-20.6 did not restore "pack propagate ."; event.test leaves the pointer over the larger windows
    • 🔧 2026-10-06 check-in 25633400a6 on trunk, in main, 9.0: Selected fix from deca13e833: test bind-34.1: fit a 800x600 screen
    • 🔧 2026-10-06 check-in ef5aab8436 on core-9-0-branch, in main, 9.0: Selected fix from deca13e833: test bind-34.1: fit a 800x600 screen
    • 🔧 2026-11-07 check-in 672afed883 on trunk, in main, 9.0: Fix deca13e833: Tests fail at high display scale (tests did not take into account the minimum width of a toplevel).
    • 🔧 2026-11-07 check-in 21892ec6c4 on core-9-0-branch, in main, 9.0, 8.6: Fix deca13e833: Tests fail at high display scale (tests did not take into account the minimum width of a toplevel).
    • 🔧 2026-11-07 check-in d6121ae3be on core-8-6-branch, in 8.6: Fix deca13e833: Tests fail at high display scale (tests did not take into account the minimum width of a toplevel).
    • 🔧 2026-11-07 check-in 834172a0bd on tests-min-width, in main: Test textDisp-33.3 failed on Windows when the 480 pixel wide frame .f and the 80-column text widget were wider than the screen: the text widget was laid out a s
    • 🔧 2026-11-08 check-in a4b00e8746 on tests-high-dpi, not merged: Test font-10.9: use TkFixedFont on Windows.
    • 💬 2026-10-02 erikleunissen: Regarding the changes to test event-c3f5c85322: I agree that they are an improvement.
    • 💬 2026-10-02 erikleunissen: And I merged that particular change into trunk and core-9-0-branch, commits 7517b79368 and 6e42783e36 respectively.
    • 💬 2026-10-05 erikleunissen: I agree with: * State left by other tests: pack-20.6 did not restore pack propagate ., event.test leaves the pointer over the larger windows of focus.test. The latter change slightly adapted to be robust against any screen size. Merged into trunk and core-9-0-branch. Thanks!
    • 💬 2026-10-06 erikleunissen: Regarding: * bind-34.1 moves a window off a small screen, e.g. 800x600. Agreed and merged into trunk and core-9-0-branch.
    • 💬 2026-10-07 marc_culler: As far as I know, the macOS failures of frame-12.3 and textImage-4.2 only occur in 8.6. Have you seen macOS failures caused by screen resolution in 9.0 or 9.1?
    • 💬 2026-10-07 serhiy.storchaka: No, I do not test on macOS. I took frame-12.3 and textImage-4.2 from your report "test failures in 8.6" on tcl-core (1 October), and assumed the same cause, since both fail on X11 and Windows at high display scale. If they fail on macOS only in 8.6, the cause there is probably different.
    • 💬 2026-10-07 erikleunissen: General considerations regarding this ticket: * Thus far, I selected three changes from this ticket for merging. My reason for agreeing to them doesn't actually involve the scaling factor (which is a subject that I'm not very familiar with). * Requiring display scaling to work as intended is one thing (and merits its own tests). However, requiring all Tk tests to PASS at any or various setting for display scaling …
    • 💬 2026-10-07 serhiy.storchaka: Thank you for your review and for merging the fixes, Erik! Most modern HiDPI monitors use a display scale above 100%. If the tests passed only at 100%, we could not run them on such computers, and the tests we run would not test what is really used. Forcing 100% in tests/main.tcl would not help either. It can only set tk scaling, which converts points to pixels. That would fix the tests that use fonts and lengths …
    • 💬 2026-10-07 erikleunissen: Here is another idea: Presumption: the width of the titlebar is core to the issue. The width of the titlebar is determined by the elements that it contains, which are buttons (for closing, iconifying and maximizing the window) and the title. A less intrusive, generic measure could be to target the title: Currently the title of the main window is set to "tktest" here: [https://core.tcl-lang.org/tk/file?ci=eec7513de…
    • 💬 2026-10-07 erikleunissen: Correction: where I wrote "the width of the title bar", I meant: "the minimum width of the title bar"
    • 💬 2026-10-08 serhiy.storchaka: Thank you for the idea. I tested it on Windows 11: the title does not matter. The minimal width of a toplevel is the same for the title "tktest", an empty title, or a long title: 120 pixels at 100% and 404 pixels at 350%. It is determined by the frame and the buttons of the title bar. It applies to the root window and to any other toplevel, but not to toolwindows (wm attributes -toolwindow 1) and override-redirect…
    • 💬 2026-10-20 serhiy.storchaka: Branch tests-min-width tries the alternative to removeTitleBar: the windows are made wider (480 pixels), or the widgets are laid out in a frame which is not resized with the toplevel, so the expected results do not change. The workaround with wm overrideredirect in textDisp.test and text.test is no longer needed. An additional argument for this way: the same tests fail on the wayland branch, where a toplevel with …
    • 💬 2026-11-07 serhiy.storchaka: Branch tests-min-width is merged into trunk 672afed883 and backported to core-9-0-branch 21892ec6c4 and core-8-6-branch d6121ae3be. On the wayland branch the notWayland constraint was removed from the tests fixed by it 17d0ac78cf. Still failing on Windows at 350%: most of pack.test, 11 grid tests, and textDisp-27.7, textDisp-27.7.1 and textDisp-27.7.2.
    • 💬 2026-11-08 serhiy.storchaka: Branch tests-high-dpi is rebased on trunk: * 73f2b9cedd: the minimal width of Windows toplevels (removeTitleBar), without the parts superseded by 672afed883. * 4d3273ea61: fonts and lengths specified in points. * 48d0850ed7: the size of scrollbar arrows on Windows. * aa57e2a1bb: the default sizes in points set in tk.tcl. * a4b00e8746: font-10.9 uses TkFixedFont on Windows. The other changes are already…
    Full verdict

    REPRODUCED (own ticket, filed 2026-10-02); proposed fix in branch tests-high-dpi, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch tests-high-dpi
    💬e050c5a2a92026Aqua: drawing hangs if idle handlers keep scheduling idle handlersREPRODUCED (own ticket, filed 2026-10-02); no fix: branch aqua-bounded-idle-loops was closed as unwanted (it would mask the issue)

    Open ticket e050c5a2a9 · 66. Aqua Window Operations · filed 2026-10-02

    💬 Since triage (updated 2026-11-11)

    Status now: Open

    • 💬 2026-10-02 erikleunissen: Regarding this issue and the commit message for a0eb4e7a8c: "Aqua: do not hang if idle handlers keep scheduling idle handlers." Ticket 591829e948, part A, questions the practice of "idle handlers keep scheduling idle handlers" more fundamentally; at least for one of the instances that the present ticket also addresses. It shows for one of the entry points where such iterative or recursive processing of idle ta…
    • 💬 2026-10-02 marc_culler: Let's focus on generateExposeEvents in Tk 8.6. The loop in generateExposeEvents is a core feature of Tk 8.6, dating to the original Aqua port. It was a way of arranging for redraws of damaged parts of a window which was forced the strange limitations of drawRect. Now (meaning in 9.X) drawRect is no longer used, and that code has been removed from Tk entirely. That function is (ab)using Expose events in a special w…
    • 💬 2026-10-02 serhiy.storchaka: I have no macOS and cannot check it myself. Could you please run the reproducer from the description on macOS with 9.0 or main and check whether it hangs: after wm deiconify, when resizing the window live, or while the File menu is open? 9.x has similar loops (in updateLayer, generateExposeEvents: during a live resize, TkpWmSetState() and _runBackgroundLoop), so I suspect there may be the same issue there. Filteri…
    • 💬 2026-10-02 marc_culler: When I run this code interactivley in wish9.1 a window opens with disabled titlebar buttons, so it cannot be iconified or closed, and the counter increases steadily. The window can be moved with the mouse. The Quit button and the menus are not responsive. I never see the window getting iconified so it is not possible to deiconify it. When I run it as a script with wish9.1, no window appears at all. The menus appea…
    • 💬 2026-10-02 marc_culler: LiveResize is a completely different animal. I think that should be treated as an isolated issue. Getting it to work reliably (e.g with the "A text widget with embedded windows and other features" demo) was a big struggle.
    • 💬 2026-10-02 marc_culler: _runBackgroundLoop is also a completely different animal. That exists to allow timer tasks to get run while a menu is open. Tk does not process any Apple events while a menu is open. Apple manages the menu in a separate runloop. When a menu first opens, Aqua Tk starts a thread which processes timer tasks (only). Before that thread existed, for example, if you started the pendulum demo and then opened a menu the pe…
    • 💬 2026-10-02 marc_culler: I am not sure it makes sense to try to redesign Aqua Tk at this depth without having any access to a macOS computer. I think you may be putting too much trust in Claude's analytic capabilities when you do that.
    • 💬 2026-10-02 serhiy.storchaka: I agree that this exceeds Claude's capabilities: the proposed fix only kicked the can down the road. It looks to me that there still is a problem. Do you agree? Do you suggest closing the issue, or do you have any ideas how to solve the problem?
    • 💬 2026-10-02 serhiy.storchaka: I think the original fix for [1813595] was not wrong. It fixed that issue, but the hang can return if the application calls update idletasks (or the Aqua code runs an equivalent loop). That is a different issue, this one.
    • 💬 2026-10-03 marc_culler: I think it is complicated. Ideally, I think it should never be necessary to run one of those nested Tcl_DoOneEvent loops. Of course, there needs to be one of those loops, which handles all types of events, as well as idle tasks and timer tasks, and that one is the so-called mainloop. But I think the nested loops should be superfluous. However, every attempt to remove one of the remaining nested loops has broken so…
    Full verdict

    REPRODUCED (own ticket, filed 2026-10-02); no fix: branch aqua-bounded-idle-loops was closed as unwanted (it would mask the issue)

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype none
    ✅d08538fd782026Regression in 9.1: wrapped text counts spaces at end of line, splits words, breaks after leading spaces (X11 with bidi, Windows)REPRODUCED (own ticket, filed 2026-10-01); proposed fix in branch shaping-whole-words, not merged

    Open ticket d08538fd78 · 44. Generic Fonts · filed 2026-10-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-06 check-in 123cc66f98 on trunk, in main, 9.0: Fix d08538fd78: Regression in 9.1: wrapped text counts spaces at end of line, splits words, breaks after leading spaces (X11 with bidi, Windows)
    • 💬 2026-10-06 serhiy.storchaka: Merged into main 123cc66f98. Tested on X11 with bidi and on Windows. Tk 9.0 is not affected: it has neither text shaping on X11 nor Uniscribe on Windows.
    Full verdict

    REPRODUCED (own ticket, filed 2026-10-01); proposed fix in branch shaping-whole-words, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch shaping-whole-words
    ✅d1a711b5002026X11: wm geometry reports the requested position after the window manager moved the window out of a panel’s areaREPRODUCED (own ticket, filed 2026-09-30); proposed fix in branch x11-wm-position-work-area, not merged

    Open ticket d1a711b500 · 67. Unix Window Operations · filed 2026-09-30

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-11-04 check-in 8cbf5a82ae on trunk, in main, 9.0: Fix for d1a711b500: X11: wm geometry reports the requested position after the window manager moved the window out of a panel's area
    • 🔧 2026-11-04 check-in 2eca446735 on core-9-0-branch, in main, 9.0, 8.6: Fix for d1a711b500: X11: wm geometry reports the requested position after the window manager moved the window out of a panel's area
    • 🔧 2026-11-04 check-in 29451af554 on core-8-6-branch, in 8.6: Fix for d1a711b500: X11: wm geometry reports the requested position after the window manager moved the window out of a panel's area
    • 💬 2026-11-04 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits 8cbf5a82, 2eca4467, and 29451af5.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-30); proposed fix in branch x11-wm-position-work-area, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch x11-wm-position-work-area
    ✅65e299ace52026X11: a cascade menu near the right edge of the screen is posted over its parent menuREPRODUCED (own ticket, filed 2026-09-30); proposed fix in branch menu-cascade-flip, not merged

    Open ticket 65e299ace5 · 12. Unix Menus · filed 2026-09-30

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-11-04 nemethi: Merged by Jan into main and core-9-0-branch by commits 1b7b8535 and d4f89751. Backported to core-8-6-branch by commit b07e4dde. Closing the ticket.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-30); proposed fix in branch menu-cascade-flip, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch menu-cascade-flip
    ✅e8d4c8a5772026X11 (HarfBuzz/SheenBidi font backend): [font configure] of a font in use leaks the old fontpatchREPRODUCED (main 40.3 KB per font change; 9.0.5 and 8.6.18 0.0 KB, they use tkUnixRFont.c)

    Open ticket e8d4c8a577 · (unused) · filed 2026-09-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-04 check-in 7ae80eeb69 on x11-bidi-font-reconfigure-leak, in main, 9.0: Fix e8d4c8a577: X11 (HarfBuzz/SheenBidi font backend): [font configure] of a font in use leaks the old font. TkpGetFontFromAttributes re-initialized a font in
    • 🔧 2026-10-06 check-in 13c5fdc004 on trunk, in main, 9.0: Fix e8d4c8a577: X11 (HarfBuzz/SheenBidi font backend): [font configure] of a font in use leaks the old font
    • 💬 2026-10-04 serhiy.storchaka: Proposed fix in branch x11-bidi-font-reconfigure-leak. It also fixes two related bugs: the fallback to "sans" destroyed the pattern a second time, and InitFont() used the font id to detect a re-used font record, so a failed re-initialization would have freed everything twice.
    • 💬 2026-10-06 serhiy.storchaka: Merged into main 13c5fdc004. Tk 9.0 is not affected.
    Full verdict

    REPRODUCED (main 40.3 KB per font change; 9.0.5 and 8.6.18 0.0 KB, they use tkUnixRFont.c)

    • Platform X11 (only the bidi font backend, unix/tkUnixBidiFont.c, which is the configure default on main)
    • Tested the ticket's reproducer (1500 font configure of a named font used by a label) under Xvfb :101 with build-trunk, tk9.0 and tk8.6. With the prototype: 0.0 KB per change. valgrind on the patched build (150 changes): no invalid reads or writes, 512 bytes "definitely lost" not from the font code. font.test, unixFont.test and fontchooser.test pass (same counts as unpatched main).
    • Cause unix/tkUnixBidiFont.c:TkpGetFontFromAttributes() re-initializes a live TkFont in place (UpdateDependentFonts -> TkpGetFontFromAttributes(fontPtr,...)) by calling InitFont() directly, without the FinishedWithFont() that tkUnixRFont.c calls first. InitFont() then overwrites faces[], the FcFontSet, the FcPattern, the Xft/HarfBuzz objects, the charsets and the shaper's hb_buffer. Also confirmed in the code: (1) on the fallback path TkpGetFontFromAttributes() calls XftPatternDestroy(pattern) on a pattern that InitFont() has already destroyed on every failure path (directly after FcFontSort, or through FinishedWithFont via fontPtr->pattern), which is a double free; (2) InitFont() uses fontPtr->font.fid != 0 to mean "re-used record". A new font that fails after XLoadFont() is therefore never freed. Once FinishedWithFont() is added, a re-used record that fails is left with dangling faces/pattern/fontset pointers, and TkpDeleteFont() would free them a second time.
    • Prototype e8d4c8a577.patch (90 lines, no test: a leak test would need memory accounting). It calls FinishedWithFont() before the in-place InitFont(), makes FinishedWithFont() clear the pointers and fid it frees, replaces the fid test with an explicit reused flag, and drops the double XftPatternDestroy(). Font tests pass and the reproducer is flat.
    • Suggested ticket comment Confirmed on main; 9.0 and 8.6 are not affected. Besides the missing FinishedWithFont(), the fallback's XftPatternDestroy() is indeed a double free, and InitFont() must not use font.fid to detect a re-used record (a failed re-init would otherwise free everything twice). Proposed fix in branch <name> (all three).
    ✅965dae4a6d2026X11 font backend: stack buffer overflow (crash) for strings of 1024 or more bytes, e.g. 344 CJK charactersREPRODUCED (own ticket, filed 2026-09-29); proposed fix in branch x11-shaper-long-string, not merged

    Open ticket 965dae4a6d · 46. Unix Fonts · filed 2026-09-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-10-01 kevin_walzer: Checking the x11_shaper_unlimited branch, looks good.
    • 💬 2026-10-01 kevin_walzer: The branch I committed earlier today has a max_buffer of 2048 - I assume that will be updated once this is merged?
    • 💬 2026-10-01 serhiy.storchaka: Yes. I merged trunk into [x11-shaper-unlimited], so it merges cleanly now. The branch removes the fixed limit on the number of glyphs and cluster breaks, so the 2048 limit is gone after the merge (MAX_GLYPHS is now only the size of a drawing batch).
    • 💬 2026-10-13 kevin_walzer: Can this branch be merged?
    • 💬 2026-10-13 serhiy.storchaka: Yes, please. x11-shaper-long-string is the minimal fix for this crash. x11-shaper-unlimited includes it and also removes the 512-glyph limit of 4eef1fa86e, so merging x11-shaper-unlimited fixes both.
    • 💬 2026-10-13 kevin_walzer: Merged to trunk. Thank you for the fix.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-29); proposed fix in branch x11-shaper-long-string, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch x11-shaper-long-string
    ✅4a25f316f72026Embedded application exits with BadDrawable when its container is destroyedREPRODUCED (own ticket, filed 2026-09-26); proposed fix in branch embed-container-destroyed, not merged

    Open ticket 4a25f316f7 · 74. Application Embedding · filed 2026-09-26

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-26 check-in d6f016b586 on embed-container-destroyed, in main, 9.0: Fix 4a25f316f7: an embedded application exited with BadDrawable when its container was destroyed while it had pending drawing. ErrorProc() now ignores BadDraw
    • 🔧 2026-10-09 check-in 8bb13f5c3c on embed-container-destroyed, in main, 9.0: Fix 5f7ba8c5a3: Restore the fix for 4a25f316f7, which was backed out because the table of deleted windows grew in loops which created and destroyed windows
    • 🔧 2026-10-09 check-in 54ffe89f65 on trunk, in main, 9.0, 8.6: Fix 5f7ba8c5a3: Restore the fix for 4a25f316f7
    • 🔧 2026-10-14 check-in d5effc94df on core-9-0-branch, in main, 9.0: Fix 5f7ba8c5a3: Restore the fix for 4a25f316f7
    • 🔧 2026-10-14 check-in 474ec9d09d on core-8-6-branch, in main, 9.0, 8.6: Fix 5f7ba8c5a3: Restore the fix for 4a25f316f7
    • 💬 2026-09-26 serhiy.storchaka: Proposed fix in branch embed-container-destroyed: ErrorProc() ignores BadDrawable errors like BadWindow errors, also for windows deleted after the failed request. Recently deleted windows are recorded in a new field at the end of TkDisplay. unixEmbed-9.2 no longer sporadically fails. The new test unixEmbed-9.3 adds a more reliable reproducer.
    • 💬 2026-10-07 jan.nijtmans: Fixed in dd18288b38. Backported to core-9-0-branch and core-8-6-branch. Many thanks! Closing
    • 💬 2026-10-09 jan.nijtmans: It looks like this fix is causing a memory leakage. See 5f7ba8c5a3 ticket. So, what shall we do?
    • 💬 2026-10-09 serhiy.storchaka: The growth is fixed in branch embed-container-destroyed, see 5f7ba8c5a3: deleted windows are now recorded only after a toplevel of the application has been embedded. The new check-in is on top of main and restores the backed out fix, so a plain merge brings it back. The fix was also backed out in core-9-0-branch and core-8-6-branch, where the backout needs to be reverted too.
    • 💬 2026-10-14 nemethi: Restored the backed out fix in main, core-9-0-branch, and core-8-6-branch by commits 54ffe89f (Jan), d5effc94, and 474ec9d09. Closing the ticket.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-26); proposed fix in branch embed-container-destroyed, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch embed-container-destroyed
    ✅434c8dd8042026Crash in Tk_ConfigureWidget when a default value of a canvas item option is invalidREPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch canvas-default-error-crash, not merged

    Open ticket 434c8dd804 · 05. Canvas Items · filed 2026-09-25

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-25 check-in dc71571171 on canvas-default-error-crash, in main, 9.0: Fix 434c8dd804: crash in Tk_ConfigureWidget when a default value of a canvas item option is invalid. The error message used specPtr->dbName, which is NULL for
    • 🔧 2026-10-08 check-in a2791cb144 on trunk, in main, 9.0: Fix 434c8dd804: crash in Tk_ConfigureWidget when a default value of a canvas item option is invalid
    • 🔧 2026-10-08 check-in 388b4fee03 on core-9-0-branch, in main, 9.0, 8.6: Fix 434c8dd804: crash in Tk_ConfigureWidget when a default value of a canvas item option is invalid
    • 🔧 2026-10-08 check-in 38c2e30faa on core-8-6-branch, in main, 9.0, 8.6: Fix 434c8dd804: crash in Tk_ConfigureWidget when a default value of a canvas item option is invalid
    • 💬 2026-09-25 serhiy.storchaka: Proposed fix in dc71571171 (branch canvas-default-error-crash): use the option name in the error message if the option has no database name. Now .c create text reports the error with (default value for "-font" in widget ".c") instead of crashing.
    • 💬 2026-10-08 jan.nijtmans: Fixed in a2791cb144, backported to core-9-0-branch and core-8-6-branch Many thanks!
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch canvas-default-error-crash, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch canvas-default-error-crash
    ✅b9d85fc1002026X11 accessibility: root accessible is not registered with ATK; does not work where atk_get_root() cannot be interposed (illumos)REPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch atk-util-get-root, not merged

    Open ticket b9d85fc100 · 67. Unix Window Operations · filed 2026-09-25

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-25 check-in df6cd07ca4 on atk-util-get-root, not merged: Fix b9d85fc100: X11 accessibility: root accessible is not registered with ATK. Tk defined its own global atk_get_root(), relying on symbol interposition to ov
    • 💬 2026-09-25 serhiy.storchaka: Proposed fix in df6cd07ca4 (branch atk-util-get-root): set get_root in the AtkUtil class before initializing the bridge, instead of defining atk_get_root(). Tested with the AT-SPI registry on OpenIndiana: without the fix no application is registered; with the fix "Tk Application" is, and the warning is gone. On Linux, where the interposition worked, the accessible tree is unchanged.
    • 💬 2026-10-05 jan.nijtmans: Kevin, can you have a look?
    • 💬 2026-10-05 kevin_walzer: Yes, I will review.
    • 💬 2026-10-06 kevin_walzer: Merged into alt_x11_accessibility_fixes. I see no adverse impact on Linux, but cannot test elsewhere.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch atk-util-get-root, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch atk-util-get-root
    ✅1dfc71ae4a2026configure fails without HarfBuzz although --enable-bidi was not givenREPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch configure-bidi-requires-xft, configure-bidi-without-harfbuzz, not merged

    Open ticket 1dfc71ae4a · 84. Unix Build · filed 2026-09-25

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-25 check-in b83a809314 on configure-bidi-without-harfbuzz, in main, 9.0: Fix 1dfc71ae4a: configure fails without HarfBuzz although --enable-bidi was not given. Unless --enable-bidi is given explicitly, a missing, too old or unusabl
    • 🔧 2026-09-29 check-in 603e625351 on configure-bidi-requires-xft, in main, 9.0: Bidi requires Xft: disable it if Xft is disabled or not found 1dfc71ae4a
    • 🔧 2026-10-05 check-in 427bb09f52 on trunk, in main, 9.0: Fix 1dfc71ae4a: configure fails without HarfBuzz although --enable-bidi was not given
    • 🔧 2026-09-30 check-in ae18cffecd on configure-bidi-requires-xft, in main, 9.0: Merge trunk
    • 💬 2026-09-25 serhiy.storchaka: Proposed fix in b83a809314 (branch configure-bidi-without-harfbuzz): without --enable-bidi, a missing, too old or unusable HarfBuzz only disables Bidi with a warning; with --enable-bidi it is an error which mentions --disable-bidi. This also fixes the too old case: configure disabled Bidi but still defined HAVE_BIDI, so Tk was built with the non-Bidi font code but with TK_DRAW_IN_CONTEXT and TK_LAYOUT_WITH_BASE_…
    • 💬 2026-09-29 serhiy.storchaka: Xft has the same problem as HarfBuzz: Bidi requires Xft, but --disable-xft is overridden by the default Bidi (Xft is forced on again), and if Xft is not found, configure says "falling back to classic fonts" but still selects the Bidi backend, which cannot be built without Xft. Alternative branch configure-bidi-requires-xft, based on configure-bidi-without-harfbuzz, treats Xft like HarfBuzz: --disable-xft or a miss…
    • 💬 2026-10-05 jan.nijtmans: This looks like the right approach to me. Thanks for doing this! Merged to 427bb09f52 Since Tk 9.0 doesn't have bidi, there's no backporting to be done.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-25); proposed fix in branch configure-bidi-requires-xft, configure-bidi-without-harfbuzz, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch configure-bidi-requires-xft, configure-bidi-without-harfbuzz
    💬97a7ddac252026X11 accessibility (ATK) is not thread-safe: crash when Tk is used in several threadspatch atk-threadREPRODUCED (own ticket, filed 2026-09-24); no fix yet

    Open ticket 97a7ddac25 · 67. Unix Window Operations · filed 2026-09-24

    💬 Since triage (updated 2026-11-11)

    Status now: Open

    • 💬 2026-09-24 kevin_walzer: Tk should only be run on the main thread. The accessibility design follows this principle. Is what you are testing here a realistic scenario? Shouldn’t we just document this instead?
    • 💬 2026-09-24 serhiy.storchaka: On X11 and Windows, Tk has supported one Tk interpreter per thread for a long time (each thread has its own display connection); only macOS requires the main thread. It is used with the Thread package, and there are open tickets about it (6e3c4fb23f, b298b92ba7). 9.0 runs the same test without crashing. Documenting would leave a heap corruption in a configuration that worked before. The fix does not need to ma…
    • 💬 2026-09-24 kevin_walzer: Can you provide a patch that activates the empty commands on non-main threads?
    • 💬 2026-09-24 serhiy.storchaka: Correction: 9.0 crashes too when Tk is initialized in many threads at once (heap corruption detected in Xlib's XOpenIM). Only the crashes in TkAtkAccessibility_Init and GLib are new in 9.1. A patch that enables ATK only in the first thread removes the ATK crashes, but the older crash prevents proper testing, so I will look into it first.
    • 💬 2026-09-24 kevin_walzer: Ok, thank you for the update. The advice I’ve often followed with Tk and threads is Don’t Do That, because of cases like this.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-24); no fix yet

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype none
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main ported. Main restructured init: TkAtkAccessibility_Init now only calls TkWaylandAccessibility_Init, and there is a new ::tk::accessible::speak command backed by global speechd state (spd_conn, pending message, Tcl timer). The thread check now lives in TkAtkAccessibility_Init before that call. Non-ATK threads also get a no-op ::tk::accessible::speak, because the speechd state is process-wide too. As a result, SpeechExitProc is registered only once.
    • Configuration main configures with accessibility enabled by default ("checking for ATK and Atk-bridge... yes", HAVE_ATK=1, HAVE_SPEECHD=1).
    • Repro tclsh loads Thread 3.0.7 (a copy of work-a1's build). N threads each run load libtcl9tk9.1.so Tk, then create, pack and destroy 50 frames, then destroy ..
      • Unpatched main, 4 threads: 2/6 runs crashed (SIGSEGV after "GLib-GObject-CRITICAL ... invalid (unclassed) pointer"; abort "double free or corruption").
      • Patched, 4 threads: 0/25 and 0/6 runs crashed. 8 threads: 0/40.
      • One earlier patched run (1/20, 4 threads) aborted with "double free or corruption (!prev)". The old non-ATK multi-thread crash from the ticket explains it: a --disable-accessibility build of the same main crashes 3/40 at 8 threads ("malloc(): unaligned tcache chunk detected"). That bug is separate and not addressed here.
      • Under gdb, one run hung for 20 minutes (vfork children). I killed it and did not get a backtrace.
    • Test none. It needs the Thread package, which the Tk test suite does not depend on. The crash is probabilistic, and the older non-ATK heap corruption would make such a test flaky anyway.
    • Test file run tk.test 38/0, safe.test 16/0 (same as main; there is no dedicated accessibility test file)
    • Patch changes ported as described above (+speak no-op). Backup: 97a7ddac25-atk-thread.patch.1024.
    ✅54c2ca0e942026photo: pixels made semi-transparent with "transparency set ... -alpha" are drawn opaqueREPRODUCED (own ticket, filed 2026-09-24); proposed fix in branch photo-transparency-set-alpha, not merged

    Open ticket 54c2ca0e94 · 41. Photo Images · filed 2026-09-24

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-24 check-in 3faec47822 on photo-transparency-set-alpha, in main, 9.0: Fix 54c2ca0e94: photo: pixels made semi-transparent with "transparency set ... -alpha" are drawn opaque.
    • 🔧 2026-10-15 check-in cc79468d5e on photo-transparency-set-alpha, in main, 9.0: Rebase to main
    • 💬 2026-09-24 serhiy.storchaka: Proposed fix in branch photo-transparency-set-alpha (3faec47822): transparency set ... -alpha now sets COMPLEX_ALPHA for a partial alpha value, so such pixels are blended. Test imgPhInstance-6.1 checks the rendered pixel with testpixel; it fails without the fix.
    • 💬 2026-10-19 oehhar: The ticket says, that 9.1 is not effected. But the patch is for 9.0. Nevertheless, this looks reasonable for both. May look-in again tomorrow. Thanks, Harald
    • 💬 2026-10-19 serhiy.storchaka: Thank you, Harald. 9.1 is affected on X11 too (the ticket says: on X11 in 9.0.5 and 9.1b1, and on Windows in 9.0.5); only Windows 9.1 is not affected, because it draws the RGBA data directly. The branch photo-transparency-set-alpha is against main (rebased on 2026-10-15), and the fix also applies to 9.0.
    • 💬 2026-10-20 oehhar: Tried the example with 9.0 on MS-Win. Fix is effective. So, committed to: * main: 59fba600 * core-9-0-branch: 33904f03 but the test is missing, as there is no file "tests/imgPhInstance.test". Bug closed. Thank you, Harald
    • 💬 2026-10-20 oehhar: Also verified that tk9.1.0 release does not show the issue on MS-Win. Nevertheless, the image should be flagged as having transparent pixels. Thanks for all, Harald
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-24); proposed fix in branch photo-transparency-set-alpha, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch photo-transparency-set-alpha
    ✅5d983ebf482026Text widget: RTL auto-detection makes mapping and editing slowpatchREPRODUCED (own ticket, filed 2026-09-24); proposed fix in branch text-rtl-detect-speed, not merged

    Open ticket 5d983ebf48 · 18. [text] · filed 2026-09-24

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-24 check-in f8c5c81cb4 on text-rtl-detect-speed, in main, 9.0: Speed up the RTL auto-detection of the text widget and fix its per-widget state.
    • 🔧 2026-09-29 check-in 28f9b9c3c4 on text-rtl-detect-speed, in main, 9.0: Add tests textTag-19.5 to 19.7 for all RTL character ranges, characters outside them, and punctuation and whitespace in RTL lines 5d983ebf48.
    • 🔧 2026-10-08 check-in c47ce655fd on text-rtl-detection, in main, 9.0: Fix 5d983ebf48 and 3d6de81648: speed up the RTL auto-detection of the text widget and fix its per-widget state. Combines branches text-rtl-detect-speed and
    • 🔧 2026-10-13 check-in 3a6ffd3b98 on trunk, in main, 9.0: Fix 5d983ebf48 and 3d6de81648: speed up the RTL auto-detection of the text widget and keep its pending state per widget. Merge text-rtl-detection (Serhiy St
    • 🔧 2026-10-08 check-in 5aae2fcb27 on nab-text-rtl-scan, in main, 9.0: Text RTL detection: one regexp search of the text instead of a Tcl loop on every line at each keystroke; tag changed only where needed.
    • 🔧 2026-10-09 check-in f893f7e716 on text-rtl-detection, in main, 9.0: Merge trunk
    • 💬 2026-09-24 serhiy.storchaka: Fix on text-rtl-detect-speed: skip the scan if the text contains no RTL characters, remove the tag once instead of line by line, and check lines with regular expressions. The pending and auto-detection flags are now per widget and are removed on <Destroy>. Tests textTag-19.1 to 19.4. RTL characters now include all characters with Unicode Bidi_Class R or AL (Unicode 18): besides U+0590–U+08FF, also U+200F, the Hebr…
    • 💬 2026-09-29 serhiy.storchaka: Attached a sample script. Results on Linux/X11 (Xvfb): 9.09.19.1 with the fix map 900 texts with 300 lines708 ms5351 ms662 ms map 1 text with 100000 lines1 ms1079 ms14 ms keystroke in 5000 Hebrew lines0 ms30 ms12 ms two texts changed in the same idle cycle are tagged-only the firstboth recreated text is tagged on map-noyes
    • 💬 2026-10-08 serhiy.storchaka: Combined with Nicolas's nab-text-rtl-scan (from 3d6de81648) in branch text-rtl-detection, c47ce655fd: * the text is searched once by the widget (search -all -regexp, with Unicode white space listed so that a match does not run into the next lines), and the tk_rtl_right tag is changed only where it differs; * RTL characters are all characters with Bidi_Class R or AL (::tk::RtlChars), used by the search, the lin…
    • 💬 2026-10-12 nab: Hi, it would be very nice to have this one on trunk. many thanks, nicolas
    • 💬 2026-10-13 nab: Merged into trunk, 3a6ffd3b98. Kevin Walzer, who wrote the RTL auto-detection, gave his green light. Thanks Serhiy for combining the two branches.
    Full verdict

    REPRODUCED (own ticket, filed 2026-09-24); proposed fix in branch text-rtl-detect-speed, not merged

    • Platform see title
    • Tested filed after reproducing it; see the ticket
    • Prototype branch text-rtl-detect-speed
    ✅56d4d132fd2026Unsafe child interpreters cannot find Tk if wish is statically linkedREPRODUCED (main). Fix on branch bug-56d4d132fd-child-interp (db2fc66a5f and its follow-ups by kjnash, including the review changes), not merged.

    Open ticket 56d4d132fd · 84. Unix Build · filed 2026-09-22

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-28 check-in 31317f1a09 on bug-56d4d132fd-child-interp, in main, 9.0: Merge trunk
    • 💬 2026-09-28 kjnash: Hi Serhiy and Andreas, Thanks for your review, I agree with all your suggestions, and I have implemented them with one further change. It appears that the settings in tests/constraints.tcl cannot be overridden by a -constraints option in the "test" command. So setting testConstraint useInstalledWish 0 appears to mean that tests with this constraint will not be executed even if "-constraints useInstalledWish" is sp…
    • 💬 2026-09-28 serhiy.storchaka: Thank you, Keith. The command line options are processed before tests/constraints.tcl is sourced, so the constraint can be set from them, and the tests still run only on request: [code] With it the tests are skipped by default and run with -constraints useInstalledWish.
    • 💬 2026-09-28 kjnash: Thanks, I have now edited tests/constraints.tcl to use this form.
    • 💬 2026-09-28 kjnash: Should I commit this to 9.1.0?
    • 💬 2026-10-30 dgp: The commit to the 9.0 branch 19723732ec breaks my --disable-shared build on an Ubuntu Linux system. The failure shows up attempting to compile tkStubLib.c gcc: warning: {X11_INCLUDES}: linker input file unused because linking not done gcc: error: {X11_INCLUDES}: linker input file not found: No such file or directory gcc: warning: {SHLIB_CFLAGS}: linker input file unused because linking not done gcc: error: {SHLI…
    • 💬 2026-10-30 dgp: Here is the failing compile line. Some of it just looks wrong. gcc -c -I/home/dgp/fossil/tk9.0/unix/../unix -I/home/dgp/fossil/tk9.0/unix/../generic -I/home/dgp/fossil/tk9.0/unix/../bitmaps {X11_INCLUDES} -O2 -pipe -finput-charset=UTF-8 -Wall -Wextra -Wshadow -Wundef -Wwrite-strings -Wpointer-arith -Wc++-compat -fextended-identifiers {SHLIB_CFLAGS} -I/home/dgp/fossil/tcl9.0/generic -I/home/dgp/fossil/tcl9.0/unix -…
    • 💬 2026-10-30 jan.nijtmans: Got .... missing '$'. Don't know how that happened :-( Should be corrected now.
    • 💬 2026-10-30 jan.nijtmans: @kjnash: Sorry it took so much time to have a look at it. This definitely should be in Tk 8.0.5
    • 💬 2026-10-30 dgp: All better on the branch tips
    • 💬 2026-10-30 jan.nijtmans: 👍
    Full verdict

    REPRODUCED (main). Fix on branch bug-56d4d132fd-child-interp (db2fc66a5f and its follow-ups by kjnash, including the review changes), not merged.

    • Platform build (all platforms)
    • Tested main built with --disable-shared and installed to a scratch prefix; tk_library is //zipfs:/app/tk_library and no pkgIndex.tcl is installed for the static Tk. interp create foo; foo eval package require tk did not fail outright. Instead it found main/lib/tk9.1/pkgIndex.tcl through auto_path and loaded a second, shared libtcl9tk9.1.so into the process (info loaded in the child shows both {} Tk and the .so), so the child silently uses a different Tk build. Shared builds of main and 9.0 work.
    • Cause unix/Makefile.in, win/Makefile.in and win/makefile.vc do not install a pkgIndex.tcl with load {} Tk for static builds.
    • Prototype none (already on branch bug-56d4d132fd-child-interp; kjnash asks whether to commit it).
    • Suggested ticket comment Confirmed on main with a static build: with another Tk installed on auto_path, the child interpreter even loads that shared Tk next to the static one, which is worse than failing. The branch has been reviewed and the review changes are in. I think it can go into main (it is a build-only change), and the 9.0 part can be backported.
    ✅586f722cdd2026New test winfo-4.8 failspatchREPRODUCED (main, 9.0, 8.6) as a timing race. winfo-4.8 itself passes here with no WM and with twm/fvwm/icewm/openbox/wmaker/i3/xfwm4, but it fails…

    Open ticket 586f722cdd · 65. Generic Window Operations · filed 2026-05-27

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-24 check-in 8a0667c7fc on x11-coords-to-window, in main: Fix 586f722cdd and 220856ffff in Tk_CoordsToWindow() on X11. A toplevel that was just withdrawn was still found while the window manager had not yet unmappe
    • 🔧 2026-11-10 check-in d25d2eeb12 on trunk, in main, 9.0: Fix for 586f722cdd: New test winfo-4.8 fails
    • 🔧 2026-11-10 check-in 2431c82bf4 on core-9-0-branch, in 9.0, 8.6: Fix for 586f722cdd: New test winfo-4.8 fails
    • 🔧 2026-11-10 check-in 5a03b015c3 on core-8-6-branch, in 8.6: Fix for 586f722cdd: New test winfo-4.8 fails
    • 💬 2026-10-24 serhiy.storchaka: This is a race: wm withdraw unmaps the toplevel at once, but its window manager frame stays mapped until the window manager handles the unmap, and Tk_CoordsToWindow() still returned the toplevel. With a slow window manager, after 10 is not enough. Fixed on branch x11-coords-to-window (8a0667c7fc): a toplevel which is not mapped is not returned. New test winfo-4.9 has no delay; without the fix it failed under ope…
    • 💬 2026-11-10 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits d25d2eeb, 2431c82b, and 5a03b015.
    Full verdict

    REPRODUCED (main, 9.0, 8.6) as a timing race. winfo-4.8 itself passes here with no WM and with twm/fvwm/icewm/openbox/wmaker/i3/xfwm4, but it fails whenever the WM is slower than after 10.

    • Platform X11 (the #591829e948-B fix in check-in 5261b9cd44 changed only macosx/tkMacOSXWm.c)
    • Tested main, 9.0, 8.6 on Xvfb. A probe calls winfo containing right after wm withdraw .one; update. It returns .one on 9.0+fvwm, main/9.0+icewm, and main/9.0/8.6+openbox. After 10 ms it returns "", and winfo ismapped .one is already 0.
    • Cause unix/tkUnixWm.c Tk_CoordsToWindow. The toplevel wrapper is unmapped at once, but the WM's decoration frame (wmPtr->reparent) stays mapped until the WM handles the UnmapNotify. XTranslateCoordinates still finds the frame, and the code accepts that toplevel without checking whether it is mapped.
    • Prototype 586f722cdd.patch (23 lines). After gotToplevel it returns NULL if !Tk_IsMapped(winPtr). It adds test winfo-4.9, which uses no delay after the withdraw. winfo-4.9 fails without the fix under openbox/icewm and passes with it. winfo.test has no new failures: with no WM, winfo-9.4/9.7 fail both before and after the patch; under wmaker, winfo-4.5 fails both before and after.
    • Suggested ticket comment On X11 this is a race. After wm withdraw the WM frame stays mapped until the WM reacts, and Tk_CoordsToWindow accepts the frame without checking that the toplevel is mapped. A fix and a test that doesn't depend on timing (winfo-4.9) are ready.
    ⬜3249f7d4c62026MS-Win: canvas does not rotate text if font replacedpatchREPRODUCED (9.1 main); NOT-REPRODUCED (8.6 tip)

    Open ticket 3249f7d4c6 · 05. Canvas Items · filed 2026-04-21

    Full verdict

    REPRODUCED (9.1 main); NOT-REPRODUCED (8.6 tip)

    • Platform Windows
    • Tested code review. Repro script: win/3249f7d4c6.tcl (Courier and Courier New side by side, -angle 90, with bbox rectangles).
    • Cause (likely): c0dbdd3ff3 (3f19aebf80, 2017-01-28) was merged to 9.0 but not 8.6. It removed the hack in win/tkWinFont.c FamilyExists that rejected "Courier", "Times" and "Helvetica". In 9.0, "Courier" can therefore resolve to the GDI raster font (coure.fon). GDI ignores lfEscapement/lfOrientation for raster fonts, so GetScreenFont(angle) returns an unrotated font while the bbox is rotated. Possible fix: make WinFontExistProc reject fontType & RASTER_FONTTYPE, so the fallback picks a TrueType font. Alternatively, create angled fonts with OUT_TT_ONLY_PRECIS.
    • Prototype none (Windows-only; confirm first)
    • Suggested ticket comment Probably "Courier" now maps to the raster Courier font, which GDI cannot rotate. That is a side effect of c0dbdd3ff3 no longer rejecting Courier/Times/Helvetica. Rejecting raster fonts in WinFontExistProc (or using OUT_TT_ONLY_PRECIS for angled fonts) should fix it. Harald, please run the attached script.
    Windows VM, run 1
    • Platform Windows
    • Verdict REPRODUCED (9.1 main); NOT-REPRODUCED (8.6 tip)
    • Tested automated screen-capture repro (work-win/fontrot.tcl + fontrot.ps1: measures the ink bbox of "ABCDEF" at -angle 90).

    9.1: font actual gives Courier, and the text is drawn unrotated (ink 53x10). Courier New, Times and Helvetica are rotated.

    8.6: Courier resolves to Courier New and is rotated.

    • Cause 3f19aebf80 (2017-01-28, fix for c0dbdd3ff3) removed the Courier/Times/Helvetica rejection from win/tkWinFont.c:FamilyExists. On Windows, "Courier" now matches the raster coure.fon, and GDI cannot rotate raster fonts (lfEscapement is ignored). Explicitly chosen raster fonts such as Fixedsys and Terminal are also unrotated. That is inherent to GDI.
    • Prototype 3249f7d4c6.patch (16 changed lines). FamilyExists passes a flag to WinFontExistProc that skips RASTER_FONTTYPE entries, but only for families that have a Tk alias list (TkFontGetAliasList: Courier/Times/Helvetica/...). "Courier" therefore falls back to "Courier New", while a real TrueType "Courier" (the case in c0dbdd3ff3) is still found. I relaxed test winfont-2.8 to accept {Courier New}. Verified: the Courier text is now rotated, and font.test, winFont.test and canvText.test pass (0 failures). Before the test change, winfont-2.8 failed as expected.
    • Suggested ticket comment On Windows, "Courier" is the raster font coure.fon, which GDI cannot rotate. Since c0dbdd3ff3, Tk no longer replaces it with Courier New. A fix is to ignore raster fonts in FamilyExists for the aliased families (Courier/Times/Helvetica), so Courier falls back to Courier New again. A TrueType "Courier" is still used when installed.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before: Courier text with -angle 90 is not drawn rotated (testpixel finds only one pixel of ink) / after: a tall narrow ink box
    • Test winfont-13.1 (new; ink bounding box of ABCDEFGH in {Courier 12} at -angle 90 must be taller than twice its width; fails before: yes; passes after: yes). The relaxed winfont-2.8 in the patch passes either way.
    • Test file run winFont.test 57/0, font.test 304/0, canvText.test 100/0 (pre-existing failures: none)
    • Patch changes added test winfont-13.1
    ⬜596ca7f5302026label text constraint by -width/-height paints into -padx/pady and border spaceREPRODUCED (main, 8.6, 8.4). This is long-standing behaviour.

    Open ticket 596ca7f530 · 47. Win Fonts · filed 2026-04-17

    Full verdict

    REPRODUCED (main, 8.6, 8.4). This is long-standing behaviour.

    • Platform generic (reproduced on X11, so not Windows-only)
    • Tested Xvfb screenshots on main, 8.6 and 8.4 of a label with -width 3 -height 1 -bd 4 -relief flat -padx 6 -pady 6 and 3-line text. The glyphs fill the whole 56x40 window, including the padding and the border.
    • Cause TkpDisplayButton (unix/win/macosx *Button.c) draws the text layout into the full-window pixmap with no clip to the inner area. Only the window bounds clip it.
    • Prototype none. A fix would touch all 3 platform files (clip region, or re-filling pad and border before drawing the relief), and it changes long-standing behaviour that chw calls a feature. It would be over 60 lines.
    • Suggested ticket comment This is the same on X11 and has been since at least 8.4: button/label text is clipped only by the window, not by -padx/-pady/-bd. Fixing it needs clipping in each platform's TkpDisplayButton. It's a behaviour change rather than a regression, so it could be recategorised as an RFE.
    ✅11bd4a03a02026MS-Win toplevel top menu does not honor configured font and colorREPRODUCED (9.1 main, 8.6 tip), limitation / RFE

    Open ticket 11bd4a03a0 · 13. Win Menus · filed 2026-04-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-28 check-in ad39c64432 on win-menubar-colors-ownerdraw, not merged: Fix 11bd4a03a0: the Windows menu bar ignored the colors and fonts of the menu. The items of a menu bar with non-default colors or fonts are now drawn by Tk (o
    • 🔧 2026-09-28 check-in 0204f94e45 on win-menubar-colors-uah, in main, 9.0: Fix 11bd4a03a0: the Windows menu bar ignored the colors and fonts of the menu. A menu bar with non-default colors or fonts is now drawn by Tk, handling the un
    • 🔧 2026-09-28 check-in 80663c877e on win-menubar-colors-ownerdraw, not merged: The default font of the menu is the system menu font specified by its name and size, not TkMenuFont: do not draw a menu bar with the default font by Tk.
    • 🔧 2026-10-07 check-in b9d8f1fa27 on win-menubar-colors-uah, in main, 9.0: Merge main
    • 💬 2026-09-28 oehhar: Serhiy, if you are looking for a challenge, this is be my favorite ;-) (as all other challenges like text eliding performance has been solved). Thanks for all, Harald
    • 💬 2026-09-28 serhiy.storchaka: Thank you for the challenge, Harald. I already looked at this ticket, but held back, because it looked rather like a complex feature request. There are two alternative proposed fixes. In both, a menu bar with non-default colors or fonts of the menu or its entries is drawn by Tk, including the active and disabled entries and images; otherwise it is drawn by the system as before. * Branch win-menubar-colors-ownerdra…
    • 💬 2026-09-28 oehhar: Serhiy, thanks for the light speed solution branches. For me personally, the font size of the menu is a critical feature. So, the undocumented version would be great for me. The tk scaling factor should be respected independently of any MS-Win system settings. I will test and come back and attach some images. And yes, the dark mode has to be configured. This is intentional, as menus are not themed. Nevertheless, t…
    • 💬 2026-09-30 oehhar: Great! My test case: [code] The results of the two branches are attached by images with the branch name. IMHO, the win-menubar-colors-uah branch should be merged yesterday. I would really love to have it. Are there any other opinions on this? I would love to merge it at least in 9.1.1. Thanks for all and take care, Harald
    • 💬 2026-10-01 oehhar: The example for Tk9.1.0 is attached as a picture. Here is a message by Marc on the core list: I see that the fix involves using Windows functions which are "undocumented". With macOS we got into trouble for using "private APIs", which I think may be the same as "undocumented". As I recall, the main sort of trouble we got into was that Apple would reject any app submitted to the App Store which used a "private API"…
    • 💬 2026-10-07 nemethi: IMHO, Serhiy's solution implemented in the branch win-menubar-colors-uah is simply brilliant. And Harald's screenshot win- menubar-colors-uah.png is absolutely convincing. I, too, am strongly in favor of adding this great improvement to the forthcoming Tk 9.1.1 release.
    • 💬 2026-10-07 oehhar: Great! Merged branch win-menubar-colors-uah to main by a07f27dd. Bug closed. Thanks Sergey and Csaba! Take care, Harald
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip), limitation / RFE

    • Platform Windows
    • Tested X11 main: the menubar clone .#m gets -bg black and the font, so X11 honors them. Windows repro script: win/11bd4a03a0.tcl (interactive; press a button to record the verdict).
    • Cause win/tkWinMenu.c:ReconfigureWindowsMenu inserts MENUBAR entries as native MF_STRING (only other menus use MF_OWNERDRAW). DrawMenuEntry* returns early for MENUBAR, and there is no SetMenuInfo(hbrBack). So Windows draws the bar with system colors and fonts. A fix needs owner-drawn or UAH-custom-drawn menubar items, as in the links in the ticket.
    • Prototype none (native menubar custom-draw is well over 60 lines and is a design change)
    • Suggested ticket comment Confirmed by code: menubar entries are added as native MF_STRING items (tkWinMenu.c), so -bg/-fg/-font are ignored on Windows by design. Fixing it means owner-drawing the menubar (UAH messages); I'd treat this as a feature request.
    Windows VM, run 1
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip), limitation / RFE
    • Tested work-win/menubar.tcl + menubar.ps1. A menubar with -bg red -fg yellow -font {Courier 24 bold}, captured with CopyFromScreen, has 0 red and 0 yellow pixels on 9.1 and 8.6: system font, default colors.
    • Cause win/tkWinMenu.c:ReconfigureWindowsMenu adds MENUBAR entries as MF_STRING, not MF_OWNERDRAW, and DrawMenuEntry* returns early for MENUBAR.
    • Prototype none (needs owner-drawn/UAH menubar painting; a design change of more than 60 lines)
    • Suggested ticket comment Confirmed: menubar items are native MF_STRING items (tkWinMenu.c), so -bg/-fg/-font are ignored on Windows. Supporting them needs owner-drawing of the menubar (UAH messages). Changing this to a feature request.
    ✅9ab6e37e1a2026[wm forget] segfaultpatchREPRODUCED (main fb166f9f73, 9.0.5, 8.6.18: "wm forget .; pack ." → SIGSEGV)

    Open ticket 9ab6e37e1a · 08. [(label)frame] and [toplevel] · filed 2026-01-13

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-11-07 check-in 694e33a172 on trunk, in main, 9.0: 9ab6e37e1a disallow "wm forget ."
    • 🔧 2026-11-07 check-in b7c97fb76a on core-9-0-branch, in main, 9.0: 9ab6e37e1a disallow "wm forget ."
    • 🔧 2026-11-10 check-in a985934f08 on wayland, not merged: Fix wm-forget-3: on Wayland, wm forget . forgot the main window instead of raising an error, as on other platforms (bug 9ab6e37e1a).
    • 💬 2026-11-05 oehhar: Thanks Serhiy for the test. Merged main with conflicts on Mac and Unix by 4f5a466cdc. Will merge to main if CI is ok. Thanks for all, Harald
    • 💬 2026-11-05 serhiy.storchaka: 4f5a466cdc left a >>>>>>> END MERGE CONFLICT line at the end of tests/wm.test, which makes the test file fail. Otherwise the branch works on Linux and Windows.
    • 💬 2026-11-05 jan.nijtmans: Thanks, Serhiy! Should be fixed now
    • 💬 2026-11-06 oehhar: Ok, thanks. From the CI run [https://github.com/tcltk/tk/actions?query=branch%3Acore-9ab6e37e-wm-forget-dot] there is nothing serious. Only the end break which only shows-up on Linux and MacOS. On MacOS, some other tests fail: [code] This does not looks to be related. But who knows, what the patch does no MacOS... Nobody tested it since one year... So, a test will be forced by merging tomorrow, if CI is clean. Tha…
    • 💬 2026-11-06 serhiy.storchaka: I merged main into the branch (e7e7a342c4) and moved the core-* tag there, so the break on Linux and in the macOS build with X11, 9427fb312d, should be gone in the next run. The wm-stackorder failures on macOS are not related to this branch: main fails the same tests with the same results since 2026-11-04.
    • 💬 2026-11-07 oehhar: Great, thanks ! At the moment, on Linux, we always have those warnings which let many tests fail: [code] On MacOS, we still have the stackorder issue, which is unrelated: [code] I tend to merge the branch now... * Main: 694e33a1 * core-9-0-branch: b7c97fb7 Merge to 8.6 gave a merge conflict. May happen later. Thanks for all, Harald
    Full verdict

    REPRODUCED (main fb166f9f73, 9.0.5, 8.6.18: "wm forget .; pack ." → SIGSEGV)

    • Platform generic (X11 tested)
    • Tested X11 with no WM: installed main (after the reference rebuild), wt main build, 9.0, 8.6. The 8c3620752c fixes (2026-03-27) only stop "wm forget ." itself from crashing on Windows and Aqua; the geometry managers still crash afterwards.
    • Cause after "wm forget ." the root window becomes a non-toplevel with no parent, and every geometry manager dereferences the parent.
    • Prototype 9ab6e37e1a.patch. This is branch [9ab6e37e-wm-forget-dot] (8ad08cf160) rebased onto main, since the branch no longer applies after a9aa5d6552. "wm forget ." raises 'window "." can not be forgotten' on unix/win/macosx, plus the man page sentence and test wm-forget-3 (31+/6- lines). Repro now gives an error and no crash; wm.test 256 passed, 0 failed.
    • Suggested ticket comment Still crashes on main/9.0/8.6 with "pack ." after "wm forget .". Branch 9ab6e37e-wm-forget-dot no longer applies after a9aa5d6552. A rebased version passes wm.test on X11; I propose to merge it to trunk and 9.0.
    ⬜b0249861182025Tcl_IsEmpty called through pointer to incorrect function typepatchREPRODUCED (main, 9.1, 9.0; confirmed by reading the code; 8.6 and 8.7 do not have this code)

    Open ticket b024986118 · 81. Portability · filed 2025-12-20

    Full verdict

    REPRODUCED (main, 9.1, 9.0; confirmed by reading the code; 8.6 and 8.7 do not have this code)

    • Platform generic
    • Tested main tip source. generic/tkConfig.c casts stub slot 690 to bool (*)(Tcl_Obj *), but Tcl (tclmain and tclDecls.h) declares int Tcl_IsEmpty(Tcl_Obj *). 9.0 has the same cast.
    • Cause generic/tkConfig.c, the #define Tcl_IsEmpty above TkObjIsEmpty (from 50a8805c).
    • Prototype b024986118.patch (2 changed lines: cast to int (*)(Tcl_Obj *) and return Tcl_IsEmpty(objPtr) != 0;. It builds warning-free, and config.test passes: 241 passed, 1 skipped. I did not rerun UBSan.) If TIP #748 changes the Tcl return type to bool, the cast must follow it.
    • Suggested ticket comment Until TIP #748 is decided, the stub cast in tkConfig.c should match Tcl's current int prototype. That is a two-line change, which I can commit to 9.0 and main.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED (code review and build only; UBSan not rerun)
    • Applies to main clean (offset 5)
    • Repro Tcl main (tcl9.1/generic/tclDecls.h:1899) still declares int Tcl_IsEmpty(Tcl_Obj *), and tkConfig.c still casts the stub slot to bool (*)(Tcl_Obj *). After the patch the cast matches. It builds without warnings.
    • Test none. The mismatch is only visible with -fsanitize=function, which needs Tcl itself built with clang UBSan.
    • Test file run config.test 242/0 (same as main)
    • Patch changes none
    Prototype verified 2026-10-24, Windows VM, generic patches
    ✅3ac12ff34c2025configure may not find cups.hpatchREPRODUCED (main, 9.0; 8.6 has no CUPS code)

    Open ticket 3ac12ff34c · 84. Unix Build · filed 2025-12-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-27 check-in f99d3c2cbf on configure-cups-cflags, in main, 9.0: Fix 3ac12ff34c: configure may not find cups.h. Add the output of "cups-config --cflags" to CUPS_CFLAGS, so cups.h is found in a non-default prefix, e.g. /usr/
    • 🔧 2026-09-27 check-in bd2f2ed155 on trunk, in main, 9.0: Fix 3ac12ff34c: configure may not find cups.h
    • 🔧 2026-09-27 check-in 94362ce151 on core-9-0-branch, in main, 9.0: Fix 3ac12ff34c: configure may not find cups.h
    • 💬 2026-09-27 serhiy.storchaka: Reproduced on OpenBSD 7.9 with the cups-libs package: "checking for cups/cups.h... no", and Tk is silently built without CUPS support. With cups-config --cflags added to CUPS_CFLAGS, cups.h is found and Tk links with libcups. Proposed fix in branch configure-cups-cflags. The OpenBSD port applies the same patch.
    • 💬 2026-09-27 jan.nijtmans: Fixed bd2f2ed155. Also backported to core-9-0-branch. Thanks for the report and the fix. And thanks for reminding!
    Full verdict

    REPRODUCED (main, 9.0; 8.6 has no CUPS code)

    • Platform build
    • Tested a fake cups-config whose --cflags returns -I<dir>, with cups/cups.h only in that dir. Unpatched main configure prints "checking for cups/cups.h... no" and leaves CUPS_CFLAGS empty. Patched configure prints "yes" and sets CUPS_CFLAGS = -DHAVE_CUPS -I<dir>. (Real libcups-dev is not installed here.)
    • Cause unix/configure.ac: CUPS_CFLAGS does not include cups-config --cflags, so AC_CHECK_HEADER and the build of tkUnixPrint.c miss non-default include dirs, e.g. /usr/local/include on BSD.
    • Prototype 3ac12ff34c.patch (configure.ac plus configure regenerated with autoconf 2.73; 1 line changed in each)
    • Suggested ticket comment Confirmed: CUPS_CFLAGS lacks cups-config --cflags, so cups.h in a non-default prefix is not found. The fix is to append it as suggested, for 9.0 and main.
    ✅042da6013e2025Pendulum demo text not visible under dark modepatchREPRODUCED by code reading (main and 8.6): on Aqua, DEF_CANVITEM_OUTLINE is NORMAL_FG = systemTextColor, which is white in dark mode, while the can…

    Open ticket 042da6013e · 76. Widget Tour · filed 2025-11-24

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-26 check-in 72d8eb4751 on demo-colors-dark-mode, in main, 9.0: Fix for 042da6013e: give the demos colours which go with their background.
    • 🔧 2026-10-28 check-in d005489a6b on trunk, in main, 9.0: Fix for 042da6013e: Pendulum demo text not visible under dark mode
    • 🔧 2026-10-28 check-in 02bf07d147 on core-9-0-branch, in main, 9.0, 8.6: Fix for 042da6013e: Pendulum demo text not visible under dark mode
    • 🔧 2026-10-28 check-in e751424387 on core-8-6-branch, in 8.6: Fix for 042da6013e: Pendulum demo text not visible under dark mode
    • 💬 2026-10-26 serhiy.storchaka: Proposed fix in branch demo-colors-dark-mode 72d8eb4751: the three text items get a fill of their own, so they no longer depend on a default foreground which a dark appearance inverts. This is the conservative half of the question Jan raised: the white background stays, which is what the phase space graph needs for the reason Christopher gave, and only the items which had no colour of their own are given one. …
    • 💬 2026-10-28 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits d005489a, 02bf07d1, and e7514243.
    Full verdict

    REPRODUCED by code reading (main and 8.6): on Aqua, DEF_CANVITEM_OUTLINE is NORMAL_FG = systemTextColor, which is white in dark mode, while the canvases have -background white. Not tested on macOS.

    • Platform Aqua (demo script; on X11 the default canvas text fill is black, so X11 is not affected)
    • Tested patched demo on main under Xvfb :124; the text items report -fill black
    • Cause library/demos/pendulum.tcl creates the text items without -fill on white canvases.
    • Prototype 042da6013e.patch (3 text items get -fill black, about 6 lines; applies to main; the same change is needed on 8.6, whose text lines differ slightly)
    • Suggested ticket comment The canvases have a fixed white background, so the text items now get an explicit black -fill too. The white background is kept because the phase graph fades from black to white.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED (by running the demo on X11; the dark-mode effect is Aqua-only and untested)
    • Applies to main clean
    • Repro ran the widget demo + pendulum.tcl with the patch: no errors, and all 3 text items (Click to Adjust..., θ, δθ) have fill=black and valid coordinates. On X11 the default text fill is already black, so there is no visible change.
    • Test none (demo script)
    • Test file run n/a
    • Patch changes none

    Notes:

    • Backups of the original patches: prototypes/{2805928fff,3119824fff,3515522fff,3611174fff}.patch.1024
    • Scripts and logs work-a2/
    • The checkout wt-a2 has a local fossil setting diff-command = gdiff-command, so a plain fossil diff > file fails. I used fossil diff -i.
    • wt-a2/unix is now configured against tcl9.1. The checkout is reverted and clean, and Xvfb :112 has been killed.
    ✅b900dc75522025Compilation using installed TCL fails if TCL source is not accessiblepatchREPRODUCED (main; 8.6 has the same Makefile.in). With TCL_SRC_DIR pointing to a directory that is not readable, cc1 fails with "Permission denied" …

    Open ticket b900dc7552 · 84. Unix Build · filed 2025-10-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-02 check-in ac6f167d30 on unix-tcl-installed-headers, in main, 9.0: Fix b900dc7552: compilation using installed Tcl fails if Tcl source is not accessible.
    • 🔧 2026-10-02 check-in 885bd22979 on win-tcl-installed-headers, in main, 9.0: Fix b900dc7552 for MinGW: compilation using installed Tcl fails if Tcl source is not accessible.
    • 🔧 2026-10-04 check-in 2e03afb4d0 on trunk, in main, 9.0: Fix b900dc7552: compilation using installed Tcl fails if Tcl source is not accessible.
    • 🔧 2026-10-04 check-in c0549801a3 on core-9-0-branch, in main, 9.0: Fix b900dc7552: compilation using installed Tcl fails if Tcl source is not accessible.
    • 🔧 2026-10-05 check-in 6b825fa561 on core-8-6-branch, in main, 9.0, 8.6: Fix b900dc7552: compilation using installed Tcl fails if Tcl source is not accessible.
    • 🔧 2026-10-05 check-in d80a5f1b14 on trunk, in main, 9.0: Fix b900dc7552 for MinGW: compilation using installed Tcl fails if Tcl source is not accessible
    • 🔧 2026-10-05 check-in 0f27f7212f on core-9-0-branch, in main, 9.0: Fix b900dc7552 for MinGW: compilation using installed Tcl fails if Tcl source is not accessible
    • 💬 2026-10-02 serhiy.storchaka: TCL_SRC_DIR is used because Tk normally builds against the Tcl source tree, and it never used TCL_INCLUDE_SPEC. Tk needs only the public Tcl headers. Proposed fixes in branches unix-tcl-installed-headers and win-tcl-installed-headers (MinGW): configure uses the installed Tcl headers if the Tcl source tree is not readable. The MSVC build (makefile.vc) still requires the Tcl sources.
    • 💬 2026-10-05 serhiy.storchaka: Backported to 8.6 in 6b825fa561.
    • 💬 2026-10-05 serhiy.storchaka: The Unix part was merged into main in 2e03afb4d0 and 9.0 in c0549801a3. The MinGW part in win-tcl-installed-headers is not merged yet.
    • 💬 2026-10-07 jan.nijtmans: The d80a5f1b14 is merged now too. Thanks! Closing
    Full verdict

    REPRODUCED (main; 8.6 has the same Makefile.in). With TCL_SRC_DIR pointing to a directory that is not readable, cc1 fails with "Permission denied" on -I$TCL_SRC_DIR/generic. When the directory does not exist, the build finds tcl.h only if the headers are on a default or CPATH include path, because Tk never uses TCL_INCLUDE_SPEC.

    • Platform build (unix)
    • Tested main, configured against a copy of an installed tclConfig.sh with TCL_SRC_DIR changed, in both cases (unreadable directory, missing directory)
    • Cause unix/Makefile.in uses -I${TCL_GENERIC_DIR} -I${TCL_PLATFORM_DIR} (from TCL_SRC_DIR) unconditionally and ignores TCL_INCLUDE_SPEC.
    • Prototype b900dc7552.patch (configure.ac + Makefile.in, 16 changed lines; unix/configure must be regenerated with autoconf 2.73). configure sets TCL_INCLUDES to the source-tree -I flags if $TCL_SRC_DIR/generic/tcl.h is readable, and to TCL_INCLUDE_SPEC otherwise. Verified: a full build succeeds with the unreadable source directory, and a build against tclmain/unix still uses the source tree.
    • Suggested ticket comment Tk always added the -I flags for the Tcl source tree and never used TCL_INCLUDE_SPEC. Proposed fix: configure uses the installed Tcl headers (TCL_INCLUDE_SPEC) when the Tcl source tree is not readable.
    ⬜280189e35d2025Duplicate interp names with [tk appname], and empty interp list with [winfo interps]patchREPRODUCED (main, 9.0, 8.6, 8.5) on X11. Marc's "correct" output was from macOS mac_send only.

    Open ticket 280189e35d · 99. Other · filed 2025-08-13

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5) on X11. Marc's "correct" output was from macOS mac_send only.

    • Platform X11 (Windows: send/winfo interps are disabled under TK_SEND_ENABLED_ON_WINDOWS, bug 858822, so the Windows output is expected)
    • Tested exercise script on Xvfb with main, 9.0, 8.6, 8.5: "before: interps=|foo| parent=foo child=foo, after: ||". It passes if the parent runs update after tk appname foo.
    • Cause unix/tkUnixSend.c:Tk_SetAppName() calls UpdateCommWindow() (XChangeProperty of TK_APPLICATION on the comm window) *after* RegClose(), which does the XFlush. The property stays in the Xlib output buffer until the next event-loop flush. Meanwhile another process's ValidateName() sees no TK_APPLICATION property, treats "foo" as stale, deletes it and takes the name.
    • Prototype 280189e35d.patch (64 lines: 7 lines of C that set riPtr->name and call UpdateCommWindow() before RegClose(), plus new test send-17.2). Verified: the repro gives "foo #2", and send-17.2 fails without the fix and passes with it. send.test otherwise unchanged (send-8.16 fails before and after, pre-existing). The workaround comment in tests/testutils.tcl childTkProcess could be dropped later.
    • Suggested ticket comment On X11, Tk_SetAppName() updated the TK_APPLICATION property of the comm window only after the registry was closed and flushed, so another process could consider the name stale and take it over. Fix: update the property before RegClose(), with test send-17.2. On Windows, send is disabled (bug 858822).
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before send-17.2 gives the duplicate name (4 of 4 runs) / after {t_s_dup #2} {t_s_dup {t_s_dup #2}} t_s_dup (4 of 4 runs)
    • Test send-17.2 (existing in patch; fails before: yes; passes after: yes)
    • Test file run send.test 78/0 (pre-existing failures: none; main 77/0)
    • Patch changes none
    ⬜a58cda72fe2025Threaded Window Creation Leads to Geometry Corruption and Crashes windows 10patchREPRODUCED (9.1 main). The crash is GDI handle exhaustion.

    Open ticket a58cda72fe · 68. Win Window Operations · filed 2025-07-24

    Full verdict

    REPRODUCED (9.1 main). The crash is GDI handle exhaustion.

    • Platform Windows
    • Tested not on Linux (no Thread package installed). Repro: win/a58cda72fe.tcl (driver runs child processes with 50 to 500 threads and prints a VERDICT line).
    • Cause xlib/ximage.c:XCreateBitmapFromData() does not check for Tk_GetPixmap() returning None, which win/tkWinPixmap.c does when CreateBitmap and CreateDIBSection both fail. TkPutImage() then gets drawable 0, which is a NULL TkWinDrawable in TkWinGetDrawableDC(). A fix would return None from XCreateBitmapFromData() when pix == None, and make GetBitmap() (tkBitmap.c) fail with an error instead of caching a None bitmap. Graceful degradation only: running out of GDI handles is an OS limit.
    • Prototype none (Windows-only, cannot verify)
    • Suggested ticket comment The crash is running out of GDI handles: Tk_GetPixmap() returns None and XCreateBitmapFromData() passes it on without checking. Tk could report an error instead of crashing, but the 10000-handle limit remains.
    Windows VM, run 2
    • Platform Windows
    • Verdict REPRODUCED (9.1 main). The crash is GDI handle exhaustion.
    • Tested the VM has no Thread package, so I built tcltest91.exe/tcltest86.exe (the tcltest target in tcl91 and tcl8.6; only new build outputs were added) and used testthread (work-win/a58tt.tcl + a58.cmd). With the ref91 build, 50–500 threads were OK. 600 threads crashed with 0xC0000005 in TkWinGetDrawableDC(NULL): the per-process GDI count reached its 10000 limit.

    After fixing that NULL case, a second crash appeared: 0xC0000094 (integer divide by zero) in Tk_GetPixmap. A thread whose TkpOpenDisplay found GetDC(NULL) failing got planes=0 and depth=0 in TkWinDisplayChanged.

    With the prototype, 600, 800 and 1000 threads run without a crash. The extra threads get the Tcl error couldn't connect to display ":0" (721 of 1000).

    • Cause after the GDI quota runs out, Tk_GetPixmap() returns None. The crash I captured was widget double-buffering (TkpDisplayButton → XCopyArea → TkWinGetDrawableDC(None)): the NULL twdPtr is dereferenced. XCreateBitmapFromData also passes None on to TkPutImage. GetBitmap() caches a None bitmap, so a second failure would panic with "bitmap already registered". Also, TkWinDisplayChanged ignores a failed GetDC(NULL).
    • Prototype a58cda72fe.patch (about 60 changed lines incl. comments; same patch as f96c6eb36b). The changes:
      • TkWinGetDrawableDC/TkWinReleaseDrawableDC treat None as a NULL DC, so drawing into it does nothing.
      • XCreateBitmapFromData returns None if Tk_GetPixmap fails.
      • GetBitmap raises cannot create bitmap "gray25" (errorCode TK BITMAP CREATE).
      • TkWinDisplayChanged keeps the old data if GetDC(NULL) fails, and TkpOpenDisplay then returns NULL.
      • Photo: see f96c6eb36b.

    Verified on 9.1: bitmap, button, canvas, entry, font, frame, imgPhoto, imgBmap, image and listbox tests all pass (0 failures).

    • Suggested ticket comment Confirmed on Windows 11. With ~600 threads the process hits the 10000 GDI object limit. Tk_GetPixmap() then returns None, which is later dereferenced (TkWinGetDrawableDC), and a thread opening its display gets depth 0 (a divide by zero in Tk_GetPixmap). A patch that makes these cases fail gracefully (drawing into None does nothing, bitmaps give a Tcl error, new threads get "couldn't connect to display") is ready. The GDI limit itself remains.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED for the panic, with a performance caveat
    • Applies to main clean. It conflicts textually with 1567354fff in TkWinDisplayChanged.
    • Repro (f96c6eb36b.tcl, 12000 photo labels): before: the process aborts (0x80000003) after ~4750 images / after: all 12000 are created without a panic. But once the GDI quota is exhausted, every 250 images take 25–40 s instead of under 1 s. The instances retry Tk_GetPixmap on every redraw, which looks quadratic. The final update+exit did not finish within 20 min, and process teardown then took several minutes. The 600-thread repro (needs a tcltest build with testthread) was not re-run.
    • Test none (needs GDI quota exhaustion; this would break the test process)
    • Test file run B3: bitmap 7/0, imgPhoto 248/0, canvas 175/0, button 416/0, frame 207/0, winfo 76/0 (pre-existing failures: none)
    • Patch changes none. Suggest throttling the retry in TkImgPhotoInstanceSetSize/ImgPhotoDisplay, e.g. retry only after the size changes or after another image is freed.
    ⬜d94a3c7b422025Selection by words: difference between mouse selection & keyboard selectionREPRODUCED (main, 9.0); not in 8.6, 8.5 or 8.7b1 (no ICU word procs there)

    Open ticket d94a3c7b42 · 18. [text] · filed 2025-06-06

    Full verdict

    REPRODUCED (main, 9.0); not in 8.6, 8.5 or 8.7b1 (no ICU word procs there)

    • Platform generic (X11 and Windows; any build with ICU available)
    • Tested X11 with "foo.bar baz". In text on main/9.0: Control-Shift-Left from after "bar" selects {foo.bar}, Control-Shift-Right from 1.0 selects {foo.bar}, double-click on "foo" selects {foo}. The entry behaves the same way. On 8.6/8.5/8.7b1 the keys select {bar} and {foo}, which matches double-click.
    • Cause generic/tkIcu.c replaces ::tk::startOfNextWord, ::tk::startOfPreviousWord and ::tk::endOfWord with ICU word-break iterators. ICU follows UAX#29, where "." between letters (MidNumLet) does not break a word. ::tk::wordBreakBefore/After, which double-click uses (text.tcl TextSelectTo, entry.tcl), are still the Tcl regex procs from library/tk.tcl that use tcl_wordchars (\w). The two sets of procs therefore disagree about what a word is.
    • Prototype none (this is a design choice: either add ICU-based ::tk::wordBreakBefore/After in tkIcu.c so double-click also uses UAX#29, or stop using ICU for word motion)
    • Suggested ticket comment Reproduced on X11 with 9.0/main but not 8.6. Keyboard word motion uses ICU (UAX#29, which does not split "foo.bar"), while double-click uses the regex-based ::tk::wordBreakBefore/After. We need to decide which definition to use and apply it to both.
    ✅c81e0ccee22025Named font late creationpatchREPRODUCED (main, 9.0, 8.6, 8.5)

    Open ticket c81e0ccee2 · 44. Generic Fonts · filed 2025-06-06

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-26 check-in 9fc3317cbb on named-font-late-creation, in main, 9.0: Fix c81e0ccee2: a named font was not used by widgets that referred to its name before it was created. Fonts cached under the name are now converted to the new
    • 🔧 2026-10-24 check-in f89023e3ee on trunk, in main, 9.0: Fix for c81e0ccee2: Named font late creation
    • 🔧 2026-10-24 check-in 3e71cad966 on core-9-0-branch, in main, 9.0, 8.6: Fix for c81e0ccee2: Named font late creation
    • 🔧 2026-10-24 check-in a344fe53e9 on core-8-6-branch, in 8.6: Fix for c81e0ccee2: Named font late creation
    • 🔧 2026-10-15 check-in bd3b75148a on named-font-late-creation, in main, 9.0: Rebase to main
    • 💬 2026-09-26 serhiy.storchaka: The patch looks good, but has a few conflicts with main, so I created a rebased branch named-font-late-creation with two more low-level tests, one of them for a regression from 8da7af2f8e. The examples from the ticket (including ttk) work on Linux, and the font and ttk tests pass. The regression: if a font name was used before font create with that name (e.g. by font actual), the unused font, which is now kept i…
    • 💬 2026-10-24 marc_culler: On macOS, * With the first script I see two labels appear in a small Helvetica font. Then one second later the third appears, also in the small Helvetica font. * With the second script I see the first two labels appear in a large Helvetica font. Then after a second they change to Courier in the same size. Then after another second a third label appears in Courier. I think that means that the problem is the same on…
    • 💬 2026-10-24 serhiy.storchaka: Thank you, Marc. Which check-in did you test? The fix was merged into main in f89023e3ee and into core-9-0-branch in 3e71cad966 today. If your build includes it, the first script should show the labels in Helvetica 30 after a second.
    • 💬 2026-10-24 nemethi: Christian, Harald, and Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits f89023e3ee, 3e71cad9, and a344fe53.
    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5)

    • Platform generic
    • Tested X11, main/9.0.5/8.6.18/8.5.19. A label with -font F2 stays at the fallback size after font create F2 -size 40 (reqwidth 55 instead of 192). The ttk::style case is the same. The fix from branch bug-c81e0cce-named-font-late-creation (ed0d5d2335/560de1252d, based on 8.6) was never merged.
    • Cause generic/tkFont.c TkCreateNamedFont only calls UpdateDependentFonts when an existing deletePending named font is recreated. TkFont cache entries that were made earlier from the same string (parsed as a font description, so namedHashPtr == NULL) are never bound to the new named font.
    • Prototype c81e0ccee2.patch (45 lines, test font-47.4). This ports chw's approach to main: UpdateDependentFonts gets an isNew flag and, for a new named font, binds cache entries with the same key (namedHashPtr, refCount++) before recomputing them. Verified: both the label and the ttk::label pick up the new font, and font.test passes (257 passed, 0 failed).
    • Suggested ticket comment Still present in main. The fix in branch bug-c81e0cce-named-font-late-creation was never merged. A port to main with test font-47.4 is ready for review.
    ⬜e94c8bc8452025X11 clipboard clear command has no effect when xclipboard is runningREPRODUCED (main, 8.6). This is largely inherent to X11 clipboard managers.

    Open ticket e94c8bc845 · 52. [clipboard] · filed 2025-05-23

    Full verdict

    REPRODUCED (main, 8.6). This is largely inherent to X11 clipboard managers.

    • Platform X11
    • Tested Xvfb, main and 8.6.18. xclipboard is not installed, so it was emulated by a Tk app that re-owns CLIPBOARD after fetching its content, which is what xclipboard does. After clipboard append aa followed by clipboard clear, both xsel -b and Tk's own clipboard get still return "aa"; a later append ("bb") is seen correctly. Without a manager, another client that asks for STRING after clipboard clear gets "CLIPBOARD selection doesn't exist or form STRING not defined".
    • Cause generic/tkClipboard.c Tk_ClipboardClear re-owns CLIPBOARD but has no targets. The manager then gets a conversion failure and keeps (and re-asserts) the old content. X11 has no protocol for "clipboard is now empty". One option is for clipboard clear to serve an empty STRING/UTF8_STRING, but real xclipboard may still ignore zero-length data (could not verify; xclipboard not installed).
    • Prototype none (behaviour change of questionable value; can't test against real xclipboard)
    • Suggested ticket comment Reproduced on X11 with an xclipboard-like manager. After clipboard clear, Tk owns CLIPBOARD with no targets, so the manager's fetch fails and it re-asserts the old content. X11 has no way to say "the clipboard is empty", so this looks like a limitation of the protocol rather than a Tk bug.
    ✅89436bd2442025Aqua/Windows : problem with ttk::progressbar in -mode indeterminatepatchREPRODUCED (main, 9.0). Regression from check-in ea571a1307 (2025-03-06, the fix for ef1bfef57e). 8.6 is not affected.

    Open ticket 89436bd244 · 70. Event Loop · filed 2025-04-09

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-02 check-in b4222a1f44 on ttk-progressbar-stopped-animation, in main, 9.0: Fix 89436bd244: stopped indeterminate ttk::progressbar redraws continuously.
    • 🔧 2026-10-09 check-in 414d5d8863 on trunk, in main, 9.0: Fix 89436bd244: Stop indeterminate ttk::progressbar redraws continuously
    • 🔧 2026-10-09 check-in b827c4f356 on core-9-0-branch, in main, 9.0: Fix 89436bd244: Stop indeterminate ttk::progressbar redraws continuously
    • 💬 2026-10-02 serhiy.storchaka: This is a regression from ea571a1307 (for ef1bfef57e): it dropped the check of the value in AnimationEnabled(), so an indeterminate progressbar is animated even when it is not started, if the style sets -period. Proposed fix in branch ttk-progressbar-stopped-animation: animate an indeterminate progressbar only while its value is positive, as before. Determinate progressbars are still not animated. New test pro…
    • 💬 2026-10-07 marc_culler: I tested this patch by running the "Progress Bar" demo while checking CPU usage with Activity Monitor. The patch made no difference. With or without the patch the %CPU value was: < 1% with no activity ~2% while the mouse is moving ~5% when running the Progress Bar demo with no mouse activity. It is not clear to me that this is a regression nor that the patch does anything to reduce CPU usage while a progress bar i…
    • 💬 2026-10-07 serhiy.storchaka: The patch only stops the animation of an indeterminate progressbar which is not started, the case Nicolas described ("you don't have to start the ttk::progressbar"). Since ea571a1307 such a progressbar is redrawn every -period milliseconds (100 in the aqua theme) forever. This can be checked without measuring CPU: the timer increments -phase. [code] Without the patch it prints about 20, with the patch 0. On Linu…
    • 💬 2026-10-07 marc_culler: The 5% CPU is with the demo activated. When the demo window is open but the progress bars are not activated and the mouse is not moving CPU usage is well under 1%.
    • 💬 2026-10-07 marc_culler: While the patch doesn't show measurable improvement It also doesn't do any harm and it does make sense and apparently it helps on Linux. So I will go ahead and merge it when I am able to do that later.
    • 💬 2026-10-09 jan.nijtmans: Fixed 414d5d8863, backported to core-9-0-branch
    Full verdict

    REPRODUCED (main, 9.0). Regression from check-in ea571a1307 (2025-03-06, the fix for ef1bfef57e). 8.6 is not affected.

    • Platform generic code, only visible where the theme sets -period (Aqua)
    • Tested X11 with the default theme shows no extra CPU (-period is 0). With a style that sets -period 10 (as aquaTheme.tcl sets -period 100), a stopped indeterminate bar with -value 0 keeps incrementing -phase and redrawing on main. Windows themes set no -period, which matches fvogel's observation.
    • Cause generic/ttk/ttkProgress.c AnimationEnabled now returns period > 0 && mode == indeterminate. ea571a1307 dropped the value > 0.0 condition, so every indeterminate bar on Aqua redraws 10 times per second even when it has not been started ([$pb stop] sets -value 0).
    • Prototype 89436bd244.patch (24 lines, test progressbar-5.1). It restores value > 0.0 in AnimationEnabled and removes the unused maximum. The test fails on unpatched main and passes with the patch; progressbar.test 21/21.
    • Suggested ticket comment Caused by ea571a1307, which dropped the value > 0 check in AnimationEnabled. Now any indeterminate progressbar animates on Aqua (period 100) even when stopped. A fix restoring the check is ready, with a test.
    ✅7c51624a962025The mouse in the room (paned windows).REPRODUCED (main, 8.6). The grid bookkeeping breaks; I did not see the pixel blow-up (0 or 32483 px) on X11.

    Open ticket 7c51624a96 · 21. [panedwindow] · filed 2025-03-17

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-11 check-in d1c2d15f2e on panedwindow-sash-keep-request, in main, 9.0: Fix 7c51624a96: moving a sash in a panedwindow stretched by the user made it request its current size, so a gridded toplevel (e.g. with text -setgrid 1 panes)
    • 💬 2026-10-11 serhiy.storchaka: The cause is not -setgrid itself: after the window was resized by the user, moving a sash made the panedwindow request its current size (the pane sizes were set to their stretched sizes). For a gridded toplevel this changes the base size, so the window grew by the stretched amount on Windows, and [wm geometry] reported wrong grid sizes on X11. ttk::panedwindow does not change its requested size on sash moves. Prop…
    • 💬 2026-10-15 jan.nijtmans: Fix merged to 94e351aa. Not yet to core-9-0-branch
    • 💬 2026-10-16 marc_culler: The demo works fine for me on macOS 26 now.
    • 💬 2026-10-17 nemethi: Backported to core-9-0-branch and core-8-6-branch by commits 0459709f and 1511f237. Closing the ticket.
    Full verdict

    REPRODUCED (main, 8.6). The grid bookkeeping breaks; I did not see the pixel blow-up (0 or 32483 px) on X11.

    • Platform generic (tested on X11)
    • Tested main and 8.6 tip, with no WM, openbox and icewm. Panedwindow with 2 text -setgrid 1 panes. The toplevel was resized from outside with XResizeWindow on the wrapper, then the sash was moved with sash place or a B1 drag.

    After the resize, every sash move makes the reported grid geometry wrong: at 350 px, wm geometry goes from 26x11 to 15x11; later, 4x11 at 329 px. The toplevel keeps its pixel size here because openbox/icewm don't re-apply the changed hints. The WMs on Aqua/Win evidently do (Marc's report).

    Without -setgrid everything stays consistent. Main's widget demo no longer uses -setgrid: 0e800a5026 (main) and cafa154041 (9.0), 2025-03-20/21. So the demo itself no longer shows the bug.

    • Cause the sash code (tkPanedWindow.c) sets the panedwindow's requested width to its current width after a sash move. This changes the toplevel's reqWidth, but the gridded text's reqGridWidth stays the same. So the grid base (reqWidth - reqGridWidth*inc) in tkUnixWm.c UpdateGeometryInfo/UpdateSizeHints moves by the sash delta. The WM's width in grid units and the size hints then no longer describe the same pixel size. This is a -setgrid design problem, not a panedwindow bug; see the unmerged getrid_setgrid branch.
    • Prototype none (needs a decision on -setgrid, i.e. the getrid_setgrid branch or a TIP, not a small fix)
    • Suggested ticket comment Reproduced on X11 (main, 8.6): after a WM resize, each sash move shifts the grid base, and [wm geometry] then reports nonsense grid sizes. It only happens with a gridded pane (-setgrid 1). The widget demo no longer uses -setgrid (0e800a5026). The real fix depends on the getrid_setgrid branch.
    ✅b298b92ba72025Possible Tk regression when loading into threadspatchREPRODUCED (main, 9.1b1, 9.0.5, 8.7b1, 8.6 tip, 8.6.13, 8.6.9)

    Open ticket b298b92ba7 · (unused) · filed 2025-02-25

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-23 check-in 423012d795 on x11-render-before-xft, in main: Fix b298b92ba7: when Tk was loaded into one thread after another on X11, a later thread died with a random X error. Xft's close-display hook frees glyphs thro
    • 🔧 2026-11-10 check-in fa8a84bb35 on trunk, in main, 9.0: Fix for b298b92ba7: Possible Tk regression when loading into threads
    • 🔧 2026-11-10 check-in d497451e32 on core-9-0-branch, in 9.0, 8.6: Fix for b298b92ba7: Possible Tk regression when loading into threads
    • 🔧 2026-11-10 check-in 9ef9af85b0 on core-8-6-branch, in 8.6: Fix for b298b92ba7: Possible Tk regression when loading into threads
    • 🔧 2026-11-10 check-in 3092453590 on xft-link-xrender, in 9.0: Fix 09ee8c51fe: static build failed with undefined reference to XRenderQueryExtension. tkUnixEvent.c calls it since b298b92ba7 was fixed, but -lXrender was
    • 🔧 2026-11-10 check-in ad54674e94 on xft-link-xrender-86, in 8.6: Fix 09ee8c51fe: static build failed with undefined reference to XRenderQueryExtension. tkUnixEvent.c calls it since b298b92ba7 was fixed, but -lXrender was
    • 💬 2026-10-23 serhiy.storchaka: This is not a regression: it happens in all versions, 8.6.9 included, once each thread creates a widget with a font. The script in the report passes in 8.6 and 9.1 only because no font is used there. When a display is closed, the close hook of Xft frees its cached glyphs through RENDER after the close hook of RENDER has already run. That registers RENDER again for the dying display, and a new display opened later …
    • 💬 2026-11-10 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits fa8a84bb, d497451e, and 9ef9af85.
    Full verdict

    REPRODUCED (main, 9.1b1, 9.0.5, 8.7b1, 8.6 tip, 8.6.13, 8.6.9)

    • Platform X11 (Xft builds)
    • Tested Thread 3.0.7 / 2.8.13, sequential thread::create {package require Tk; ...} + thread::release, without a WM.

    With the reporter's exact script, 9.0.5 and 8.7b1 fail but main, 9.1b1 and 8.6 pass. Bisect says main started passing at ea378fca79 (2026-05-28, TIP 740 merge), but that only moved the first Xft call out of the init path.

    Once each thread creates a widget with a font (pack [label .l -text hi]; update), every version fails within 3–5 iterations with a random X error ("BadRequest", "BadLength") and the process exits. This happens even when each thread has fully exited before the next one starts, so it is not a thread race.

    • Cause gdb shows that Xft's close-display hook runs after RENDER's hook. When closing its cached fonts it calls XRenderFreeGlyphSet → XRenderFindDisplay. That re-registers RENDER for the dying Display*, with codes that are then freed. The next thread's XOpenDisplay gets the same address (confirmed) and inherits the stale entry, so RENDER requests go out with a garbage major opcode.

    Close hooks run in LIFO order. Workaround in unix/tkUnixEvent.c TkpOpenDisplay: call XRenderQueryExtension() right after opening the display, so RENDER is registered before Xft and its hook runs last.

    • Prototype b298b92ba7.patch (17 changed lines, no test because the test needs the Thread package). With it, 40 iterations × {label, ttk::label+text+canvas}, with and without waiting for thread exit, run clean. The unpatched control fails. unixFont/font/unixEmbed tests pass.

    Separate issue: the process often segfaults at exit (rc 139) in every version, including 8.6, if threads with Tk were released just before. Not investigated.

    • Suggested ticket comment Reproduced on X11 in all versions (8.6.9 too) once a thread creates a widget with a font. The cause is Xft's close hook calling RENDER after RENDER's own hook, which leaves a stale entry for the freed Display*; a later display at the same address then sends garbage opcodes. Initializing RENDER in TkpOpenDisplay before Xft fixes it (prototype available).
    ✅a0c1a9c6c62024The fn RemoveFromBaseChunk is not thread safepatchREPRODUCED (main on X11). 8.6 and 9.0 on X11 don't build with base chunks and are not affected there; Windows/macOS 8.6/9.0 are affected per the code.

    Open ticket a0c1a9c6c6 · 18. [text] · filed 2024-12-12

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-24 check-in 3f22f896dc on text-basechunk-threads, in main, 9.0: Make the base chunk of the text layout thread-specific.
    • 🔧 2026-10-06 check-in 2320e60103 on trunk, in main, 9.0: Fix a0c1a9c6c6: Make the base chunk of the text layout thread-specific
    • 🔧 2026-10-06 check-in 97d5c4b3f8 on core-9-0-branch, in main, 9.0, 8.6: Fix a0c1a9c6c6: Make the base chunk of the text layout thread-specific
    • 🔧 2026-10-06 check-in 7d292334dc on core-8-6-branch, in main, 9.0, 8.6: Fix a0c1a9c6c6: Make the base chunk of the text layout thread-specific
    • 💬 2026-09-24 serhiy.storchaka: Confirmed, and it is now worse on 9.1: since 8ce823d20c base chunks are also used on Windows and X11, not only on macOS. On X11, four threads laying out text widgets crash in TkTextCharLayoutProc every time; 9.0 does not crash there. Fix on text-basechunk-threads: keep baseCharChunkPtr in thread-specific data. Disabling base chunks in threaded builds is not needed. The same change should apply to 8.6 and 9.0 for…
    • 💬 2026-09-29 serhiy.storchaka: The attached script a0c1a9c6c6_sample.tcl reproduces the crash. It needs the Thread package, and on X11 a build with bidi rendering. Several threads relayout text widgets at the same time, and it prints "ok" if it survives: wish a0c1a9c6c6_sample.tcl ?nthreads? ?seconds? On Linux/X11 with 4 threads, trunk crashed in all 40 runs, in TkTextCharLayoutProc. Branch text-basechunk-threads passed all 40 runs. Xft on its …
    • 💬 2026-10-26 oehhar: Merged to main by Jan by 2320e601. Closed. Thanks for all, Harald
    • 💬 2026-10-26 serhiy.storchaka: Also in 97d5c4b3f8 (9.0) and 7d292334dc (8.6).
    Full verdict

    REPRODUCED (main on X11). 8.6 and 9.0 on X11 don't build with base chunks and are not affected there; Windows/macOS 8.6/9.0 are affected per the code.

    • Platform generic (Win/macOS in 8.6 and 9.0; since 8ce823d20c 2026-05-20 also X11 on main)
    • Tested 4 threads, each with its own text widget doing insert/yview/update for 8 s. Main (installed) crashes with SIGSEGV in TkTextCharLayoutProc ← LayoutDLine. 8.6 tip is OK.

    Note: in main, unix/tkUnixInt.h defines TK_LAYOUT_WITH_BASE_CHUNKS as 0 when there is no bidi. tkTextDisp.c only tests #ifdef, so base chunks are enabled even with --disable-bidi.

    • Cause generic/tkTextDisp.c keeps the stretch being laid out in the process-global static TkTextDispChunk *baseCharChunkPtr (marked "TODO: Thread safety"). Concurrent layouts in different threads overwrite it.
    • Prototype a0c1a9c6c6.patch (66 changed lines, mostly the mechanical rename to tsdPtr->baseCharChunkPtr). It makes the pointer thread-specific (ThreadSpecificData + Tcl_GetThreadData in the 5 functions that use it). The 4-thread stress test passes twice; text*.test is 1759 passed / 0 failed.
    • Suggested ticket comment Confirmed: concurrent text layout in several threads crashes in TkTextCharLayoutProc (main on X11 now uses base chunks too). Making baseCharChunkPtr thread-specific fixes it (prototype available). The workaround proposed in the ticket (disable base chunks in threaded builds) is not needed.
    ✅3f5867c9972024Inconsistency between grid info, place info and pack infoREPRODUCED (main, 8.6, 8.4): for an unmanaged window, grid info and place info return "" but pack info raises "window ".f" isn't packed". Beh…

    Open ticket 3f5867c997 · 49. [grid] · filed 2024-09-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-28 check-in 320e533d8d on bug-3f5867c997, in main, 9.0: Fix 3f5867c997: Inconsistency between grid info, place info and pack info
    • 🔧 2026-11-02 check-in ebe87fbfd1 on trunk, in main, 9.0: Fix for 3f5867c997: Inconsistency between grid info, place info and pack info
    • 💬 2026-09-28 jan.nijtmans: Proposed fix 320e533d8d Anyone who would like to review?
    • 💬 2026-09-28 serhiy.storchaka: Returning an empty string from grid info and place info for an unmanaged window was not documented, so an alternative solution to make all three commands return an error should also be considered. Either way it is an incompatible change, which I would not backport to 8.6 and 9.0. Review of the branch: * doc/place.n: an unfinished sentence "If the " is left after the new text. * place-9.13 tests a window that was n…
    • 💬 2026-11-02 nemethi: Many thanks to Jan for the fix and to Serhiy for the review! Merged to main by commit ebe87fbf. Removed the unfinished sentence from doc/place.n and added a test for the case "content without a container", as suggested by Serhiy. Due to the change in the functionality of pack, no backport to core-9-0-branch or core-8-6-branch is planned.
    • 💬 2026-11-03 oehhar: Great. As complementary info to this ticket: all three cases now return the empty string. This is now documented. Thanks to the team! Harald
    Full verdict

    REPRODUCED (main, 8.6, 8.4): for an unmanaged window, grid info and place info return "" but pack info raises "window ".f" isn't packed". Behaviour has always been this way and is enforced by tests pack-17.2 and grid-5.4.

    • Platform generic
    • Tested main, 8.6 tip, 8.4.20
    • Cause tkPack.c PackInfo returns an error when the window has no packer container; the grid and place versions return an empty result. None of the three man pages documents the unmanaged case.
    • Prototype none. Making pack info return "" is a small change, but scripts may catch it to test whether a window is packed. That needs a decision (TIP/changes note) rather than a bug fix. Cheap alternative: document the current behaviour in pack.n, grid.n and place.n.
    • Suggested ticket comment Still the same in all versions, and pack-17.2 tests it. Changing pack info to return an empty list is a small change but visible to scripts. I suggest either a TIP-level decision for 9.1 or documenting the current behaviour in the man pages.
    ⬜957190f13e2024Handle writing photo image with more than 256 colors to GIF properREPRODUCED (main, 8.6). Writing a photo with 1024 colors using -format gif still fails with "too many colors". This is really an RFE.

    Open ticket 957190f13e · 41. Photo Images · filed 2024-07-25

    Full verdict

    REPRODUCED (main, 8.6). Writing a photo with 1024 colors using -format gif still fails with "too many colors". This is really an RFE.

    • Platform generic
    • Tested main, 8.6 tip
    • Cause generic/tkImgGIF.c CommonWriteGIF has no color reduction
    • Prototype none. Branch 957190f13-gif-export-more-colors (dd4205f486, 2024-07-25) already has chw's patch. Still to decide: the option name and its default (Paul wants it off by default for compatibility with 8.x and Img).
    • Suggested ticket comment Still open on main. The patch is in branch [957190f13-gif-export-more-colors]. It needs a format suboption (off by default, as Paul suggested) and a test before it can be merged.
    ⬜b3a1b9852d2024Enter key works differently in Windows and LinuxREPRODUCED (main, 8.6)

    Open ticket b3a1b9852d · 99. Other · filed 2024-06-27

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main, 8.6 tip (ttk::treeview, event generate). <Return> runs ActivateInvoke (main) or ToggleFocus (8.6), and <KP_Enter> does nothing.
    • Cause on X11 the keypad Enter is keysym KP_Enter. On Windows it is Return with the Extended modifier (TIP #158). The library bindings use only <Return>, and Entry/Text/Spinbox/TEntry even bind <KP_Enter> to "# nothing". Jan's branch bug-b3a1b9852d (aa26b81418) adds only a Treeview binding and was never merged.
    • Prototype none. Adding KP_Enter bindings widget by widget would only fix part of it. Jan's idea of mapping <KP_Enter> to Extended-Return on X11 changes behaviour and needs a TIP.
    • Suggested ticket comment Still present on main. This is a duplicate of [3119824]. A general fix (mapping KP_Enter to Extended-Return on X11, as on Windows) needs a TIP.
    ⬜f96c6eb36b2024Tk panics with large number of images on WindowspatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket f96c6eb36b · 41. Photo Images · filed 2024-06-10

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested main on Xvfb
    • Cause (likely): the Windows per-process GDI object quota (10000 by default). Each photo instance holds an HBITMAP. The failure count does not depend on image size, which points to a handle limit rather than memory. In win/tkWinPixmap.c:Tk_GetPixmap, CreateBitmap fails, then the GetDC(NULL) fallback fails too (ERROR_INVALID_PARAMETER). The NULL result makes TkImgPhotoInstanceSetSize panic. Tk cannot raise the quota, so the realistic fix is to fail gracefully (a Tcl error or an empty instance) instead of calling Tcl_Panic.
    • Prototype none (Windows only). Repro script: win/f96c6eb36b.tcl. It writes its progress and the verdict to f96c6eb36b.log next to the script. After a panic, the last "created N" line shows how far it got.
    • Suggested ticket comment This is most likely the default limit of 10000 GDI objects per process, since the count does not depend on image size. Tk should report an error instead of panicking when Tk_GetPixmap fails.
    Windows VM, run 2
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/f96c6eb36b.tcl under a watchdog (work-win/watch.ps1, which samples GetGuiResources and records message boxes). In both versions the GDI count reaches 10000 after about 4750–5000 photo+label pairs.

    9.1: the process aborts (0x80000003 = Tcl_Panic; nothing on stderr).

    8.6: the panic dialog "Fatal Error: Fail to allocate bitmap" appears (win/tkWinDraw.c:TkPutImage, a panic that main no longer has).

    With the prototype, 9.1 creates all 12000 images and exits 0: images beyond the quota are simply not displayed. work-win/bmerr.tcl: after 12000 photo labels, label -bitmap gray25 gives the Tcl error "cannot create bitmap", and a new toplevel+button+update does not crash.

    • Cause the per-process GDI quota (default 10000). Each photo instance holds an HBITMAP. When Tk_GetPixmap fails, generic/tkImgPhInstance.c:TkImgPhotoInstanceSetSize calls Tcl_Panic.
    • Prototype f96c6eb36b.patch (same file as a58cda72fe.patch). TkImgPhotoInstanceSetSize no longer panics: it leaves the instance without a pixmap, frees the old one, and retries when the instance is next used. TkImgDitherInstance returns early when there is no pixmap, and ImgPhotoDisplay already skips such an instance. The patch also covers the widget, bitmap and new-thread paths described under a58cda72fe. A Tcl error from image create/-image is not possible without an API change, because Tk_GetImage has no error path from getProc.

    Not changed: Tk_GetPixmap still tries a one-time MessageBox ("Error from CreateDIBSection"). I never saw it appear, probably because MessageBox itself needs GDI objects. It could be removed.

    • Suggested ticket comment Confirmed on Windows 11 (9.1 and 8.6): after about 4800 photo labels the process reaches the 10000 GDI object limit, and Tk panics. A patch is ready: images that cannot get a pixmap are not displayed (no panic), bitmaps give a Tcl error, and drawing into a failed pixmap does nothing. The GDI quota itself cannot be raised from Tk.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED for the panic, with a performance caveat
    • Applies to main clean. It conflicts textually with 1567354fff in TkWinDisplayChanged.
    • Repro (f96c6eb36b.tcl, 12000 photo labels): before: the process aborts (0x80000003) after ~4750 images / after: all 12000 are created without a panic. But once the GDI quota is exhausted, every 250 images take 25–40 s instead of under 1 s. The instances retry Tk_GetPixmap on every redraw, which looks quadratic. The final update+exit did not finish within 20 min, and process teardown then took several minutes. The 600-thread repro (needs a tcltest build with testthread) was not re-run.
    • Test none (needs GDI quota exhaustion; this would break the test process)
    • Test file run B3: bitmap 7/0, imgPhoto 248/0, canvas 175/0, button 416/0, frame 207/0, winfo 76/0 (pre-existing failures: none)
    • Patch changes none. Suggest throttling the retry in TkImgPhotoInstanceSetSize/ImgPhotoDisplay, e.g. retry only after the size changes or after another image is freed.
    ⬜47fbfc11392024text's cursor width on 0th columnpatchREPRODUCED (main, 8.6; the same code goes back to 8.4)

    Open ticket 47fbfc1139 · 18. [text] · filed 2024-05-27

    Full verdict

    REPRODUCED (main, 8.6; the same code goes back to 8.4)

    • Platform generic (tested on X11)
    • Tested main, 8.6 tip on Xvfb. Pixels measured with -insertwidth 4 -padx 3: at 1.0 only 2 columns are drawn (cols 3-4), at 1.1 all 4 are drawn (cols 11-14).
    • Cause generic/tkTextDisp.c:DisplayDLine copies the line pixmap to the window starting at dInfoPtr->x (and clips there when there is no double buffering). TkTextInsertDisplayProc draws the cursor starting at x - insertWidth/2, so at column 0 the left half falls into the -padx area and is lost. The existing comment ("Another special trick: copy the padding area to the left of the line ... insertion cursor") describes the intended behaviour, but the code never did it (checked back to 8.4.2).
    • Prototype 47fbfc1139.patch (30 changed lines, no new test because the check is pixel-level). When the view is not scrolled horizontally, DisplayDLine now:
      • includes the left -padx strip in the clip rectangle and in XCopyArea;
      • refills that strip with the widget background before drawing the cursor.

    When the view is scrolled, the old behaviour is kept, so glyphs scrolled out of view do not bleed into the padding. Verified: the cursor at 1.0 is 4 px wide and is erased when the cursor moves. The textDisp, text, textTag and textIndex tests pass (1438 passed, 0 failed).

    • Suggested ticket comment Confirmed on 8.6 and main. At column 0 the left half of the cursor is clipped away with the -padx area. A small fix in DisplayDLine (also copy the left padding when the view is not scrolled horizontally) is being prepared.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main clean (offset 47 lines)
    • Repro -insertwidth 4 -padx 3, pixel columns of the cursor at 1.0: before [3,4] / after [1,2,3,4]. At 1.1 it is [11..14] in both, and the strip at 1.0 is erased again when the cursor moves.
    • Test none (needs a pixel-level screen grab; Tk has no portable way to read back window pixels)
    • Test file run textDisp.test 409/0, text.test 703/0
    • Patch changes none. Review: the padding strip is refilled with the widget border before the cursor pass and the copy is widened only when curXPixelOffset == 0, which is correct. Small nit: it calls Tk_GetPixelsFromObj(padXObj) once per displayed line, which is cheap.
    Prototype verified 2026-10-24, Windows VM, generic patches
    ✅b7ded5749c2024Tk text line breaks Thai wordsREPRODUCED (main, 9.0, 8.6); really an RFE (dictionary-based line breaking)

    Open ticket b7ded5749c · 18. [text] · filed 2024-03-28

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-01 check-in e3ef73ae0b on trunk, in main, 9.0: Implements line breaks (UAX #14) and word/sentence breaks (UAX #29) for Unicode extended grapheme clusters utilizing the mojibake Unicode library, and implement
    • 💬 2026-10-01 kevin_walzer: Addressed through commit e3ef73ae to trunk.
    Full verdict

    REPRODUCED (main, 9.0, 8.6); really an RFE (dictionary-based line breaking)

    • Platform generic
    • Tested main, 9.0, 8.6 on Xvfb with Noto Sans Thai; -wrap char and -wrap word, -width 4/5/6/8
    • Cause The text layout only breaks at whitespace (-wrap word) or at any grapheme cluster (-wrap char). Thai has no spaces, so both modes break inside words, e.g. -width 5 gives ...รบ้านก|ารบ้าน... (KO KAI | SARA AA). Combining marks do stay with their base. At -width 6 the lines break correctly only because the word happens to be 6 glyphs wide. Main's HarfBuzz/SheenBidi shaping does not change where lines break.
    • Prototype none (needs UAX #14 plus dictionaries for Thai, Lao, Khmer and Burmese; Kevin's work is on branch tk_text_grapheme, "Add mojibake library" a70082422a)
    • Suggested ticket comment Still reproducible on trunk and 8.6 with both -wrap char and -wrap word, because Tk breaks only at spaces or grapheme clusters. Needs dictionary-based line breaking, which is being worked on in branch tk_text_grapheme. This is really a feature request.
    ⬜168f3ef1302024Going on words with Ctrl+arrow in text widgetpatchREPRODUCED (main, 9.0); 8.6 and 8.7 (no ICU) behave the old way

    Open ticket 168f3ef130 · 18. [text] · filed 2024-01-25

    Full verdict

    REPRODUCED (main, 9.0); 8.6 and 8.7 (no ICU) behave the old way

    • Platform generic (Linux with ICU)
    • Tested main, 9.0, 8.7, 8.6 on Xvfb, with the ticket's snippet, pressing Control-Right/Left via event generate. Stops on main and 9.0:

    Ctrl-Right |set |::var1 |value!!|!!! / |set |::var2 |"$value |+ |$value2"

    Ctrl-Left |set |::|var1 |value|!!|!!|! / |set |::|var2 |"$|value |+ $|value2|"

    • Cause generic/tkIcu.c startEndOfCmd uses the ICU word iterator, which ends a segment after every punctuation character. It only checks whether the character after a boundary is whitespace (FLAG_SPACE), not the segment's rule status (ubrk_getRuleStatus, UBRK_WORD_NONE). Runs of punctuation such as !!!!! and :: therefore produce unpredictable stops, and the TextNextWord (endOfWord then startOfNextWord) and TextPrevPos steps pair up inconsistently. Since b4a14597, Ctrl+Right goes to the start of the next word on every platform.
    • Prototype none (behaviour choice: treat a run of punctuation as one word, or skip UBRK_WORD_NONE segments; this ties into the -locale "legacy" discussion; the text -locale option now exists on main, default "C")
    • Suggested ticket comment Still present on trunk. The ICU-based word search in tkIcu.c only special-cases whitespace, so every punctuation character counts as a word (!!|!!|!, $|value). Using the ICU rule status to group punctuation runs would give the consistent stops editors have.
    2026-10-25 prototype, Linux
    • Status PROTOTYPED
    • Cause generic/tkIcu.c startEndOfCmd uses the ICU word iterator, which puts a boundary after every punctuation character. The code only checks whether the character after a boundary is whitespace, so every !, : or $ counts as its own word. The single extra icu_next()/icu_previous() step then lands in the middle of a punctuation run (value!!|!!!, value!|!!|!!, :|=, y-|>z).
    • Fix new helper SkipPunctuation(). In word mode it is applied after every iterator move (following/preceding/next/previous). It skips a boundary when the segment before it has rule status UBRK_WORD_NONE (<100, via ubrk_getRuleStatus, a newly loaded symbol) and does not end in a space, and the character after it (surrogate pairs decoded) is neither a space nor Tcl_UniCharIsWordChar. A run of punctuation then counts as one word. Checking the rule status, and not only the characters, keeps the boundary after a letter plus a combining mark. If the symbol is missing, the old behaviour is kept. The ICU used is the system libicuuc.so.78, loaded at run time by Icu_Init; the normal --disable-bidi build already loads it (::tk::startOfNextWord is the C command, and cluster.test's needsICU constraint is true). The text widget's default -locale C goes through ICU; -locale tcl8 still uses the Tcl 8 procs and is unchanged (|set ::|var1 |value!!!!! before and after). 168f3ef130.patch (54 changed lines, 42 C + 12 test)
    • Repro stops found with event generate <Control-Right>/<Control-Left> on the build in wt-a1.

    Before: R set |::var1 |value!!|!!!|, L |set |::|var1 |value!|!!|!!; R x |:= |y->|z|, L |x :|= |y-|>z.

    After: R set |::var1 |value!!!!!|, L |set |::|var1 |value|!!!!!; R x |:= |y->z|, L |x |:= y|->z.

    The second ticket line ("$value + $value2") is unchanged and was already coherent.

    • Test cluster-9.0/9.1/9.2 (startOfNextWord / endOfWord / startOfPreviousWord at every index of ab!!!cd ::ef, constraint needsICU) (fails before: yes, all 3; passes after: yes)
    • Test files run cluster 76/0, text 703/0, textIndex 190/0, entry 306/0, spinbox 324/0 with fix; main: cluster 76/3 (the new tests), others the same as with the fix
    • Notes:
      • A separate bug already exists in 8.6 and main and is not addressed here: TextPrevPos calls startOfPreviousWord $text end, and end is the last character, not the cursor position. Ctrl+Left therefore skips a one-character word right before the cursor ("a b c": from the end it goes to b, in 8.6.18 too; + $|value2 misses the stop before $).
      • endOfWord still does one next() when the boundary is followed by non-space, so endOfWord("ab!!!cd",0) is 5 (was 3). Ctrl+Right (endOfWord + startOfNextWord, from b4a14597) thus still stops only after whitespace, now consistently.
    • Windows Windows loads icu.dll/icuuc.dll (unversioned symbols). Check that ubrk_getRuleStatus resolves there (it is exported by the system icu.dll), then run cluster.test (needsICU must be true) and text.test/entry.test, and try the ticket snippet with Ctrl+Left/Right.
    • Suggested ticket comment Prototype in tkIcu.c: when moving by words, boundaries between two punctuation characters (ICU rule status UBRK_WORD_NONE) are skipped, so !!!!! or -> is one word for Ctrl+Left/Right, as in editors. With -locale tcl8 the 8.6 behaviour is still available.
    ⬜b54adf47962024-validatecommand of ttk::entry returns true, still the entry is invalidpatchREPRODUCED (main, 9.0, 8.6, 8.6.8, 8.5)

    Open ticket b54adf4796 · 07. [entry] · filed 2024-01-19

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.6.8, 8.5)

    • Platform generic
    • Tested a minimal script with focus .e1; focus .e2 (comV1 calls update; comV2 sets ::v1) leaves .e1 state = invalid on all versions listed
    • Cause generic/ttk/ttkEntry.c EntryValidateChange returns TCL_BREAK whenever the value changed during validation (VALIDATION_SET_VALUE), even when it is only revalidating (focusin, focusout, validate), where there is no edit to reject. The man page says this rejection applies only "during prevalidation". The update in comV1 runs e2's focus validation, which sets ::v1 while e1 is still VALIDATING.
    • Prototype b54adf4796.patch (21 lines; the VALIDATION_SET_VALUE rejection is limited to VALIDATE_INSERT/VALIDATE_DELETE; new test validate-2.1.1, which fails before and passes after; validate.test and entry.test pass)
    • Suggested ticket comment Confirmed. When the value changes during revalidation (focus or validate), the entry is marked invalid even though -validatecommand returned true. The documented "edit is rejected" rule should only apply to prevalidation. Prototype fix limits it to insert/delete validation.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro validate-2.1.1. Before: 0 newdata invalid. After: 1 newdata {}.
    • Test validate-2.1.1 (existing in patch; fails before: yes; passes after: yes)
    • Test file run validate.test 33/0 (main 32/0), entry.test 306/0 + 44/0 (same as main)
    • Patch changes none. Review: the reasons are INSERT, DELETE, FOCUSIN, FOCUSOUT and FORCED. Only INSERT and DELETE have a pending edit to reject, so limiting the VALIDATION_SET_VALUE rejection to those two is correct.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • b54adf4796: OK. validate-2.1.1 fails before, passes after.
    ⬜1d926db5662024Can't toggle visibility of Tk app main menu with main window maximisedpatchREPRODUCED (main, 8.6)

    Open ticket 1d926db566 · 12. Unix Menus · filed 2024-01-03

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main and 8.6 on Xvfb with openbox, icewm and fvwm; maximised with wm attributes . -zoomed 1, then toggling . configure -menu 4 times. After the menu is removed, the toplevel stays at y=menuHeight with the reduced height (e.g. 968 instead of 1003), leaving an empty strip.
    • Cause unix/tkUnixWm.c UpdateGeometryInfo. When the menubar is removed, the requested wrapper height equals the current wrapper height, because the WM refused the earlier "+menuHeight" request. It then returns early ("already just the size we want"), so no ConfigureNotify arrives and ConfigureEvent never re-positions the toplevel inside the wrapper. When not maximised, the wrapper really resizes, so the toplevel does get re-positioned.
    • Prototype 1d926db566.patch (43 lines; the toplevel/menubar layout code at the end of ConfigureEvent is moved into a new LayoutWrapper(), which the early-return branch also calls when the toplevel height does not match wrapper height minus menuHeight. The repro is fixed on all 3 WMs; the non-maximised behaviour is unchanged; wm.test, unixWm.test, menu.test and unixMenu.test give the same results as unpatched. No test added, because it needs a WM that refuses the resize.)
    • Suggested ticket comment Reproduced on X11 with openbox, icewm and fvwm. When the WM keeps the wrapper size fixed, UpdateGeometryInfo sees the wrapper already has the requested size and returns without laying out the toplevel under the (removed) menubar. Prototype fix re-lays out the wrapper in that case.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (patch was incomplete; fixed)
    • Applies to main clean (offsets)
    • Repro (zoomed .t, toggling -menu 6 times, fvwm/openbox/icewm): before, the toplevel height stays the same in every state, so adding the menu does not shrink it. With the original patch, the first removal and the first add are right, but the second removal leaves the toplevel at y=menuHeight with the reduced height. Cause: the requested height then equals configHeight (the size the WM refused earlier), so UpdateGeometryInfo takes the final else { return; } branch, where the patch had no layout. After my fix, all 6 toggles are correct on all 3 WMs.
    • Test unixWm-64.1 (new, constraints unix withWindowManager, uses pause; fails before: yes (main, and also the original patch); passes after: yes (fvwm, openbox, icewm); skipped without a WM)
    • Test file run unixWm.test 285/0, wm.test 302/0, menu.test 574/0, unixMenu.test 122/0, with fvwm and with no WM. Unpatched main under fvwm: unixWm-64.1 fails, and unixWm-51.7 failed once (flaky; it passed in other fvwm runs).
    • Patch changes added the same winPtr->changes.height != wrapper height - menuHeight → LayoutWrapper() check to the final else branch of UpdateGeometryInfo. Added test unixWm-64.1.
    ⬜af31a84c912023Replacing selection handler is ineffective.patchREPRODUCED (main, 9.0, 8.6, 8.5)

    Open ticket af31a84c91 · 53. [selection] · filed 2023-12-13

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5)

    • Platform generic (X11 selection code)
    • Tested selection handle . h1; selection handle . h2 then selection get -type UTF8_STRING returns h1's value on all versions listed (STRING returns h2)
    • Cause generic/tkSelect.c Tk_CreateSelHandler. When replacing a STRING handler it leaves an existing UTF8_STRING handler alone, even when that handler was created implicitly from the old STRING handler. The proposal on branch bug-af31a84c91 793effdb8e always replaces it instead, which clobbers an explicit UTF8_STRING handler (its own test select-14.2.2 fails).
    • Prototype af31a84c91.patch (95 changed lines, about 30 of them tests and much of the C just restructured). It remembers the old STRING proc and replaces the UTF8_STRING handler only if it looks implicit (format == UTF8_STRING and proc == old proc, the same heuristic Tk_DeleteSelHandler uses). An explicit UTF8_STRING handler is kept. New tests select-14.2 and select-14.3; select.test and clipboard.test pass.
    • Suggested ticket comment Still present on trunk. Prototype fix: when a STRING handler is replaced, also replace the implicitly created UTF8_STRING handler (detected the same way Tk_DeleteSelHandler does), but keep a UTF8_STRING handler the user set explicitly.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro selection handle . h1; selection handle . h2; selection get -type UTF8_STRING. Before: h1's value ("Replaced Original"). After: h2's ("Replaced Replaced").
    • Test select-14.2 (existing in patch; fails before: yes; passes after: yes). select-14.3 (explicit UTF8_STRING handler is kept; passes before and after, a regression guard).
    • Test file run select.test 113/0 (main 111/0), clipboard.test 42/0 (same as main)
    • Patch changes none. Note: an explicit handler created with both -type UTF8_STRING -format UTF8_STRING and the same Tcl handler proc would be treated as implicit. Tk_DeleteSelHandler uses the same heuristic.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • af31a84c91: CHANGED. select-14.2 fails on Windows before and after (form "UTF8_STRING" not defined). Windows has no implicit UTF8_STRING handler (utf8Atom is 0 under _WIN32), so the fix does not apply there. Added -constraints unix to select-14.2; select-14.3 passes on Windows. Backup: af31a84c91.patch.1024.
    ✅fde9dc23922023X error BadDrawable when using ttk::style element create with multiple displayspatchREPRODUCED (main, 9.0, 8.6)

    Open ticket fde9dc2392 · 88. Themed Tk · filed 2023-12-05

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-28 check-in 63631a59f2 on ttk-image-multiple-displays, in main, 9.0: Fix fde9dc2392: X error BadDrawable when using an image element created by "ttk::style element create" on another display. The images of an image specificatio
    • 🔧 2026-10-25 check-in 8542867141 on trunk, in main, 9.0: Fix for fde9dc2392: X error BadDrawable when using ttk::style element create with multiple displays
    • 🔧 2026-10-25 check-in 07e021e21a on core-9-0-branch, in main, 9.0, 8.6: Fix for fde9dc2392: X error BadDrawable when using ttk::style element create with multiple displays
    • 🔧 2026-10-25 check-in 5ace036554 on core-8-6-branch, in 8.6: Fix for fde9dc2392: X error BadDrawable when using ttk::style element create with multiple displays
    • 💬 2026-09-28 serhiy.storchaka: Still reproducible with 8.6, 9.0 and 9.1 (also with Xft). The fix in branch bug-fde9dc2392 was never merged. Proposed alternative fix in branch ttk-image-multiple-displays, with tests image-3.1 and image-3.2 (need TK_ALT_DISPLAY). Unlike bug-fde9dc2392, it does not add new functions to tkImage.c and does not create a new image instance on every drawing: the images are got once for every other display, for a window…
    • 💬 2026-10-26 nemethi: Serhiy, many thanks for your alternative fix, which is now merged into main, core-9-0-branch, and core-8-6-branch by commits 85428671, 07e021e2, and 5ace0365. Schelte, Christian, and Christopher, can you please confirm that this solution works for you as expected?
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform X11
    • Tested main tip, 9.0, 8.6 tip on Xvfb :122 with two screens (toplevel -screen :122.1). A ttk::entry whose layout uses an image element crashes with "X Error BadMatch, X_CopyArea". This is the same bug as BadDrawable, but on a second screen of the same display instead of a second display.
    • Cause generic/ttk/ttkImage.c:Ttk_CreateImageElement gets the element's images for Tk_MainWindow(interp). ImageElementDraw then redraws those instances into any widget's drawable, including widgets on another display, screen or visual. The fix from Christian Werner (6de380f344, 2023-12-09, +141 lines incl. TkImageForWindow/TkImageEqual in tkImage.c) and test image-3.1 (d7784f6081) sit only on branch bug-c23f79ef96 (last merged from 8.6 on 2025-11-03). They are not in main, 9.0 or 8.6.
    • Prototype fde9dc2392.patch (51 changed lines, test ttk image-3.1). A smaller, local fix: ImageData remembers the window its images were acquired for. When the widget's display, screen or visual differs, ImageElementDraw gets a temporary instance with Tk_GetImage(NULL, tkwin, name, ...) and frees it after drawing. Photo instances are cached, so this is cheap. With the patch the repro passes and all ttk tests pass (2043 tests, run with TK_ALT_DISPLAY=:122.1). The new test crashes the installed main. Only image elements are covered: ttkCache (TILE_07_COMPAT -map) still uses the main window. The rest of c23f79ef96 is out of scope.
    • Suggested ticket comment Still crashes on main/9.0/8.6 (reproducible with a second screen, e.g. <tt>toplevel .t -screen :0.1</tt>): the fix is still only on branch bug-c23f79ef96. A smaller alternative is to acquire a per-window image instance in ImageElementDraw when the widget's display/screen/visual differs from the main window's.
    ⬜b8fcb265292023some inconsistencies in text entry widgets and textpatchREPRODUCED (main, 8.6): partly by design, partly a real bug

    Open ticket b8fcb26529 · 07. [entry] · filed 2023-09-12

    Full verdict

    REPRODUCED (main, 8.6): partly by design, partly a real bug

    • Platform generic (bindings and docs)
    • Tested main tip and 8.6 tip on Xvfb, using event generate with the ticket's 5 examples plus BackSpace and Control-h.
      • Typed characters: entry and text replace the selection only when the insert cursor is inside or adjacent to it. ttk::entry always replaces it. This is intentional per the code comments, but the man pages don't say so.
      • BackSpace: entry and ttk::entry delete any selection (ttk also moves the cursor to sel.first). text deletes the selection only when the cursor is inside or adjacent to it, which text.n contradicts.
      • A real bug: Text <Control-h> ignores the selection completely and deletes insert-1c. Starting from "abc def ghi" with sel 1.0-1.3 and the cursor at 1.3, you get "ab def ghi". It also doesn't respect grapheme clusters, although text.n says Control-h behaves like Backspace.
    • Cause library/text.tcl, the Text <Control-h> binding is a separate simplistic script and not the BackSpace logic. Docs: doc/text.n, doc/entry.n and doc/ttk_entry.n describe insertion and deletion imprecisely.
    • Prototype b8fcb26529.patch (46 changed lines, test text-39.1). Control-h now does the same as BackSpace (when !tk_strictMotif). text.n/entry.n now say that typed text replaces the selection only when the cursor is inside or adjacent to it. ttk_entry.n now says that typed text always replaces the selection. text.n's Backspace/Control-h item gets the same inside-or-adjacent condition. text.test and entry.test pass.
    • Suggested ticket comment The differences in typed-text replacement are intentional; the man pages now describe them. Text <Control-h> really was inconsistent (it ignored the selection and grapheme clusters) and now matches BackSpace; test text-39.1.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main ported. The tests/text.test hunk was rejected because main gained text-39.1 (bug 991849) and text-46.x (RTL). The docs, library and ttk_entry.n hunks apply (offset -1).
    • Repro "abc def ghi", sel 1.0-1.3, insert 1.3, <Control-h>. Before: "ab def ghi". After: " def ghi".
    • Test text-41.1 (renumbered from text-39.1, which main now uses; text-40.1 is taken by the 3119824fff prototype). Placed after text-39.1 and before the RTL block. Fails before: yes; passes after: yes.
    • Test file run text.test 704/0 (main 703/0), entry.test unchanged
    • Patch changes test renumbered to text-41.1 and moved. Backup: b8fcb26529.patch.1024.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • b8fcb26529: OK. text-41.1 fails before, passes after. Repro: Control-h with a selection in text used to delete one character (aaa bb ccc); now it deletes the selection (aaa ccc). entry already did.
    ⬜35f86988a92023Changing background color of labelframe causes all widgets to redrawpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

    Open ticket 35f86988a9 · 08. [(label)frame] and [toplevel] · filed 2023-03-20

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

    • Platform generic (reported on Windows, same on X11)
    • Tested main, 9.0.5, 8.6.18, 8.5.19, 8.4.20 on Xvfb, no WM. .f configure -background X on a gridded labelframe with -width/-height and a child produces 2 <Configure> per reconfigure (50x50 then back to the propagated 22x48); patched build: 0.
    • Cause generic/tkFrame.c:FrameWorldChanged calls Tk_GeometryRequest(-width, -height) on every configure (and on labelwidget/font changes), overriding the size propagated by grid/pack; the geometry manager then shrinks it back, so the frame and all siblings get resized twice.
    • Prototype 35f86988a9.patch (86 +/- lines incl. new helper FrameGeometryRequest with header comment and test; ~45 code lines). Adds a GEOMETRY_MASK typeMask to -width/-height, passes a mask to Tk_SetOptions in ConfigureFrame and only requests -width/-height when they are specified (plus once at creation). New test frame-6.4 (fails on unpatched main: 0 150 0 150, passes patched). frame, grid, pack, place tests: no new failures (pack-18.x, place-8.x and wm failures are pre-existing without a WM).
    • Suggested ticket comment The frame re-requests its -width/-height on every configure, overriding the size propagated from its content, so the geometry manager resizes it twice. Reproduced on X11 too. Fix: only request -width/-height when these options are given (as suggested by fvogel), test frame-6.4.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro work-a2/35f8.tcl (gridded labelframe -width/-height 50 with a child; 3 background changes). Before: 6 <Configure> events. After: 0.
    • Test frame-6.4 (existing in patch; fails before: yes (0 150 0 150); passes after: yes)
    • Test file run frame.test 208/0, grid.test 206/0, pack.test 199/0, place.test 56/0 (pre-existing failures: none; main has the same totals minus frame-6.4)
    • Patch changes none. Review: the -width/-height request is now made only at creation and when those options are set, not on font or worldchanged events. That matches the intent, and I found no regression in the geometry-manager tests.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 35f86988a9: OK. frame-6.4 fails before, passes after; frame.test 208/0.
    ⬜be6cf582412023menu across top of window always same fontREPRODUCED (9.1 main, 8.6 tip), limitation / RFE (same as 11bd4a03a0)

    Open ticket be6cf58241 · 47. Win Fonts · filed 2023-02-17

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip), limitation / RFE (same as 11bd4a03a0)

    • Platform Windows
    • Tested code reading only. win/tkWinMenu.c:ReconfigureWindowsMenu adds MENUBAR entries as MF_STRING (system-drawn), only non-menubar entries are MF_OWNERDRAW, so -font cannot apply to the menubar. Repro: win/be6cf58241.tcl (compares menubar height for 8pt vs 40pt -font; equal => reproduced).
    • Cause by design, Windows draws the menubar itself (as oehhar said).
    • Prototype none (would need owner-drawn menubar = behaviour redesign). Possibly document in menu.n that menubar fonts/colors are platform-controlled on Windows and macOS.
    • Suggested ticket comment The menubar items are drawn by Windows (MF_STRING, not owner-drawn), so -font and colours are ignored there by design. This could be documented in menu.n; otherwise close as Works as designed.
    Windows VM, run 2
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip), limitation / RFE (same as 11bd4a03a0)
    • Tested win/be6cf58241.tcl measures the top decoration height with a menubar at -font {Helvetica 8} and at {Helvetica 40}. It is 51 px for both, on 9.1 and 8.6. Run 1 also showed that menubar.tcl draws Courier 24 bold in the system font.
    • Cause win/tkWinMenu.c:ReconfigureWindowsMenu adds menubar entries as MF_STRING items, which Windows draws itself, so -font, -bg and -fg are ignored.
    • Prototype none (would need owner-drawn/UAH menubar painting, see 11bd4a03a0)
    • Suggested ticket comment Confirmed: the Windows menubar is drawn natively (MF_STRING items), so -font is ignored, like the colours in 11bd4a03a0. Supporting it would need owner-drawing of the menubar. Changing this to a feature request, or documenting it in menu.n.
    ✅f458d442042022--disable-shared build fails against --enable-shared build of TclpatchREPRODUCED (main; same Makefile.in code in 9.0 and 8.7; 8.6 has no zipfs)

    Open ticket f458d44204 · 84. Unix Build · filed 2022-09-19

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-20 check-in 1ba9e09bd0 on static-tk-shared-tcl, in main, 9.0: Fix f458d44204: --disable-shared build failed against --enable-shared build of Tcl: the zip file of Tk was created from tclsh, which only contains the zip arc
    • 🔧 2026-10-20 check-in a08fcbc978 on trunk, in main, 9.0: Fix f458d44204 to build --disable-shared Tk against --enable-shared Tcl
    • 🔧 2026-10-20 check-in 1e956ebe17 on core-9-0-branch, in main, 9.0: Fix f458d44204 to build --disable-shared Tk against --enable-shared Tcl
    • 💬 2026-10-20 serhiy.storchaka: Fixed in branch static-tk-shared-tcl (1ba9e09bd0). With --disable-shared the zip file of Tk was created from a copy of tclsh, so that it contains the zip archive of Tcl too. But with a shared Tcl, the zip archive of Tcl is in libtcl9.1.so, not in tclsh. Now tclsh is only used if Tcl is not shared. wish finds tk_library in its own zip archive and tcl_library in libtcl9.1.so. The build against a static Tcl works a…
    • 💬 2026-10-20 dgp: Confirmed this branch fixes the reported build failure.
    • 💬 2026-10-20 dgp: Fix accepted onto the 9.0 and 9.1 branches.
    Full verdict

    REPRODUCED (main; same Makefile.in code in 9.0 and 8.7; 8.6 has no zipfs)

    • Platform build (Unix)
    • Tested main Tk --disable-shared against shared tclmain: zip error: Zip file structure invalid (../libtk9.1b1.zip), make fails.
    • Cause unix/Makefile.in ${TK_ZIP_FILE} rule: for ZIPFS_BUILD=2 (static Tk) it starts the zip from a copy of ${TCL_BIN_DIR}/tclsh expecting an embedded tcl_library zip; a shared-Tcl tclsh has none (the zip is in libtcl.so), so zip -A and then zip -rq fail.
    • Prototype f458d44204.patch (8 lines: AC_SUBST(TCL_SHARED_BUILD) in configure.ac + regenerated configure with autoconf 2.73, and the tclsh copy in Makefile.in is done only when TCL_SHARED_BUILD=0). Verified: static wish builds, mounts //zipfs:/app/tk_library and works; normal shared build unaffected.
    • Suggested ticket comment The static Tk build starts its zip from a copy of tclsh, which only contains the Tcl library zip when Tcl is static too. Fix: skip that copy when Tcl is a shared build; wish then gets only tk_library appended and libtcl.so provides its own zip.
    ✅290ce5f4602022TK performance on X11 with Forwarding (via SSH) decreased when upgrading from 8.6.10 to 8.6.11patchREPRODUCED (8.6 tip, 8.6.13, 8.6.15, 8.7, 9.0, main built with --disable-bidi). Regression came in 8.6.11 with check-in 7f20966df9 (2020-04-14, "Wo…

    Open ticket 290ce5f460 · (unused) · filed 2022-07-28

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-02 check-in 2b118379ee on x11-error-handler-sync, in main, 9.0: Fix 290ce5f460: Tk performance on X11 with forwarding decreased in 8.6.11.
    • 💬 2026-10-02 serhiy.storchaka: The slowdown comes from 7f20966df9 (in 8.6.11), not from b101d48433: text measurement with Xft is now wrapped in an X error handler, and every tenth deletion of an error handler does a round trip with the X server to catch errors for the outstanding requests. For the example in the report it adds about 4 times more round trips in a build without bidi. Proposed fix in branch x11-error-handler-sync: skip the rou…
    • 💬 2026-10-16 jan.nijtmans: Fixed in 3d1d80d012, core-9-0-branch and core-8-6-branch Many thanks!
    Full verdict

    REPRODUCED (8.6 tip, 8.6.13, 8.6.15, 8.7, 9.0, main built with --disable-bidi). Regression came in 8.6.11 with check-in 7f20966df9 (2020-04-14, "Workaround for 3767882e06: BadLength..."). The installed main build does not show it.

    • Platform X11
    • Tested counted blocking X round trips (poll calls, strace) for the ticket's 20x20 ttk::button grid on Xvfb. 8.6.8: 1716, 8.6.9: 1864, 8.6.13: 2653, 8.6 tip: 2597, 8.7: 2586, 9.0: 2628, main --disable-bidi: 2528, installed main: 1735. 8.6.10 is not installed. Stack traces put the extra ~800 round trips in Tk_MeasureChars -> Tk_DeleteErrorHandler -> XSync.
    • Cause 7f20966df9 wraps every Tk_MeasureChars (and some other Xft calls) in unix/tkUnixRFont.c in Tk_CreateErrorHandler/Tk_DeleteErrorHandler. Tk_DeleteErrorHandler (generic/tkError.c) calls XSync on every 10th deletion when requests are outstanding (added 2018 in 3233521135). So about 1 of every 10 text measurements costs a full round trip. This is not related to b101d48433 (Xft locking), which Jan suggested.
    • Prototype 290ce5f460.patch (tkError.c, +6/-2). Skips the XSync when the handler was active for no requests at all (lastRequest < firstRequest), which is the usual case for Xft text extents once glyphs are cached. Round trips drop from 2528 to 1735. font, unixFont and canvText tests pass. send-8.16, winfo-9.4 and winfo-9.7 fail with and without the patch (no WM). No automated test (round-trip count).
    • Suggested ticket comment The slowdown comes from 7f20966df9: Tk_MeasureChars is now wrapped in an X error handler, and Tk_DeleteErrorHandler does an XSync every 10th call, which adds about 40% more round trips for this example. The fix is to skip that XSync when no request was sent while the handler was active.
    ⬜3c4435c9ae2022tk entry on linux a paste does not replace a selectionpatchREPRODUCED (main, 8.6, 8.4). This is intentional old Motif-style behaviour.

    Open ticket 3c4435c9ae · 07. [entry] · filed 2022-05-10

    Full verdict

    REPRODUCED (main, 8.6, 8.4). This is intentional old Motif-style behaviour.

    • Platform X11
    • Tested event generate <<Paste>> with a selection on Xvfb. On main and 8.6, entry, spinbox and text insert after the selection ("aaa bbbXYZ ccc"), while ttk::entry replaces it ("aaa XYZ ccc"). 8.4 has the same behaviour for entry and spinbox.
    • Cause library/entry.tcl <<Paste>>, library/spinbox.tcl <<Paste>> and library/text.tcl tk_textPaste delete the selection only when [tk windowingsystem] ne "x11". ttk::entry::Paste always calls PendingDelete.
    • Prototype 3c4435c9ae.patch (19+/13-). Drops the x11 exception in all three places so that paste replaces the selection everywhere, as ttk::entry and modern X11 toolkits do. New test entry-25.4. entry, spinbox and text tests pass. This is a behaviour change on X11 (policy decision).
    • Suggested ticket comment The X11 exception in the entry, spinbox and text <<Paste>> bindings is deliberate (Motif), but ttk::entry and current X11 toolkits replace the selection on paste. The proposed change removes the exception for consistency.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main ported (the entry.test hunk was rejected because entry-26.1 now follows entry-25.3; entry-25.4 now goes right after entry-25.3)
    • Repro before: <<Paste>> with "bbb" selected gives entry/spinbox/text "aaa bbb cccXYZ" (ttk::entry gives "aaa XYZ ccc") / after: the selection is deleted and the text is inserted at the insert cursor (entry-25.4 with icursor at the selection end gives "aaa XYZ ccc")
    • Test entry-25.4 (from the patch, ported; fails before: yes; passes after: yes)
    • Test file run entry.test 307/0, spinbox.test 324/0, text.test 703/0 (also ttk/entry 44/0, ttk/spinbox 37/0); no failures before or after apart from entry-25.4 before
    • Patch changes the Entry and Spinbox <<Paste>> bindings now read the CLIPBOARD first (if {![catch {::tk::GetSelection %W CLIPBOARD} tk::Priv(data)]}, the same pattern as <<Cut>>/<<Copy>>) and only then delete the selection. Without this, now that X11 also deletes, pasting from an empty clipboard deleted the selection and inserted nothing. Checked: an empty clipboard now leaves "aaa bbb ccc". tk_textPaste already reads the clipboard first. Remaining (unchanged) design point: like the existing Windows/macOS behaviour, the selection is deleted even when the insert cursor is elsewhere, and the text goes to the cursor (ttk::entry moves the cursor to sel.first first). Original kept as 3c4435c9ae.patch.1024.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 3c4435c9ae: OK. entry-25.4 passes before and after on Windows (Windows already deleted the selection on paste). Repro: paste replaces the selection in entry/spinbox/text before and after. With an empty clipboard, entry and spinbox used to delete the selection (aaa ccc); now they keep it (aaa bbb ccc). text kept it either way. entry 307/0, spinbox 324/0.
    ✅f9d316a5e02022linux menu problem if menu near the bottom of screenpatchREPRODUCED (main, 8.6, 8.4)

    Open ticket f9d316a5e0 · 12. Unix Menus · filed 2022-05-10

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-29 check-in c81290221a on menu-post-under-pointer, in main, 9.0: Fix [470331], [680660], 1596662fff, f9d316a5e0: a menu posted over its menubutton or menubar entry invoked the entry under the pointer on the release of the
    • 🔧 2026-09-30 check-in 0f86666370 on trunk, in main, 9.0: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in 5c47ae3e9c on core-9-0-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in 3aacabb15f on core-8-6-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in a33031d2f8 on menu-post-under-pointer, in main, 9.0: Add a widget demo for popup (context) menus posted with tk_popup, with menus of different sizes.
    • 💬 2026-09-30 serhiy.storchaka: Duplicate of [470331]. Proposed fix in branch menu-post-under-pointer: as in Firefox, a menu which does not fit below the menubar is posted above it. A menu which still covers the pointer does not invoke an entry on the release of the click which posted it. See [470331] for details.
    • 💬 2026-09-30 nemethi: Serhiy, many thanks! The fix is now in main, core-9-0-branch, and core-8-6-branch; see commits 0f866663, 5c47ae3e, and 3aacabb1.
    Full verdict

    REPRODUCED (main, 8.6, 8.4)

    • Platform X11
    • Tested Xvfb, no WM. Toplevel with a 20-item menubar cascade near the bottom of the screen. Press on the menubar entry (warped pointer + event generate), then release without moving: "Item 15" is invoked on main, 8.6 and 8.4. A menubutton with -direction below has the same problem: the menu is placed at 522..1024 on top of the button at 964.
    • Cause generic/tkMenuDraw.c AdjustMenuCoords always places the menubar dropdown below the entry. unix/tkUnixMenu.c TkpPostTearoffMenu then clamps it up onto the menubar, under the pointer. For menubuttons, the "show as above" fallback in library/menu.tcl tk::PostMenubuttonMenu has a wrong formula (vrooty + vrootheight + rooty - mh), which is then clamped back over the button.
    • Prototype f9d316a5e0.patch (74+/3-, most of it tests). AdjustMenuCoords posts the menubar cascade above the menubar when it does not fit below but fits above. The menubutton fallback is fixed to rooty - mh (only when it fits). Tests menuDraw-17.3 and menubutton-10.1, which fail on unpatched main. The dropdown now opens above and stays posted after release. menu, menuDraw and menubut tests pass.
    • Suggested ticket comment Reproduced on X11. The menubar dropdown and -direction below menubutton menus are shifted up over the button when they don't fit below, so the button release invokes an item. The fix is to post them above the button when they fit there.
    ⬜1bddba50f82022<Double-1> delivery difference between linux and windowspatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 1bddba50f8 · 69. Events · filed 2022-04-21

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows (also Aqua)
    • Tested not reproducible on Linux/X11 by design (X server timestamps); Windows repro script win/1bddba50f8.tcl (manual double-click, prints VERDICT)
    • Cause win/tkWinX.c → TkWinPointerEvent → generic/tkPointer.c:InitializeEvent sets ButtonPress time = TkGetMS() (GetTickCount at *processing* time), not the time the click happened. When a <<ListboxSelect>> handler blocks 1.5 s, the queued second click gets a timestamp ~1.5 s later than the first, so it fails the double-click interval check in tkBind.c. X11 uses server timestamps, so the double click survives. macOS has the same pattern (TkGetMS in tkPointer), matching fvogel's observation.
    • Prototype none (needs Windows to verify; fix = pass the native message time — GetMessageTime() on Windows, [NSEvent timestamp] on macOS — down to InitializeEvent, e.g. via a platform hook/static "current pointer event time" set by TkWinPointerEvent before Tk_UpdatePointer)
    • Suggested ticket comment On Windows/macOS, button events are timestamped when Tk processes them (TkGetMS in tkPointer.c InitializeEvent) instead of when the click occurred, so a blocking handler makes the second click look too late for a double click. Using GetMessageTime() for pointer events should make Windows behave like X11.
    Windows VM, run 2
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/1bddba50f8.tcl, driven by work-win/w1bdd.tcl plus dblclick.ps1 (SetCursorPos + mouse_event, a real double click 100 ms apart in the interactive session).

    9.1: ListboxSelect=2 Double-1=0, press times differ by 1516 ms. 8.6: same, 1547 ms.

    With the prototype: ListboxSelect=1 Double-1=1. Control run: clicks 700 ms apart (Windows double-click time is 500) are correctly not a double click. bind, event, listbox, entry, text, button, scale, scrollbar and winfo tests: 2841 tests, 0 failed.

    • Cause generic/tkPointer.c:InitializeEvent stamps pointer events with TkGetMS() (GetTickCount) at processing time. After the handler blocks for 1.5 s, the queued second click looks 1.5 s late. win/tkWinX.c:GenerateXEvent does the same for key and wheel events.
    • Prototype 1bddba50f8.patch (10 changed lines): uses (Time)(DWORD)GetMessageTime() in InitializeEvent (under #ifdef _WIN32; tkPointer.c is only built on Windows and macOS) and in GenerateXEvent. There is no automated test, because it needs real queued input messages.

    Caveat: for Tk_UpdatePointer calls that do not come from a message (the mouse timer, warp), GetMessageTime() returns the time of the last retrieved message. That time is slightly older, which is harmless for crossing events. macOS needs the equivalent ([NSEvent timestamp]); not done.

    • Suggested ticket comment Confirmed on Windows (9.1, 8.6): button events are timestamped when Tk processes them (TkGetMS in tkPointer.c), not when the click happened. A patch that uses GetMessageTime() for pointer and key events on Windows is ready; with it the double click survives a blocking handler.
    Prototype verified 2026-10-24, Windows VM
    • Status FIXED-ON-MAIN (2535a940fe, 2026-09-25: merge of 6498c68552 "Fix [1954237]", new TkpGetEventTime() = GetMessageTime() used in tkPointer.c and tkWinX.c)
    • Applies to main no (both hunks are rejected; superseded by main)
    • Repro main e1bf: a real double click 100 ms apart with the 1.5 s blocking handler gives ListboxSelect=1 Double-1=1, OK
    • Test none
    • Test file run not run
    • Patch changes none. The patch is obsolete, and the ticket is a duplicate of [1954237].
    🟢de6ef8276c2022imageName configure -gamma - has no effect on macOSpatchREPRODUCED on X11 (main, 9.0, 8.6) for images with partial transparency; on Aqua (and Windows on main since 3bbe54bc74, 2026-06-15, which defined T…

    Open ticket de6ef8276c · 41. Photo Images · filed 2022-02-14

    🟢 Since triage (updated 2026-11-11)

    Status now: Open (Fixed)

    • 🔧 2026-10-18 check-in 8316cf936d on aqua-photo-gamma-palette, in main, 9.0: Fix de6ef8276c: photo -gamma and -palette had no effect on macOS. Apply the fix for 32f2e19de8 to all platforms which render RGBA images, not only Windows:
    • 🔧 2026-10-18 check-in dee0685394 on trunk, in main, 9.0: Fix de6ef8276c: photo -gamma and -palette had no effect on macOS.
    • 💬 2026-10-17 chrstphrchvz: fe942372 and f215a9fd were resolved: color channel order assumptions have been corrected for some time, and XGetImage()/XCopyArea() should now work again on Tk Aqua. I thought the remaining step to fix this ticket's issue would simply be to remove #define TK_CAN_RENDER_RGBA from tkMacOSXPort.h, which would make Aqua go back to drawing instance pixmaps, rather than drawing directly from model data (which I prev…
    • 💬 2026-10-18 serhiy.storchaka: Proposed fix in branch aqua-photo-gamma-palette, as suggested by Christopher: the fix for 32f2e19de8 now applies to macOS too, not only Windows, so images without partial transparency and with a non-default -gamma or -palette are drawn from the pixmap of the instance. Tests imgPhInstance-7.1 and 7.2 now run on Aqua too. Marc, could you check them on macOS?
    • 💬 2026-10-18 marc_culler: Yes, the fix works. I attached a screenshot which shows different colors in the two rectangles, and the imgPhInstance tests all pass. Serhiy, it would be better to send me an email. It was more or less an accident that I noticed your request.
    • 💬 2026-10-18 serhiy.storchaka: Thank you for testing, Marc! Merged into main in dee0685394. I will send an email next time.
    • 💬 2026-10-18 chrstphrchvz: Reopening to ask: will this be fixed for 9.0?
    • 💬 2026-10-18 serhiy.storchaka: I do not know. The other fixes for -gamma and -palette (32f2e19de8 and [481255]) were not backported to 9.0 either.
    • 💬 2026-10-19 chrstphrchvz: Rather than backporting changes prompted by TIP 754, my suggestion for 9.0 remains to disable the code in TkImgPhotoDisplay() currently being enabled by TK_CAN_RENDER_RGBA; probably the simplest way would be to disable the definition in tkMacOSXPort.h and leave TkpPutRGBAImage() unused. Is this acceptable? ``` --- macosx/tkMacOSXPort.h +++ macosx/tkMacOSXPort.h @@ -138,12 +138,14 @@ #define TK_HAS_DYNAMIC_CO…
    • 💬 2026-10-19 marc_culler: The original reason for the flag TK_CAN_RENDER_RGBA was to allow macOS to use the GPU to blend an RGBA image instead of doing that with slow code on the CPU. Are we proposing to go back to doing all blending with inefficient CPU code?
    • 💬 2026-10-19 kevin_walzer: I agree with Marc. The flag TK_CAN_RENDER_RGBA is a key part of current Tk on macOS, and Wayland borrows this design.
    • 💬 2026-10-19 chrstphrchvz: What I have read elsewhere is that while Apple has other frameworks which are more GPU-focused, Core Graphics mostly uses CPU. On Aqua, alpha blending in TkpPutRGBAImage() is done by CGContextDrawImage(), which reportedly only uses CPU. It is likely just far more optimized (possibly using CPU SIMD) than Tk's BlendComplexAlpha(). If using BlendComplexAlpha() on Aqua is unacceptable, then an alternative which does n…
    • 💬 2026-10-20 marc_culler: I think you are correct that CGContextDrawImage may not use the GPU. And I also think you are correct that its code for alpha blending is likely to be much more highly optimized than BlendComplexAlpha(). The Wayland port does use the GPU, and having the TK_CAN_RENDER_RGBA flag available was a big help there. I tried to think of a use for the -palette option. The only use I could come up with was converting a color…
    Full verdict

    REPRODUCED on X11 (main, 9.0, 8.6) for images with partial transparency; on Aqua (and Windows on main since 3bbe54bc74, 2026-06-15, which defined TK_CAN_RENDER_RGBA in tkWinPort.h) for all images

    • Platform generic (Aqua, Windows main, and X11 for images with partial alpha)
    • Tested X11 main/9.0/8.6: opaque image with -gamma 2 → (255,209,0) correct; same image with one semi-transparent pixel → (255,170,0), gamma ignored. Both XRender and software-blend (Xvfb -extension RENDER) paths affected.
    • Cause generic/tkImgPhInstance.c:TkImgPhotoDisplay — gamma is only applied in the instance's color table/pixmap, but the TK_CAN_RENDER_RGBA path (macOS, Windows main), the X11 XRender path and BlendComplexAlpha draw the model's pix32 directly.
    • Prototype de6ef8276c.patch (70 lines incl. comment header; new GetGammaPixels() returns a gamma-corrected RGBA copy when gamma != 1.0, used in all three direct-RGBA paths). Verified on X11 with and without RENDER; imgPhoto.test passes. No test added (no pixel read-back available in the test suite). Needs checking on macOS/Windows.
    • Suggested ticket comment The direct RGBA drawing paths (macOS, Windows on main, and X11 for images with partial alpha) use the model data and ignore the instance gamma; the X11 part reproduces on 8.6 too. Prototype fix applies the gamma to a copy of the RGBA data before drawing.
    ⬜9fd94c63b02021Text containing "wide" character not correctly refreshedpatchREPRODUCED (main, 8.6 on X11)

    Open ticket 9fd94c63b0 · 05. Canvas Items · filed 2021-12-29

    Full verdict

    REPRODUCED (main, 8.6 on X11)

    • Platform generic
    • Tested X11 main, 8.6 with Liberation Serif Italic 200: text item bbox x=161..240 but glyph ink extends to x=270; after itemconfigure -fill red 561 black pixels remain at x=240..270.
    • Cause generic/tkCanvText.c:ComputeTextBbox uses the text layout's advance width, which excludes glyph overhang (right/left side bearing); the redraw area (item bbox) therefore misses the overhanging ink. Same limitation exists on all platforms that report advance widths.
    • Prototype none (a proper fix needs ink extents from the font backend, e.g. XftTextExtents x/width, plus separating the redraw region from the reported bbox, which many tests depend on; not small)
    • Suggested ticket comment Reproduced on X11 with an italic font. The canvas redraws only the item bbox, computed from advance widths, so glyph overhang is not refreshed. Fixing it needs ink-extent metrics from the font code; workaround: append a thin space (U+2009/U+202F).
    2026-10-25 prototype, Linux
    • Status PROTOTYPED
    • Cause generic/tkCanvText.c:ComputeTextBbox builds the item bbox from the layout's advance width, so ink that sticks out (italic or overhanging glyphs) lies outside it. tkCanvas.c uses that bbox for the damage area (EventuallyRedrawItem and the "old area" redraws in dchars/insert/rchars/imove), and DisplayCanvas uses it to choose which items to redraw. So the overhang is (a) never erased on recolor/move/delete, and (b) clipped away when the item is first drawn after the window is mapped. Point (b) was not in the ticket.
    • Fix new MODULE_SCOPE TkCanvTextRedrawMargin() in tkCanvText.c returns (ascent+descent)/4 for a visible text item. New static GetItemRedrawArea() in tkCanvas.c returns the item bbox grown by that margin on all four sides (this also covers -angle). It is used in EventuallyRedrawItem, in the 4 old-area sites and in the DisplayCanvas item-selection loop. bbox, find and the item header are unchanged. Measured overhang at -100px is at most 17px (DejaVu/Noto/Liberation/Free italic); the margin there is 25-34px. Not changed: the image subcommand (DrawCanvas) still selects items by bbox. 9fd94c63b0.patch (162 changed lines: about 60 are function header comments, 20 are the test, and the EventuallyRedrawItem change is mostly renaming itemPtr->x1 to local x1. That is over the 60-line guideline. It could be made smaller, e.g. by inlining the margin in EventuallyRedrawItem only, but then the display loop and the old-area sites stay wrong.)
    • Repro ticket script with Liberation Serif italic -150, recolor black to red. Before: 150 black pixels remain at x=222..237 (bbox x2=222). After: 0.
    • Test canvText-23.1 "glyph overhang is redrawn". It uses {Times -100 italic} "f", creates the item before the first update so the overhang is drawn, deletes the item, and counts non-white pixels in bbox±30 with testpixel (constraint testpixel). Fails before: yes (main: non-zero). Passes after: yes. With fontconfig all of Times/Helvetica/DejaVu/Liberation italic "f" overhang. If a platform's Times italic f has no overhang, the test passes trivially.
    • Test files run canvText 101/0 with fix; main 101/1 (only the new test fails). canvas 175/0, canvImg 84/0, canvRect 57/0, canvWind 5/0, canvPs 14/0, canvMoveto 8/0: identical on main and with the fix.
    • Windows run canvText.test (23.1) and canvas.test. Check that Times New Roman italic "f" overhangs (GDI reports advance widths), so 23.1 fails without the fix and passes with it. macOS: run 23.1 too (nab could not reproduce on Aqua; there it may pass either way).
    • Suggested ticket comment The canvas redraws only the item bbox, which is computed from advance widths, so overhanging ink is neither erased nor fully drawn. The prototype keeps bbox unchanged but enlarges the redraw area of text items by a quarter of the line height (test canvText-23.1).
    ⬜bf65b83d432021Tk crash: XIO: fatal IO error 88 (Socket operation on non-socket)REPRODUCED (main): verified from the main source, no live test this round

    Open ticket bf65b83d43 · 67. Unix Window Operations · filed 2021-11-20

    Full verdict

    REPRODUCED (main): verified from the main source, no live test this round

    • Platform X11
    • Tested no live run, because the scenario needs a second X server besides my assigned display. The main source of 2026-10-22 still keeps the display close in Tk_DestroyWindow disabled (generic/tkWindow.c:1747, NOT_YET), so displays stay open until exit. This matches serhiy's runtime confirmation on 2026-09-23. The same XIO death on server loss was confirmed live on :145 (see 923174ffff).
    • Cause generic/tkWindow.c:Tk_DestroyWindow never calls TkCloseDisplay when the display's refCount reaches 0. When the server goes away, DisplayCheckProc's XFlush hits EOF and Xlib exits. Enabling the close requires TkCloseDisplay to also free fonts, colors, cursors, borders and bitmaps; otherwise the X_SetFontPath BadLength error appears when a window is created on that display again.
    • Prototype none (needs a per-display resource cleanup rework, which is over the size limit; the deferred-close experiment is described in serhiy's comment).
    • Suggested ticket comment Still present in main (2026-10-22). This is the same root cause as 923174ffff, and the fix depends on making all per-display resources releasable in TkCloseDisplay.
    ⬜58418fbd372021Incorrect pointer window reported upon <Leave>'ing toplevel after [wm withdraw]patchREPRODUCED (main, 8.6)

    Open ticket 58418fbd37 · 48. Geometry Management · filed 2021-08-12

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested the attached exercise.tcl with N=20 on Xvfb. With icewm, 18-20/20 <Leave> events reported .t on main and 19/20 on 8.6. With openbox, 12-15/20 on main (0/20 on 8.6 in one run, so the race is timing dependent). fvwm and no WM: 0. twm: no Leave events generated.
    • Cause unix/tkUnixWm.c: Tk_CoordsToWindow. When XTranslateCoordinates finds the WM decoration frame (child == wmPtr->reparent), it returns the toplevel without checking that it is still mapped. After wm withdraw the WM destroys/unmaps the frame asynchronously, so for a short time the frame still covers the point.
    • Prototype 58418fbd37.patch (9 added lines: skip the reparent match if !Tk_IsMapped(wmPtr->winPtr)). After the patch: icewm 0/20 and openbox 0/20 (2 runs each). winfo/unixWm/wm tests show the same results as unpatched (no WM), and unixWm.test passes fully under icewm. No new test (needs a reparenting WM).
    • Suggested ticket comment Reproduced on X11 with icewm and openbox (main and 8.6). Tk_CoordsToWindow() accepts the WM frame of a toplevel that is already withdrawn but whose frame the WM has not yet removed. Skipping unmapped toplevels there fixes it.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main clean (offset 80 lines)
    • Repro work-a4/ex20.tcl under icewm: before 20/20 and 20/20 <Leave> reported .t / after 0/20 and 0/20 (20 Leave events still delivered)
    • Test none (needs a reparenting WM; the race does not occur with no WM, which is how the test suite normally runs)
    • Test file run no WM: unixWm.test 284/0, winfo.test 76/0, identical to unpatched. Under icewm: unixWm.test 284/0 in 3 of 4 patched runs. One run failed unixWm-51.7 (raise/lower plus winfo containing right after after 200); 3 reruns passed and unpatched passed 2/2, so I count it as a WM timing flake, but it could be re-checked.
    • Patch changes none. Review: after continue the search descends into the stale frame, where the unmapped wrapper is not found, so it returns NULL. That is correct. Pre-existing, unrelated: the return NULL in the title-bar branch leaks the error handler.
    ⬜03fcf233e12021Apply zoom before subsample when copying photospatchREPRODUCED (main, 9.0, 8.6), as designed. This is a behaviour change or enhancement rather than a bug. The documentation does not specify the order.

    Open ticket 03fcf233e1 · 41. Photo Images · filed 2021-07-06

    Full verdict

    REPRODUCED (main, 9.0, 8.6), as designed. This is a behaviour change or enhancement rather than a bug. The documentation does not specify the order.

    • Platform generic
    • Tested 8-pixel row copied with -zoom 2 -subsample 3 on 9.1b1, 9.0.5 and 8.6.18 gives 000 000 333 333 666 666 in 6x2 (subsample first). With the patch it gives 000 111 333 444 666 777 in 6x1.
    • Cause generic/tkImgPhoto.c: Tk_PhotoPutZoomedBlock subsamples the source and then repeats each pixel zoom times. The result size is ceil(w/sub)*zoom instead of ceil(w*zoom/sub).
    • Prototype 03fcf233e1.patch (the reporter's 2026-04-03 tkImgPhoto.c.diff rebased onto main, plus a new test imgPhoto-4.119; 35+/25-). imgPhoto.test passes (244 tests). This is an incompatible change: the output size changes, e.g. the height 1 → zoom 2/subsample 3 case now gives 1 row instead of 2, and existing pixel results change. It probably needs a TIP or at least a changes.md "incompatibility" note, and should go to 9.1 only.
    • Suggested ticket comment The attached patch applies to main and passes imgPhoto.test (tested with a new zoom+subsample test). It changes the result size and pixels of copy -zoom -subsample, so it should go only to 9.1 and be documented as an incompatibility (or via a TIP).
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (behaviour change: copy with zoom+subsample gives a different size and different pixels. Needs a changes.md incompatibility note, 9.1 only)
    • Applies to main clean (offsets only)
    • Repro before -zoom 2 -subsample 3 of 8 pixels gives 000 000 333 333 666 666 / after 000 111 333 444 666 777
    • Test imgPhoto-4.119 (in patch; fails before: yes; passes after: yes)
    • Test file run imgPhoto.test 249/0
    • Patch changes style only. The mid-block declarations int yProgress = 0; and int xProgress = 0; are now declared at the top of the function. The pixel loop body is re-indented one level, because the removed xRepeat loop left it over-indented. The long srcPtr =/srcLinePtr = lines are wrapped. I checked the logic, including the negative subsample mirror (C truncation gives the correct mirrored index) and tiling. No functional bugs found. The per-pixel division replaces the old repeat loop, which may be slower for large zooms.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 03fcf233e1: OK. imgPhoto-4.119 (photo copy) fails before, passes after on Windows; imgPhoto.test 249/0.
    ⬜da5f2266df2021update idletasks services queued eventsREPRODUCED (main, 9.0, 8.6), but NOT-A-BUG in Tk (the Thread package's event proc ignores the event mask). The reporter withdrew the ticket in thei…

    Open ticket da5f2266df · 69. Events · filed 2021-04-26

    Full verdict

    REPRODUCED (main, 9.0, 8.6), but NOT-A-BUG in Tk (the Thread package's event proc ignores the event mask). The reporter withdrew the ticket in their last comment.

    • Platform X11 (the event-queue side is generic)
    • Tested main, 9.0 (Thread 3.0.7) and 8.6 (Thread 2.8.13) on Xvfb, no WM, with the attached exercise.tcl. The test script is racy as posted, so I added after 300 after thread::send -async. With that change: when wm geometry runs before update idletasks, the result is {B A} every time; without it, {A B} (as expected).
    • Cause update idletasks runs the idle UpdateGeometryInfo. That calls unix/tkUnixWm.c:WaitForEvent → tkUnixEvent.c:TkUnixDoOneXEvent, which calls Tcl_ServiceEvent(TCL_WINDOW_EVENTS) to wait for the ConfigureNotify. Thread's ThreadEventProc ignores its mask argument and so runs even for a TCL_WINDOW_EVENTS-only service. Event procs are supposed to return 0 for flags they don't handle.
    • Prototype none (the fix belongs in the Thread package: ThreadEventProc should return 0 unless e.g. mask & TCL_FILE_EVENTS; changing Tk's nested wait would be risky)
    • Suggested ticket comment Reproducible: the nested ConfigureNotify wait in the X11 wm code services the event queue with TCL_WINDOW_EVENTS only, but Thread's ThreadEventProc ignores the mask and runs anyway. This belongs in the Thread package; the reporter has withdrawn the request. Proposing to close here.
    ⬜1f1273f7392021Checkin #3ccd19e0a7 breaks expected operation of text window on linuxpatchREPRODUCED (main, 8.6). This is intentional behaviour since 8.6.11 (3ccd19e0a7 for d91e05bf20), so it is a design dispute.

    Open ticket 1f1273f739 · 18. [text] · filed 2021-04-14

    Full verdict

    REPRODUCED (main, 8.6). This is intentional behaviour since 8.6.11 (3ccd19e0a7 for d91e05bf20), so it is a design dispute.

    • Platform generic
    • Tested main and 8.6 tip (Xvfb, no WM). Clicking a disabled text moves the focus to it, while clicking a disabled entry does not (tk entry.tcl and ttk entry.tcl only focus when not disabled).
    • Cause library/text.tcl:tk::TextButton1 calls focus $w unconditionally, whatever the state.
    • Prototype 1f1273f739.patch (31 lines, test text-39.1). This is the compromise sbron proposed: a disabled text still takes the focus (so <<Copy>> keeps working), unless the application set -takefocus 0. This matches the recent ttk::notebook fix 3de1b72157. Normal-state behaviour is unchanged. text.test passes (689/0 failed).
    • Suggested ticket comment Keeping the d91e05bf20 behaviour but honouring -takefocus 0: a disabled text widget with -takefocus 0 no longer grabs the focus on click, as with ttk::notebook 3de1b72157. Patch and test text-39.1 on a branch.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (design decision: it honours -takefocus 0 on a disabled text)
    • Applies to main ported (text-39.1 is now taken on main; the test became text-38.3, next to the other mouse-click tests 38.1/38.2)
    • Repro before, a click on a disabled text with -takefocus 0 still takes the focus / after, it does not. Normal state is unchanged.
    • Test text-38.3 (fails before: yes; passes after: yes)
    • Test file run text.test 704/0
    • Patch changes test renumbered/moved
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 1f1273f739: OK. text-38.3 fails before, passes after.
    🔧130949dafe2021Linux : no rule to make minizip needed to build libtk_8_7_a4.zipREPRODUCED (main, 9.0); 8.6 not affected (no zipfs)

    Open ticket 130949dafe · (unused) · filed 2021-02-02

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-09-25 check-in d399f15e5f on unix-zipfs-mkzip, not merged: Fix 130949dafe: build fails without zip: no rule to make minizip. The autoconf builds (unix and MinGW) used an external zip program to create the Tk library a
    • 💬 2026-09-25 serhiy.storchaka: Reopening: this was never fixed. Tk's Makefiles never had a rule to build minizip, and Tk has no minizip sources, so without zip the build fails in 9.0 and 9.1 (8.6 has no zipfs and is not affected). I encountered this problem when building Tk on various platforms, such as FreeBSD, NetBSD, OpenBSD, DragonFly and MinGW. Also reported as Tcl 406bba1baf. One possible solution would be to copy the Makefile rules and t…
    • 💬 2026-10-05 serhiy.storchaka: Updated in 314d29097c: if no Tcl 9 shell runs on the build host, e.g. when cross-compiling, the archive is created with the zip program instead. The native tclsh is now set with NATIVE_TCLSH, as in Tcl's autoconf Makefiles, which also use NATIVE_ZIP for the zip program; TCLSH_NATIVE is the nmake name, and the MinGW build also accepts it.
    Full verdict

    REPRODUCED (main, 9.0); 8.6 not affected (no zipfs)

    • Platform Unix/MinGW autoconf builds
    • Tested out-of-tree ac_cv_path_zip=no configure --with-tcl=... and then make tkzipfile. Trunk: No rule to make target "minizip", needed by "libtk9.1.1.zip"; 9.0.5: same, for libtk9.0.5.zip. User's branch unix-zipfs-mkzip (d399f15e5f), same setup: full make succeeds. wish with TK_LIBRARY unset gets tk_library = //zipfs:/lib/tk/tk_library; msgs/ttk/images are directories with 18/26/13 entries. The archive differs from the zip-made one only in missing explicit directory entries (harmless).
    • Cause unix/tcl.m4 SC_ZIPFS_SUPPORT falls back to ZIP_PROG=./minizip with ZIP_INSTALL_OBJS=minizip, but unix/Makefile.in has no rule or sources for minizip (they exist only in Tcl).
    • Prototype none (fix already in branch unix-zipfs-mkzip; verified above). Minor leftover: configure still prints "No zip found on PATH. Building minizip", which is now misleading.
    • Suggested ticket comment Verified: without zip, trunk and 9.0 stop at "no rule to make minizip", and the unix-zipfs-mkzip build succeeds and loads the library from //zipfs:. (configure's "Building minizip" message could be reworded with it.)
    ✅3da0592db42020Filesystem dialogs conflict with auto_reset (under Linux)patchREPRODUCED (main, 9.0, 8.6); 8.5 OK (no TclOO iconlist)

    Open ticket 3da0592db4 · 37. [tk_get*File] · filed 2020-12-17

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-23 check-in a2611c4418 on dialogs-after-auto-reset, in main: Fix 3da0592db4: after [auto_reset], the Tk file dialogs failed with "invalid command name .__tk_filedialog.contents.icons" (X11), or with "Duplicate element C
    • 🔧 2026-11-09 check-in 5926b175cb on trunk, in main, 9.0: Fix for 3da0592db4: Filesystem dialogs conflict with auto_reset (under Linux)
    • 🔧 2026-11-09 check-in 835c26cda9 on core-9-0-branch, in 9.0, 8.6: Fix for 3da0592db4: Filesystem dialogs conflict with auto_reset (under Linux)
    • 🔧 2026-11-09 check-in 02e5e9b5f2 on core-8-6-branch, in 8.6: Fix for 3da0592db4: Filesystem dialogs conflict with auto_reset (under Linux)
    • 💬 2026-10-23 serhiy.storchaka: auto_reset removes every command listed in auto_index. This includes the TclOO classes ::tk::Megawidget and ::tk::IconList, and deleting a class deletes its objects, so the icon list of the cached dialog is gone. Then megawidget.tcl could not be sourced again, because ::tk::MegawidgetClass survives. The commands of tk.tcl are removed too, so tk.tcl and ttk.tcl get sourced again, and on Windows (and in 8.6 and 9.0 …
    • 💬 2026-11-09 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits 5926b175, 835c26cd, and 02e5e9b5.
    • 💬 2026-11-10 serhiy.storchaka: Test choosedir-6.1 failed in CI on Linux and macOS: package require tk in a new interpreter could not find Tk, because Tk is not installed there. Where Tk is installed, the child interpreter could load the installed Tk instead of the tested one. Fixed in 4aaad335cc (main, using childTkInterp), fdbba3c0a8 (9.0) and 2d73d468a9 (8.6, using loadTkCommand).
    Full verdict

    REPRODUCED (main, 9.0, 8.6); 8.5 OK (no TclOO iconlist)

    • Platform X11 (the Tk file dialogs, which are also used on Windows only when requested explicitly)
    • Tested main, 9.0, 8.6, 8.5 on Xvfb with no WM
    • Cause auto_reset renames every command listed in auto_index, including the TclOO classes ::tk::Megawidget and ::tk::IconList. That destroys all their instances, including the cached dialog's $w.contents.icons. The dialog reuse code in choosedir.tcl and tkfbox.tcl still assumes the icons widget exists. When megawidget.tcl is autoloaded again, it fails with "can't create object ::tk::MegawidgetClass" because that class is not in tclIndex and survives. In 8.6/9.0 there is an extra failure: tk.tcl is re-sourced through the autoloader, and ttk.tcl then fails with "Duplicate element Combobox.downarrow" when it reloads the themes. The prototype does not address that; main does not show it.
    • Prototype 3da0592db4.patch (38 lines, test choosedir-6.1). It recreates the dialog when $w.contents.icons no longer exists, and megawidget.tcl destroys a leftover ::tk::MegawidgetClass so the file can be sourced again. The new test runs in a child interp and fails without the fix; choosedir.test and filebox.test pass.
    • Suggested ticket comment auto_reset destroys the TclOO classes ::tk::Megawidget/::tk::IconList and thus the cached dialog's icon list, and megawidget.tcl cannot be sourced twice. A fix recreates the dialog in that case and makes megawidget.tcl re-sourceable (test choosedir-6.1).
    ⬜1600dc73a92020Application window opens between ttk::combobox and its pop-down listpatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 1600dc73a9 · 88. Themed Tk · filed 2020-10-09

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform X11 + Windows (separate causes)
    • Tested main and 8.6 with fvwm: after another app maps over the window, the popdown stays mapped and is stacked above the new window, which is above the combobox (checked with xwininfo). With no WM the new window goes on top and hides the still-posted popdown.
    • Cause library/ttk/combobox.tcl: the popdown is an override-redirect window that stays posted, so the WM keeps it above normal windows. On Windows it additionally has wm attributes -topmost 1, added for bug 1991930. The combobox does get <Visibility> VisibilityFullyObscured events on X11.
    • Prototype 1600dc73a9.patch (27 lines, test combobox-1600dc73a9). On X11, a class binding on TCombobox <Visibility> unposts a mapped popdown when the state is not VisibilityUnobscured; this is a tidier version of bll's closed bug-1600dc73 branch. The test fails without the fix, and all ttk tests pass. Windows would still need -topmost replaced by taskbar-aware placement, as fvogel suggested.
    • Suggested ticket comment On X11 the popdown can be unposted when the combobox gets obscured (<Visibility> binding, test combobox-1600dc73a9). Windows needs a different fix: drop -topmost and account for the taskbar work area in PlacePopdown.
    Windows VM, run 3
    • Platform Windows (plus X11, prototyped earlier)
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested work-win/1600_auto.tcl posts the popdown, then starts another wish process whose 500x400 window covers the combobox. On both versions the popdown stays mapped and is the window under the pointer at its centre (WindowFromPoint), above the other application's window, which is itself above the combobox. The screenshot confirms it. Removing -topmost from the posted popdown gives the same picture, but this is inconclusive: in the scheduled-task session the other process cannot take the foreground (Program Manager stays the foreground window), so it is not a real activation.
    • Cause library/ttk/combobox.tcl, the popdown stays posted (with -topmost 1 on Windows); nothing unposts it when another application appears
    • Prototype none for Windows (the X11 <Visibility> patch does not help here, because Windows does not generate Visibility events)
    • Suggested ticket comment Confirmed on Windows 11 (9.1, 8.6): the posted popdown stays above a newly opened window of another application. On Windows the popdown should probably be unposted when the application is deactivated.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (X11 only; the Windows part has no fix, see the triage)
    • Applies to main clean
    • Repro before the popdown stays posted when a new toplevel covers the combobox / after it is unposted (no WM and fvwm)
    • Test combobox-1600dc73a9 (in patch; fails before: yes, both with no WM and with fvwm; passes after: yes, both)
    • Test file run combobox.test 21/0. All ttk tests: 2064/0. A full-suite run that happened along the way had 1 failure, image-11.1, which passes alone on main (flaky).
    • Patch changes test uses pause 200 instead of after 200 {set done 1}; vwait done
    • Review note any partial obscuring of the combobox, even by another window of the same app, unposts the popdown. That is acceptable, but more aggressive than strictly needed.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 1600dc73a9: test combobox-1600dc73a9 is x11-only, skipped; ttk/combobox.test 21/0 before and after.
    ⬜b358779bd42020[place] places at wrong position inside menubarsREPRODUCED (main, 8.6, 8.4). Arguably NOT-A-BUG (cannot work).

    Open ticket b358779bd4 · 12. Unix Menus · filed 2020-09-28

    Full verdict

    REPRODUCED (main, 8.6, 8.4). Arguably NOT-A-BUG (cannot work).

    • Platform X11
    • Tested main, 8.6, 8.4.20 on Xvfb: place .marker -in .#m -relx .5 -rely .5 puts the marker a menubar height too low (y 153 instead of 117 on main); x is correct.
    • Cause the menubar clone .#m is a Tk child of "." but its X window is a child of the wrapper, above "."'s window. Tk_MaintainGeometry (generic/tkGeometry.c) adds up Tk_Y from the container to the parent and gets 0 for the menubar instead of -menubarheight. Even with correct coordinates, a child of "." cannot be drawn over the menubar because X clips it to "."'s window.
    • Prototype none (no useful fix; the most that could be done is to unmap such content)
    • Suggested ticket comment The X11 menubar is outside the area of ".", so no child of "." can be placed over it. Suggest closing as "Won't fix".
    ⬜4476fd61442020X11 clip region ignored for Xft text drawingpatchREPRODUCED (main, 8.6)

    Open ticket 4476fd6144 · 46. Unix Fonts · filed 2020-08-22

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main (9.1b1, bidi/HarfBuzz backend) and 8.6 tip, Xvfb, no WM. A small stubs extension (work-a2/clip/clip.c) sets XSetClipRectangles(gc, 0,0,30x40) and calls Tk_DrawChars. Text is drawn out to x=252, so the 30px clip is ignored on both versions.
    • Cause unix/tkUnixRFont.c (and tkUnixBidiFont.c) only call XftDrawSetClip() from the thread-global tsdPtr->clipRegion, which is set by the private TkUnixSetXftClipRegion() (used only by ttkLabel/ttkEntry). The GC clip mask is never used, and Xlib gives no way to read it back from a GC. Also, Tk_ClipDrawableToRect (stub 284) is NULL/no-op on X11 and Windows.
    • Prototype none (needs a public API, i.e. a TIP-level change). Options: implement Tk_ClipDrawableToRect on X11 so it sets/clears the Xft clip region (this fills the NULL stub slot and removes the #undef/#define ... 0 in tkDecls.h/tkStubInit.c for non-Mac), or export TkUnixSetXftClipRegion through the platform stubs.
    • Suggested ticket comment Still present in 8.6 and main: Xft text drawing ignores the clip mask of the GC, and Xlib cannot read that clip mask back. A fix needs a public API. The candidate is Tk_ClipDrawableToRect(), which is a no-op on X11 today and could set the Xft clip region.
    2026-10-25 prototype, Linux
    • Status PROTOTYPED (core change only; no test in the suite)
    • Cause unix/tkUnixRFont.c and tkUnixBidiFont.c apply only the thread-global Xft clip set by the private TkUnixSetXftClipRegion() (used by ttkLabel/ttkEntry). The GC clip mask cannot be read back from Xlib. The public Tk_ClipDrawableToRect (stub 284) is NULL on X11: tkStubInit.c has #define Tk_ClipDrawableToRect 0 for every non-Mac build, and tkDecls.h has #undef for non-Mac. A stubs extension that calls it on main segfaults (verified).
    • Fix implement Tk_ClipDrawableToRect in unix/tkUnixDraw.c (X11 only). It keeps a per-thread Region, sets it with TkUnixSetXftClipRegion() under HAVE_XFT || HAVE_BIDI, and clears and frees it when width or height is negative. Without Xft it is a no-op, because core X fonts already honour the GC clip. The drawable argument is ignored: the clip applies to all Xft text drawing in the thread until reset. Only Xft text is clipped, not other Xlib drawing; the docs should say so (doc/3DBorder.3 is not updated yet). Stub table: no regeneration. tk.decls and the generated part of tkDecls.h are unchanged. In tkStubInit.c #else #define ... 0 became #elif defined(_WIN32). In the hand-written tail of tkDecls.h (below !END!) #ifndef MAC_OSX_TK became #ifdef _WIN32. So the slot is filled on X11, and Windows keeps the NULL slot. 4476fd6144.patch (63 changed lines, about 25 of them comments).
    • Repro work-a2/clip/clip2.c and t2.tcl (stubs extension: Tk_ClipDrawableToRect(0,0,30,40), Tk_DrawChars "WWWWWWWW", reset, draw another W). Main (bidi): SIGSEGV through the NULL stub. Patched, both --disable-bidi (Xft/RFont) and --enable-bidi (HarfBuzz) builds: ink stops at x=29 (unclipped: 252), and text drawn after the reset is not clipped.
    • Test none. Tk_ClipDrawableToRect cannot be called from Tcl. A test would need a tkTest.c helper (e.g. testfont drawclipped font window text x y w h, ~30 lines) plus a testpixel check. Left out to keep the change under ~80 lines.
    • Test files run (with fix, --disable-bidi): font 304/0, unixFont 50/0, canvText 100/0, textDisp 409/0, frame 207/0, entry 306/0, ttk entry 44/0. The fix only adds a function, so main results are the same. The bidi build compiles without warnings.
    • Windows nothing changes there (slot stays NULL). Separately, decide whether Windows should also implement it, e.g. with SelectClipRgn.
    • Suggested ticket comment Prototype: implement Tk_ClipDrawableToRect() on X11 (stub slot 284 was NULL there) so that it sets the Xft clip region, as TkUnixSetXftClipRegion() does for ttk. Extensions like Tktable can then call it around Tk_DrawChars()/Tk_DrawTextLayout(). On X11 it affects only Xft text; other drawing still uses the GC clip.
    ⬜1e9d7439972020Right-to-left language in Entry WidgetREPRODUCED (main, 8.6)

    Open ticket 1e9d743997 · 79. L10N · filed 2020-05-28

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform generic (tested X11)
    • Tested main 9.1b1 with bidi (SheenBidi+HarfBuzz) and 8.6 tip, Xvfb. Main shapes and reorders Arabic correctly, but after icursor end the cursor is drawn at the right edge of the word (screenshot work-a2/1e9d743997-main.png). .e bbox 0 is at x=3, i.e. index-to-x mapping stays logical LTR. 8.6 does not shape or reorder at all.
    • Cause generic/tkEntry.c and ttkEntry map indices to x using logical-order text layout (Tk_CharBbox/Tk_PointToChar). There is no visual-order (bidi) caret or selection mapping.
    • Prototype none (bidi-aware caret, hit-testing and selection is a larger design task, beyond 60 lines)
    • Suggested ticket comment Still present in main. The new X11 bidi backend renders RTL text correctly, but entry indices, the insertion cursor and bbox still use logical left-to-right positions. This needs bidi-aware index-to-x mapping in the entry widgets.
    ✅e1d86b048b2019-cursor options: documentationpatchREPRODUCED (main, 8.6), but it is not what the ticket assumes. The forms are documented in GetCursor.3 (and options.n refers to it). On X11 `{name …

    Open ticket e1d86b048b · 31. Cursor Names · filed 2019-12-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-08 check-in 428d9b1bb2 on cursor-file-tests, in main, 9.0: Fix e1d86b048b: -cursor options: documentation. Fix the corrupted reference to the X cursor font (cursorfont.h), document that the cursor colors are only supp
    • 🔧 2026-10-22 check-in 250b737fb0 on trunk, in main, 9.0: Fix e1d86b048b: -cursor options: documentation improvements
    • 🔧 2026-10-22 check-in e281821ba7 on core-9-0-branch, in main, 9.0: Fix e1d86b048b: -cursor options: documentation improvements
    • 🔧 2026-10-21 check-in 0615181c88 on cursor-file-tests, in main, 9.0: Rebase to trunk
    • 💬 2026-10-08 serhiy.storchaka: Fixed in branch cursor-file-tests, 428d9b1bb2. The colors of a named cursor are only supported on X11, and if a cursor theme is used (with the Xcursor library, the default on most current desktops), a cursor with two colors is shown with the colors of the theme; a cursor with only fgColor is still drawn in that color. Windows and macOS ignore the colors. This is now documented. The files of the @sourceName forms…
    • 💬 2026-10-09 jan.nijtmans: I'll have to look at "make dist", making sure the additional cursors will be included into the distribution tar.gz.
    • 💬 2026-10-23 jan.nijtmans: Fixed in 250b737f and core-9-0-branch Many thanks!
    Full verdict

    REPRODUCED (main, 8.6), but it is not what the ticket assumes. The forms are documented in GetCursor.3 (and options.n refers to it). On X11 {name fg bg} colors are ignored because libXcursor replaces a glyph cursor from the cursor font with the themed ARGB cursor whenever source font == mask font and mask = source+1, and it ignores the colors (with XCURSOR_CORE=1 the colors work). {name fg} and @src fg / @src mask fg bg do work. @src alone is documented as Windows-only (the ticket's failing .xbm test). There is also a doc typo in GetCursor.3: "standard X cursor cursor", "cursorcursor.h" (should be cursor font / cursorfont.h).

    • Platform X11 (+docs)
    • Tested main, 8.6 on Xvfb (checked the cursor image with XFixesGetCursorImage), with and without XCURSOR_THEME / XCURSOR_CORE
    • Cause unix/tkUnixCursor.c:CreateCursorFromTableOrFile() calls XCreateGlyphCursor(cursorFont, cursorFont, shape, shape+1, ...), and XcursorTryShapeCursor() replaces that call.
    • Prototype e1d86b048b.patch (72 lines). When colors other than black/white are given, it loads a second font ID for the mask so that Xcursor does not replace the cursor. It also fixes the GetCursor.3 typo and documents the theme behaviour. {watch red blue} now shows red/blue, while watch and {watch black white} keep the themed cursor. cursor.test passes. No automated test (it would need XFixes).
    • Suggested ticket comment On X11 the Xcursor library replaces cursor-font cursors with themed ones and ignores the colors given as {name fg bg}. The branch uses a separate mask font ID when colors are given, so they are honoured again. The @file forms work as documented (@file alone is Windows-only).
    ✅340006f5d02019X11 is very slow to create widgetsREPRODUCED (main, 8.6 source: generic/tkEvent.c still creates an XIC for every window on its first event (TK_CHECKED_IC)). A fix already exists on …

    Open ticket 340006f5d0 · 79. L10N · filed 2019-11-22

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-10-08 serhiy.storchaka: Fixed by the fix for 5d52ad92fa: the X input context is now created only for a toplevel and for a window which gets the focus, as proposed in branch bug-xim, instead of for every window. Merged into main e978f0d457 (Tk 9.1.0), core-9-0-branch 7345531fbe (9.0.5) and core-8-6-branch 0003874f7a (8.6.19).
    Full verdict

    REPRODUCED (main, 8.6 source: generic/tkEvent.c still creates an XIC for every window on its first event (TK_CHECKED_IC)). A fix already exists on branch x11-lazy-xic 894c1665c5 (2026-09-23, fixes the duplicate 5d52ad92fa); not merged yet.

    • Platform X11
    • Tested source reading. The timings (ibus: 470 ms -> 20-29 ms for 480 labels) are in the x11-lazy-xic check-in.
    • Cause generic/tkEvent.c:Tk_HandleEvent() calls CreateXIC() for every window that gets an event, and each call is a round trip to the input method.
    • Prototype none (branch x11-lazy-xic)
    • Suggested ticket comment Duplicate of 5d52ad92fa. It is fixed on branch x11-lazy-xic: the input context is created only for toplevels and for the window that gets the focus.
    ⬜162257a8272019Tk should move away from [bgerror] and [tkerror]patchREPRODUCED (main): still an open design item. library/bgerror.tcl still defines ::tk::dialog::error::bgerror, falls back to ::tkerror and is im…

    Open ticket 162257a827 · 71. Error Handling · filed 2019-11-03

    Full verdict

    REPRODUCED (main): still an open design item. library/bgerror.tcl still defines ::tk::dialog::error::bgerror, falls back to ::tkerror and is imported as global bgerror. library/console.tcl:55 and library/fontchooser.tcl:347 still call bgerror directly.

    • Platform generic
    • Tested source inspection of main
    • Cause library/bgerror.tcl predates TIP 221 (interp bgerror); Tk never switched to installing a handler with interp bgerror.
    • Prototype none (this is an RFE/redesign that needs compatibility decisions for user-defined bgerror/tkerror). A small first step: make console.tcl and fontchooser.tcl call {*}[interp bgerror {}] $msg $opts (or after 0 [list error ...]) instead of bgerror.
    • Suggested ticket comment Still valid on main: bgerror.tcl still uses the bgerror/tkerror protocol, and console.tcl and fontchooser.tcl call bgerror directly. Moving to interp bgerror (TIP 221) is an RFE, so maybe retype it.
    2026-10-25 prototype, Linux
    • Status PROTOTYPED (only the small first step; bgerror.tcl unchanged)
    • Cause library/fontchooser.tcl ::tk::fontchooser::Apply and library/console.tcl ::tk::ConsoleInit call bgerror directly, which bypasses a handler installed with interp bgerror (TIP 221). A grep of library/ (outside bgerror.tcl and msgs) finds no other direct calls to bgerror/tkerror, and generic/ and unix/ have none either.
    • Fix both places now use catch {...} err opts followed by {*}[interp bgerror {}] $err $opts. This is synchronous, like before. With the default handler (::tcl::Bgerror) it still ends up in Tk's bgerror dialog, so nothing changes for users who did not install a handler. 162257a827.patch (33 changed lines, 4 library + 29 test)
    • Repro before: a fontchooser -command error goes to ::bgerror even with an interp bgerror handler installed / after: the installed handler gets it (msg and options dict)
    • Test fontchooser-6.1 "-command error goes to interp bgerror handler" (scriptImpl; installs interp bgerror and a recording proc bgerror, then clicks Apply with -command {error oops}) (fails before: yes, result {bgerror oops}; passes after: yes, {handler oops}). The console.tcl change has no test: the error path needs menu .menubar to fail inside the console interp at init.
    • Test files run fontchooser 20/0, bgerror 3/0 with fix; main fontchooser 20/1 (the new test), bgerror 3/0
    • Windows Windows uses the native font dialog, so fontchooser-6.1 is skipped there (scriptImpl). Nothing else to check beyond a console startup smoke test.
    • Suggested ticket comment First step prototyped: console.tcl and fontchooser.tcl now report errors through {*}[interp bgerror {}] instead of calling bgerror directly, so TIP 221 handlers are honoured. Moving bgerror.tcl itself to an interp bgerror handler (and dropping the tkerror fallback) is a compatibility decision, closer to an RFE.
    🔧a0109510a82019make not rebuilding after header files changepatchREPRODUCED (main). In unix/Makefile.in, object rules list only their .c file (plus a few tkUuid.h/tkImgPhoto.h cases). Touching generic/tkButton.h …

    Open ticket a0109510a8 · None · filed 2019-08-13

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-20 check-in 4d8ef8c7c3 on header-dependencies, not merged: Fix a0109510a8: make did not rebuild object files after a header file was changed. If the compiler supports -MMD -MP (gcc, clang) and make supports -include (
    • 🔧 2026-10-20 check-in a2feb92f50 on header-dependencies, not merged: Header dependencies on Windows: the same for win/Makefile.in (MinGW, MSYS2, Cygwin), and pass the compiler options to mkdepend.tcl in the depend target of win/m
    • 💬 2026-10-20 serhiy.storchaka: Fixed in branch header-dependencies. If the compiler supports -MMD -MP (gcc and clang) and make supports -include, the compiler writes the header dependencies of every object file into a .d file, which the Makefile includes. This is done in unix/ and win/ (MinGW, MSYS2, Cygwin). Solaris make and dmake do not support -include, so configure checks it with ${MAKE-make}; nothing changes with them, nor with MSVC. For w…
    Full verdict

    REPRODUCED (main). In unix/Makefile.in, object rules list only their .c file (plus a few tkUuid.h/tkImgPhoto.h cases). Touching generic/tkButton.h or unix/tkUnixDefault.h rebuilds nothing. win/Makefile.in has the same problem.

    • Platform build
    • Tested main wt build on Linux: touch a header, then make -n, gives 0 compiles before the patch and 137 after.
    • Cause unix/Makefile.in: objects have no header dependencies.
    • Prototype a0109510a8.patch (6 changed lines, unix only). It adds a coarse rule $(OBJS) $(WISH_OBJS) $(TKTEST_OBJS) $(STUB_LIB_OBJS): $(DEP_HDRS), where DEP_HDRS is generic/*.h, ttk/*.h and unix/*.h (X11) or macosx/*.h + xlib (Aqua). Any header change rebuilds all objects. The rule sits after all:, so the default goal is unchanged. A finer alternative is -MMD -MP with -include *.d (gcc/clang only). win/Makefile.in would need the same rule.
    • Suggested ticket comment Still true on main. A coarse fix is to make all objects depend on the source headers (prototype: rebuild everything when any .h changes). Per-file dependencies would need -MMD (gcc/clang).
    ⬜f436621c1a2019tk_getSaveFile: -defaultextension should work on windows as it does on unix, wrt adding a suffixpatchf436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.

    Open ticket f436621c1a · 37. [tk_get*File] · filed 2019-08-07

    Full verdict

    f436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.

    • Platform Windows
    • Tested source inspection. The repro script's cases were checked only against the Unix tkfbox logic. Note that on Unix an explicit -defaultextension "" does NOT disable the guess, which is the opposite of what the 2026 anonymous comment asks for.
    • Cause win/tkWinDialog.c:GetFileNameVista(): no default extension is set when -defaultextension is omitted.
    • Prototype f436621c1a.patch (19 changed lines, NOT compiled: no mingw here). For a save dialog without -defaultextension, it takes the extension from the first *.ext pattern of the initially selected filter and calls SetDefaultExtension. The Common Item Dialog is expected to update it when the user changes the type; this needs checking on Windows. An explicit -defaultextension "" still disables it. Repro: win/f436621c1a.tcl (interactive; type "foo" in 4 cases; prints OK/DIFFERS per case).
    • Suggested ticket comment Still present on main from the code: the Windows dialog only calls SetDefaultExtension when -defaultextension is given. A prototype that guesses it from -filetypes like tkfbox.tcl does needs testing on Windows. It is a behaviour change, so it belongs in 9.1.
    Windows VM, run 3
    • Platform Windows
    • Verdict f436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.
    • Tested dialogs driven with act.exe (filename set in Edit 1001, file type changed by a real click on the type combobox plus Down/Enter, then IDOK). Results:
      • ref91 and 8.6: no -defaultextension gives foo (case 1), and also foo after switching to Text (case 2). -defaultextension "" gives the type's extension (foo.tcl, or foo.txt with Text selected). 2045 script: "" gives myfile.xyz, absent gives myfile.
      • Additional check (work-win/dx.tcl, ref91): -defaultextension .dat with first type {"All files" *} gives foo.dat; with first type {Text .txt} gives foo.txt; with {"Tcl or Text" {.tcl .txt}} gives foo.tcl. On Windows the extension of the selected type wins over -defaultextension (on Unix an explicit -defaultextension wins, see tkfbox.tcl SetFilter).
      • Prototype: case 1 foo.tcl, case 2 foo.txt (follows the type selector). 2045 script: both give myfile.xyz. The dx.tcl results do not change.
      • New tests winDialog-5.7.9 (first type .txt gives bar.txt) and 5.7.10 (-typevariable picks .tcl) fail on ref91 and pass with the prototype. winDialog-5.7.* (47af) still pass.
    • Cause win/tkWinDialog.c GetFileNameVista calls SetDefaultExtension only when -defaultextension is given. With an empty default extension, IFileDialog uses the selected type's extension and follows type changes.
    • Prototype f436621c1a.patch (57 lines, reworked). When -defaultextension is absent for a save dialog, it now just calls SetDefaultExtension(L"") (7 code lines), replacing the earlier filter-parsing version. It includes getOpenFile.n text for the patched behaviour and the two tests. 2045553fff.patch (doc only, for unpatched behaviour) is updated with the rule that the selected type's extension wins on Windows. Use one of the two doc texts, depending on whether f436 goes in.
    • Suggested ticket comment f436621c1a: On Windows, IFileDialog uses the selected file type's extension only if some default extension is set. Setting an empty one when -defaultextension is not given makes Windows behave like Unix, including following the type selector. Patch and tests are ready. 2045553fff: confirmed that "" and an absent option differ on Windows; this goes away with the f436621c1a fix.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before: winDialog-5.7.9/5.7.10 give bar / after: bar.txt and bar.tcl
    • Test winDialog-5.7.9, winDialog-5.7.10 (existing in patch; fail before: yes; pass after: yes)
    • Test file run winDialog.test 71/0, filebox.test 63/0 (pre-existing failures: none)
    • Patch changes none
    🟢f3e40e98d42017outlinestipple not workingREPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple

    Open ticket f3e40e98d4 · 05. Canvas Items · filed 2017-11-04

    🟢 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-23 check-in 5b6f94abcd on win-canvas-arc-stipple, in main, 9.0: Update changes.md
    • 💬 2026-10-22 serhiy.storchaka: Re-tested on Windows 11 with today's main: stippled oval fills and the outlines of ovals, rectangles and arcs are still drawn solid (polygon outlines are fine). Branch win-canvas-arc-stipple fixes all of them, and the canvas tests pass with it.
    Full verdict

    REPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple

    • Platform Windows (macOS part remains, CANNOT-TEST)
    • Tested X11 main, all stipple cases correct. Wrote win/f3e40e98d4.tcl. It covers stippled oval fill and oval/rectangle/arc/polygon outlines, checks for a gray50 checkerboard of the right color, and prints PASS/FAIL lines and a VERDICT. It reads pixels with testpixel if tktest is loaded, otherwise with a DPI-aware PowerShell screen grab, and exits by itself. On X11 with tktest all cases PASS (VERDICT: FIXED).
    • Cause win/tkWinDraw.c drew arcs, chords, pies and rectangle outlines with a plain pen/brush, ignoring the GC stipple (fixed on the branch via RenderStippled()).
    • Prototype none (exists on branch win-canvas-arc-stipple)
    • Suggested ticket comment Run win/f3e40e98d4.tcl on main and on win-canvas-arc-stipple; it should print BUG PRESENT on main and FIXED on the branch. Then merge. The macOS part stays open.
    Windows VM, run 6
    • Platform Windows
    • Verdict REPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple
    • Tested testpixel: stippled oval fill and oval, rectangle and arc outlines are drawn solid; the polygon outline is stippled. Branch win-canvas-arc-stipple: all 5 cases PASS, VERDICT FIXED.
    • Prototype none
    • Suggested ticket comment Still present on main on Windows 11. Branch win-canvas-arc-stipple fixes it and is waiting for review.
    ⬜cb21f579dd2017Embedded windows leak over border and highlight areas of [text]REPRODUCED (main, 8.6)

    Open ticket cb21f579dd · 18. [text] · filed 2017-10-28

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform generic
    • Tested main 9.1b1, 8.6.18 on Xvfb, no WM. With the ticket script, .b ends up at x=6 (main) / x=11 (8.6), while the inner area starts at x=45 (bd 20 + highlight 25). The ghosts/vapor trails only seen in revised_text were fixed in 94547cbef7 (closed).
    • Cause generic/tkTextWind.c: TkTextEmbWinDisplayProc places the child with Tk_MaintainGeometry/Tk_MoveResizeWindow at the dline position and never clips it to the text's inner area. Canvas windows have the same limitation. This is a duplicate of the 2003 ticket 842504ffff "embedded windows (in text/canvas) are not cropped", which is still open.
    • Prototype none. Real clipping needs a clipping wrapper window, or unmapping windows that go past the inner area (which changes behaviour). Too big for a prototype.
    • Suggested ticket comment Still present in 9.1 and 8.6 on X11: the embedded window is drawn over the border and highlight ring. Duplicate of 842504ffff. The revised-text ghosts were fixed in 94547cbef7.
    ⬜964af13db02017tk_chooseDirectory : linuxpatchREPRODUCED (main, 8.6). It is closer to an RFE: the current behaviour is deliberate fallback code.

    Open ticket 964af13db0 · 36. [tk_chooseDirectory] · filed 2017-10-02

    Full verdict

    REPRODUCED (main, 8.6). It is closer to an RFE: the current behaviour is deliberate fallback code.

    • Platform X11 (library/choosedir.tcl)
    • Tested main 9.1b1 on Xvfb. With -initialdir /nonexistent/sub -mustexist 0 the dialog opens in [pwd] and the entry shows [pwd].
    • Cause library/choosedir.tcl, Config: when -initialdir is not an existing directory, selectPath is set to [pwd] whatever -mustexist says.
    • Prototype 964af13db0.patch (45 changed lines, including tests choosedir-4.4 and 4.5). With -mustexist 0 and a non-existent -initialdir, the dialog now browses the nearest existing ancestor and pre-fills the entry with the normalized requested path. Pressing OK twice returns that path, as the existing flow for non-existent dirs already does. -mustexist 1 keeps the old [pwd] fallback. choosedir.test passes 15 tests, 1 skipped.
    • Suggested ticket comment Confirmed: a non-existent -initialdir falls back to [pwd]. Proposed change: with -mustexist 0, open the nearest existing parent and pre-fill the entry with the requested path. Prototype with tests is available.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before, with -initialdir <fake>/sub -mustexist 0, the dialog starts in [pwd] and the entry shows [pwd]. After, it browses the nearest existing ancestor and the entry shows <fake>/sub; OK twice returns that path.
    • Test choosedir-4.4 (existing in patch, changed; fails before: yes; passes after: yes). choosedir-4.5 (-mustexist 1, guards the unchanged behaviour; passes before and after).
    • Test file run choosedir.test 16/0 (main: 14/0)
    • Patch changes choosedir-4.4 now saves the delayed second OK press as set afterId [after 50 ...] and cancels it in a new -cleanup. On unpatched main the dialog closes after the first OK, and the pending press then hit 4.5's dialog, which cascaded into failures of 4.5 and 5.1. Now only 4.4 fails without the fix. Backup: 964af13db0.patch.1024.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 964af13db0: choosedir-4.4/4.5 are unix-only (native dialog on Windows), skipped; choosedir.test 16/0.
    ⬜c6978a9ec72017Revised [text]: Linux : spacemode trim reports wrong character as hidden (24.11.1)REPRODUCED (revised_text 8b6e012e5d, 2026-09-15). Not applicable to main/8.6 (no -spacemode option there).

    Open ticket c6978a9ec7 · 18. [text] · filed 2017-07-04

    Full verdict

    REPRODUCED (revised_text 8b6e012e5d, 2026-09-15). Not applicable to main/8.6 (no -spacemode option there).

    • Platform generic (revised_text branch only)
    • Tested revised_text tip built in wt-a1 on Xvfb, no WM. Main 9.1 gives "unknown option -spacemode".
    • Cause revised text layout (tkTextDisp.c) with -spacemode trim. The result is now the same with or without the insert mark: space {143 3 0 14}, newline {148 3 0 14}. That matches neither expectation in the ticket (bll: space 5 px wide and newline at 148; gcramer: space 0 px wide and newline at 143 with 5 px). There is an unassigned 5 px gap between the trimmed space and the newline. Test textDisp-24.11.1 is no longer in the branch's textDisp.test, and no test uses -spacemode.
    • Prototype none (revised_text is out of scope for main, and the expected semantics were never agreed)
    • Suggested ticket comment Still differs on current revised_text: the trimmed space has bbox {143 3 0 14} and the newline {148 3 0 14}, leaving a 5 px gap that belongs to no character. The branch owner needs to decide the intended bboxes and restore a -spacemode test; not relevant to main.
    ⬜a34b49f8c62017Revised [text]: (textDisp-2.26 and 24.24) tag -tabs does not work properlypatchREPRODUCED (main, 8.6) for part (b), the "arrows" drawn with tags that combine -tabs and -justify right. Fixed on revised_text. Parts (a) and (c) w…

    Open ticket a34b49f8c6 · 18. [text] · filed 2017-06-25

    Full verdict

    REPRODUCED (main, 8.6) for part (b), the "arrows" drawn with tags that combine -tabs and -justify right. Fixed on revised_text. Parts (a) and (c) were script errors or are fixed on revised_text. On revised_text, textDisp-2.26 and textDisp-24.24 are still marked knownBug and still fail if forced (2.26: 48 78 108 138; 24.24: {14 7 7 13} {28 20 7 13}), because the branch intentionally applies tabs under right/center justify.

    • Platform X11 (the drawing part); generic (the revised_text test part)
    • Tested main 9.1 and 8.6 tip (the bug shows: HT glyphs are drawn and "d e" are pushed out of the window), revised_text tip (correct tab layout). All on Xvfb, no WM, Liberation Mono.
    • Cause generic/tkTextDisp.c. For -justify right/center, LayoutDLine does not split chunks at tabs, and MeasureChars gives an embedded tab zero width. CharDisplayProc strips only a trailing tab, so embedded tabs are passed to Tk_DrawCharsInContext. Xft draws them as the U+2409-like control glyph and measures that glyph into the prefix width, so every later character is shifted right and clipped. Windows (the #else branch) uses the same path, but GDI draws nothing for a tab.
    • Prototype a34b49f8c6.patch (tkTextDisp.c, +45/-23, only the TK_DRAW_IN_CONTEXT branch of CharDisplayProc). It draws each run between tabs separately, positioned from the layout width (tabs count as 0) and in the context of the text since the previous tab; underline and overstrike are handled the same way. With no tabs it makes exactly the same call as before. After the fix the window shows "abcde" right-aligned and underlined; textDisp, text and textTag tests pass. No automated test, because only rendering changes and bbox output was already correct. It is independent of the a0c1a9c6c6 prototype, which touches other hunks of the same file.
    • Suggested ticket comment On main/8.6 X11, a tag with -tabs and -justify right/center gives embedded tabs zero width in the layout, but CharDisplayProc still passes them to Xft, which draws a control glyph and shifts the rest of the line out of view. Fix prototyped: draw only the runs between tabs (branch TBD). The revised_text part is fixed there.
    Prototype verified 2026-10-24, Linux
    • Status PATCH-BROKEN as written (fixed). The original patch only changed the TK_DRAW_IN_CONTEXT branch of CharDisplayProc. On main, unix/tkUnixInt.h defines TK_DRAW_IN_CONTEXT only #ifdef HAVE_BIDI, so a --disable-bidi build (and every non-bidi build) still drew the HT glyphs. I added the same run-between-tabs drawing to the #else (Tk_DrawChars) branch. VERIFIED in both builds.
    • Applies to main clean (offset 264 lines), plus the new hunk
    • Repro right-justified tag with -tabs {2c 4c}, text "a\tb\tc\td e", compared pixel by pixel with "abcd e".
      • --disable-bidi build: before, control glyphs are drawn and the text is shifted ("DIFFERENT"); with the original patch, still broken; with the updated patch, identical to the version without tabs.
      • Out-of-tree bidi build (HarfBuzz): before "DIFFERENT", after identical.
    • Test none (rendering only; bbox was already correct)
    • Test file run textDisp.test 409/0, text.test 703/0, textTag.test 182/0 (non-bidi); textDisp.test 409/0, textTag.test 182/0 (bidi build). Same as main.
    • Patch changes added the #else /* !TK_DRAW_IN_CONTEXT */ hunk, which loops over runs separated by tabs and advances x with MeasureChars, treating tabs as 0 width. Underline and overstrike are drawn per run. Backup: a34b49f8c6.patch.1024.
    Prototype verified 2026-10-24, Windows VM, generic patches
    🟢a4fa3ad2d62017event misbehavior with certain widgets when embeddingREPRODUCED (main). The fix exists on branch x11-embed-focus-return (eb2b41d5b3, test 63c88e8a6d, f98ef68b49, 11bfb3794d) and is not merged …

    Open ticket a4fa3ad2d6 · 69. Events · filed 2017-04-20

    🟢 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-08 check-in 1dea0b9739 on x11-embed-focus-return, in main, 9.0: Fix a4fa3ad2d6: event misbehavior with certain widgets when embedding
    • 🔧 2026-10-08 check-in cb8992a33d on mistake, not merged: Fix a4fa3ad2d6: event misbehavior with certain widgets when embedding
    • 🔧 2026-10-08 check-in 75525b8851 on mistake, not merged: Fix a4fa3ad2d6: event misbehavior with certain widgets when embedding
    • 🔧 2026-10-09 check-in f98ef68b49 on x11-embed-focus-return, not merged: unixEmbed tests a4fa3ad2d6: make the focus tests independent of the previous tests. Restart the child process in the cleanup of unixEmbed-5.4 before destroyin
    • 🔧 2026-10-19 check-in 11bfb3794d on x11-embed-focus-return, not merged: Merge main. Main contains the earlier check-ins of this branch in its history, but not the fix for a4fa3ad2d6 and test unixEmbed-5.4, which the merge removed;
    • 🔧 2026-10-24 check-in a9a328bb7d on embed-container-destroyed-pixmap, not merged: Merge main.
    • 🔧 2026-10-12 check-in ec6cd2f9fd on x11-implicit-focus-release, in main, 9.0: Adapt regression test to use new utility proc "movePointerAwayFromTestArea" and move test constraint noWindowManager to a dedicated LOCAL TEST CONSTRAINTS secti
    • 🔧 2026-10-24 check-in abb3ccc33e on x11-embed-focus-return, not merged: Merge main.
    • 💬 2026-10-08 jan.nijtmans: I'm seeing those test failures on X11 in the fix branch, although I cannot explain it: ==== unixEmbed-6.1 EmbedGeometryRequest procedure, window changes size FAILED ==== Contents of test case: frame .f1 -container 1 -width 200 -height 50 pack .f1 childTkProcess eval "set w1 [winfo id .f1]" childTkProcess eval { destroy {*}[winfo children .] toplevel .t1 -use $w1 update bind .t1 {lappend x {configure .t1 %w %h}} se…
    • 💬 2026-10-08 serhiy.storchaka: I cannot reproduce these failures on X11: with branch x11-embed-focus-return, unixEmbed.test passes alone (6 runs, with and without a window manager) and in two full test suite runs, all under Xvfb; unixEmbed-5.4 runs and passes. The "broken pipe" errors mean that the child process used by the tests had exited, probably already during unixEmbed-5.4 or 6.1. Jan, how did you run the tests: on which X server and wind…
    • 💬 2026-10-08 jan.nijtmans: I cannot explain, but you can see the failed run here: [https://github.com/nijtmans/tk/actions/runs/37835555887/job/113511546477]
    • 💬 2026-10-09 serhiy.storchaka: This was a hard one: the failures were difficult to reproduce, and making the tests independent exposed further bugs. Two new tickets: * f1fe739a4f embedded application exits with an X error when its container is destroyed while it draws or claims the focus; proposed fix in branch embed-container-destroyed-pixmap. * 0c3bb74e37 without a window manager, an application keeps the implicit focus after another appl…
    Full verdict

    REPRODUCED (main). The fix exists on branch x11-embed-focus-return (eb2b41d5b3, test 63c88e8a6d, f98ef68b49, 11bfb3794d) and is not merged into main yet. Main's generic/tkFocus.c still ignores FocusIn/NotifyInferior, and so do 9.0 and 8.6.

    • Platform X11
    • Tested main (fb166f9f73) with the branch's test unixEmbed-5.4 copied into tests/unixEmbed.test: it FAILS on unmodified main. With the branch fix (eb2b41d5b3) applied to main, unixEmbed.test passes twice in a row (59 tests, 0 failed) and focus.test passes. The ticket's two-process script driven by XTest (click into the child entry, then type) did not reproduce, on any version back to 8.6.8, under twm (ClickToFocus), icewm, openbox or no WM: none of these WMs moves the X focus back to the container's toplevel after the click.
    • Cause generic/tkFocus.c:TkFocusFilterEvent drops FocusIn with detail NotifyInferior. When a click-to-focus WM sets the X focus back to the container's toplevel after the embedded application claimed it, Tk never passes the focus on to the embedded application again.
    • Prototype none (already on branch x11-embed-focus-return, and it still applies to current main with an offset). As the 2026-10-09 comment says, it must be merged with --baseline or cherry-picked, because the branch is already an ancestor of main.
    • Suggested ticket comment Still present on main as of fb166f9f73: unixEmbed-5.4 from x11-embed-focus-return fails there. With eb2b41d5b3 applied to current main, unixEmbed.test and focus.test pass. Ready to cherry-pick (or merge with --baseline) to main, and then to core-9-0-branch and core-8-6-branch, which have the same code.
    ⬜e7ac02121f2017bad reqheight after update idletask and strange borderpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket e7ac02121f · 88. Themed Tk · filed 2017-01-15

    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested main, 9.0, 8.6 on X11 with the default, clam, alt and classic themes. reqheight of the ttk::frame equals the child's reqheight, so the groove is drawn under the child.
    • Cause generic/ttk/ttkFrame.c:FrameMargins only uses the *widget* -borderwidth option. Frame.border draws with the style's value (default/classic set "." -borderwidth 1) or with the element default (clam/alt use 2), so the margins are 0 while a border is drawn. The same mismatch exists in Labelframe, which does query the style.
    • Prototype e7ac02121f.patch (30 lines, test ttk-10.6). When -relief is not flat (widget option or style), the margins come from Ttk_LayoutNodeInternalPadding of the "border" element. -relief is now GEOMETRY_CHANGED. Flat frames keep their old size, which limits the layout impact that bll mentioned. ttk.test, labelframe, notebook, panedwindow and nodecache pass. The new test fails without the fix.
    • Suggested ticket comment Still present in 8.6/9.x: FrameMargins() ignores the border width that Frame.border actually draws with (style or theme default). A fix that takes the margin from the border element when the relief is not flat is prototyped (test ttk-10.6). Flat frames keep their current size.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean (test offset 43)
    • Repro ttk::frame -relief groove holding a label, after update idletasks. Before, the frame's reqsize equals the label's in every theme (default 36x22, clam/alt 38x24). After: default/classic 38x24 vs label 36x22, clam/alt 42x28 vs 38x24.
    • Test ttk-10.6 (existing in patch; fails before: yes (10 10); passes after: yes (16 16))
    • Test file run ttk.test 114/0 (main 113/0), labelframe 21/0, notebook 64/0, panedwindow 412/0 + 41/0, nodecache 20/0, frame.test 207/0 (all the same as main)
    • Patch changes none
    Prototype verified 2026-10-24, Windows VM, generic patches
    • e7ac02121f: OK. ttk-10.6 fails before, passes after.
    ✅557441302b2016wm geometry returns wrong position for zoomed toplevelspatchREPRODUCED (9.1 main, 8.6 tip); prototype verified

    Open ticket 557441302b · 62. Geometry Functions · filed 2016-07-14

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-11-04 check-in e5a95ef737 on win-zoomed-geometry, in main, 9.0: Fix 557441302b: wm geometry returns wrong position for zoomed toplevels. On Windows the position of the restored window was reported with the size of the zoom
    • 🔧 2026-11-04 check-in 3c2445afc1 on win-zoomed-geometry, in main, 9.0: Fix 557441302b also for a toplevel iconified from zoomed: on Windows wm geometry reported the position of the restored window while it was iconified. Tests wm
    • 🔧 2026-11-05 check-in b362a6982f on trunk, in main, 9.0: Fix for 557441302b: wm geometry returns wrong position for zoomed toplevels
    • 🔧 2026-11-05 check-in 59adf2caf4 on core-9-0-branch, in main, 9.0, 8.6: Fix for 557441302b: wm geometry returns wrong position for zoomed toplevels
    • 🔧 2026-11-05 check-in 18879eb247 on core-8-6-branch, in 8.6: Fix for 557441302b: wm geometry returns wrong position for zoomed toplevels
    • 💬 2026-11-04 serhiy.storchaka: Fixed on branch win-zoomed-geometry: on Windows, wm geometry of a zoomed toplevel reports its actual position, as on X11, for example +-8+-8 (the frame border is off the screen). This includes a toplevel iconified from zoomed, which will be restored zoomed. The restored position is kept, and restoring the window still goes back to it. A wm geometry with a new position or size while the window is zoomed is still ne…
    • 💬 2026-11-05 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits b362a698, 59adf2ca, and 18879eb2. Tested that the bug is not present on macOS.
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip); prototype verified

    • Platform Windows
    • Tested X11 main and 8.6 under openbox and icewm: after -zoomed 1, wm geometry reports the zoomed position (+0+0 / -4-4).
    • Cause win/tkWinWm.c: UpdateGeometry updates wmPtr->x/y only in NormalState, so in ZoomState WmGeometryCmd reports the zoomed size (winPtr->changes) with the restored position (wmPtr->x/y).
    • Prototype 557441302b.patch (18 lines, not compiled, no Windows build): in ZoomState WmGeometryCmd reports winPtr->changes.x/y. wmPtr->x/y are left alone, so restoring still works. Repro win/557441302b.tcl.
    • Suggested ticket comment On X11 [wm geometry] of a zoomed window reports the actual position. On Windows the restored position is kept in wmPtr->x/y and reported together with the zoomed size. A prototype that reports the zoomed position is available but needs testing on Windows.
    Windows VM, run 3
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip); prototype verified
    • Tested repro (with the regexp fixed to accept +-8, work-win/557441302b.tcl): ref91 and 8.6 report zoomed 1023x697+303+175 (the restored position). The prototype reports 1023x697+-8+-8 (the real frame position; rootx=0, rooty=23), and restore gives back 200x200+303+175. The new test winWm-11.1 fails on ref91 and passes with the prototype. wm, winWm, winDialog, menu and winMenu tests: 1096 tests, 0 failed.
    • Cause win/tkWinWm.c WmGeometryCmd reports wmPtr->x/y, which hold the restored position in ZoomState
    • Prototype 557441302b.patch (35 lines: the original 18 plus test winWm-11.1). Note: the result has the X11-style +-8+-8 for the off-screen frame border. Tk parses this back as an absolute position.
    • Suggested ticket comment Confirmed on Windows 11 (9.1, 8.6). A patch makes wm geometry report the position of the maximized window (e.g. +-8+-8) while the restored geometry is kept. Test winWm-11.1 is included.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean (code). The test was renumbered winWm-11.1 → winWm-14.1, because 11.1 now exists.
    • Repro before: winWm-14.1 fails (the zoomed geometry reports the restored position) / after: passes (+-8+-8, and restore gives 200x200+303+175)
    • Test winWm-14.1 (existing in patch; fails before: yes; passes after: yes)
    • Test file run winWm.test 39/0, wm.test 306/0 (pre-existing failures: none)
    • Patch changes the test was renumbered, and unset res x y was added to its cleanup. 557441302b.patch is the current file. 557441302b.patch.new is the older code-only version (the same code, without the test) and can be deleted.
    ⬜5d5c43acad2016linux: when window is maximized, menu replacement failspatchREPRODUCED (main, 8.6)

    Open ticket 5d5c43acad · 12. Unix Menus · filed 2016-06-07

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main and 8.6 under openbox, icewm and fvwm (bug) and under twm and no WM (no maximize, OK). After . configure -menu .n in a maximized window, the new menubar is 1 px high and the toplevel is placed at y=1 in the wrapper.
    • Cause unix/tkUnixWm.c. TkUnixSetMenubar sets menuHeight=1 because the clone's size is not computed yet. UpdateGeometryInfo asks for wrapper height +1, which the WM refuses. When MenubarReqProc later sets menuHeight=35, the wanted wrapper size equals the current one, so UpdateGeometryInfo returns early ("already just the size we want") and never rearranges the menubar or toplevel inside the wrapper.
    • Prototype 5d5c43acad.patch (38 lines, no automated test: needs a WM that refuses resizing). Factors the tail of ConfigureEvent into UpdateWrapperContents() and calls it on that early return when the menubar height differs from menuHeight. Verified under openbox, icewm and fvwm. unixWm, wm, menu and unixMenu tests show the same results as before, with and without openbox.
    • Suggested ticket comment Still present in 8.6/9.x. UpdateGeometryInfo() skips re-laying out the wrapper when the maximized window's size doesn't change, so the new menubar keeps its initial 1-pixel height. Prototype fix available.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main clean (offsets 8-11 lines)
    • Repro work-a3/5d.tcl under openbox (zoomed, then . configure -menu .n): before no mapped menubar and toplevel rooty=22 / after menubar .#n mapped with h=35 and toplevel rooty=56 (same after switching back to .m)
    • Test none (needs a WM that refuses to resize a maximized window)
    • Test file run no WM: unixWm.test 284/0, wm.test 302/0, menu.test 574/0, unixMenu.test 122/0 (same as unpatched)
    • Patch changes none. Review: only the size-only branch of UpdateGeometryInfo gets the new call. The move+resize branch's "already this size" early return does not, but the maximized case goes through the size-only branch.
    ✅fecd6bd2fa2016linux only: window -zoomed attribute is restored when it should not be.patchREPRODUCED (main, 8.6)

    Open ticket fecd6bd2fa · 48. Geometry Management · filed 2016-04-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-03 check-in 1fd4e8ee60 on x11-keep-wm-state, in main, 9.0: Fix fecd6bd2fa: linux only: window -zoomed attribute is restored when it should not be. CheckNetWmState only updated the current attributes, so mapping the wi
    • 🔧 2026-10-07 check-in 9b3a47eb22 on trunk, in main, 9.0: Fix for fecd6bd2fa: linux only: window -zoomed attribute is restored when it should not be.
    • 🔧 2026-10-07 check-in 78765cad24 on core-9-0-branch, in main, 9.0, 8.6: Fix for fecd6bd2fa: linux only: window -zoomed attribute is restored when it should not be.
    • 🔧 2026-10-07 check-in 129fd4e2cb on core-8-6-branch, in main, 9.0, 8.6: Fix for fecd6bd2fa: linux only: window -zoomed attribute is restored when it should not be.
    • 💬 2026-10-03 serhiy.storchaka: Simpler reproducer: wm attributes . -zoomed 1 button .b -text "withdraw + deiconify" -command {wm withdraw .; after 500 wm deiconify .} pack .b Un-maximize the window with the title bar button, then click the button: the window is maximized again. Tk only updated the current attributes from _NET_WM_STATE, not the requested ones, so mapping the window again restored the stale requested state. Proposed fix in branch…
    • 💬 2026-10-07 nemethi: Serhiy, many thanks for the fix! Merged into main, core-9-0-branch, and core-8-6-branch by commits 9b3a47eb, 78765cad, and 129fd4e2.
    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main, 8.6 tip × openbox, icewm (reproduced; the WM un-maximize was simulated by sending a _NET_WM_STATE remove client message the way a WM button does); fvwm maximizes only partly and is not usable for this.
    • Cause unix/tkUnixWm.c: CheckNetWmState() updates only wmPtr->attributes when the WM changes _NET_WM_STATE, but UpdateNetWmState() (called before every map) rewrites the property from wmPtr->reqState, which still holds the old zoomed=1 from wm attributes -zoomed 1. Same for -topmost and -fullscreen.
    • Prototype fecd6bd2fa.patch (44 lines, no test). CheckNetWmState() copies topmost/zoomed/fullscreen into reqState unless the window is withdrawn (WMs delete the property on withdraw). Fixes the repro; a window that stays maximized is still maximized after withdraw/deiconify; wm.test and unixWm.test pass under openbox. No test added because Tcl alone cannot send a WM client message.
    • Suggested ticket comment Still present in 8.6 and 9.x. CheckNetWmState() does not update the requested state (reqState) when the user changes it through the WM, so UpdateNetWmState() restores the stale -zoomed state on remap. Fix prototyped: sync reqState from _NET_WM_STATE while the window is not withdrawn.
    ✅bffa794b1b2016BadAlloc (insufficient resources for operation) in the grid commandpatchREPRODUCED (main, 8.6)

    Open ticket bffa794b1b · 49. [grid] · filed 2016-04-08

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-02 check-in c91110556f on x11-pixmap-size-limit, in main, 9.0: Fix bffa794b1b: BadAlloc (insufficient resources for operation) in the grid command.
    • 🔧 2026-10-03 check-in bc5e8ae560 on trunk, in main, 9.0, 8.6: Fix for bffa794b1b: "BadAlloc (insufficient resources for operation)" in the grid command.
    • 🔧 2026-10-03 check-in f3849ef4c1 on core-9-0-branch, in main, 9.0: Fix for bffa794b1b: "BadAlloc (insufficient resources for operation)" in the grid command.
    • 🔧 2026-10-03 check-in dee12bec14 on core-8-6-branch, in main, 9.0, 8.6: Fix for bffa794b1b: "BadAlloc (insufficient resources for operation)" in the grid command.
    • 🔧 2026-10-04 check-in 2cf953f38c on trunk, in main, 9.0, 8.6: Test unixbutton-3.1: place the huge widgets instead of gridding them. The gridded toplevel exceeded the 16-bit limit of X window sizes, and if it was already ma
    • 💬 2026-10-02 serhiy.storchaka: Proposed fix in branch x11-pixmap-size-limit: Tk_GetPixmap() clips the size to 32767, the largest size the X server accepts (32768 to 65535 fail with BadAlloc, larger sizes with BadValue). Drawing beyond it is lost, but there is no crash. New test unixbutton-3.1.
    • 💬 2026-10-03 nemethi: Serhiy, many thanks for the ticket and the fix! Merged into main, core-9-0-branch, and core-8-6-branch by commits bc5e8ae5, f3849ef4, and dee12bec.
    • 💬 2026-10-04 nemethi: Unfortunately, the new test case unixbutton-3.1 in unixButton.test failed in the CI run of last night. Before committing the merge, I had tested it successfully via make test TESTFLAGS="-verbose bps -file ../tests/unixButton.test" but it failed in the CI and also in the complete manual test started via make test I have spent a lot of time to find out why the two methods behave differently, but without success. By …
    • 💬 2026-10-04 serhiy.storchaka: The test gridded the three widgets, so the toplevel became about 120112x80023, beyond the 16-bit limit of X window sizes. When . is already mapped by previous tests, as in a full test run, the X server truncates its size and grid shrinks the widgets. Fixed in 2cf953f38c: the widgets are now placed, so the toplevel is not resized. Without the fix the test still crashes with an X error.
    • 💬 2026-10-06 serhiy.storchaka: In 0309359d58 Don disabled unixbutton-3.1 because it locked up his system. I ran the test multiple times in Xvfb and on a real display in Xwayland, and did not notice any issues. My guess is that the cause is the button, which was huge in both directions and needed a 32767x32767 pixmap: 4 GB in Xvfb, 8 GB of GPU buffers in Xwayland here. With less memory, or with a GPU with less video memory, this can stall the …
    • 💬 2026-10-06 dgp: The revised test does not lock up my system.
    • 💬 2026-10-07 nemethi: Jan merged the improved test into core-9-0-branch by commit 3fca5ed4 and into main by commit 909a5f45. Now I performed the missing merge into core-8-6-new by commit 2d437e24. Closing the ticket (again), with resolution "Fixed".
    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested main (0x1000 zeros crashes), 8.6 tip (0x1400 crashes; its default font is narrower) on Xvfb.
    • Cause unix/tkUnixXId.c:Tk_GetPixmap() passes the widget size straight to XCreatePixmap(). X servers reject dimensions > 32767 with BadAlloc, and values > 65535 wrap in the 16-bit protocol field. The first caller to hit this is TkpDisplayButton (double buffering).
    • Prototype bffa794b1b.patch (74 lines incl. test, 15 in C, test unixbutton-3.1). Tk_GetPixmap() clamps width/height to 32767, so drawing past that is lost but nothing crashes. The repro no longer crashes. unixbutton-3.1 crashes the test run without the fix. unixButton, button, frame and canvas tests pass.
    • Suggested ticket comment Still crashes on 8.6 and 9.x (X11 only). Tk_GetPixmap() passes sizes above 32767 to XCreatePixmap(), and the server fails with BadAlloc. Prototype: clamp the pixmap size in Tk_GetPixmap() (unix), test unixbutton-3.1.
    ✅131adc6b772016bound buttonpress is passed on to other programpatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 131adc6b77 · 01. Bindings · filed 2016-03-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-23 check-in 27801b131f on pointer-dead-restrict-window, in main: Fix 131adc6b77: on Windows and macOS, when the window that got a ButtonPress was destroyed while the button was down, the following ButtonRelease and Motion e
    • 🔧 2026-11-10 check-in d1af9365de on trunk, in main, 9.0: Fix for 131adc6b77: bound buttonpress is passed on to other program
    • 🔧 2026-11-10 check-in 5f5bd4feee on core-9-0-branch, in 9.0, 8.6: Fix for 131adc6b77: bound buttonpress is passed on to other program
    • 🔧 2026-11-10 check-in f84582b421 on core-8-6-branch, in 8.6: Fix for 131adc6b77: bound buttonpress is passed on to other program
    • 💬 2026-10-23 serhiy.storchaka: Confirmed on Windows 11 (8.6.18, 9.1): when a binding on the ButtonPress destroys the window, the ButtonRelease goes to the window below. Tk had released its mouse grab when the window died, so it sent the remaining button and motion events to the window then under the pointer. X11 is not affected. Fixed on branch pointer-dead-restrict-window (27801b131f): these events are discarded until all buttons are release…
    • 💬 2026-11-10 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits d1af9365, 5f5bd4fe, and f84582b4.
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested X11 main, 8.6 tip, with no WM and with openbox. Real XTEST clicks on Xvfb: a toplevel destroyed from <1> or <Double-1> leaks neither the press nor the release to the toplevel below.
    • Cause likely Windows-specific. After the window is destroyed in the WM_LBUTTONDOWN handler, the pending WM_LBUTTONUP (or the second click's messages) goes to the window now under the pointer. bll suggests draining WM_MOUSEFIRST..WM_MOUSELAST before destroy, like EatSpuriousMessageBugFix() in win/tkWinDialog.c.
    • Prototype none. Repro script win/131adc6b77.tcl (interactive single/double click; prints OK/BUG and writes 131adc6b77.out).
    • Suggested ticket comment Does not happen on X11 (checked on 9.x/8.6). Needs checking on Windows 11 with the attached script.
    Windows VM, run 3
    • Platform Windows (generic/tkPointer.c, which is also used on macOS)
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/131adc6b77.tcl with real clicks from act.exe (a single click, and a double click 100 ms apart). On both versions . ButtonRelease-1 reaches the window below, in both cases (the press does not leak, and there is no Double-1). With the prototype: no events below in either case. bind, event, grab, winfo, frame, button, listbox, entry, text, scale, scrollbar and menu tests: 3630 tests, 0 failed.
    • Cause generic/tkPointer.c: when the implicit-grab (restrict) window is destroyed while a button is down, TkPointerDeadWindow clears restrictWinPtr. The following release (and motion) is then sent to the window now under the pointer. On X11 the server drops events for the destroyed grab window.
    • Prototype 131adc6b77.patch (23 lines). A restrictWinDead flag, set in TkPointerDeadWindow when buttons are down, makes Tk_UpdatePointer discard button and motion events until all buttons are released, then restores the grab capture. No automated test: event generate does not go through Tk_UpdatePointer. macOS gets the same change (untested).
    • Suggested ticket comment Confirmed on Windows (9.1, 8.6): the ButtonRelease of a click whose press destroys the toplevel goes to the window below. A patch in tkPointer.c discards pointer events until the buttons are released, as X11 does for a destroyed grab window.
    ⬜6e3c4fb23f2015Can crash on Tk load in separate thread after Tk loaded in main threadREPRODUCED (main, 8.6)

    Open ticket 6e3c4fb23f · 99. Other · filed 2015-03-10

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested a C driver (work-a2/thr.c: every thread does CreateInterp, Tk_Init, a canvas with text items, then destroy and Tcl_FinalizeThread). It was linked against main and 8.6 tip and run on Xvfb.
      • With Xft fonts, both versions fail after 1–3 rounds, even with a single worker thread at a time and whether or not Tk is also loaded in the main thread. The X error is BadRequest/BadLength ("major opcode 0") in XRenderQueryFormats ← XftDefaultSubstitute ← tkUnixRFont.c:InitFont. Calling XInitThreads() first does not help.
      • Without fonts (frame only), 8.6 passes serially and with 4–8 concurrent threads.
      • Without fonts on main, serial runs pass, but concurrent threads abort with a glib assertion or malloc corruption in tk_atk_accessible_finalize.
    • Cause two separate problems.

    1. The per-display cache in libXrender/libXft goes stale once one thread's display has been closed with XCloseDisplay (TkpCloseDisplay at thread exit) and another thread opens a new one. Tk's own font code has only thread-specific data.

    2. main only (a new regression): unix/tkUnixAccessibility.c keeps process-global unlocked state (tk_to_atk_map, toplevel_accessible_objects, tk_root_accessible, acc_context), so ATK is not thread-safe.

    • Prototype none. (1) is outside Tk or would need a redesign, for example never closing displays or sharing one per process. (2) needs a mutex or thread-specific data across the ATK module and should be its own ticket.
    • Suggested ticket comment Still reproducible on 8.6 and main on Linux. Creating Tk in successive threads fails in Xft/Xrender (XRenderQueryFormats) after a previous thread closed its display, even without concurrency. On main, the new ATK accessibility code also has unprotected global state that crashes when Tk runs concurrently in several threads.
    ⬜0cb85bcb482014Mysteries when all default named fonts are deletedREPRODUCED (main, 8.6); this is a design question more than a bug

    Open ticket 0cb85bcb48 · 44. Generic Fonts · filed 2014-10-24

    Full verdict

    REPRODUCED (main, 8.6); this is a design question more than a bug

    • Platform generic
    • Tested main, 8.6 on Xvfb, no WM
    • Cause font delete on a named font that is still in use only marks it deletePending. Widgets that hold it (and new widgets that ask for the same name while it is held) keep the old attributes. After the last reference goes away, "TkFixedFont" etc. are parsed as a font description with an unknown family. They then fall back to the system default (here Noto Sans 12, proportional), so the text widget metrics change. Recreating the named font updates all users. font configure TkFixedFont errors even while the font is still effectively in use.
    • Prototype none (policy decision: whether to make the Tk* standard fonts undeletable, or reset them instead of deleting)
    • Suggested ticket comment Still the same on main and 8.6. Deleted standard fonts silently fall back to the default font (proportional) once no widget holds them. Should font delete refuse, or reset instead, for the Tk* standard fonts?
    ⬜3c8a969ac12014[selection handle] doesn't replace a previously installed utf-8 handlerREPRODUCED (main, 9.0, 8.6, 8.5); duplicate of af31a84c91 / [2443069]

    Open ticket 3c8a969ac1 · 53. [selection] · filed 2014-09-22

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5); duplicate of af31a84c91 / [2443069]

    • Platform generic (X11)
    • Tested main, 9.0, 8.6, 8.5 on Xvfb: after replacing handler h1 with h2, selection get -type UTF8_STRING still returns h1
    • Cause generic/tkSelect.c Tk_CreateSelHandler: the implicit UTF8_STRING handler is only created, never replaced ("Looks like we had a utf-8 target already. Leave it alone.").
    • Prototype already exists as af31a84c91.patch (tests select-14.2/14.3). I verified that it fixes this ticket's repro on main (UTF8_STRING now returns h2).
    • Suggested ticket comment Duplicate of af31a84c91 (and [2443069]). This is still present on main and 8.6 and will be fixed there.
    ⬜d1d4286dda2013ttk::scrollbar behaves differently than tk::scrollbar on win32patchREPRODUCED (main, 8.6). Holding Button-1 in the ttk trough pages all the way to the end instead of stopping at the pointer. On X11 the classic tk::…

    Open ticket d1d4286dda · 88. Themed Tk · filed 2013-08-03

    Full verdict

    REPRODUCED (main, 8.6). Holding Button-1 in the ttk trough pages all the way to the end instead of stopping at the pointer. On X11 the classic tk::scrollbar does the same (it is not native there); the native win32 scrollbar stops at the pointer. The "trough turns dark while pressed" part is a theme/visual issue and not addressed.

    • Platform generic (reported for Windows)
    • Tested main, 8.6 on Xvfb, no WM, with tk and ttk scrollbars (event generate press at 40% of height, hold 3 s): both end at 0.9–1.0
    • Cause library/ttk/scrollbar.tcl Press: ttk::Repeatedly Scroll $w ±1 pages repeats unconditionally until release.
    • Prototype d1d4286dda.patch (about 22 code lines + 19-line test scrollbar-10.3). A new ScrollPages proc pages only while the thumb has not yet reached the pointer, and Drag tracks the pointer position during a trough press. The test fails without the fix and passes with it, and scrollbar.test passes. The change applies on all platforms. It could be limited to win32, but stopping at the pointer matches GTK, Qt and Windows.
    • Suggested ticket comment ttk::scrollbar auto-repeats page scrolling until button release. A prototype makes it stop once the thumb reaches the pointer, as native scrollbars do.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean (test offset 1)
    • Repro scrollbar-10.3 (press at 40% of the trough, hold 3 s). Before, the scrollbar pages to the end and the result is 1 0. After, it stops at the pointer and the result is 1 1.
    • Test scrollbar-10.3 (existing in patch; fails before: yes; passes after: yes)
    • Test file run scrollbar.test 169/0 (core) + 13/0 (ttk; main 12/0)
    • Patch changes in the test, after 3000 {set done 1} ; vwait done is replaced with pause 3000. Backup: d1d4286dda.patch.1024.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • d1d4286dda: OK. scrollbar-10.3 fails before, passes after.
    ⬜3611174fff2013Empty text items aren't found by "find enclosed"patchREPRODUCED (main, 9.0, 8.6); not in 8.5, 8.4

    Open ticket 3611174fff · 05. Canvas Items · filed 2013-04-17

    Full verdict

    REPRODUCED (main, 9.0, 8.6); not in 8.5, 8.4

    • Platform generic
    • Tested 9.1b1, 9.0.5, 8.6.18 give find enclosed = "" after -text "". 8.5.19 and 8.4.20 give 1. find overlapping works on every version, and angled empty text is found.
    • Cause generic/tkFont.c:Tk_IntersectTextLayout. Since 6c007ba101 (2009-07-20, fix for 2328657, the trailing-newline case of canvText-18.1) it skips empty chunks. For empty text the only chunk is empty, so the function always returns 0 ("overlaps") and never 1 ("inside").
    • Prototype 3611174fff.patch (27 changed lines). Empty chunks are skipped only if the layout also has a non-newline, non-empty chunk. New test canvText-18.2 covers "" and "\n". canvText, canvas, font and canvWind tests pass.
    • Suggested ticket comment Still present in 8.6 and 9.x. 6c007ba101 made Tk_IntersectTextLayout ignore empty chunks, and for empty text that is the only chunk. Prototype fix: use empty chunks when the layout has nothing else; test canvText-18.2.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main ported (the test hunk was rejected: main's canvText-19.1 no longer has the aqua constraint or a blank-line context. canvText-18.2 is inserted after 18.1 as before. tkFont.c applied clean at offset +157.)
    • Repro before find enclosed on "" and "\n" items gives {} {} / after {1 2} {}
    • Test canvText-18.2 (existing in patch; fails before: yes; passes after: yes)
    • Test file run canvText.test 101/0, font.test 304/0, canvas.test 175/0 (pre-existing failures: none)
    • Patch changes rebased the test hunk only
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 3611174fff: OK. canvText-18.2 fails before, passes after; font.test 304/0.
    ✅3604346fff2013Checkin a7eee0499a broke Ttk docs.patchREPRODUCED (main, 8.6). unix/installManPage still maps every non-alphanumeric sequence to _. ttk::treeview is installed as ttk_treeview.n alt…

    Open ticket 3604346fff · 84. Unix Build · filed 2013-02-12

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-20 check-in dc7a50073b on man-colon-names, in main, 9.0: Fix 3604346fff: "::" in manual page names was replaced with "_" on installation, so "man ttk::treeview" did not find the page. Keep ":" in the names, and inst
    • 🔧 2026-10-23 check-in 974d5f0485 on man-colon-names, in main, 9.0: Merge-mark
    • 💬 2026-10-20 serhiy.storchaka: Fixed in branch man-colon-names (dc7a50073b), and the same in Tcl (Tcl ticket 0ef68e99e1). This is the change suggested in 7de29bd83e: unix/installManPage keeps : in the names, so man ttk::treeview and man oo::class find the pages. The pages are also installed under the old names with _ (ttk_treeview.n), so man ttk_treeview still works. None of these 40 names conflicts with a name of another page. Tested with …
    • 💬 2026-10-23 jan.nijtmans: I like this solution, and it's fully upwards compatible. Fixed in a248516e5b and core-9-0-branch.
    Full verdict

    REPRODUCED (main, 8.6). unix/installManPage still maps every non-alphanumeric sequence to _. ttk::treeview is installed as ttk_treeview.n although its .TH/NAME says ttk::treeview. Tcl's installManPage (tclmain) does the same.

    • Platform build
    • Tested ran installManPage from main and from 8.6 on doc/*.n into a scratch directory
    • Cause unix/installManPage, the sed s/[^ A-Za-z0-9][^ A-Za-z0-9]*/_/g from dkf's a7eee0499a (2012)
    • Prototype 3604346fff.patch (2-line change, no test). This is Jan's 7de29bd83e from branch bug-3604346: : is kept, so pages install as ttk::treeview.n, tk::mac.n, safe::loadTk.n, etc. I checked all of doc/*: only :: names change. It is a policy decision, and Tcl's installManPage would need the same change.
    • Suggested ticket comment Still present. installManPage maps :: to _, so man ttk::treeview fails. Branch bug-3604346 (7de29bd83e) keeps : in page file names; this needs a decision (together with Tcl) whether to merge it or close as won't-fix.
    ⬜3603397fff2013ttk::sizegrip should be hidden for maximised windowspatchREPRODUCED (9.1 main, 8.6 tip); prototype verified

    Open ticket 3603397fff · 88. Themed Tk · filed 2013-02-05

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip); prototype verified

    • Platform Windows (generic Tcl code)
    • Tested main, 8.6 × openbox, icewm (X11): no visible effect. Windows repro: win/3603397fff.tcl
    • Cause library/ttk/sizegrip.tcl ttk::sizegrip::Press does not check whether the toplevel is maximized
    • Prototype 3603397fff.patch (7 lines, no test: needs a WM). Press returns early when [wm state $top] eq "zoomed" (Windows/macOS) or [wm attributes $top -zoomed] (X11). Verified on X11/openbox that a press no longer starts a drag on a zoomed window; ttk.test passes. The patch does not hide the grip or change the cursor; hiding is an RFE.
    • Suggested ticket comment The sizegrip bindings ignore the zoomed state. A small fix makes ttk::sizegrip::Press do nothing on a maximized toplevel; hiding the grip would be a separate enhancement.
    Windows VM, run 3
    • Platform Windows (library)
    • Verdict REPRODUCED (9.1 main, 8.6 tip); prototype verified
    • Tested win/3603397fff.tcl: ref91 and 8.6 give "state after drag=zoomed, restored size=923x597 (expected 300x200)". With the prototype, restored size=300x200. ttk tests on the prototype build: 2045 tests, 0 failed.
    • Cause library/ttk/sizegrip.tcl Press ignores the zoomed state
    • Prototype 3603397fff.patch (unchanged, 7 lines)
    • Suggested ticket comment Confirmed on Windows 11: dragging the sizegrip of a maximized window changes its restored size. A 7-line fix in sizegrip.tcl ignores the grip while the toplevel is zoomed.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main clean
    • Repro work-a2/3603.tcl (200x150 window, wm attributes -zoomed 1, then press and drag the sizegrip by -100,-100, then unzoom). Unlike the triage note, this does reproduce on X11. openbox before: restored size 1180x903 (wrong), after: 200x150. icewm before: 1180x878, after: 200x150. Without a WM, -zoomed is never set, so the result is the same before and after.
    • Test none (needs a WM that honours _NET_WM_STATE_MAXIMIZED)
    • Test file run ttk/ttk.test 113/0 (pre-existing failures: none)
    • Patch changes none
    🔧3598564fff2012wm maxsize does not take into account task bar on Windows 7REPRODUCED (9.1 main, 8.6 tip)

    Open ticket 3598564fff · 62. Geometry Functions · filed 2012-12-26

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-11-05 check-in 9daa5310c5 on default-maxsize-workarea, not merged: Fix 3598564fff: the default maximum size of a toplevel did not take into account the taskbar or panels, so a window whose natural size was larger than the scr
    • 🔧 2026-11-05 check-in 7ab79e1ad8 on default-maxsize-workarea, not merged: Fix 3598564fff: the default maximum size of a toplevel did not take into account the taskbar or panels, so a window whose natural size was larger than the scr
    • 🔧 2026-11-06 check-in a0e386c4bb on default-maxsize-workarea, not merged: Rebase to main
    • 💬 2026-11-05 serhiy.storchaka: Still reproducible on Windows 11: the window extends under the taskbar and past the bottom of the screen. X11 has the same problem: the default maximum size is the size of the display less fixed 15 and 30 pixels for the decorations, which are wrong for most modern window managers, and Tk did not limit the natural size of a window by it at all. Fixed on branch default-maxsize-workarea: the default wm maxsize is the…
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested code reading; Windows repro: win/3598564fff.tcl
    • Cause win/tkWinWm.c UpdateGeometryInfo/GetMaxSize. defMaxWidth/Height come from WM_GETMINMAXINFO ptMaxTrackSize, which is the whole screen and not the work area, so the natural size is clamped only to the screen height. A fix would clamp the default (not the user-set) maxsize to the monitor's rcWork (GetMonitorInfo). That changes behaviour and could also restrict manual resizing, so it needs discussion.
    • Prototype none (behaviour change; untestable here)
    • Suggested ticket comment The default maxsize comes from ptMaxTrackSize, which covers the whole screen and does not exclude the task bar. Clamping the default maximum to the monitor work area (rcWork) would fix this. X11 behaves the same way with panels.
    Windows VM, run 3
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/3598564fff.tcl on Windows 11 at 1023x768 with a 48 px task bar (work area bottom 720). Both versions give wm maxsize=1027x749, and the window bottom is at 780: below the task bar and even below the screen.
    • Cause win/tkWinWm.c default max size from ptMaxTrackSize (whole screen, and bigger than it), not the monitor work area
    • Prototype none (behaviour change, see result-29)
    • Suggested ticket comment Confirmed on Windows 11 (9.1, 8.6): the default maxsize (1027x749 on a 1023x768 screen with a 48 px task bar) lets the natural size extend under the task bar. Clamping the default maximum to the monitor work area would fix it.
    ✅3515522fff2012Crash when adding system menu to undocked windowpatchREPRODUCED (9.1 main, 8.6 tip, 8.5.19): win/3515522fff.tcl crashes with an access violation (exit code -1073741819)

    Open ticket 3515522fff · 13. Win Menus · filed 2012-04-06

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-25 check-in 0a7452d86d on wm-forget-release-menubar, in main, 9.0: Fix 3515522fff: [wm forget] did not release the menubar of the toplevel, so destroying it or the menu later crashed on Windows (TkWinGetWrapperWindow() derefe
    • 💬 2026-10-25 serhiy.storchaka: Confirmed on Windows 11 with 8.6 and 9.1: the script crashes in TkWinGetWrapperWindow(). As you found, wm forget does not release the menubar, so the frame stays registered as its owner. On X11 this does not crash, but a new menu with the same name then creates a menubar clone inside the frame. Fixed on branch wm-forget-release-menubar 0a7452d86d: the menubar is released when a toplevel becomes a frame, and TkWi…
    • 💬 2026-10-30 nemethi: Merged by Jan into main and core-9-0-branch by commits 42b7b7e9 and c2f2b683. Backported to core-8-6-branch by commit [9d8564a29. Closing the ticket.
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip, 8.5.19): win/3515522fff.tcl crashes with an access violation (exit code -1073741819)

    • Platform Windows (a related X11 crash exists in 8.6)
    • Tested X11 main, 9.0, 8.6 tip, 8.6.8, 8.5. The ticket script survives on X11 everywhere. A variant (menubar, wm forget, wm manage, destroy, then exit) crashes 8.6 with SIGBUS in TkUnixSetMenubar from TkDestroyMenu at exit, but not main/9.0. Windows repro: win/3515522fff.tcl
    • Cause generic/tkFrame.c TkMapTopFrame. On wm forget the toplevel becomes a frame but keeps its menubar registration: the menu's topLevelList entry and the MENUBAR clone with parentTopLevelPtr set. On Windows, later destroying the clone's .system cascade calls TkWinGetWrapperWindow (win/tkWinMenu.c TkpDestroyMenu), which dereferences the now-NULL wmInfoPtr. On 8.6 X11, forget+manage leaves a stale duplicate entry that is used after the window is freed.
    • Prototype 3515522fff.patch (35 lines, test wm-forget-2.2). It releases the menubar in TkMapTopFrame on the toplevel-to-frame transition, and makes TkWinGetWrapperWindow return NULL when wmInfoPtr is NULL. The 8.6 backport is work-a1/3515522fff-8.6.patch: the tkFrame.c part only, verified to fix the 8.6 X11 exit crash. wm/frame/menu/menubut/unixMenu/menuDraw tests: no new failures on X11 main. The new test does not detect the bug on X11 (it passes before and after the fix); it is meant to catch the crash on Windows.
    • Suggested ticket comment Still present: wm forget leaves the menubar registered to the former toplevel, so the Windows system-menu code later dereferences a NULL wmInfoPtr. 8.6 on X11 also crashes after forget/manage/destroy. Fix prototype: release the menubar in TkMapTopFrame when a toplevel becomes a frame, and guard TkWinGetWrapperWindow. Needs testing on Windows.
    Windows VM, run 4
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip, 8.5.19): win/3515522fff.tcl crashes with an access violation (exit code -1073741819)
    • Tested the repro on all three builds, then the prototype built in tk-wt91. With the prototype the repro ends with "OK - no crash". The new test wm-forget-2.2 crashes tktest91 from ref91 (access violation), so it detects the bug on Windows. wm.test, frame.test, menu.test, winMenu.test, menubut.test with the prototype: 1308 tests, 0 failed.
    • Cause as in result-30. After wm forget the menubar stays registered, and TkpDestroyMenu → TkWinGetWrapperWindow dereferences a NULL wmInfoPtr.
    • Prototype 3515522fff.patch (unchanged, 85 lines including test wm-forget-2.2). Compiles cleanly with MSVC.
    • Suggested ticket comment Confirmed on Windows 11 with 8.5, 8.6 and 9.1: the crash is a NULL wmInfoPtr in TkWinGetWrapperWindow after wm forget. Fix: release the menubar when a toplevel becomes a frame, and guard TkWinGetWrapperWindow; test wm-forget-2.2.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (generic part on X11) / NEEDS-WINDOWS (win/tkWinWm.c part)
    • Applies to main clean (tkFrame.c -1, wm.test +63, win/tkWinWm.c +99 lines offset; the win hunk still applies)
    • Repro work-a2/3515b.tcl. On X11 main, the stale menubar registration is visible: toplevel .t -menu .m; wm forget .t; destroy .m; menu .m creates a menubar clone .t.#m inside the frame .t. After wm manage .t there are two clones (.t.#m, .t.#m1). With the patch: no clone while forgotten, and exactly one after manage.
    • Test wm-forget-2.3 (new; fails before: yes; passes after: yes). wm-forget-2.2 (existing in patch) passes before and after on X11; it targets the Windows crash. I added a -cleanup to it.
    • Test file run wm.test 304/0, frame.test 207/0, menu.test 574/0, unixMenu.test 122/0 (pre-existing failures: none)
    • Patch changes added test wm-forget-2.3 and the -cleanup for wm-forget-2.2. Win part reviewed: with the NULL return, the callers in tkWinMenu.c are fine (line 695 and line 772 check for NULL; at line 402 GetSystemMenu(NULL, TRUE) is harmless). tkWinWm.c:6956 is only reached for toplevels.
    • Windows tests to run wm.test (wm-forget-2.2, 2.3 and the whole file), winWm.test, winMenu.test, menu.test, frame.test, plus win/3515522fff.tcl
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean (the 21:43 version)
    • Repro before: wm-forget-2.2 crashes tktest at once / after: passes
    • Test wm-forget-2.2 (existing; fails before: yes, crash; passes after: yes)
    • Test file run wm.test 303/0 (B3) and 306/0 (B6), frame.test 207/0, menu.test 574/0, winMenu.test 148/0, menubut.test 117/0 (pre-existing failures: none)
    • Patch changes none
    ⬜3466588fff2011panedwindow resize problem with text -setgrid 1patchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 3466588fff · 48. Geometry Management · filed 2011-12-28

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested main, 8.6.18, 8.5.19 on X11: no WM, openbox, icewm. The toplevel's pixel size never changes when the sash is moved. The gridded size reported by wm geometry does change (64x30 -> 64x24).
    • Cause the panedwindow's requested height depends on the sash position, so the grid base size (reqHeight - gridHeight*inc) changes when the sash moves. On X11 the pixel size is kept, so only the reported grid units change. tkWinWm apparently keeps the grid units instead, so the pixel size changes. The fix would be in tkWinWm.c.
    • Prototype none (Windows only). Repro: win/3466588fff.tcl
    • Suggested ticket comment Not reproducible on X11: the toplevel keeps its pixel size, and only the grid units reported by wm geometry change. Probably a tkWinWm.c issue.
    Windows VM, run 4
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/3466588fff.tcl on both. Each sash move makes the toplevel grow: 518x490 → 586 → 682 → 746 px high (the last move is -40 px, and the growth stops at the screen size). wm geometry stays 64x30 and then becomes 64x28.
    • Cause not pinned down. After a sash move the panedwindow requests its current pane sizes, so the grid base (reqHeight - reqGridHeight*inc) grows. win/tkWinWm.c UpdateGeometryInfo keeps wmPtr->height in grid units (30), so the pixel height grows by (30-24)*16 px each time. X11 converts back from the actual pixel size and does not grow. The tkUnixWm.c and tkWinWm.c formulas are the same; the difference is probably in the ConfigureNotify/ConfigureTopLevel feedback.
    • Prototype none (needs more study of the Windows grid feedback)
    • Suggested ticket comment Still present in 8.6.18 and 9.1 on Windows 11: every sash move grows the gridded toplevel by (gridHeight - text -height) lines. X11 is not affected.
    2026-10-25 prototype, Windows VM
    • Status PROTOTYPED
    • Cause win/tkWinWm.c:TopLevelReqProc schedules UpdateGeometryInfo for every geometry request. After a sash move, the panedwindow requests its current pane sizes, so the toplevel's reqHeight grows. With a gridded explicit size (wmPtr->height = 30, reqGridHeight 24), UpdateGeometryInfo then computes reqHeight + 6*inc and the window grows. unix/tkUnixWm.c:TopLevelReqProc ignores geometry requests when explicit dimensions are set, so on X11 the pixel size stays and wm geometry reports 64x24. The triage guess about ConfigureTopLevel was wrong: both platforms convert the size back in the same way.
    • Fix TopLevelReqProc returns early when wmPtr->width >= 0 && wmPtr->height >= 0, as on X11. The TK_GEOMETRYREQ message for embedded windows is still sent first. For non-gridded windows nothing changes, because the size then does not depend on the request. 3466588fff.patch (32 changed lines: 8 tkWinWm.c, 24 test)
    • Repro work-win/p1025/pw.tcl. Before: 518x490 becomes 518x586 after the first sash move, wm geometry stays 64x30. After: 518x490 throughout, wm geometry 64x24, identical to X11 (Xvfb, no WM: 520x522, 64x24).
    • Test panedwindow-31.2 "sash move does not resize a gridded toplevel" (fails before: yes, on Windows; passes after: yes; passes on X11 as before)
    • Test files run panedwindow 413/0, winWm 37/0, wm 302/0, grid 206/0, pack 199/0, place 56/0, geometry 14/0, text 703/0, frame 207/0 with fix; main panedwindow 412/0, the others identical
    • Windows done. Worth checking by hand: interactive resizing of a gridded toplevel, and menubar add/remove on a window with an explicit size (TopLevelReqProc no longer triggers UpdateGeometryInfo there, but menubar changes call it directly).
    • Suggested ticket comment On Windows every geometry request recomputes a gridded toplevel's size from its new requested size, and the panedwindow's request grows after a sash move. X11 ignores geometry requests once the size is set explicitly. A prototype does the same in tkWinWm.c (test panedwindow-31.2), and the toplevel now keeps its size as on X11.
    ✅3464070fff2011MenuNewEntry() return NULL -> no error msgpatchREPRODUCED (9.1 main, 8.6 tip); prototype verified and extended

    Open ticket 3464070fff · 10. Generic Menus · filed 2011-12-22

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-25 check-in 18d8a385e1 on menu-new-entry-error, in main, 9.0: Fix 3464070fff: when a menu entry could not be created, MenuNewEntry() freed it but left it in the entries array, and on Windows no error message was set. The
    • 💬 2026-10-25 serhiy.storchaka: The only case without a message is on Windows, when all menu command IDs are in use. On that path MenuNewEntry() also frees the new entry but leaves it in the menu, so using or destroying the menu later crashes. Fixed on branch menu-new-entry-error 18d8a385e1: the slot is removed again, and the error is "No more menu entries can be allocated." with error code TK MENU SYSTEM_RESOURCES, as for the existing CreateP…
    • 💬 2026-10-30 jan.nijtmans: Fixed on 1512c574, core-9-0-branch and core-8-6-branch (thanks @scaba for doing the last one) Many thanks!
    • 💬 2026-10-30 nemethi: Merged by Jan into main and core-9-0-branch by commits 1512c574 and [7fef988d ]. Backported to core-8-6-branch by commit 6db26499. Closing the ticket.
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip); prototype verified and extended

    • Platform generic + Windows
    • Tested code review. On X11/macOS, TkpMenuNewEntry cannot fail and Tk_InitOptions sets its own message. menu.test passes with the prototype (549/553, 4 skipped).
    • Cause win/tkWinMenu.c:TkpMenuNewEntry returns TCL_ERROR without a message when GetNewID has used up all command IDs (0xF000). In addition, generic/tkMenu.c:MenuNewEntry frees mePtr on failure but leaves it in menuPtr->entries[index] with numEntries already incremented, so the menu holds a dangling entry.
    • Prototype 3464070fff.patch (23 lines: MenuNewEntry removes the slot again on failure; Windows sets "No more menu entries can be allocated." with errorcode TK MENU SYSTEM_RESOURCES, matching the existing CreatePopupMenu error). No test (the failure cannot be triggered portably). Windows repro: win/3464070fff.tcl (creates about 61k entries).
    • Suggested ticket comment The only case without a message is running out of Windows menu command IDs. On that path MenuNewEntry also leaves a freed entry in the menu. A prototype fixes both.
    Windows VM, run 3
    • Platform generic + Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip); prototype verified and extended
    • Tested work-win/3464_run.tcl (the original script misread the non-empty result of $m add on 9.1 as an error). ref91: after 61405 entries the error message is empty, errorcode NONE, and the menu is broken afterwards (entrycget end -label gives "unknown option -label", from the dangling freed entry). 8.6: the same after 61403 entries. Prototype: "No more menu entries can be allocated." / TK MENU SYSTEM_RESOURCES, and the menu stays consistent. Also found: after deleting the last entry, adding one still failed, because GetNewID stopped the search before trying the last allocated ID. Fixed as well, and adding after a delete now works. menu.test and winMenu.test pass (part of the 1096).
    • Cause win/tkWinMenu.c TkpMenuNewEntry sets no message; generic/tkMenu.c MenuNewEntry leaves the freed entry in the array; win/tkWinMenu.c GetNewID has an off-by-one in its wrap-around check
    • Prototype 3464070fff.patch (34 lines; adds the GetNewID fix to the earlier 23-line version). No test (61k entries takes about 5 s, so a test would be possible but slow).
    • Suggested ticket comment Confirmed on Windows: when menu command IDs run out, $m add fails with an empty message and leaves a dangling entry. The patch sets an error message, removes the slot, and also fixes GetNewID never reusing the most recently freed ID.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before: the new test crashes tktest. The failed entry stays in the array after being freed, and the menu is destroyed in cleanup. / after: "No more menu entries can be allocated." with TK MENU SYSTEM_RESOURCES, a consistent menu, and add-after-delete works
    • Test winMenu-36.1 (new; fills the menu command IDs with 100 menus of 1000 entries. The consistency checks run only if the error message is right, because otherwise the menu is corrupt; an earlier version hung. Fails before: yes, crash; passes after: yes)
    • Test file run winMenu.test 148/0, menu.test 574/0, menubut.test 117/0 (pre-existing failures: none)
    • Patch changes added test winMenu-36.1
    ✅3414693fff2011tk_dialog: (mis)placementpatchREPRODUCED (main, 8.6)

    Open ticket 3414693fff · 35. [tk_dialog] · filed 2011-09-27

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-01 check-in e34eec51d5 on tk-dialog-over-parent, in main, 9.0: Fix 3414693fff: tk_dialog is centered over its parent window, like tk_messageBox and other Tk dialogs, not on the screen.
    • 💬 2026-10-01 serhiy.storchaka: All other Tk dialogs (tk_messageBox, tk_chooseColor, tk_chooseDirectory, tk_getOpenFile, the font chooser) are centered over their parent, and tkinter's simpledialog, which reimplements tk_dialog, does the same. Proposed fix in branch [tk-dialog-over-parent]: fvogel's patch, plus test dialog-2.4. The parent is the parent of the dialog window (usually "."). The Motif-style placement can still be obtained: if the pa…
    • 💬 2026-10-14 jan.nijtmans: Fixed in 7607b7988e, core-9-0-branch and core-8-6-branch Many thanks!
    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11 (any platform using the script dialog; tk_dialog is always scripted)
    • Tested main, 8.6.18 on X11 without a WM. Parent at +900+700: the dialog is centered on the screen and not over the parent.
    • Cause library/dialog.tcl calls ::tk::PlaceWindow $w with no anchor, i.e. screen center ("Motif style"). tk_messageBox, tk_chooseColor, tk_chooseDirectory, tkfbox and fontchooser all use PlaceWindow $w widget $parent.
    • Prototype 3414693fff.patch (26 lines, fvogel's one-line change plus a comment update; new test dialog-2.4 fails before the fix and passes after; dialog.test 7/7). PlaceWindow still falls back to screen center when the parent is not mapped.
    • Suggested ticket comment Reproduced. A prototype centers tk_dialog over its parent like tk_messageBox does (fvogel's patch plus test dialog-2.4).
    ✅3313032fff2011ttk::spinbox doesn't expand its height on LinuxpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket 3313032fff · 88. Themed Tk · filed 2011-06-07

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-30 check-in e077521308 on spinbox-fill-height, in main, 9.0: Fix [3313032]: ttk::spinbox does not expand its height. The field of the default TSpinbox layout fills the whole height of the spinbox, as in the classic and vi
    • 🔧 2026-10-01 check-in c548a23a81 on trunk, in main, 9.0, 8.6: Fix for 3313032fff: ttk::spinbox doesn't expand its height on Linux.
    • 🔧 2026-10-01 check-in ece6a4fe34 on core-9-0-branch, in main, 9.0: Fix for 3313032fff: ttk::spinbox doesn't expand its height on Linux.
    • 🔧 2026-10-01 check-in 0335f9259b on core-8-6-branch, in main, 9.0, 8.6: Fix for 3313032fff: ttk::spinbox doesn't expand its height on Linux.
    • 💬 2026-09-30 serhiy.storchaka: The default TSpinbox layout, used by the default, alt, clam and winnative themes, did not stretch Spinbox.field vertically, so it kept its requested height at the top of the spinbox. The winnative theme is used on Windows with the Basic and High Contrast themes; the vista and classic themes have their own layouts in which the field fills the spinbox. Just making the field fill the spinbox is not enough: the arrows…
    • 💬 2026-10-01 nemethi: Serhiy, many thanks for the fix! Merged into main, core-9-0-branch, and core-8-6-branch by commits c548a23a, ece6a4fe, and 0335f925.
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic (every theme that uses the default TSpinbox layout: default, clam, alt on X11; winnative/xpnative on Windows)
    • Tested main, 8.6 tip on Xvfb (no WM), all X11 themes. When packed with -fill both -expand 1 at 100 px height, identify element in the middle of the spinbox gives the background, not textarea, in default/clam/alt. The classic theme is fine because its layout uses FILL_BOTH. ttk::entry and ttk::combobox are fine.
    • Cause generic/ttk/ttkEntry.c SpinboxLayout: Spinbox.field is TTK_PACK_TOP|TTK_FILL_X, so the field keeps its requested height. The classic, vista and Wide.TSpinbox layouts use -sticky nswe.
    • Prototype 3313032fff.patch (1-line layout change + test spinbox-6.1 in tests/ttk/spinbox.test; the test fails before and passes after; the ttk spinbox.test and spinbox.test files pass)
    • Suggested ticket comment The default TSpinbox layout packs Spinbox.field with -sticky ew only, so the field keeps its requested height. Changing it to FILL_BOTH (as the classic and vista layouts already do) fixes it. Prototype with test spinbox-6.1.
    ⬜3307625fff2011scroll region of text limited to 32767 pixelsREPRODUCED (main, 8.6) for embedded windows only. NOT-A-BUG for canvas items.

    Open ticket 3307625fff · 04. Canvas Basics · filed 2011-05-25

    Full verdict

    REPRODUCED (main, 8.6) for embedded windows only. NOT-A-BUG for canvas items.

    • Platform generic (X11 and Windows)
    • Tested main 9.1.1 and 8.6.18 on X11.
      • Canvas text item of 5000 lines: scrolls to canvasy 99519 (main) / 84591 (8.6), well past 32767.
      • Frame of 60000 px height as an embedded window, scrolled to the end: winfo viewable is 0.
    • Cause generic/tkCanvWind.c / Tk_CanvasWindowCoords() hold the position in a short, and X11 ConfigureWindow uses INT16 for it. As serhiy noted, this is a platform limit.
    • Prototype none (platform limitation)
    • Suggested ticket comment Canvas items are not limited: a text item of 100000 px scrolls correctly. Only embedded windows are limited, because window positions are 16-bit on X11 and Windows. Suggest closing as Won't Fix, with the workaround of scrolling the embedded widget itself.
    ✅3301552fff2011ttk::spinbox -font does not resize buttonsREPRODUCED (main, 8.6), cosmetic

    Open ticket 3301552fff · 88. Themed Tk · filed 2011-05-13

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-28 check-in 6ff1ea8099 on vista-spinbox-arrows, in main, 9.0: Fix [3301552]: ttk::spinbox -font does not resize buttons
    • 🔧 2026-09-29 check-in e7c81329e6 on trunk, in main, 9.0, 8.6: Fix for 3301552fff: ttk::spinbox -font does not resize buttons.
    • 🔧 2026-09-29 check-in 8ec01a629c on core-9-0-branch, in main, 9.0: Fix for 3301552fff: ttk::spinbox -font does not resize buttons.
    • 🔧 2026-09-29 check-in 98f2f50ea1 on core-8-6-branch, in main, 9.0, 8.6: Fix for 3301552fff: ttk::spinbox -font does not resize buttons.
    • 💬 2026-09-28 serhiy.storchaka: The wiki page with the example is now [https://wiki.tcl-lang.org/22156]. Still reproducible on Windows 11 with 8.6.18 and 9.1 in the vista theme: the vista layout of TSpinbox packs the arrows to the top and bottom of the field, so with a large font there is a gap between them. The other themes group the arrows in the middle; there the arrows keep their size, but -arrowsize can be used in clam and alt. A simple fix…
    • 💬 2026-09-28 serhiy.storchaka: Proposed fix in branch vista-spinbox-arrows: if the layout places the arrows at the top and at the bottom of the same column, the spinbox extends them to fill the space between them. So in the vista theme each arrow button takes half of the field height, as in the native up-down control. Other themes group the arrows in the middle and are not affected. The layout in the vista_theme_improvements branch also places …
    • 💬 2026-09-29 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits e7c81329, 8ec01a62, and 98f2f50e. Closing the ticket.
    Full verdict

    REPRODUCED (main, 8.6), cosmetic

    • Platform generic (X11 confirmed; the original report was about Vista)
    • Tested main, 8.6 tip, all X11 themes, font size 30. Widget height is 62–64 px, but the up and down arrows are about 10 px each, centered (for example up=21–30 and down=31–40). The combobox arrow fills the height.
    • Cause generic/ttk/ttkEntry.c SpinboxLayout (and classic/aqua layouts): the "null" group holding the arrows is packed right without fill, and the arrows keep their natural -arrowsize height. The pack-based layout engine can't split the height into two equal halves (TTK_EXPAND takes the whole cavity). Vista solves this with the HALF_HEIGHT element flag in win/ttkWinVistaTheme.c.
    • Prototype none. The fix needs either a new element or a layout-engine feature (split the parcel evenly), or making the arrow elements size from the font. That is design work, not a small patch. The Wide.TSpinbox layout in 9.1 is a workaround.
    • Suggested ticket comment Still true on X11 with 9.1: the spin arrows keep their -arrowsize and don't follow the font/widget height. Fixing it needs an element that takes half the field height (as vista's HALF_HEIGHT does), not only a layout tweak.
    ⬜3152603fff2011Cascade entry stops working if reconfigured with same -menupatchREPRODUCED (main 9.1b1, 9.0.5, 8.6.18, 8.5.19, 8.4.20)

    Open ticket 3152603fff · 12. Unix Menus · filed 2011-01-06

    Full verdict

    REPRODUCED (main 9.1b1, 9.0.5, 8.6.18, 8.5.19, 8.4.20)

    • Platform generic (tested on X11)
    • Tested script from the ticket (checks [.#menu entrycget end -menu] before and after) on all listed versions, no WM needed
    • Cause generic/tkMenu.c:ConfigureMenuCloneEntries. When the cascade name does not change (cascadeEntryChanged == false), the same objv is still passed to ConfigureMenuEntry for every clone. So -menu .menu.sub overwrites the clone's -menu .#menu.#menu#sub with the original, non-clone menu.
    • Prototype 3152603fff.patch (+47 lines: 32 in tkMenu.c, 15 in the test). For each clone of a cascade entry it saves the clone's -menu when the cascade did not change, and restores it after ConfigureMenuEntry. New test menu-12.5. menu.test passes and valgrind is clean.
    • Suggested ticket comment Still present in 8.6 and 9.x. ConfigureMenuCloneEntries passes the unchanged -menu value to the clones too, so they end up pointing to the original submenu instead of its clone. Prototype fix with a test (menu-12.5) is ready.
    Prototype verified 2026-10-24, Linux
    • Status FIXED-ON-MAIN (5f1de6fb06, 2026-10-24, "Fix for 1723059fff: entryconfigure destroys submenus", from branch menu-clone-cascade-destroy ef4c358706. Also on core-9-0-branch a3f03a589c and core-8-6-branch edff8f5386.)
    • Applies to main no (the tkMenu.c hunk was rejected: ConfigureMenuCloneEntries now drops -menu from the clone objv when the cascade does not change)
    • Repro the patch's test menu-12.5 already passes on unpatched main
    • Test menu-12.5 (passes on main without the C change). The ticket can be closed. menu-12.5 could optionally be added as an extra regression test, but main's menu-45.1 covers the same code path.
    • Test file run n/a
    • Patch changes none (the patch is obsolete)
    ⬜3141377fff2010[winfo screenwidth/height] fail to track resolution changesREPRODUCED on main and 8.6 by code inspection: main still only reads the Xlib Screen struct (no RandR code in tkmain/unix). The fix is on branch …

    Open ticket 3141377fff · 65. Generic Window Operations · filed 2010-12-21

    Full verdict

    REPRODUCED on main and 8.6 by code inspection: main still only reads the Xlib Screen struct (no RandR code in tkmain/unix). The fix is on branch x11-randr-screen-size (ae3b099be2, 2026-09-18), not merged.

    • Platform X11
    • Tested could not change the resolution at run time. Xvfb has a single fixed RandR mode and refuses xrandr --fb/--newmode, and Xephyr/Xorg-dummy would need another display. I merged ae3b099be2 into a main checkout:
      • It merges cleanly.
      • configure finds XRRUpdateConfiguration.
      • It builds, and winfo.test passes (74/69/5 skipped/0 failed).
    • Cause generic/tkCmds.c uses WidthOfScreen()/HeightOfScreen(), which Xlib only updates after XRRUpdateConfiguration().
    • Prototype none (exists on branch x11-randr-screen-size).
    • Suggested ticket comment Branch x11-randr-screen-size (ae3b099be2) still merges cleanly into main and passes the winfo tests. It needs a review/merge decision (new optional libXrandr dependency).
    ⬜3119824fff2010Enter doesn't work in text input boxpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4). bind Text <KP_Enter> is still # nothing

    Open ticket 3119824fff · 18. [text] · filed 2010-11-26

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5, 8.4). bind Text <KP_Enter> is still # nothing

    • Platform X11 (on Windows, keypad Enter arrives as Extended-Return, so it already works there)
    • Tested bind Text <KP_Enter> on main/9.0/8.6/8.5/8.4.20. With the patch, event generate .t <KeyPress> -keysym KP_Enter inserts a newline on the wt build. Without it the # nothing binding fires instead
    • Cause library/text.tcl explicitly binds <KP_Enter> to # nothing (so that the \r character is not inserted), and no binding makes it act like Return. Jan marked this as a dup of b3a1b9852d (the same problem for all widgets) and proposes a TIP to map KP_Enter to Extended-Return on X11.
    • Prototype 3119824fff.patch (58-line diff; the code change is 2 lines: bind Text <KP_Enter> [bind Text <Return>] replaces the # nothing binding). New test text-40.1. text.test passes. This fixes only Text; the generic fix proposed in b3a1b9852d would make it unnecessary.
    • Suggested ticket comment Still true: the Text class binds KP_Enter to nothing on X11. A one-line fix for Text is to bind it like Return. The generic fix (KP_Enter treated as Extended-Return) is tracked in b3a1b9852d.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main ported (the test hunk was rejected because main has a new text-39.1 at the end of text.test; I added text-40.1 after it, since 40-45 are free. library/text.tcl applied clean.)
    • Repro before KP_Enter inserts nothing / after it inserts "\n"
    • Test text-40.1 (existing in patch, moved; fails before: yes; passes after: yes)
    • Test file run text.test 704/0 (pre-existing failures: none; main 703/0)
    • Patch changes moved the test after text-39.1
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 3119824fff: FAILS ON WINDOWS: text-40.1 errors before and after with no keycode for keysym "KP_Enter" (event generate -keysym KP_Enter); the repro also gives no newline. Windows has no KP_Enter keysym until 787402ffff (phase 1, bind-38.1) is applied. The patch is unchanged; commit it after 787402ffff, or give text-40.1 a constraint.
    ✅3009885fff2010tk_messageBox can return empty string on WindowspatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 3009885fff · 38. [tk_messageBox] · filed 2010-06-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-02 check-in b015fbc9e3 on win-transient-keep-wrapper, in main, 9.0: Fix 3009885fff: tk_messageBox can return empty string on Windows.
    • 🔧 2026-10-14 check-in 3b64795c37 on trunk, in main, 9.0: Fix 3009885fff: tk_messageBox can return empty string on Windows
    • 🔧 2026-10-14 check-in 3a4343ddd5 on core-9-0-branch, in main, 9.0, 8.6: Fix 3009885fff: tk_messageBox can return empty string on Windows
    • 🔧 2026-10-14 check-in 802fe5c0c3 on core-8-6-branch, in main, 9.0, 8.6: Fix 3009885fff: tk_messageBox can return empty string on Windows
    • 💬 2026-10-02 serhiy.storchaka: Still present in 8.6 and 9.1 on Windows 11. When the master is deiconified, the wrapper window of the transient is recreated, and this destroys the message box owned by it. Proposed fix in branch win-transient-keep-wrapper: recreate the wrapper only if it is not yet owned by the master's wrapper. New test winWm-9.4.
    • 💬 2026-10-14 jan.nijtmans: Fixed in 3b64795c, core-9-0-branch and core-8-6-branch. Many thanks!
    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested X11 main + 8.6 under icewm: the Tk messagebox stays open after its transient parent's master is iconified and deiconified, and it returns "ok".
    • Cause Suspected (not verified): win/tkWinDialog.c:MessageBoxObjCmd passes the parent wrapper HWND as the MessageBoxW owner. When the master is iconified, the transient's wrapper is hidden or re-created (the "updated the wrapper of the parent" comment after MessageBoxW hints at this). Destroying the owner closes the owned message box, so winCode is 0, which maps to "".
    • Prototype none (needs Windows to confirm cause); repro win/3009885fff.tcl (iconifies/deiconifies automatically, writes 3009885fff.result)
    • Suggested ticket comment Not reproducible on X11. Needs re-checking on Windows 10/11 with current Tk. A possible cause is the owner HWND of the native MessageBox being destroyed or re-created when the transient's master is iconified.
    Windows VM, run 4
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested 3009885_auto.tcl (automatic). The message box, with parent a transient of ., returns {} right after wm deiconify . (3528 ms), with no click. wm frame .t changes (0xd50102 → 0xaf1b32), so the transient's wrapper was recreated.
    • Cause win/tkWinWm.c WmWaitVisibilityOrMapProc calls UpdateWrapper(transient) on every VisibilityNotify of the container. That destroys the transient's wrapper HWND, and with it the MessageBox it owns, so MessageBoxW returns 0 and the result is "".
    • Prototype 3009885fff.patch (25 lines): on VisibilityNotify, recreate the wrapper only if it does not exist yet or is not already owned by the container's wrapper (GetWindow(GW_OWNER) against GetAncestor(container, GA_ROOT)). With the prototype the message box survives iconify/deiconify and is still visible afterwards. wm.test, winWm.test, winDialog.test, grab.test, focus.test: 483 tests, 0 failed. No automated test (it needs a native message box).
    • Suggested ticket comment Confirmed on Windows 11 with 8.6 and 9.1. Deiconifying the master recreates the transient's wrapper window (WmWaitVisibilityOrMapProc → UpdateWrapper), which destroys the message box it owns. Fix: only recreate the wrapper when its owner is wrong.
    ✅2995339fff2010ttk::sizegrip strange behaviourpatchREPRODUCED (main, 8.6, 8.4 core cause) with metacity. Not reproduced with no WM, fvwm, icewm, openbox, wmaker, xfwm4, mutter.

    Open ticket 2995339fff · 88. Themed Tk · filed 2010-05-01

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-30 check-in 48524db843 on wm-geometry-size-only, in main, 9.0: Fix [2995339]: setting only the size with wm geometry moves the window under some window managers. ParseGeometry no longer re-sends the position recorded from t
    • 🔧 2026-10-13 check-in a0496a5855 on trunk, in main, 9.0: Fix for 2995339fff: ttk::sizegrip strange behaviour.
    • 🔧 2026-10-13 check-in 6873ef21c0 on core-9-0-branch, in main, 9.0, 8.6: Fix for 2995339fff: ttk::sizegrip strange behaviour.
    • 🔧 2026-10-13 check-in 2067f3d7f4 on core-8-6-branch, in main, 9.0, 8.6: Fix for 2995339fff: ttk::sizegrip strange behaviour.
    • 💬 2026-09-30 serhiy.storchaka: Reproduced on X11 with metacity, mutter/gnome-shell and KWin, even without the sizegrip: wm geometry . 400x300 moves the window. It moves by (-22,-14) under metacity, (-14,-12) under gnome-shell and 42 pixels down under KWin (Xwayland). It does not move without a window manager, or under fvwm, openbox, xfwm4, icewm or compiz 0.9. Windows is not affected. The cause is in Tk: a size-only geometry still sends back th…
    • 💬 2026-10-13 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits a0496a58, 6873ef21, and 2067f3d7.
    Full verdict

    REPRODUCED (main, 8.6, 8.4 core cause) with metacity. Not reproduced with no WM, fvwm, icewm, openbox, wmaker, xfwm4, mutter.

    • Platform X11
    • Tested main and 8.6 × {no WM, fvwm, icewm, openbox, wmaker, metacity, xfwm4, mutter}, windows with and without an explicit +x+y position. With metacity and a WM-placed window, the first sizegrip drag moves the window by (-22,-14). A plain wm geometry . 400x300 does the same (8.4 too).
    • Cause Two parts:

    1. unix/tkUnixWm.c:ParseGeometry always sets WM_MOVE_PENDING, even for a size-only "WxH". UpdateGeometryInfo then sends XMoveResizeWindow with the frame position recorded from the WM. Metacity (invisible borders) interprets this differently, so the window moves.

    2. library/ttk/sizegrip.tcl Drag re-applies +x+y from wm geometry, which has the same effect.

    • Prototype 2995339fff.patch (13 changed lines): WM_MOVE_PENDING only when a position was given or the geometry is right/bottom-relative (unixWm-48.13 still passes); sizegrip sets only WxH. Verified under metacity (no move), fvwm and openbox. wm.test and unixWm.test give the same results before and after (only the no-WM baseline failures). No new test: it needs a reparenting WM with frame extents.
    • Suggested ticket comment Reproduced on X11 with metacity. wm geometry . WxH (and the sizegrip, which re-sets +x+y) re-sends the frame position, and the WM moves the window. Fix: only move when a position is given, and make the sizegrip set only the size.
    ⬜2980137fff2010support return optionsREPRODUCED (main, 8.6). RFE-like.

    Open ticket 2980137fff · 58. [send] · filed 2010-03-31

    Full verdict

    REPRODUCED (main, 8.6). RFE-like.

    • Platform generic (X11 and Windows send)
    • Tested X11 main + 8.6, two wish processes. send app {return -code error -errorcode X -foo bar msg} gives code 1, the message, -errorcode and -errorinfo. The custom option -foo is lost, and -errorline/-errorstack are the local ones.
    • Cause The unix/tkUnixSend.c protocol only has -c/-r/-i/-e reply fields (errorInfo/errorCode). The local-interp path in Tk_SendObjCmd and win/tkWinSend.c also copy only errorInfo/errorCode.
    • Prototype none (protocol extension across unix/win/macOS plus a design choice on which options to transport). Possible fix: add a "-o <dict>" reply field carrying Tcl_GetReturnOptions minus -code/-level, and apply it with Tcl_SetReturnOptions. Old peers ignore unknown options.
    • Suggested ticket comment Still true: only -errorinfo and -errorcode go through [send]. Could be done with an extra "-o" reply field (ignored by old peers) carrying the return options dict.
    ⬜2896605fff2009Change to grab on windows breaks application behaviourREPRODUCED (case 1, 9.1 main); case 2 NOT-REPRODUCED

    Open ticket 2896605fff · 56. [grab] · filed 2009-11-12

    Full verdict

    REPRODUCED (case 1, 9.1 main); case 2 NOT-REPRODUCED

    • Platform Windows
    • Tested the script runs on X11 (8.6: both cases pass). Windows not tested here.
    • Cause win/tkWinWm.c grab handling from [1847002] (raises or reactivates the grab window on activation). Current Windows no longer raises a window when it is activated.
    • Prototype none. Windows check: win/2896605fff.tcl. Case 1: busy grab on a frame of ., then wm deiconify of a withdrawn toplevel; it must end up above .. Case 2: raise . while a dialog holds the grab. Both checked with wm stackorder; prints PASS/FAIL and a VERDICT line.
    • Suggested ticket comment The script in the ticket no longer shows the problem on Windows 11. Please run the attached check on other Windows versions; if it passes there too, close (together with [3414089]).
    Windows VM, run 6
    • Platform Windows
    • Verdict REPRODUCED (case 1, 9.1 main); case 2 NOT-REPRODUCED
    • Tested Case 1: while ._busy holds a grab, wm deiconify of a withdrawn toplevel leaves it below . (stackorder .app1 .). Case 2: raise . above a dialog holding the grab is refused as expected (stackorder .d .).
    • Prototype none
    • Suggested ticket comment The script in the ticket no longer misbehaves, but a deiconified toplevel still stays below the main window while another window holds a grab (Windows 11, 9.1).
    🔧2865578fff2009Changin menu in fullscreen mode messes with geometrypatchREPRODUCED (main, 8.6, 8.5)

    Open ticket 2865578fff · 12. Unix Menus · filed 2009-09-24

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-11-05 check-in bb1cb9dec3 on x11-menubar-height, not merged: Fix 2865578fff: X11: the toplevel shrank when its menubar was set again while it was going fullscreen or zoomed. The size requested for the toplevel was compu
    • 🔧 2026-11-05 check-in bd8005bcdb on x11-menubar-height, not merged: Fix 2865578fff: X11: the toplevel shrank when its menubar was set again while it was going fullscreen or zoomed. The size requested for the toplevel was compu
    • 💬 2026-11-05 serhiy.storchaka: Still reproducible on X11 under Openbox: after the four commands of sample.tk the toplevel shrinks to 1 pixel high. The same happens with wm attributes . -zoomed instead of -fullscreen. When the menubar is set again, Tk requested the size of the toplevel before the height of the new menubar was computed. Openbox applied that request after restoring the window from fullscreen, and Tk took it for a resize by the use…
    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform X11
    • Tested script "menu + label; fullscreen 1; -menu {}; -menu .menu; fullscreen 0" on Xvfb with no WM, openbox, fvwm, icewm. No WM: OK. With update between steps: icewm leaves the window at 1280x955 (main)/1280x961 (8.6) instead of 40x24; openbox/fvwm OK. Without update (the ticket's sample): openbox gives 40x1 or 40x52, fvwm 40x-10 (negative height from wm geometry), icewm 1280x1023 — timing dependent, all versions. Fullscreen-only or menu-only toggles are fine.
    • Cause unix/tkUnixWm.c. While fullscreen, ConfigureEvent treats the WM's fullscreen resize as a user resize and stores it in wmPtr->width/height; a menubar change then makes UpdateGeometryInfo XResizeWindow the wrapper to that size (minus/plus the menu height), and some WMs (icewm) record this as the size to restore. In the no-update race, a resize issued with the not-yet-sized new menubar clone (menuHeight 1) is applied by the WM after unfullscreen and taken as user size (height = 25 - 35 = -10).
    • Prototype 2865578fff.patch (104 lines diff, ~45 changed: factor the toplevel/menubar layout tail of ConfigureEvent into LayoutWrapper(); UpdateGeometryInfo only calls LayoutWrapper while attributes.fullscreen; ConfigureEvent does not record user size while reqState/attributes.fullscreen). Fixes the realistic case (steps separated by event processing) on icewm/fvwm/openbox and keeps natural size tracking after a fullscreen round trip; does NOT fix the all-in-one-idle race of sample.tk on fvwm/icewm. No new test (WM-dependent); wm.test, unixWm.test, menu.test, unixMenu.test pass under fvwm.
    • Suggested ticket comment Still reproducible on X11 (icewm, fvwm, openbox) with 8.6 and 9.1. While fullscreen, Tk records the fullscreen size as a user-requested size and resizes the wrapper on menubar changes; icewm then restores to that size. A prototype that leaves the wrapper size alone while fullscreen fixes the case with event processing between steps.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main clean (offsets 8-11 lines)
    • Repro work-a2/2865578.tcl (menu+label; fullscreen 1; -menu {}; -menu .menu; fullscreen 0, with update/pause between the steps). icewm before: 37x24 → 1280x954 (BAD, 2 of 2 runs). After: 37x24 → 37x24 (OK, 2 of 2 runs, also after a clean rebuild). fvwm and openbox: OK both before and after (no regression).
    • Test none (needs a real WM: without a WM main is already OK)
    • Test file run wm.test 302/0, unixWm.test 284/0, menu.test 574/0, unixMenu.test 122/0 (pre-existing failures: none)
    • Patch changes none. Review: while fullscreen, the early return in UpdateGeometryInfo leaves WM_MOVE_PENDING set, so a pending move is applied after leaving fullscreen. That is acceptable. As noted at triage, the all-in-one-idle race in the ticket's sample is not fixed.
    ⬜2847319fff2009Undocumented -help (or -h) option causes init failureREPRODUCED (main, 8.6, 8.5)

    Open ticket 2847319fff · 75. wish · filed 2009-08-30

    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform generic
    • Tested wish script.tcl -h and wish -help on main, 8.6, 8.5 (Xvfb): prints "application-specific initialization failed: Command-specific options: ..." and then continues: the script runs without Tk ("package Tk is not present") or the interactive shell starts; exit status 0 for wish -help </dev/null.
    • Cause generic/tkWindow.c Initialize() uses Tcl_ParseArgsObjv with TCL_ARGV_AUTO_HELP (a62e3e93d0, 2016), which returns TCL_ERROR with the usage text; Tk_MainEx reports the failed app init and continues. -help is intentional and covered by main-3.1..3.3 (they expect the error). Note: -h after the script name also matches because wish parses all of argv unless -- is used. Separately, Tk_ParseArgv still ignores TK_ARGV_NO_DEFAULTS when matching (nijtmans' 2009 patch never applied), though Tk no longer uses it for init.
    • Prototype none (behaviour decision: exit after printing help in Tk_MainEx, or drop TCL_ARGV_AUTO_HELP; either changes main-3.x)
    • Suggested ticket comment Still present in 8.6 and 9.1: -help makes Tk_Init fail with the usage text and wish continues without Tk. Proposal: have wish exit after printing the -help usage (Tk_Init keeps returning the usage as an error), and document -help in wish.1.
    ⬜2825888fff2009Off-by-one errors in overlap computation of text itemspatchREPRODUCED (main, 8.6); not in 8.5/8.4. Introduced by 9dc3d444b7 (2008-11-22, patthoyts: replaced C99 round() with (int)(x + 0.5) in the angled-t…

    Open ticket 2825888fff · 44. Generic Fonts · filed 2009-07-23

    Full verdict

    REPRODUCED (main, 8.6); not in 8.5/8.4. Introduced by 9dc3d444b7 (2008-11-22, patthoyts: replaced C99 round() with (int)(x + 0.5) in the angled-text TextToArea).

    • Platform generic
    • Tested ticket script on main, 8.6, 8.5, 8.4 (Xvfb). .c find overlapping 0 0 99 300 returns the text item at x=100 on main/8.6, not on 8.5/8.4; top edge not affected in the ticket example but same code.
    • Cause generic/tkCanvText.c TextToArea: (int) ((rectPtr[0] + 0.5) - textPtr->drawOrigin[0]) truncates toward zero for negative values (-99.5 -> -99 instead of -100), shifting the area one pixel right/down when it starts left/above the text origin. Tk_IntersectTextLayout's comparisons are fine (8.5 semantics restored without touching them).
    • Prototype 2825888fff.patch (11+/2-, use the existing ROUND() macro (floor-based); test canvText-11.3, fails before, passes after; canv*.test all pass)
    • Suggested ticket comment Regression from 9dc3d444b7 (round() replaced by a truncating (int)(x+0.5) in TextToArea, wrong for negative values). Using floor-based rounding restores the 8.5 behaviour; fix and test canvText-11.3 prepared.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST
    • Applies to main clean (test hunk offset -2)
    • Repro before find overlapping 0 0 99 300 finds the text at x=100 / after it does not
    • Test canvText-11.3 (existing in patch; fails before: yes; passes after: yes)
    • Test file run canvText.test 101/0 (pre-existing failures: none)
    • Patch changes none
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 2825888fff: OK. canvText-11.3 fails before, passes after.
    ⬜2822368fff2009Can't generate binary selection formatsREPRODUCED (main, 8.6) (by code and a partial run). Really an RFE.

    Open ticket 2822368fff · 53. [selection] · filed 2009-07-16

    Full verdict

    REPRODUCED (main, 8.6) (by code and a partial run). Really an RFE.

    • Platform X11
    • Tested main, 8.6 (selection handle -type image/png -format image/png with a handler returning "0x89 0x50 ..."; another Tk client reads it back as 32-bit values. xclip/xwd are not installed, so the property format was checked in the code)
    • Cause unix/tkUnixSelect.c ConvertSelection/TkSelPropProc: every type other than STRING/UTF8_STRING/COMPOUND_TEXT goes through SelCvtToX and is always written with format 32. Handlers pass NUL-terminated char * chunks (INCR code uses strlen), so raw bytes cannot be carried either. Retrieval already handles format 8 (SelCvtFromX8).
    • Prototype none (needs API design, e.g. a way to say "8-bit" for -format, or a bytearray handler result. Also needs NUL-safe chunking in the INCR path)
    • Suggested ticket comment Still true on main and 8.6: non-text selection targets are always sent as 32-bit items (SelCvtToX), and handler chunks are NUL-terminated strings. Fixing this needs a new API (8-bit/binary -format), so this is effectively an RFE.
    🟢2817531fff2009"wm manage" is deficient and flickeringpatchREPRODUCED (main, 8.6, 8.5)

    Open ticket 2817531fff · 08. [(label)frame] and [toplevel] · filed 2009-07-06

    🟢 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-02 check-in 9f64c9486a on x11-wm-manage-no-flicker, in main, 9.0: Fix 2817531fff: "wm manage" is deficient and flickering.
    • 🔧 2026-10-27 check-in 5e0e41adca on trunk, in main, 9.0: Fix for 2817531fff: "wm manage" is deficient and flickering
    • 🔧 2026-10-27 check-in e03850a345 on core-9-0-branch, in main, 9.0, 8.6: Fix for 2817531fff: "wm manage" is deficient and flickering
    • 🔧 2026-10-27 check-in 0f26cf41e1 on core-8-6-branch, in 8.6: Fix for 2817531fff: "wm manage" is deficient and flickering
    • 💬 2026-10-02 serhiy.storchaka: Still present. On X11, WmManageCmd created the wrapper by mapping it (TkWmMapWindow) and unmapping it at once, so the window flashed at +0+0 even if wm withdraw followed. These calls are needed to create the wrapper, which is why removing them did not work. Proposed fix in branch x11-wm-manage-no-flicker: only create the wrapper and leave mapping to the idle handler. Then wm manage, wm withdraw, wm overrideredirec…
    • 💬 2026-10-27 nemethi: Serhiy, many thanks! Merged into main, core-9-0-branch, and core-8-6-branch by commits 5e0e41ad, e03850a3, and 0f26cf41. Note that the macOS case is still open. Simply removing the TkWmMapWindow invocation causes troubles (tested on macOS). For this reason I added the notAqua constraint to the new wm-manage-1.9 test, and left the ticket open.
    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform X11
    • Tested main, 9.1b1, 8.6.18, 8.5.19, no WM (xev -root substructure) and openbox. wm manage .f; wm withdraw .f still maps the new wrapper at +0+0 and unmaps it again at once (flicker). The frame gets Map/Unmap events.
    • Cause unix/tkUnixWm.c WmManageCmd calls TkWmMapWindow() + Tk_UnmapWindow() only to create the wrapper, which maps it for real. TkMapTopFrame already maps at idle time. Note: in the reporter's script wm overrideredirect $w has no value, so it is only a query. With the fix, wm manage; wm withdraw; wm overrideredirect 1; wm geometry; wm deiconify maps once, override-redirect, at the right place.
    • Prototype 2817531fff.patch (30 lines, test wm-manage-1.9). Calls CreateWrapper() in place of map+unmap. wm.test (openbox and no WM), unixWm.test and frame.test show no new failures. Repeated manage/forget with -menu works.
    • Suggested ticket comment Still present. WmManageCmd maps the wrapper right away (TkWmMapWindow + Tk_UnmapWindow), so the window flashes at +0+0 even if <tt>wm withdraw</tt> follows. Creating only the wrapper and leaving the mapping to the idle handler fixes it, so manage/withdraw/configure/deiconify works without flicker. Note that <tt>wm overrideredirect $w</tt> without a value is only a query.
    ⬜2816150fff2009Rotated text displays incorrectly on WinXPpatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 2816150fff · 05. Canvas Items · filed 2009-07-03

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested main, 8.6 on X11 (Courier, Arial, Helvetica). After each font shrink step, the screen grab equals the grab after a forced full redraw, so no traces are left.
    • Cause per dkf, the bbox of rotated text (ComputeTextBbox/TkIntersectAngledTextLayout) disagrees with GDI glyph placement, which rounds cumulatively. Windows-specific.
    • Prototype none. Repro: win/2816150fff.tcl (compares PowerShell screen grabs before and after a forced full redraw, prints BUG PRESENT/NOT PRESENT)
    • Suggested ticket comment Not reproducible on X11 (main, 8.6). Needs a check on current Windows with win/2816150fff.tcl-style repro (Arial is the font reported to show it).
    Windows VM, run 5
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested win/2816150fff.tcl, capturing with act.exe (deterministic; a repeated capture shows no noise). 8.6: Arial 28→22 leaves a few pixels of the previous size's last "t" beyond the new bbox (3–16 px per step). 9.1: Arial is clean in this run, but "Courier" (a raster font that GDI cannot rotate, so it is drawn unrotated) leaves traces at 6 sizes, because the bbox is computed for rotated text.
    • Cause ComputeTextBbox (generic/tkCanvText.c) computes the bbox from Tk's unrotated metrics. GDI places the rotated glyphs with its own cumulative rounding, and does not rotate raster fonts at all, so the drawn ink can fall outside the bbox.
    • Prototype none. Padding the bbox of angled text by a few pixels would hide the Arial case. The raster-font case needs a check in win/tkWinFont.c that the angled font really rotates.
    • Suggested ticket comment Still present on Windows 11 (8.6.18, 9.1): shrinking the font of angled text leaves pixels of the previous size outside the new bbox. Raster fonts such as "Courier" are drawn unrotated, which leaves larger traces.
    2026-10-25 prototype, Windows VM
    • Status PROTOTYPED
    • Cause win/tkWinFont.c:MultiFontTextOut draws angled text with GetScreenFont(angle). GDI ignores lfEscapement for raster fonts ("Courier" = coure.fon on 9.x, Fixedsys, Terminal), so their text is drawn horizontally while the item's bbox (generic/tkCanvText.c:ComputeTextBbox) is rotated. The ink falls outside the bbox, so when the font shrinks or the item changes it is not erased. For TrueType fonts the problem is now small: a probe of Arial, Courier New and Times New Roman at sizes 9–31 and angles 15/30/90/200 found ink outside the bbox only once (2 px, Times New Roman 9 at 90°). Arial is clean in 9.1.
    • Fix a new helper DrawUnrotatableRun() is called for angled runs. If the selected angled font is not a vector font (!(tmPitchAndFamily & TMPF_VECTOR)), it creates the same LOGFONT with OUT_TT_ONLY_PRECIS and the raster font's pitch and family (Courier gives Courier New), and draws the run's characters with ExtTextOutW. The lpDx widths are summed from the shaped glyph advances through logClust, so the text length matches Tk's measurement. RTL runs and TrueType fonts take the old path unchanged. Bbox padding is not included: 9fd94c63b0's redraw margin already covers the remaining 1–2 px TrueType rounding and the slightly wider substitute glyphs. The patch is independent of 3249f7d4c6.patch: both apply on top of each other in either order, since 3249 only changes FamilyExists/WinFontExistProc and winfont-2.8/13.1. 2816150fff.patch (101 changed lines: 69 in tkWinFont.c, of which about 25 are the function header and declaration, plus 32 test lines. That is over the 60-line guideline.)
    • Repro ink pixels outside the bbox (work-win/p1025/rot.tcl, "Angled text", sizes 9/14/22/31). Before: Courier 90° 193/189/171/153, Courier 200° 103/100/31, Fixedsys 90° 275/275/242/208, Fixedsys 15°/30°/200° 5–145. After: everything 0, except Courier 90° at sizes 9 and 14 (10 and 4 px), where the raster size snapping makes Tk's metrics differ from Courier New's. A screenshot shows correctly rotated Courier, Fixedsys and Terminal text inside its bbox, including at 200°. testwingdiobjects shows no GDI leak after 600 redraws.
    • Test winfont-6.2 "Tk_DrawAngledChars procedure: raster font is rotated" (Fixedsys -angle 90, full redraw, counts non-white pixels outside the bbox with testpixel) (fails before: yes, 51; passes after: yes)
    • Test files run winFont 57/0, canvText 100/0, font 304/0 with fix; main winFont 56/0, canvText 100/0, font 304/0
    • Windows done on Windows 11. Not checked: vertical (@-prefixed) fonts, and a raster subfont used as a fallback for missing characters.
    • Suggested ticket comment On Windows 11 the traces now come mainly from raster fonts (e.g. "Courier", Fixedsys): GDI cannot rotate them, so the text is drawn horizontally outside the rotated bbox. The prototype draws such runs with a rotated TrueType substitute, spaced like the original glyphs (test winfont-6.2). The 1–2 px TrueType rounding is covered by the redraw margin of 9fd94c63b0.
    ⬜2805928fff2009wm iconphoto -default in slave interp affects other interpspatchREPRODUCED (main, 8.6, 8.5)

    Open ticket 2805928fff · 65. Generic Window Operations · filed 2009-06-13

    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform X11 (Windows has a similar effect by design: class icon)
    • Tested main, 8.6.18, 8.5.19: a child interp with Tk loaded (same library) runs wm iconphoto . -default p, and a new toplevel in the parent interp then gets _NET_WM_ICON (checked with xprop).
    • Cause unix/tkUnixWm.c stores the default icon in TkDisplay (dispPtr->iconDataPtr), which all interps using that display share. On Windows -default sets the icon of the shared window class (SetClassLongPtrW), which is process-wide.
    • Prototype 2805928fff.patch (28 lines, no test: Tk tests cannot read _NET_WM_ICON). Moves iconDataPtr/iconDataSize from TkDisplay to TkMainInfo (per application). It is freed with the main info. Adds a doc tweak in wm.n. Verified with xprop: the child's default no longer leaks and the parent's own default still works. Interp delete does not crash. wm.test/unixWm.test show no new failures.
    • Suggested ticket comment Still present: on X11 the default icon is stored per display, so it is shared by all interps. Moving it to the per-application TkMainInfo fixes this. On Windows <tt>-default</tt> sets the icon of the shared window class, so there it stays process-wide unless reworked.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED
    • Applies to main ported (the tkInt.h TkDisplay hunk was rejected because main added fields after ximGeneration; I removed the same two lines by hand. Only the context changed: the +/- lines are identical to the original.)
    • Repro work-a2/2805928.tcl (reads _NET_WM_ICON on the wrapper with xwininfo/xprop). Before: a new parent toplevel gets the child interp's default icon ("set", expected "none"). After: "none". The parent's own default still works after destroying a toplevel and after interp delete.
    • Test none (Tk tests cannot read the _NET_WM_ICON property; it needs xprop)
    • Test file run wm.test 302/0, unixWm.test 284/0 (pre-existing failures: none)
    • Patch changes rebased tkInt.h context only. Note: the patch changes the TkDisplay and TkMainInfo layout, so after applying it run make clean first (Tk's Makefile does not track tkInt.h, and a partial rebuild segfaults in the tests).
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 2805928fff: no new test (unix/tkUnixWm.c + generic); builds, wm.test 302/0, window.test 19/0.
    ✅2790175fff2009Wrong directory name is returned by tk_chooseDirectorypatchREPRODUCED (main, 9.0, 8.6). Regression since 8.5 (b12f6d6484, 2007-11-02, "file chooser now uses Ttk widgets")

    Open ticket 2790175fff · 36. [tk_chooseDirectory] · filed 2009-05-11

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Duplicate)

    • 💬 2026-10-23 serhiy.storchaka: Duplicate of [600313]: fixed by 1994866fa3.
    Full verdict

    REPRODUCED (main, 9.0, 8.6). Regression since 8.5 (b12f6d6484, 2007-11-02, "file chooser now uses Ttk widgets")

    • Platform X11 (generic Tk dialog, unix)
    • Tested main, 9.0.5, 8.6.18 (event generate: click on "child" in the icon list, then press/release OK) → returns the parent dir. 8.5/8.4 not run (different IconList API). 8.4 used a classic button, which does not take focus.
    • Cause clicking the ttk::button OK takes focus. IconList FocusOut (library/iconlist.tcl) clears the selection, so chooseDir::OkCmd step 1 finds no selection and returns the entry (the current dir). chooseDir::ListBrowse, which should copy the browsed dir into the entry, is never called: tkfbox ListBrowse ignores directories.
    • Prototype 2790175fff.patch (37 lines, test choosedir-7.1). tkfbox ListBrowse calls chooseDir::ListBrowse for TkChooseDir. That puts the path in the entry and remembers it (data(browsedPath)). OkCmd ends the dialog when the entry equals that path. choosedir.test and filebox.test pass. The new test fails without the fix.
    • Suggested ticket comment Still present in 8.6/9.x (regression from switching the dialog to ttk::button in 8.5: pressing OK moves the focus, and the icon list clears its selection on FocusOut). The fix makes a single click on a directory put it into the entry and lets OK return it.
    ⬜2473895fff2008TIP141 not yet implementedpatchREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 2473895fff · 37. [tk_get*File] · filed 2008-12-29

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows (Unix works; macOS unknown)
    • Tested main and 8.6 tip on X11 (tkfbox). With -multiple 1 -initialfile {a.txt {b c.txt}} the entry shows the list, and pressing OK returns both files, so the list semantics work. Repro script: win/2473895fff.tcl (the user just presses Open).
    • Cause tkfbox.tcl puts -initialfile into the entry, and in multiple mode the entry is parsed as a list, so Unix complies. win/tkWinDialog.c (FILE_INITFILE) passes the whole string verbatim to IFileDialog::SetFileName. It should convert a Tcl list to the Windows "a" "b" form when -multiple is set (options can come in any order, so this must happen after parsing). macosx/tkMacOSXDialog.c (OPEN_INITFILE) also uses the string as one file name.
    • Prototype none (Windows-only code, cannot be verified here).
    • Suggested ticket comment On X11 (tkfbox) -initialfile is already treated as a list with -multiple. Windows (tkWinDialog.c) and macOS still pass it as a single name.
    Windows VM, run 4
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested 2473895_auto.tcl presses Open without changes. With ref91/ref86 the file-name box shows a.txt {b c.txt} verbatim; pressing Open brings up an error box, and the dialog never returns. With the prototype the box shows "a.txt" "b c.txt" and the result is both files.
    • Cause win/tkWinDialog.c ParseOFNOptions translates -initialfile as one file name and passes it to IFileDialog::SetFileName.
    • Prototype 2473895fff.patch (71 lines): -initialfile is parsed after the options loop. With -multiple it is taken as a list, and 2 or more names are quoted as "n1" "n2". winDialog.test and filebox.test: 120 tests, 0 failed. Compatibility caveat: with -multiple, an unbraced single name containing spaces is now split. The doc (getOpenFile.n) still says "filename", so the doc needs updating too.
    • Suggested ticket comment Still present: with -multiple the list is shown verbatim and Windows rejects it. A prototype passes the names as "a" "b" to SetFileName. It changes the meaning of -initialfile with -multiple (as TIP 141 intended), so docs/compatibility need a decision.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main ported. Main now frees optsPtr->utf16FileName for a repeated -initialfile, which made hunk 2 conflict. It is replaced by initFileObj = valuePtr;, so the last -initialfile wins and the result is converted once after the loop.
    • Repro before: the file name box shows a.txt {b c.txt} / after: "a.txt" "b c.txt"
    • Test winDialog-5.28 (new; reads the file name combobox, id 0x47c, and cancels. 0x3e9 matched the Address bar in the Open dialog. Fails before: yes; passes after: yes)
    • Test file run winDialog.test 71/0, filebox.test 63/0 (pre-existing failures: none)
    • Patch changes port and new test. Still open: doc/getOpenFile.n says "filename" for -initialfile with -multiple. With -multiple, an unbraced single name that contains spaces is now split.
    ⬜2472990fff2008PNG: format handler doesn't do -gammaREPRODUCED (main, by code inspection). The PNG reader ignores the gAMA chunk: generic/tkImgPNG.c lists CHUNK_gAMA among the "not used by this imple…

    Open ticket 2472990fff · 41. Photo Images · filed 2008-12-28

    Full verdict

    REPRODUCED (main, by code inspection). The PNG reader ignores the gAMA chunk: generic/tkImgPNG.c lists CHUNK_gAMA among the "not used by this implementation" ancillary chunks. The photo's -gamma option is applied only at display time, the same way for every format.

    • Platform generic
    • Tested source inspection (main).
    • Cause generic/tkImgPNG.c ReadChunkHeader skips gAMA. Nothing combines the file gamma with the photo -gamma.
    • Prototype none. The semantics need to be decided first: whether to correct the pixels by fileGamma*displayGamma when loading, or to set the image's -gamma. This is effectively an RFE.
    • Suggested ticket comment Still true: tkImgPNG.c ignores gAMA. It needs a decision on semantics (apply the file gamma to the pixel data on read?) before anyone implements it. We suggest converting this to an RFE.
    ⬜2443069fff2008Cannot replace [selection handle] commandREPRODUCED (main, 9.0, 8.6, 8.5, 8.4). Result is 1 1 2 1 instead of 1 1 2 2. This is a duplicate of af31a84c91.

    Open ticket 2443069fff · 53. [selection] · filed 2008-12-17

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5, 8.4). Result is 1 1 2 1 instead of 1 1 2 2. This is a duplicate of af31a84c91.

    • Platform generic (tested on X11)
    • Tested main, 9.0, 8.6, 8.5, 8.4 on Xvfb without a WM.
    • Cause generic/tkSelect.c Tk_CreateSelHandler: when a STRING handler is replaced, it finds the existing implicit UTF8_STRING handler and leaves it alone ("Looks like we had a utf-8 target already").
    • Prototype the existing af31a84c91.patch covers this ticket. It replaces the UTF8_STRING handler when its proc matches the replaced STRING handler, and adds tests select-14.2/14.3. I applied it to main: the repro gives 1 1 2 2 and select.test passes (97 passed, 0 failed). I then reverted it. The old branch bug-2443069 (2012) was reverted from trunk because of [3543050]; since then the current code copies CommandInfo per handler.
    • Suggested ticket comment Still present in 8.6 and 9.x. Duplicate of af31a84c91. The fix prepared there also fixes this repro.
    ✅2358545fff2008spinbox formatting has side effectpatchREGRESSED on 8.6 tip. dgp fixed it in 8.5.10 (Tcl_GetDouble replaced by sscanf("%lf")). Check-in e0dd3fdb59 (2026-09-22, core-8-6-branch, "Fix [9…

    Open ticket 2358545fff · 17. [spinbox] · filed 2008-11-29

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-24 check-in 8ebff84d76 on core-8-6-branch, in main, 9.0, 8.6: Fix 2358545fff (again). Thanks, Serhiy!
    • 🔧 2026-09-24 check-in 8ac67e76a4 on spinbox-octal-86, in main, 9.0, 8.6: Fix the bug reference in the spinbox-26.2 test title. [Bug 9e89324a0a]
    • 💬 2026-09-24 serhiy.storchaka: The same octal problem came back on core-8-6-branch with the fix for 9e89324a0a, which replaced sscanf() with Tcl_GetDouble() again. That regression is tracked in 9e89324a0a and fixed on spinbox-octal-86. The remaining reason to keep this ticket open, a public Tcl API for decimal-only parsing, no longer applies to 9.x, where leading zeros are not octal. I suggest closing it as Fixed.
    • 💬 2026-09-27 jan.nijtmans: Fixed (again) 8ebff84d76
    Full verdict

    REGRESSED on 8.6 tip. dgp fixed it in 8.5.10 (Tcl_GetDouble replaced by sscanf("%lf")). Check-in e0dd3fdb59 (2026-09-22, core-8-6-branch, "Fix 9e89324a0a: spinbox parsing depends on the LC_NUMERIC locale") switched back to Tcl_GetDouble, so the octal problem is back on 8.6: with -format %02.0f the value goes 07 → 08 → 00. It is also silently wrong: "010" is read as 8. main and 9.x are not affected because Tcl 9 has no octal parsing.

    • Platform generic
    • Tested main 9.1 OK. 8.6 tip (8.6.18) FAIL: 06 07 08 00 01 02 11 10 09 00 00. 8.6.15, 8.6.13, 8.6.8, 8.5.19 and 8.4 OK.
    • Cause generic/tkEntry.c ConfigureSpinbox and SpinboxInvoke call Tcl_GetDouble, and Tcl 8.6 treats leading zeros as octal.
    • Prototype 2358545fff.patch (against core-8-6-branch, 62 changed lines, test spinbox-26.2). It adds a helper GetSpinboxValue() that skips leading zeros after an optional sign and then calls Tcl_GetDouble, so the value is still parsed independently of the locale. Verified with 8.6 tip plus the patch: the repro is fixed, spinbox.test (293 passed) and entry.test pass, and the new test fails without the patch. It is only needed for 8.6 (harmless on 9.x).
    • Suggested ticket comment The fix for 9e89324a0a (e0dd3fdb59) reintroduced this on core-8-6-branch: Tcl_GetDouble("08") fails and "010" becomes 8 in Tcl 8.6. We propose stripping leading zeros before Tcl_GetDouble (8.6 only). Test spinbox-26.2.
    ✅2182522fff2008Ttk combobox fails to restore grab.patchREPRODUCED (main, 8.6, 8.5)

    Open ticket 2182522fff · 88. Themed Tk · filed 2008-10-20

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-08 check-in 03e1d1af7c on x11-grab-unviewable-button, in main, 9.0: Fix [2182522]: Ttk combobox fails to restore grab. A local grab made while a mouse button is down is turned into a temporary global grab, which fails if the win
    • 🔧 2026-10-09 check-in 80bb803853 on trunk, in main, 9.0: Fix for [2182522]: Ttk combobox fails to restore grab.
    • 🔧 2026-10-09 check-in efac46c2fa on core-9-0-branch, in main, 9.0, 8.6: Fix for [2182522]: Ttk combobox fails to restore grab.
    • 🔧 2026-10-09 check-in 3b02fe5f23 on core-8-6-branch, in main, 9.0, 8.6: Fix for [2182522]: Ttk combobox fails to restore grab.
    • 💬 2026-10-08 serhiy.storchaka: Still reproducible with 8.6.18 and 9.1 if the mouse button is still down when the grab is restored, e.g. when the list is closed by clicking another window of the application. ttk::RestoreGrab runs grab .tree1 again. A plain grab made while a button is down is temporarily promoted to a global grab until the button is released, and this fails because .tree1 is not viewable (it is a frame in the withdrawn .): grab r…
    • 💬 2026-10-09 nemethi: Serhiy, many thanks for fixing this very long-standing bug! Merged into main, core-9-0-branch, and core-8-6-branch by commits 80bb8038, efac46c2, and 3b02fe5f.
    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform X11 (generic tkGrab.c)
    • Tested Xvfb without a WM and under twm, main/8.6/8.5. Clicking outside the app with the button held about 1s: afterwards grab current is empty and top2's Quit works. With a quick click (release before the Unmap is processed) the grab is restored, so the bug depends on timing.
    • Cause generic/tkGrab.c Tk_Grab. A local grab made while a button is down becomes a GRAB_TEMP_GLOBAL XGrabPointer. That fails with GrabNotViewable for an unmapped grab window (.tree1 under a withdrawn .), so the whole local grab fails. ttk::RestoreGrab hides the error with catch. TkPointerEvent already ignores a failed temporary global grab silently.
    • Prototype 2182522fff.patch (30 lines). If the temporary global grab fails, it falls back to a plain local grab. No test, because a real held button is needed (XQueryPointer). Verified with an XTest repro (work-a3/cb.tcl, xbtn). grab/menu/bind tests pass. The focus.test failures (4) also happen without the patch.
    • Suggested ticket comment Still reproducible. When the grab window is unviewable and a button is down, Tk_Grab's temporary global grab fails and so does the whole local grab. The fix falls back to a local grab, as TkPointerEvent does.
    ⬜2095675fff2008fedora 9 tk.spec failure and fixpatchREPRODUCED (main, 8.6): unix/tk.spec still has Buildrequires: XFree86-devel

    Open ticket 2095675fff · 84. Unix Build · filed 2008-09-05

    Full verdict

    REPRODUCED (main, 8.6): unix/tk.spec still has Buildrequires: XFree86-devel

    • Platform build
    • Tested inspected the spec on main and 8.6 (rpmbuild not available)
    • Cause obsolete XFree86-devel dependency. No current distribution provides it.
    • Prototype 2095675fff.patch (1 line: XFree86-devel -> libX11-devel). Removing tk.spec entirely, as dgp suggested, is an alternative.
    • Suggested ticket comment tk.spec still requires XFree86-devel. Either replace it with libX11-devel or drop the unmaintained spec file.
    ⬜2064914fff2008Ttk::menubutton -width overrides its real widthpatchREPRODUCED (main, 8.6, 8.5)

    Open ticket 2064914fff · 88. Themed Tk · filed 2008-08-21

    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform generic (ttk layouts)
    • Tested main/8.6/8.5, default theme, Xvfb. The label element with -width 1 spans only 7px (x 12..19) of a 300px menubutton. A ttk::button spans the full width. The style's -anchor has no effect for the same reason.
    • Cause the TMenubutton layouts place Menubutton.label with TTK_PACK_LEFT and no fill (default, generic/ttk/ttkButton.c), or with 0 flags (classic and vista). The label parcel therefore gets only its requested width, which -width sets, and the text is clipped to it.
    • Prototype 2064914fff.patch (25 changed lines, test layout-3.1). It makes Menubutton.label TTK_FILL_BOTH in the default, classic and vista layouts. It adds -anchor center to TMenubutton in classic and vista to keep their centered text; the default layout already anchors w. Aqua was left unchanged. The new test fails without the fix, and all ttk tests pass.
    • Suggested ticket comment Still present. Menubutton.label is packed -side left without -sticky, so it gets only the width requested by -width. Making it fill the padding cavity, like Button.label, fixes this, and style -anchor then works as well.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST for the generic/classic part on X11; NEEDS-WINDOWS for the vista part
    • Applies to main clean, all 6 files including win/ttkWinVistaTheme.c and library/ttk/vistaTheme.tcl
    • Repro screenshots of ttk::menubutton in the default/alt/clam/classic themes. The natural-size and pack -fill x versions are pixel-identical before and after in all four themes. Only the -width 1 + -fill x case changes (the label is no longer clipped).
    • Test layout-3.1 (existing in the patch; fails before: yes; passes after: yes)
    • Test file run ttk/layout.test 3/0. Full suite with the patch: 9877 total, 0 failed (ttk dir 2064/0).
    • Patch changes none. Aqua is unchanged on purpose (its layout already uses PACK_LEFT and its style already has -anchor center). xpnative no longer exists on main, so vista is the only Windows layout affected. winnative uses the generic layout and has no -anchor, so it stays left-aligned as before.
    • Windows tests to run tests/ttk/layout.test (layout-3.1 under the default theme), tests/ttk/ttk.test, tests/ttk/vsapi.test, and the whole tests/ttk dir. Also a visual check in the vista and winnative themes: ttk::menubutton .a -text Hello at natural size, packed -fill x (the vista text must stay centered, winnative left), and -width 1 with -fill x (the text must not be clipped). Under vista, also check that the dropdown arrow is unchanged.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main clean
    • Repro before: identify element in the middle of a -width 1 menubutton gives padding with both the default and the vista theme / after: label
    • Test layout-3.1 (existing, default theme) and layout-3.2 (new, vista theme, constraint vista); fail before: yes; pass after: yes
    • Test file run ttk/layout.test 4/0, ttk/ttk.test 113/0, ttk/vsapi.test 3/0, menubut.test 117/0 (pre-existing failures: none)
    • Patch changes added layout-3.2 and testConstraint vista. layout-3.2 restores the theme. layout-3.1 still leaves the theme at "default" (harmless in this file).
    ⬜2046267fff2008Ttk::button -takefocus option influences <Button-1>patchREPRODUCED (main, 8.6, 8.5). This is intended behaviour, but it is not documented, so it is a DOCS issue.

    Open ticket 2046267fff · 88. Themed Tk · filed 2008-08-11

    Full verdict

    REPRODUCED (main, 8.6, 8.5). This is intended behaviour, but it is not documented, so it is a DOCS issue.

    • Platform generic
    • Tested main, 8.6 tip, 8.5 on X11 (Xvfb, no WM), using event generate. A click focuses ttk::button/checkbutton with -takefocus 1 and does not with -takefocus 0. ttk::frame -takefocus 1 does not take focus on click. Classic buttons never take focus on click.
    • Cause library/ttk/button.tcl binds TButton <Button-1> to ttk::clickToFocus, which checks ttk::takesFocus, i.e. -takefocus (library/ttk/utils.tcl). TCheckbutton and TRadiobutton copy this binding.
    • Prototype 2046267fff.patch (19 lines, docs only). It adds a sentence to ttk_widget.n -takefocus saying that button-like widgets also use the option to decide click-to-focus. Changing the behaviour or adding a new option (dkf objected to that) is not prototyped.
    • Suggested ticket comment Still present in 9.1/8.6. Focus-on-click is intentional in ttk (''ttk::clickToFocus'' honours -takefocus), so the proposal is to document it in ttk_widget(n) rather than change the behaviour.
    ⬜2045553fff2008tk_getSaveFile -defaultextension disagrees with its man pagepatchf436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.

    Open ticket 2045553fff · 37. [tk_get*File] · filed 2008-08-10

    Full verdict

    f436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.

    • Platform generic (X11 part tested; Windows part NEEDS-WINDOWS)
    • Tested main and 8.6 tip on X11, tk_getSaveFile filled in by script. With -defaultextension "" + -filetypes {{xyz *.xyz}} the result is myfile.xyz. With the option absent it is also myfile.xyz. With .abc it is myfile.abc.
    • Cause On Unix, tkfbox.tcl declares the option with default "", and SetFilter replaces an empty value with the filter's extension, so "" and absent cannot be told apart. On Windows (win/tkWinDialog.c), SetDefaultExtension() is called only when the option is given. IFileDialog treats an empty default extension as "use the selected filter's extension", so "" appends .xyz and absent appends nothing.
    • Prototype 2045553fff.patch (29 lines, docs only: getOpenFile.n describes the actual per-platform behaviour). The alternative is to make the behaviour uniform (danckaert, 2015): "" means no extension and absent means use the filter extension. That needs a "was specified" flag in tkfbox.tcl, and on Windows SetDefaultExtension(L"") when the option is absent but -filetypes is present. It is small (about 20 lines) but changes behaviour on Windows. Windows repro (manual input: type myfile twice): win/2045553fff.tcl
    • Suggested ticket comment Still present in 9.1/8.6. On X11, "" and an absent -defaultextension behave the same (the extension is taken from -filetypes). On Windows only "" adds it. Either the man page is fixed (prototype) or the behaviour is unified (""=none, absent=from filetype).
    Windows VM, run 3
    • Platform Windows
    • Verdict f436621c1a REPRODUCED (9.1 main, 8.6 tip), prototype verified. 2045553fff REPRODUCED ("" and absent differ); doc prototype verified and extended.
    • Tested dialogs driven with act.exe (filename set in Edit 1001, file type changed by a real click on the type combobox plus Down/Enter, then IDOK). Results:
      • ref91 and 8.6: no -defaultextension gives foo (case 1), and also foo after switching to Text (case 2). -defaultextension "" gives the type's extension (foo.tcl, or foo.txt with Text selected). 2045 script: "" gives myfile.xyz, absent gives myfile.
      • Additional check (work-win/dx.tcl, ref91): -defaultextension .dat with first type {"All files" *} gives foo.dat; with first type {Text .txt} gives foo.txt; with {"Tcl or Text" {.tcl .txt}} gives foo.tcl. On Windows the extension of the selected type wins over -defaultextension (on Unix an explicit -defaultextension wins, see tkfbox.tcl SetFilter).
      • Prototype: case 1 foo.tcl, case 2 foo.txt (follows the type selector). 2045 script: both give myfile.xyz. The dx.tcl results do not change.
      • New tests winDialog-5.7.9 (first type .txt gives bar.txt) and 5.7.10 (-typevariable picks .tcl) fail on ref91 and pass with the prototype. winDialog-5.7.* (47af) still pass.
    • Cause win/tkWinDialog.c GetFileNameVista calls SetDefaultExtension only when -defaultextension is given. With an empty default extension, IFileDialog uses the selected type's extension and follows type changes.
    • Prototype f436621c1a.patch (57 lines, reworked). When -defaultextension is absent for a save dialog, it now just calls SetDefaultExtension(L"") (7 code lines), replacing the earlier filter-parsing version. It includes getOpenFile.n text for the patched behaviour and the two tests. 2045553fff.patch (doc only, for unpatched behaviour) is updated with the rule that the selected type's extension wins on Windows. Use one of the two doc texts, depending on whether f436 goes in.
    • Suggested ticket comment f436621c1a: On Windows, IFileDialog uses the selected file type's extension only if some default extension is set. Setting an empty one when -defaultextension is not given makes Windows behave like Unix, including following the type selector. Patch and tests are ready. 2045553fff: confirmed that "" and an absent option differ on Windows; this goes away with the f436621c1a fix.
    ✅1867834fff2008ttk Toolbutton has wrong default anchorpatchREPRODUCED (main, 8.6, 8.5)

    Open ticket 1867834fff · 88. Themed Tk · filed 2008-01-09

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-30 check-in 628bdbd3dd on toolbutton-center, in main, 9.0: Fix [1867834]: ttk Toolbutton has wrong default anchor. The text of a Toolbutton is centered in all built-in themes, as for TButton and for Toolbutton in the cl
    • 🔧 2026-10-03 check-in cc851a9f6c on trunk, in main, 9.0, 8.6: Fix for 1867834fff: ttk Toolbutton has wrong default anchor.
    • 🔧 2026-10-03 check-in dd13e37a93 on core-9-0-branch, in main, 9.0: Fix for 1867834fff: ttk Toolbutton has wrong default anchor.
    • 🔧 2026-10-03 check-in 3b8f72ca6d on core-8-6-branch, in main, 9.0, 8.6: Fix for 1867834fff: ttk Toolbutton has wrong default anchor.
    • 💬 2026-09-30 serhiy.storchaka: Still present in 9.1. Toolbutton is a separate style, and ttk style settings are not inherited across themes: every theme sets -anchor center for TButton, but only clam and aqua do so for Toolbutton, so in the default, alt, classic, winnative and vista themes the text falls back to the element default "w". Multi-line text is also left-justified within the button in all themes, while native Windows buttons and the …
    • 💬 2026-10-03 nemethi: Serhiy, many thanks for fixing this very long-standing bug! Merged into main, core-9-0-branch, and core-8-6-branch by commits cc851a9f, dd13e37a, and 3b8f72ca.
    Full verdict

    REPRODUCED (main, 8.6, 8.5)

    • Platform generic
    • Tested main, 8.6 tip, 8.5 on Xvfb. ttk::style lookup Toolbutton -anchor is empty in the default, alt and classic themes and "center" only in clam (and aqua). A screen grab of a -width 10 Toolbutton in the default theme shows 6px left / 50px right text margins; in clam it is 30/28.
    • Cause the Label element's default anchor is TK_ANCHOR_W (generic/ttk/ttkLabel.c:LabelElementDraw). TButton sets -anchor center in each theme, but Toolbutton only does so in clam/aqua. Style settings do not inherit across parent themes, so every theme needs it.
    • Prototype 1867834fff.patch (28 changed lines: -anchor center for Toolbutton in defaults/alt/classic/winTheme/vistaTheme.tcl; new test ttk-16.3 in tests/ttk/ttk.test fails before and passes after; ttk.test passes 111/111). For 8.6, xpTheme.tcl also needs it. This slightly changes behaviour for apps that give a Toolbutton an explicit width.
    • Suggested ticket comment Still present. Toolbutton gets -anchor center only in clam/aqua, so the label falls back to the element default "w". Fix: set -anchor center for Toolbutton in every theme, as TButton already does (test ttk-16.3).
    ⬜1850888fff2007Dead code: [testfont subfont], TkpGetSubFontsREPRODUCED (main, 8.6). Still true, but this is a cleanup decision rather than a bug.

    Open ticket 1850888fff · 44. Generic Fonts · filed 2007-12-14

    Full verdict

    REPRODUCED (main, 8.6). Still true, but this is a cleanup decision rather than a bug.

    • Platform generic
    • Tested testfont subfonts works on the main tktest build (it lists the fontconfig fallback fonts). No test in tests/ uses it. TkpGetSubFonts is still internal stub slot 110 and is implemented in tkUnixFont.c, tkUnixRFont.c, tkUnixBidiFont.c (newly added), tkWinFont.c and tkMacOSXFont.c.
    • Cause debug-only helper used only by generic/tkTest.c:TestfontObjCmd.
    • Prototype none (removal touches the stub table plus 5 platform files and is a policy call; the helper is still useful for font-fallback debugging)
    • Suggested ticket comment testfont subfonts still works and is a useful debugging aid (it was even implemented in the new tkUnixBidiFont.c). Suggest keeping it and closing as Won't Fix. Alternatively, add a smoke test that calls it.
    ⬜1847926fff2007case of [font actual] resultspatchREPRODUCED (main, 8.6): the tests still compare case-sensitively. unixfont-2.9 was already removed on main/9.0 in 186f54c998/73db542952 (Bug 4e…

    Open ticket 1847926fff · 46. Unix Fonts · filed 2007-12-10

    Full verdict

    REPRODUCED (main, 8.6): the tests still compare case-sensitively. unixfont-2.9 was already removed on main/9.0 in 186f54c998/73db542952 (Bug 4eb4c07dae) but is still present in 8.6.

    • Platform X11
    • Tested main tktest with a private FONTCONFIG_FILE that renames "Misc Fixed" to "Fixed" (as older fontconfig/xorg did). font actual {-family fixed} gives "Fixed". The haveFixedFamilyFont constraint lowercases the name, so it passes, while unixfont-2.6/2.8 compare exactly against "fixed" and fail.
    • Cause tests/unixFont.test. With Xft, [font actual] returns the fontconfig family name with its own capitalisation. The constraint in tests/constraints.tcl uses string tolower but the tests do not.
    • Prototype 1847926fff.patch (4 changed lines: string tolower in unixfont-2.6 and 2.8; they fail before and pass after in the reproduced setup). For 8.6 the same change is needed, plus 2.9 (or backport its removal).
    • Suggested ticket comment Font family names come from fontconfig and keep their case ("Fixed"). The tests should compare case-insensitively, like the haveFixedFamilyFont constraint already does. unixfont-2.9 was removed in 186f54c998.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED (test-only patch, still needed)
    • Applies to main clean
    • Repro with FONTCONFIG_FILE=work-a1/fonts.conf (renames "Misc Fixed" to "Fixed"), unixfont-2.6 and 2.8 fail on main and pass with the patch. With the default fontconfig both are skipped (haveFixedFamilyFont).
    • Test unixfont-2.6/2.8 (existing, modified; fail before: yes in that setup; pass after: yes)
    • Test file run unixFont.test, default config 50/0 (patched). With the renamed-Fixed fonts.conf: main 6 failed, patched 4 failed. The remaining unixfont-1.2, 5.8, 5.12 and 8.5 are pre-existing and caused by that font setup.
    • Patch changes none

    Notes: the checkout was reconfigured against tcl9.1. It is reverted on main b56b1a051f with no changes left. The untracked doc/man.macros-original was already there. Xvfb :111 (my PID) is stopped. The modified patches are 03fcf233e1, 1158738fff, 1600dc73a9, 1d926db566 and 1f1273f739; the originals are kept as <id>.patch.1024. During the session a fossil diff without -i once failed with "gdiff-command: not found" and produced nothing (the patch was redone with fossil diff -i); all saved patches were made with the internal diff.

    ⬜1755313fff2007Big difference in speed between x and y scrolling in text wiREPRODUCED (main, 8.6)

    Open ticket 1755313fff · 18. [text] · filed 2007-07-17

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform generic
    • Tested attached ttext3.tcl on Xvfb. main: X 0.13 s, Y 3.1 s. 8.6.18: X 0.04 s, Y 2.2 s. yview scroll 25 units takes about 300 ms per step, compared with 2.5 ms for xview scroll. Main is about 1.4x slower than 8.6 here.
    • Cause generic/tkTextDisp.c. Y scrolling has to lay out newly exposed display lines (LayoutDLine), and with -wrap none each of these lines is 5000 chars long with a tag toggle on nearly every char. X scrolling reuses the display lines that are already laid out. Improving this means laying out only the visible part of unwrapped lines, which the revised_text branch does.
    • Prototype none (redesign of text layout, not a small fix)
    • Suggested ticket comment Still present in 8.6 and 9.1. The cost is in laying out the newly exposed, very long and heavily tagged lines, not in scrolling as such. The revised_text branch improves it a lot, and a real fix needs lazy layout of -wrap none lines.
    ✅1723059fff2007Entryconfigure destroys submenusREPRODUCED (main fb166f9f73, 9.0.5, 8.6.18, 8.5.19, 8.4.20). Fixed on branch menu-clone-cascade-destroy (ef4c358706 2026-09-23, merged with trunk…

    Open ticket 1723059fff · 10. Generic Menus · filed 2007-05-21

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-30 check-in 257a67c1c0 on menu-clone-cascade-destroy, in main, 9.0: Merge trunk. Renumber the test for [1723059] to menu-45.1, since trunk now has another menu-42.1.
    • 🔧 2026-10-24 check-in 5f1de6fb06 on trunk, in main, 9.0: Fix for 1723059fff: entryconfigure destroys submenus
    • 🔧 2026-10-24 check-in a3f03a589c on core-9-0-branch, in main, 9.0, 8.6: Fix for 1723059fff: entryconfigure destroys submenus
    • 🔧 2026-10-24 check-in edff8f5386 on core-8-6-branch, in 8.6: Fix for 1723059fff: entryconfigure destroys submenus
    • 💬 2026-10-24 nemethi: Serhiy, many thanks for the fix! Merged into main, core-9-0-branch, and core-8-6-branch by commits 5f1de6fb, a3f03a58, and edff8f53. The relevant test is menu-47.1.
    Full verdict

    REPRODUCED (main fb166f9f73, 9.0.5, 8.6.18, 8.5.19, 8.4.20). Fixed on branch menu-clone-cascade-destroy (ef4c358706 2026-09-23, merged with trunk in 257a67c1c0 2026-09-30, test menu-45.1), not merged to main.

    • Platform generic
    • Tested X11 with no WM, the menu-45.1 scenario as a script. Result: sub1 is destroyed (".bar.sub1 exists = 0") on all versions, including installed main after the reference rebuild.
    • Cause generic/tkMenu.c:ConfigureMenuCloneEntries() passes -menu to the clones even when the cascade is unchanged, so the clones' entries point at the original cascade menu, which is then deleted as a clone on the next change.
    • Prototype none (exists on branch menu-clone-cascade-destroy)
    • Suggested ticket comment Still reproducible on main, 9.0 and 8.6. The fix on branch menu-clone-cascade-destroy is ready for review and merge to trunk, 9.0 and 8.6.
    💬1687486fff2007tk is not robust in the face of bad xrdb settingsREPRODUCED (main, 8.6, 8.4)

    Open ticket 1687486fff · 57. [option] · filed 2007-03-24

    💬 Since triage (updated 2026-11-11)

    Status now: Open

    • 💬 2026-09-29 serhiy.storchaka: Still present in 8.6 and 9.1: a bad X resource for *background, *borderWidth, *relief, *cursor or *highlightThickness stops wish from starting, and a bad *foreground or *font makes most widgets fail to create. Tk_InitOptions() treats an invalid value from the option database as an error, like an invalid value given explicitly. This is intended and tested (config-3.7, button-2.9). Xt applications print a warning (C…
    Full verdict

    REPRODUCED (main, 8.6, 8.4)

    • Platform X11 (generic option-db code)
    • Tested Xvfb, RESOURCE_MANAGER set with xrdb (a holder client was needed so that Xvfb does not reset). *background: stupid, *borderWidth: abc, *relief: bogus, *cursor: nosuchcursor and *highlightThickness: x make wish fail at startup ("application-specific initialization failed: …") on 9.1b1 and 8.6.18. 8.4.20 prints the exact message from the report. *foreground: stupid and *font: {nosuch let wish start, but most widgets then fail to create. *Text.tabs: {1c q} breaks text in 9.1 only.
    • Cause generic/tkConfig.c:Tk_InitOptions. A value from the option database that fails DoObjConfig is a hard error, and it is reported as "(database entry for …)". This is intentional and tested by config-3.7 and button-2.9. The root toplevel hits it during Tk_Init.
    • Prototype none. The fix is a behaviour change: on failure of an OPTION_DATABASE (or SYSTEM_DEFAULT) value, retry with the table default and possibly emit a warning. That is about 15 lines in Tk_InitOptions, but it changes documented and tested behaviour (config-3.7, button-2.9), so it needs a decision or TIP. A narrower variant would apply the fallback only to user-level X resources (RESOURCE_MANAGER/.Xdefaults) and not to option add, but Tk_GetOption does not return the priority.
    • Suggested ticket comment Still present in 8.6 and 9.1: a bad RESOURCE_MANAGER entry for background/borderWidth/relief/cursor stops wish from starting. A possible fix is to fall back to the built-in default when a database value fails to parse, but that changes the tested error behaviour (config-3.7, button-2.9), so it needs agreement first.
    ⬜1660811fff2007Wrong rendering in transparent areaspatchREPRODUCED (main, 8.6, 8.4). This is by design: frames with -background {} are not transparent.

    Open ticket 1660811fff · 02. Appearance · filed 2007-02-15

    Full verdict

    REPRODUCED (main, 8.6, 8.4). This is by design: frames with -background {} are not transparent.

    • Platform X11 (generic)
    • Tested Xvfb, no WM, the reporter's script with a screen grab (PIL). After destroy .c.f2.f1, the padding area of .c.f2 (bg {}) keeps stale cyan pixels instead of the canvas red. The result is the same on main, 8.6 and 8.4.20.
    • Cause an empty background sets the X window background to None, and Tk never draws the frame. So the server never repaints the exposed areas. X11 has no true child-window transparency without compositing or the Shape extension (dkf's comment).
    • Prototype 1660811fff.patch (docs only, 40 lines). frame.n and toplevel.n now say that an empty -background does not make the widget transparent and that uncovered areas may show stale contents.
    • Suggested ticket comment Still the case on 8.6 and 9.x. An empty -background gives the window a None background, so Tk never repaints it. Real transparency is not possible there. Proposed documenting this in frame.n/toplevel.n and closing.
    ✅1652598fff2007Free memory read tkOption.c SetupStackspatchREPRODUCED (main; the same code is in 8.6 and 8.4)

    Open ticket 1652598fff · 10. Generic Menus · filed 2007-02-05

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-29 check-in 2d944be259 on option-stack-realloc, in main, 9.0: Fix [1652598]: Free memory read tkOption.c SetupStacks
    • 🔧 2026-10-08 check-in 6043b10570 on trunk, in main, 9.0: Fix for [1652598]: Free memory read by SetupStacks in tkOption.c.
    • 🔧 2026-10-08 check-in e047308b2b on core-9-0-branch, in main, 9.0, 8.6: Fix for [1652598]: Free memory read by SetupStacks in tkOption.c.
    • 🔧 2026-10-08 check-in 7a4d817627 on core-8-6-branch, in main, 9.0, 8.6: Fix for [1652598]: Free memory read by SetupStacks in tkOption.c.
    • 💬 2026-09-29 serhiy.storchaka: Still present. ExtendStacks() appends the children of a matching node to the stacks, and if a child has the same type as the stack being scanned (e.g. *Frame*Button*opt, both class wildcard nodes), it can reallocate that stack, and the loop continues with the pointer to the freed array. Usually this goes unnoticed because the freed memory is not reused yet. Proposed fix in branch option-stack-realloc, with test op…
    • 💬 2026-10-07 chrstphrchvz: I encountered the same problem during test ttk-coreoptions-combobox for a -DPURIFY build of recent Tcl/Tk 9.1 on macOS 15 (Intel). The fix proposed by Serhiy works for me.
    • 💬 2026-10-08 nemethi: Serhiy, many thanks for fixing this very long-standing bug! Merged into main, core-9-0-branch, and core-8-6-branch by commits 6043b105, e047308b, and 7a4d8176.
    Full verdict

    REPRODUCED (main; the same code is in 8.6 and 8.4)

    • Platform generic
    • Tested main with instrumentation. With option add *f*gN.foo for N up to 2000, followed by *zz*q.foo, creating .f makes ExtendStacks() reallocate the WILDCARD_NODE_NAME stack while SetupStacks() still loops over it with elPtr ("DANGLING stack 6 remaining 1"). valgrind does not flag it because Tcl's allocator caches the freed blocks.
    • Cause generic/tkOption.c SetupStacks step 5 walks tsdPtr->stacks[i]->els with a pointer. ExtendStacks() → ExtendArray() can Tcl_Realloc that same stack (a wildcard node's children are appended to the same wildcard stack), which leaves elPtr dangling.
    • Prototype 1652598fff.patch (34 changed lines). The loop uses an index and recomputes elPtr after each ExtendStacks() call, which is essentially the reporter's suggestion. Test option-17.1 added. option.test passes (105/105).
    • Suggested ticket comment Confirmed. ExtendStacks() can reallocate the stack that SetupStacks() is iterating over, so elPtr dangles. The fix is to iterate by index.
    ✅1596662fff2006Menu behaviour at screen bottomREPRODUCED (main, 8.6, 8.4). This is a duplicate of f9d316a5e0 ("linux menu problem if menu near the bottom of screen").

    Open ticket 1596662fff · 10. Generic Menus · filed 2006-11-14

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-09-29 check-in c81290221a on menu-post-under-pointer, in main, 9.0: Fix [470331], [680660], 1596662fff, f9d316a5e0: a menu posted over its menubutton or menubar entry invoked the entry under the pointer on the release of the
    • 🔧 2026-09-30 check-in 0f86666370 on trunk, in main, 9.0: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in 5c47ae3e9c on core-9-0-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in 3aacabb15f on core-8-6-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
    • 🔧 2026-09-30 check-in a33031d2f8 on menu-post-under-pointer, in main, 9.0: Add a widget demo for popup (context) menus posted with tk_popup, with menus of different sizes.
    • 💬 2026-09-30 serhiy.storchaka: Duplicate of [470331]. Proposed fix in branch menu-post-under-pointer: as suggested here and as other toolkits do, a menu which does not fit below the menubar is posted above it. A menu which still covers the pointer does not invoke an entry on the release of the click which posted it. See [470331] for details.
    • 💬 2026-09-30 nemethi: Serhiy, many thanks! The fix is now in main, core-9-0-branch, and core-8-6-branch; see commits 0f866663, 5c47ae3e, and 3aacabb1.
    Full verdict

    REPRODUCED (main, 8.6, 8.4). This is a duplicate of f9d316a5e0 ("linux menu problem if menu near the bottom of screen").

    • Platform X11
    • Tested Xvfb, no WM, a toplevel at the bottom of the screen with a 10-item menubar cascade. A press and release on the menubar entry (event generate) invokes "Item 8" on main, 8.6 and 8.4.20. I made the same fix independently: with the cascade posted above the menubar, nothing is invoked and the menu stays posted.
    • Cause generic/tkMenuDraw.c AdjustMenuCoords always places the dropdown below the menubar entry. unix/tkUnixMenu.c TkpPostTearoffMenu then clamps it up under the pointer. The menubutton -direction below fallback in library/menu.tcl also uses a wrong formula.
    • Prototype see f9d316a5e0.patch, which covers this ticket (menubar and menubutton, test menuDraw-17.3). No separate patch.
    • Suggested ticket comment Still present on X11. Duplicate of f9d316a5e0: post the dropdown above the menubar/menubutton when it doesn't fit below.
    ⬜1571737fff2006PRIMARY sel handlers and non-Latin-1 encodingspatchREPRODUCED (main, 8.6)

    Open ticket 1571737fff · 53. [selection] · filed 2006-10-05

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11 (generic tkSelect.c)
    • Tested main and 8.6 tip on Xvfb. The owner is a text widget with a handler returning "тест". The requestor was xsel -o -p plus a second wish that requests UTF8_STRING and STRING. Results for the ticket's cases: I (replace text's STRING handler) gives ????. II and IV (selection handle -type UTF8_STRING) give ????. V works. A STRING request returning ???? is correct per the ICCCM (Latin-1), as dkf said.
    • Cause two bugs. (1) Case I: Tk_CreateSelHandler leaves the UTF8_STRING handler that was implicitly created for the widget's original STRING handler, so UTF8_STRING requests still go to the text widget's own handler. This is the same bug as af31a84c91, and the existing prototype af31a84c91.patch fixes case I (verified). (2) Cases II and IV: selection handle -type UTF8_STRING with no -format defaults the format to STRING, so tkUnixSelect.c ConvertSelection encodes the data as iso8859-1.
    • Prototype 1571737fff.patch (25 changed lines incl. doc + test select-14.4). When -type is UTF8_STRING and no -format is given, the format now defaults to UTF8_STRING, and selection.n says so. Together with af31a84c91.patch, xsel gets "тест" in cases I, II, IV and V. select.test (96 passed) and clipboard.test pass. select-14.4 fails without the fix.
    • Suggested ticket comment STRING is Latin-1 by definition, but UTF8_STRING requests were also broken. Replacing a widget's STRING handler did not replace the implicit UTF8_STRING handler (same as af31a84c91). selection handle -type UTF8_STRING also defaulted to -format STRING and so sent Latin-1. There is a prototype fix for both.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (cases II and IV. Case I also needs af31a84c91.patch, which I did not check here)
    • Applies to main clean
    • Repro before UTF8_STRING handler without -format sends Latin-1 (????) / after "тест"
    • Test select-14.4 (in patch; fails before: yes; passes after: yes)
    • Test file run select.test 112/0, clipboard.test 42/0
    • Patch changes none
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 1571737fff: code is inert on Windows (no UTF8_STRING atom under _WIN32); select-14.4 is x11-only, skipped. select.test: no new failures.
    ⬜1567354fff2006[tk scaling] changespatchREPRODUCED (9.1 main, 8.6 tip): after tk scaling 2.0, a posted WM_DISPLAYCHANGE (same depth and size) resets it to 1.3317.

    Open ticket 1567354fff · 99. Other · filed 2006-09-28

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip): after tk scaling 2.0, a posted WM_DISPLAYCHANGE (same depth and size) resets it to 1.3317.

    • Platform Aqua (original), Windows (comment)
    • Tested X11 main, 8.6 and 8.4 on Xvfb. tk scaling 2.0 survives update and a 1.5 s after.
    • Cause tk scaling is stored in WidthMMOfScreen/HeightMMOfScreen. On Aqua, TkMacOSXDisplayChanged overwrote them (it now does so only when they are 0). On Windows, the WM_DISPLAYCHANGE handler in win/tkWinWm.c (TkWinWmProc) recomputes WidthMMOfScreen/HeightMMOfScreen from LOGPIXELS, which silently resets a user-set scaling. The same "only if unset" approach or a saved user value would fix it.
    • Prototype none (Windows repro needs a monitor change or hibernate)
    • Suggested ticket comment The Aqua reset was fixed by eefcafad13 (fe932e457). The 8.6 branch still has it. On Windows, WM_DISPLAYCHANGE in tkWinWm.c recomputes the screen mm size and so resets tk scaling.
    Windows VM, run 5
    • Platform Windows (Aqua part fixed on main, see result-44)
    • Verdict REPRODUCED (9.1 main, 8.6 tip): after tk scaling 2.0, a posted WM_DISPLAYCHANGE (same depth and size) resets it to 1.3317.
    • Tested 1567354fff.tcl (act.exe post), with the real depth and with a fake depth of 16 (the TkWinDisplayChanged path).
    • Cause both the WM_DISPLAYCHANGE handler in win/tkWinWm.c and TkWinDisplayChanged in win/tkWinX.c recompute WidthMMOfScreen/HeightMMOfScreen from LOGPIXELS.
    • Prototype 1567354fff.patch (79 changed lines incl. comments and test). A new TkWinScreenSizeChanged() keeps the current pixels-per-mm once they are known and uses LOGPIXELS only for the first initialisation. Both paths use it. testwinevent now accepts WM_DISPLAYCHANGE (sent to the window itself). New test winWm-11.1. With the patch the scaling stays at 2.005 on both paths. winWm, wm and winfo tests: 405 run, 0 failed. Caveat: an old tktest ignores the unknown message, so the test only detects the bug when tkWinTest.c has the patch.
    • Suggested ticket comment Confirmed on Windows 11 (8.6.18, 9.1): any WM_DISPLAYCHANGE resets tk scaling, because the screen mm size is recomputed from LOGPIXELS. Prototype: keep the current pixel/mm ratio on display changes; test winWm-11.1.
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST
    • Applies to main ported. tkWinInt.h: the context moved. tkWinTest.c: the table and switch changed on main, so WM_DISPLAYCHANGE was added to the message table and shares the new WM_MOUSEACTIVATE case (SendMessageW to the window). The test was renumbered winWm-11.1 → winWm-13.1, because 11.1 and 12.1 now exist, and moved to the end of the file.
    • Repro before: winWm-13.1 gives 0 (scaling reset by WM_DISPLAYCHANGE; checked with a tktest that has only the tkWinTest.c part) / after: 1
    • Test winWm-13.1 (existing in patch, renumbered; fails before: yes; passes after: yes)
    • Test file run winWm.test 39/0, wm.test 306/0, winfo.test 76/0 (B6, current main; pre-existing failures: none)
    • Patch changes port as above. Note: it conflicts textually with a58cda72fe in TkWinDisplayChanged (win/tkWinX.c); the merge is trivial.
    ⬜1556406fff2006Unloading of Tk can replace user procedurespatchREPRODUCED (main, 8.6)

    Open ticket 1556406fff · 63. Tk_Win Functions · filed 2006-09-11

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform generic
    • Tested tclsh main and 8.6 + package require Tk, on Xvfb. After rename scale tk_scale; proc scale ...; destroy ., calling scale gives "application has been destroyed". Also found: rename canvas tk_canvas; destroy .; tk_canvas .c segfaults on main (valgrind: invalid read in Tk_NameToWindow). The renamed command keeps its now-freed main window as clientData.
    • Cause generic/tkWindow.c Tk_DestroyWindow replaces every Tk command by name with TkDeadAppObjCmd. It does not check whether the command under that name is still Tk's, and it does not follow renamed commands.
    • Prototype 1556406fff.patch (32 changed lines, test window-2.13). Before replacing a command, it checks with Tcl_GetCommandInfo. If the command under that name is not Tk's objProc2, it leaves it alone. The ticket repro now keeps the user proc. window.test passes (19/19, run with NO_AT_BRIDGE=1), and window-2.13 fails without the fix. It does not fix the renamed-away case or the tk_canvas crash. That needs dkf's approach of storing the Tcl_Command tokens, plus a way to detect tokens the user has deleted, for example command delete traces.
    • Suggested ticket comment Still present in 9.x. A prototype now leaves commands the user replaced alone. Renamed Tk commands are still not disabled, and for PASSMAINWINDOW commands such as canvas, calling one after destroy . crashes. That needs token-based tracking.

    Notes: one slip during this batch. A wish8.5 started with the default DISPLAY=:0 ran for up to about 2 minutes, then I killed it. It only ran puts [info patchlevel] from stdin, with no input synthesis. All other runs used Xvfb :124, which has been stopped. wt-a4 is reverted and rebuilt on main. doc/man.macros-original is an untracked extra in wt-a4 that I did not create.

    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (partial fix, as the triage said: renamed Tk commands are still not disabled, and the tk_canvas crash after destroy . remains)
    • Applies to main clean (offset 34)
    • Repro before scale -> "application has been destroyed" / after scale -> "user scale"
    • Test window-2.13 (in patch; fails before: yes; passes after: yes)
    • Test file run window.test 19/0 (NO_AT_BRIDGE=1)
    • Patch changes none. Reviewed: Tk commands are created with Tcl_CreateObjCommand2, so the isNativeObjectProc == 2 && objProc2 == cmdPtr->objProc check is right. USEINITPROC and SAVEUPDATECMD commands keep the old path.
    Prototype verified 2026-10-24, Windows VM, generic patches
    • 1556406fff: OK. window-2.13 (unixOrWin) hangs window.test before (TIMEOUT), window.test 19/0 after.
    ⬜1539962fff2006can’t correctly mark Hebrew text in tk entry/text widgetREPRODUCED for the entry widget (main, 9.0, 8.6). For the text widget it is only partly fixed on main: the text widget has bidi layout there, but t…

    Open ticket 1539962fff · 18. [text] · filed 2006-08-14

    Full verdict

    REPRODUCED for the entry widget (main, 9.0, 8.6). For the text widget it is only partly fixed on main: the text widget has bidi layout there, but tagged RTL text is still scrambled (see 59d6d5f6c0).

    • Platform generic
    • Tested X11 main (bidi build), 9.0.5, 8.6.18. The entry lays Hebrew out left to right in logical order on every version (bbox x(0)=3, x(4)=46).
    • Cause generic/tkEntry.c has no bidi reordering. The 9.1 text widget reorders runs only for untagged lines (tkTextDisp.c).
    • Prototype none (bidi for entry is a feature-size change)
    • Suggested ticket comment In 9.1 the text widget shows Hebrew and Arabic right to left. The selection problem with tagged text is tracked in 59d6d5f6c0. The entry and ttk::entry widgets still have no bidi support. Keep this open for entry, or retitle it to that.
    ⬜1233834fff2005Restrictions on window resizing in Motif wmsREPRODUCED (main, 8.6)

    Open ticket 1233834fff · 67. Unix Window Operations · filed 2005-07-06

    Full verdict

    REPRODUCED (main, 8.6)

    • Platform X11
    • Tested my own repro, a mapped toplevel with wm resizable .t 0 0 then 1 1, after which the content grows to 480x500. Under mwm the window stays 200x150 on both main and 8.6 (2026-10-10). Without the toggle, mwm resizes it to 480x500. With no WM and with icewm, 480x500 is correct. twm needs interactive placement, so I could not test it.
    • Cause unix/tkUnixWm.c (UpdateSizeHints). While the window is not resizable, Tk sends PMaxSize equal to the current size. When it becomes resizable again, Tk drops the flag, and mwm keeps the last maximum it saw.
    • Prototype none. The workaround would be to send PMaxSize with the default maximum (screen size) instead of omitting it. That changes the hints every WM sees, which serhiy already judged not worth the risk. The script-level workaround is a wm maxsize call.
    • Suggested ticket comment Still reproducible in main and 8.6 with mwm. Only the wm resizable 0 0 / 1 1 toggle triggers it, and an explicit wm maxsize avoids it. I suggest closing as Won't Fix (window manager quirk).
    ⬜1191258fff2005text/entry become very inefficients with long linesREPRODUCED (main, 9.0, 8.6, 8.5; the slowdown is much milder than the "tens of seconds" reported in 2015)

    Open ticket 1191258fff · 18. [text] · filed 2005-04-27

    Full verdict

    REPRODUCED (main, 9.0, 8.6, 8.5; the slowdown is much milder than the "tens of seconds" reported in 2015)

    • Platform generic
    • Tested Xvfb, a single line of N×"x", -wrap char/word. N=100k: typing and cursor moves take 10–16 ms and see end 0.09 s (8.6) / 0.17 s (main), so it is usable. Scaling is quadratic: see end N=200k 0.39 s/0.45 s, N=400k 1.85 s/2.2 s (8.6 / main non-bidi build). The installed main build with bidi is about 2x slower (IsSimpleOnly in unix/tkUnixBidiFont.c rescans the rest of the string). 8.4.20 is about 2.5x faster, with the same growth.
    • Cause every display line is laid out with scans that run to the end of the logical-line segment. In generic/tkTextDisp.c, LayoutDLine's tab search (for (p = ... ; *p != 0; p++)) and MeasureChars's search for the next '\t'/'\n' are both O(rest of segment), and TkTextFindDisplayLineEnd/MeasureUp lay out from the start of the logical line. So one pass over a line of n chars costs O(n²/width).
    • Prototype none. The scans would have to be bounded (e.g. scan in windows while Tk_MeasureChars still fits) or the next-special-char position cached per segment. Doing that safely with TK_WHOLE_WORDS and the -wrap word logic is more than a small change.
    • Suggested ticket comment Still quadratic in the length of one logical line: LayoutDLine and MeasureChars in tkTextDisp.c scan to the end of the segment for every display line. It is usable at 100k chars (about 0.1–0.2 s for see end) but takes about 2 s at 400k.
    ⬜1187074fff2005Windows XDrawArc implementationREPRODUCED (9.1 main, 8.6 tip)

    Open ticket 1187074fff · 82. X11 Emulation · filed 2005-04-21

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip)

    • Platform Windows
    • Tested X11 main (Xvfb, no WM): the concentric circles form solid disks, which is the expected result.
    • Cause win/tkWinDraw.c:DrawOrFillArc still draws with GDI Arc/Chord/Pie on a +1-enlarged box. GDI's ellipse rasterisation is different from X11's, so concentric 1px rings leave gaps. The code has not changed materially since 2005.
    • Prototype none (Windows-only; a fix would be to rasterise the arcs ourselves, as in hobbs' CE code; see also 1081603). Repro: win/1187074fff.tcl (checks the pixels automatically when Img is present, otherwise shows the circles to inspect by eye).
    • Suggested ticket comment DrawOrFillArc in win/tkWinDraw.c still uses GDI Arc/Chord/Pie, so the rasterisation difference from X11 is probably still there. Needs a retest on Windows 11.
    Windows VM, run 5
    • Platform Windows
    • Verdict REPRODUCED (9.1 main, 8.6 tip)
    • Tested the same capture, with concentric ovals of diameter 1..39 and 2..40 step 2 and -width 1. White pixels inside the disks: X11 0/0; Windows 92 (odd) and 67 (even) on both versions.
    • Cause GDI Ellipse rasterisation differs from X11's arc algorithm (adjacent circles don't tile).
    • Prototype none (it would need Tk's own circle rasteriser for 1-px outlines)
    • Suggested ticket comment Still present on Windows 11 (8.6.18, 9.1): concentric 1-px circles leave 67–92 holes per disk, where X11 leaves none. This comes from GDI's ellipse rasterisation.
    ✅1185640fff2005Strange word and line selection in the text widgetpatchREPRODUCED (main with bidi/HarfBuzz enabled, the default build); NOT-REPRODUCED on 9.0, 8.6, 8.5, 8.4 and on main built with --disable-bidi

    Open ticket 1185640fff · 46. Unix Fonts · filed 2005-04-19

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 💬 2026-09-24 serhiy.storchaka: Fixed by 437a4bc8fb (lax UTF-8 parsing in Tk_MeasureChars): not reproducible with 8.6 and 9.0. The text widget still measures a partial UTF-8 character in CharBboxProc, which gives wrong results with the new bidi text rendering in 9.1: 7579c335f1.
    Full verdict

    REPRODUCED (main with bidi/HarfBuzz enabled, the default build); NOT-REPRODUCED on 9.0, 8.6, 8.5, 8.4 and on main built with --disable-bidi

    • Platform X11 (generic code)
    • Tested Greek text, comparing .t bbox width with font measure for each character: 12 of 17 wrong on main (bidi build), 0 elsewhere. No WM.
    • Cause generic/tkTextDisp.c:CharBboxProc measures the character at byteIndex as the byte range [byteIndex, byteIndex+1], which cuts a multi-byte UTF-8 character in half. This is part #1 of jenglish's 2005 analysis, which was never fixed. Older font code tolerated the half character. The HarfBuzz measuring path in the bidi build does not, and gives wrong widths.
    • Prototype 1185640fff.patch (39 changed lines incl. test, test textDisp-25.2). It ends the measured range at the next character boundary (Tcl_UtfNext) for both the LTR and RTL branches. The characters are read from the base-chunk DString because ciPtr->chars can be stale there: using it crashed textDisp-22.10. textDisp, text, textTag and textIndex tests pass in the bidi build; textDisp passes with --disable-bidi. The new test fails without the fix.
    • Suggested ticket comment CharBboxProc still measures [byteIndex, byteIndex+1] bytes. With the new bidi/HarfBuzz text measurement this gives wrong ''bbox'' widths for non-ASCII characters on main. Fix: measure up to the next character boundary; test textDisp-25.2.
    ⬜1158738fff2005transient window madnesspatchREPRODUCED (9.1 main, 8.6 tip), both examples. Ex1: .t2 never appears after both windows are deiconified (ismapped 0, no wrapper, owner 0). Ex2: …

    Open ticket 1158738fff · 68. Win Window Operations · filed 2005-03-08

    Full verdict

    REPRODUCED (9.1 main, 8.6 tip), both examples. Ex1: .t2 never appears after both windows are deiconified (ismapped 0, no wrapper, owner 0). Ex2: .t2 appears, but GetWindow(GW_OWNER) is 0 instead of .t1's wrapper. Control (both mapped): the owner is correct.

    • Platform Windows (a related bug on X11)
    • Tested X11 main, 8.6, 8.4.20 without a WM, checking WM_TRANSIENT_FOR on the wrapper. Example 1 is OK. In example 2 WM_TRANSIENT_FOR is never set, so .t2 is not transient for the WM. The patched build was also run under fvwm.
    • Cause when wm transient is set while the container is unmapped and the transient is no longer WM_NEVER_MAPPED, both unix/tkUnixWm.c:WmTransientCmd and win/tkWinWm.c:WmTransientCmd only withdraw the transient. Neither sets the hint: XSetTransientForHint on X11, UpdateWrapper (the owner HWND) on Windows. Nothing sets it later when the window is deiconified. For the never-mapped case, TkWmMapWindow already sets it on X11 ([Bug 1163496]).
    • Prototype 1158738fff.patch (X11 only, 38 changed lines incl. test, test unixWm-37.6). It always sets or deletes WM_TRANSIENT_FOR, then withdraws if the container is unmapped. unixWm.test and wm.test show no new failures with or without fvwm, and the new test fails without the fix. The Windows equivalent would call UpdateWrapper (or at least set the owner) before withdrawing; it is untested. Repro: win/1158738fff.tcl (checks the owner with twapi when available).
    • Suggested ticket comment The same flow also fails on X11: wm transient on a withdrawn window whose container is withdrawn never sets WM_TRANSIENT_FOR. Prototype fix for unix/tkUnixWm.c with test unixWm-37.6. win/tkWinWm.c has the same pattern (withdraws without UpdateWrapper) and needs a retest on Windows.
    Windows VM, run 5
    • Platform Windows (X11 counterpart too)
    • Verdict REPRODUCED (9.1 main, 8.6 tip), both examples. Ex1: .t2 never appears after both windows are deiconified (ismapped 0, no wrapper, owner 0). Ex2: .t2 appears, but GetWindow(GW_OWNER) is 0 instead of .t1's wrapper. Control (both mapped): the owner is correct.
    • Tested 1158738fff.tcl, with the owner read through act.exe.
    • Cause win/tkWinWm.c. (1) TkWmMapWindow returns early for a never-mapped transient whose container is unmapped. The window stays WM_NEVER_MAPPED without a wrapper, and nothing maps it later once it is WM_WITHDRAWN. (2) WmTransientCmd only withdraws the window when the container is unmapped and never calls UpdateWrapper, so the owner is not set.
    • Prototype 1158738fff.patch, updated: the X11 part from result-47 plus Windows and tests, 76 changed lines. Windows: TkWmMapWindow no longer returns early (the wrapper is created withdrawn). WmTransientCmd withdraws, sets initial_state to Withdrawn and then always calls UpdateWrapper. New tests wm-transient-8.2 and 8.3 (both examples; checks ismapped and wm stackorder after raise .t1). Windows: the repro gives OK/OK with the correct owners. wm.test and winWm.test: 332 run, 0 failed. ref91 tktest fails 8.2 and 8.3. X11 (Linux wt-win, fvwm): 8.1–8.3 pass with the patch; without it 8.3 fails ({.t2 .t1}).
    • Suggested ticket comment Both examples still fail on Windows 11 (8.6.18, 9.1): example 1 never shows the transient, and example 2 shows it without an owner. The fix is not to leave never-mapped transients unwrapped and to update the wrapper when wm transient is set on a withdrawn window (X11 has the same example-2 bug). Prototype with tests wm-transient-8.2/8.3.
    Prototype verified 2026-10-24, Linux
    • Status VERIFIED+TEST (X11); NEEDS-WINDOWS (win part)
    • Applies to main ported. Today's main check-in b56b1a051f (fix for 562c3a3da5, by csaba) already changed TkWmMapWindow in win/tkWinWm.c: the never-mapped transient whose container is unmapped no longer returns early. So I dropped the patch's first win hunk (the return; removal). The second win hunk (WmTransientCmd: withdraw, set initial_state Withdrawn, then always UpdateWrapper) still applies. Main also added its own wm-transient-8.2, so the patch's tests 8.2/8.3 became wm-transient-8.3/8.4.
    • Repro before, under fvwm, WM_TRANSIENT_FOR is never set (unixWm-37.6 fails, wm-transient-8.4 gives the wrong stackorder) / after correct
    • Test unixWm-37.6 (fails before: yes; passes after: yes). wm-transient-8.3/8.4 (fvwm: 8.4 fails before, 8.3 passes before, both pass after). I changed their constraint from {failsOnCILinux failsOnXQuartz} to withWindowManager, the same as 8.1: without a WM both failed even with the fix, because the stackorder is meaningless there.
    • Test file run (fvwm): unixWm.test 285/0, wm.test 304/0. Unpatched: unixWm-37.6 and wm-transient-8.4 fail. No WM: unixWm 285/0; 8.3/8.4 are skipped.
    • Windows tests to run wm.test (wm-transient-8.1 to 8.4; 8.2 is main's 562c3a3da5 test, check that the win hunk does not break it), winWm.test, and win/1158738fff.tcl (owner via twapi/act.exe, both examples). Without the WM-specific constraint, 8.3/8.4 run on Windows (withWindowManager is 1 there).
    • Patch changes dropped the obsolete win hunk 1, renumbered the wm.test tests to 8.3/8.4, and changed their constraint to withWindowManager
    Prototype verified 2026-10-24, Windows VM
    • Status VERIFIED+TEST (Windows part, the current 21:38 version of the patch)
    • Applies to main clean on current main b56b1a051f (on the VM: e1bf plus the e1bf..b56b delta). The wm.test hunk does not apply to e1bf, because it needs the new wm-transient-8.2 from b56b1a051f.
    • Repro on current main, wm-transient-8.3 (example 1) already passes, because b56b1a051f fixed it. wm-transient-8.4 (example 2, owner not set): fails before / passes after.
    • Test wm-transient-8.4 (existing in patch; fails before: yes; passes after: yes); 8.3 passes both ways on current main
    • Test file run wm.test 306/0 (3 runs), winWm.test 39/0 (3 runs), frame.test 207/0, menu.test 574/0 (pre-existing failures: wm-attributes-1.5.1/1.5.2/1.5.4 and winWm-7.3 failed once each, both with and without the patch; they passed in all later runs)
    • Patch changes none
    ⏳1144186fff2005-use -container embedded windows menu problemspatchREPRODUCED (Windows part, 9.1 main)

    Open ticket 1144186fff · 65. Generic Window Operations · filed 2005-02-19

    ⏳ Since triage (updated 2026-11-11)

    Status now: Pending (Wont Fix)

      Full verdict

      REPRODUCED (Windows part, 9.1 main)

      • Platform X11 + Windows
      • Tested Xvfb, no WM and twm. Container frame at root y=110, embedded app in a second process. xwininfo shows the real X geometry: menubar at y=110, toplevel at 142. Tk reports winfo rooty . = 174 and winfo rooty .#mb = 142, both one menubar height too low. Pulldowns are posted from these coordinates, so they open 31–32 px too low, which is the original report. A same-process container is not affected.
      • Cause unix/tkUnixWm.c Tk_GetRootCoords(), branch for an embedded toplevel whose container is in another application. The loop has already added winPtr->changes.y (= menuHeight, the toplevel's offset in its wrapper). XTranslateCoordinates(winPtr->window → root) already includes that offset, so it is counted twice.
      • Prototype 1144186fff.patch (88 lines; 2 changed C lines + comment, plus new test unixEmbed-11.3, which embeds a menubar toplevel through open "|[interpreter] … -use"). The test fails without the fix and passes with it; unixEmbed, unixWm, winfo, menu and unixMenu tests all pass. With the fix, menus post right under the embedded menubar.
      • Windows check win/1144186fff.tcl. It embeds a second wish with a menubar into a container frame and measures how far its first widget is below the container top; 0 px means FAIL (no menubar). Prints a VERDICT line.
      • Suggested ticket comment On X11 the misplaced menus are real when the container is in another process: Tk_GetRootCoords() adds the menubar height twice, so winfo rooty of the embedded toplevel and its menubar is off by the menubar height and pulldowns open too low. A fix with a test is ready. On Windows the menubar is missing because an embedded toplevel is a WS_CHILD window.
      Windows VM, run 6
      • Platform Windows
      • Verdict REPRODUCED (Windows part, 9.1 main)
      • Tested A second wish embedded with -use into a container frame: its client area starts 0 px below the container top, and its menubar has height 1 and is not mapped, so the embedded toplevel shows no menubar.
      • Prototype none
      • Suggested ticket comment On Windows the menubar of an embedded toplevel is still missing in current main (the embedded toplevel is a WS_CHILD window). The X11 part (menus posted too low) has a prototype fix.
      Prototype verified 2026-10-24, Linux
      • Status VERIFIED (X11 part). The Windows part (no menubar in an embedded toplevel) is still NEEDS-WINDOWS, with no patch.
      • Applies to main clean
      • Repro before winfo rooty of the embedded toplevel and its menubar is one menubar height too low / after correct
      • Test unixEmbed-11.3 (in patch; fails before: yes; passes after: yes). It waits on the child pipe with vwait plus a 10 s safety timer, not a time delay, so I did not convert it to pause.
      • Test file run unixEmbed.test 59/0, unixWm.test 284/0, winfo.test 76/0
      • Patch changes none
      ✅1100518fff2005Button should stay in -overrelief state after eventpatchREPRODUCED (main, 8.6, 8.5, 8.4)

      Open ticket 1100518fff · 03. [*button] and [label] · filed 2005-01-11

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-09-28 check-in f27b1cfd17 on button-overrelief-after-click, in main, 9.0: Fix [1100518]: Button should stay in -overrelief state after event
      • 🔧 2026-09-29 check-in 3aa08341e5 on trunk, in main, 9.0, 8.6: Fix for 1100518fff: Button should stay in -overrelief state after event.
      • 🔧 2026-09-29 check-in 50f72925c1 on core-9-0-branch, in main, 9.0: Fix for 1100518fff: Button should stay in -overrelief state after event.
      • 🔧 2026-09-29 check-in ee0599a59d on core-8-6-branch, in main, 9.0, 8.6: Fix for 1100518fff: Button should stay in -overrelief state after event.
      • 💬 2026-09-28 serhiy.storchaka: Still reproducible with 9.1 on X11 and Windows. tk::ButtonUp restores the saved -relief after a click even if the mouse is still in the button, so -overrelief is lost until the mouse leaves and enters the button again. Checkbuttons and radiobuttons are not affected. Proposed fix in branch button-overrelief-after-click, with test button-16.1.
      • 💬 2026-09-29 nemethi: Serhiy, many thanks for fixing this very long-standing bug! The fix is now merged into main, core-9-0-branch, and core-8-6-branch by commits 3aa08341, 50f72925, and ee0599a5. Closing the ticket.
      Full verdict

      REPRODUCED (main, 8.6, 8.5, 8.4)

      • Platform generic (X11, Windows, Aqua all share the logic in library/button.tcl)
      • Tested event generate Enter/Press/Release/Leave on main, 8.6.18, 8.5.19 and 8.4.20 (Xvfb, no WM). All give enter=raised press=sunken release=flat when release=raised is expected.
      • Cause library/button.tcl, ::tk::ButtonUp (all three variants: x11, win32, aqua). It always restores the cached -relief (Priv($w,relief)) and forgets the cache, even when the pointer is still over the button (Priv(window) eq $w).
      • Prototype 1100518fff.patch (57+/6-, 132-line diff; most of it is the same 13-line change repeated in 3 procs). If the pointer is still inside and -overrelief is set, ButtonUp switches to the overrelief and keeps the cached relief so that ButtonLeave can restore it later. New test button-16.1. button.test passes (414/415, 1 skipped). Also checked by hand: press, drag out and release gives flat. Press, drag out, drag back and release gives raised, then Leave gives flat.
      • Suggested ticket comment Still present on 8.4–9.1 (X11 too): tk::ButtonUp restores -relief even when the pointer is still over the button. Prototype fix: switch back to -overrelief in ButtonUp when Priv(window) is the button, with test button-16.1.
      ⬜1096580fff2005soft-hyphen in (legacy) text widgetREPRODUCED (main, 8.6). An unmerged fix branch already exists: bug-1096580fff (fvogel, last check-in 6c6755350a, 2017-02-07, tag soft-hyphen). …

      Open ticket 1096580fff · 18. [text] · filed 2005-01-05

      Full verdict

      REPRODUCED (main, 8.6). An unmerged fix branch already exists: bug-1096580fff (fvogel, last check-in 6c6755350a, 2017-02-07, tag soft-hyphen). I did not redo it.

      • Platform generic
      • Tested main, 8.6.18. -wrap word, Courier 12, text aaaaaaaa abcdefgh­123456 in a 20-char widget. The second display line starts at 1.9, so there is no break at SHY. The SHY bbox is 10 px wide, the same as a full character.
      • Cause tkTextDisp.c, TkTextCharLayoutProc: the break search considers only whitespace, and MeasureChars/CharDisplayProc treat U+00AD like any other character (as described in the ticket).
      • Prototype none (branch bug-1096580fff already implements it. The ticket says the fix is also in revised_text, which is not merged. The next step is to rebase or merge that branch onto main, then review.)
      • Suggested ticket comment Still present on 8.6 and 9.1 with the legacy text widget. Branch [bug-1096580fff] (2017) needs to be updated to main and reviewed.
      ⬜1081603fff2004Arc item anomalies under WindowspatchREPRODUCED (9.1 main, 8.6 tip) for the 1-px endpoint mismatch. The small-extent arcs are no longer full circles (fixed part confirmed).

      Open ticket 1081603fff · 05. Canvas Items · filed 2004-12-08

      Full verdict

      REPRODUCED (9.1 main, 8.6 tip) for the 1-px endpoint mismatch. The small-extent arcs are no longer full circles (fixed part confirmed).

      • Platform Windows
      • Tested not on X11 (the reporter and fvogel say X11 is correct). Repro script: win/1081603fff.tcl (visual yes/no check of the rounded square, fvogel's 0–90 pie, and the small arcs).
      • Cause win/tkWinDraw.c DrawOrFillArc computes the radial endpoints with integer rounding and passes them to GDI Arc/Pie/Chord. GDI's bounding-rectangle convention (right/bottom exclusive) and its endpoint-ray semantics differ from X11, so the endpoints can be off by a pixel.
      • Prototype none (needs Windows to verify pixel-exact output)
      • Suggested ticket comment The full-circle part is fixed via 6051a9fc. The remaining 1-px endpoint mismatch is still unverified; the repro script is in the triage set.
      Windows VM, run 5
      • Platform Windows
      • Verdict REPRODUCED (9.1 main, 8.6 tip) for the 1-px endpoint mismatch. The small-extent arcs are no longer full circles (fixed part confirmed).
      • Tested arcs.tcl captured on Windows and on X11 (Xvfb, main). 9.1 and 8.6 are pixel-identical to each other and differ from X11: 1-px gaps between the arcs and the lines at 9 o'clock (100,150) and 6 o'clock (150,300), the arc ends step off the lines at the other corners, and the red 0–90° pie edge has a kink where it meets the blue line.
      • Cause win/tkWinDraw.c DrawOrFillArc rounds the radial endpoints and uses GDI's Arc/Pie conventions (exclusive right/bottom), as described in result-48.
      • Prototype none
      • Suggested ticket comment The full-circle part is fixed. The 1-px endpoint mismatch is still there on Windows 11 in 8.6.18 and 9.1: gaps and steps where the arcs of a rounded rectangle meet the lines. X11 is exact.
      2026-10-25 prototype, Windows VM
      • Status PROTOTYPED
      • Cause win/tkWinDraw.c:DrawOrFillArc passes radial points to Arc/Pie/Chord that are rounded points on the ellipse, measured from x+width/2. GDI measures the rays from the centre of its rectangle, which is one pixel larger (x+(width+1)/2), so the rays are tilted by up to half a pixel. Depending on the quadrant, an end stops one pixel short (a gap next to the line) or overshoots by one pixel. Arc also never draws its last pixel, like LineTo.
      • Fix compute the radial points from GDI's centre, 32× the size away, so that rounding them no longer tilts the ray. Move both ends inwards by a quarter pixel, clamped to a quarter of the extent so that tiny arcs never become full ellipses, because GDI includes the next pixel when the ray passes exactly through a pixel centre. For solid 1-px GXcopy outlines, also draw the last pixel. 1081603fff.patch (72 changed lines: 36 tkWinDraw.c, 36 test). It applies together with 1055974fff.patch (offsets only).
      • Repro work-win/arcs.tcl (rounded square plus pie) compared with the X11 capture (shots/arcs_x11.png). Before: gaps at 9 o'clock (100,149) and 6 o'clock (149,300). After: all 8 arc/line joints are closed, and the pie start no longer sticks out under the blue line. The rest of the difference from X11 (about 110 px, where the curve leaves the axes) comes from GDI's ellipse rasterisation and is not fixable without our own rasteriser. Tiny extents (0.05°–2°, both directions, arc and pieslice) are still drawn as tiny arcs, not full circles.
      • Test canvas-27.3 "arc ends neither stop short nor overshoot" (8 quarter arcs, testpixel: at each end, the pixel before or at the end is drawn and the one after is not) (fails before: yes, 8 ends on Windows main; passes after: yes; also passes on X11 with the wt-W build)
      • Test files run canvas 176/0, canvRect 57/0, canvPs 14/0, canvText 100/0, canvImg 84/0, canvWind 5/0 with fix; main canvas 175/0, the others identical
      • Windows done. Not checked: dashed and wide arcs, which are visually unchanged in the capture but have no pixel test, and stippled arcs, which are covered by canvas-27.1/27.2 and pass.
      • Suggested ticket comment The 1-px gaps and overshoots come from DrawOrFillArc: its radial points are measured from a centre half a pixel away from the one GDI uses, and GDI also skips the last pixel of Arc. A prototype aims the rays from GDI's centre and draws the last pixel, so arcs now meet the lines of a rounded rectangle exactly (test canvas-27.3). The remaining differences from X11 are GDI's ellipse rasterisation.
      ⬜1055974fff2004-dashofset option doesnt work on ms windows buildpatchREPRODUCED (9.1 main, 8.6 tip): offset 0 and offset 8 give identical captures at widths 1 and 3. The dash lengths are also wrong: {8 8} is drawn as…

      Open ticket 1055974fff · 05. Canvas Items · filed 2004-10-28

      Full verdict

      REPRODUCED (9.1 main, 8.6 tip): offset 0 and offset 8 give identical captures at widths 1 and 3. The dash lengths are also wrong: {8 8} is drawn as the stock PS_DASH pattern at width 1, and as dots at width 3.

      • Platform Windows
      • Tested not run on X11, where -dashoffset goes through XSetDashes and works. Windows repro script: win/1055974fff.tcl. It draws {8 8} dashes with offset 0 and offset 8 at widths 1 and 3, captures the screen through PowerShell, and reports BUG PRESENT if the two captures are identical. It writes the verdict to %TEMP%\tk1055974\verdict.txt and exits.
      • Cause xlib/xgc.c:XSetDashes stores gc->dash_offset, but win/tkWinDraw.c:SetUpGraphicsPort never reads it. It also maps any dash list onto one of the fixed styles PS_DASH, PS_DOT, PS_DASHDOT or PS_DASHDOTDOT, so neither the offset nor the real dash lengths are honoured.
      • Prototype none. The proper fix is ExtCreatePen(PS_GEOMETRIC|PS_USERSTYLE) built from gc->dashes, with the pattern rotated by dash_offset (as nijtmans suggested in 2012). That changes the documented Windows dash behaviour in Tk_GetDash.3 and needs Windows to verify, so it is beyond a quick prototype.
      • Suggested ticket comment Still valid: XSetDashes stores the dash offset, but ''SetUpGraphicsPort'' in tkWinDraw.c ignores it and only picks one of the four stock pen styles. The fix is a PS_USERSTYLE pen made from the real dash list, rotated by the offset.
      Windows VM, run 5
      • Platform Windows
      • Verdict REPRODUCED (9.1 main, 8.6 tip): offset 0 and offset 8 give identical captures at widths 1 and 3. The dash lengths are also wrong: {8 8} is drawn as the stock PS_DASH pattern at width 1, and as dots at width 3.
      • Tested 1055974fff.tcl (act.exe capture) and dashes.tcl: 26 lines with various dash lists and offsets at widths 1 and 3, plus a dashed rectangle and oval, compared against X11.
      • Cause win/tkWinDraw.c SetUpGraphicsPort maps any dash list to PS_DASH/PS_DOT/PS_DASHDOT/PS_DASHDOTDOT and never reads gc->dash_offset.
      • Prototype 1055974fff.patch (76 changed lines, about 40 of them code; also doc/canvas.n and doc/GetDash.3). It builds a PS_USERSTYLE geometric pen from the real dash list, repeated twice when the list has an odd length, as in X11, and rotated by the offset. A 0-length entry is inserted when the pattern would start in an "off" segment. Dashed lines always use a geometric pen, because cosmetic pens count user styles in device style units (3 px on this display) and zero entries made them solid. Lists that would need more than 16 entries (odd lists longer than 7) fall back to the old mapping. Result: all 26 lines are pixel-identical to X11. Only the phase along the rectangle and oval outlines differs slightly. canvas, canvRect, canvPoly, canvLine, canvArc, canvText and winDraw tests: 312 run, 0 failed. No automated test (it needs screen capture).
      • Suggested ticket comment Confirmed on Windows 11 (8.6.18, 9.1). A prototype that builds a PS_USERSTYLE pen from the real dash list, rotated by -dashoffset, makes dashed lines match X11 pixel for pixel. The docs' "only the first 4 patterns on Windows" note would then be dropped.
      Prototype verified 2026-10-24, Windows VM
      • Status VERIFIED+TEST
      • Applies to main clean (offset 401 lines)
      • Repro before: canvas-31.1 gives the same stock pattern for every offset (111111111111111111000000... at width 1, dots at width 3) / after: {8 8} dashes are exact and rotated by -dashoffset, the same as X11
      • Test canvas-31.1 (new; checks the dash list and -dashoffset with testpixel at widths 1 and 3; fails before: yes; passes after: yes; also passes on X11, Xvfb build of wt-W)
      • Test file run canvas.test 176/0, canvRect.test 57/0, canvText.test 100/0 (pre-existing failures: none)
      • Patch changes added test canvas-31.1. Code reviewed: seg[20] is enough (XSetDashes keeps at most 10 entries), and the PS_USERSTYLE array stays within 16 entries.
      ✅991849ffff2004removal/optimization of scrollbar events?patchREPRODUCED (main, 8.6, 8.4). This is closer to an RFE than a bug: the final callback is always correct.

      Open ticket 991849ffff · 48. Geometry Management · filed 2004-07-15

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-09-28 check-in 48830cec09 on scroll-command-unmapped, in main, 9.0: Fix [991849]: removal/optimization of scrollbar events?
      • 🔧 2026-10-05 check-in aa35622f95 on trunk, in main, 9.0: Fix for 991849ffff: removal/optimization of scrollbar events?
      • 🔧 2026-10-05 check-in fd492feb2b on trunk, in main, 9.0: changes.md: Bug [991849]: new text: "canvas,text,ttk widgets: call scrollbar commands after window mapping (and not with 1x1 initial size)"
      • 🔧 2026-10-05 check-in 9ec29de97d on core-9-0-branch, in main, 9.0, 8.6: Fix for 991849ffff: canvas,text,ttk widgets: call scrollbar commands after window mapping (and not with 1x1 initial size)
      • 🔧 2026-10-05 check-in d23599f129 on core-8-6-branch, in main, 9.0, 8.6: Fix for 991849ffff: canvas,text,ttk widgets: call scrollbar commands after window mapping (and not with 1x1 initial size)
      • 💬 2026-09-28 serhiy.storchaka: Still reproducible with 9.1: the canvas and the text call the scroll commands before the window is mapped, when it has size 1x1 (e.g. 0.004 0.004 for a canvas whose scroll region fits in it). A ttk::treeview does the same if xview or yview is called before mapping. The listbox postpones the call until the window is mapped. Proposed fix in branch scroll-command-unmapped: the canvas, the text and ttk scrollable widg…
      • 💬 2026-10-05 nemethi: Serhiy, many thanks for fixing this very long-standing bug! Merged into main, core-9-0-branch, and core-8-6-branch by commits aa35622f, 9ec29de9, and d23599f1.
      Full verdict

      REPRODUCED (main, 8.6, 8.4). This is closer to an RFE than a bug: the final callback is always correct.

      • Platform generic
      • Tested main, 8.6 tip, 8.4.20 on Xvfb without a WM. canvas -xscrollcommand fires 0.002 0.002 while winfo width is 1 and winfo ismapped is 0, then fires the correct value after mapping. The text widget behaves the same way. Listbox and entry do not.
      • Cause generic/tkCanvas.c:DisplayCanvas. The !Tk_IsMapped early exit jumps to done:, which still calls CanvasUpdateScrollbars with the 1x1 size the canvas has before geometry management.
      • Prototype 991849ffff.patch (about 25 changed lines, test canvas-24.1). While the canvas is unmapped it keeps UPDATE_SCROLLBARS pending. The update runs on the redraw after Map/Expose. canvas, canvText, canvWind and scrollbar tests pass. Behaviour change: a canvas that is never mapped no longer calls its scroll commands. xview/yview still return correct values. The text widget (tkTextDisp.c) would need a similar change, which is not done.
      • Suggested ticket comment Still happens in 9.1 and 8.6: the canvas reports its scroll position while it is still unmapped and 1x1. Branch prototype: skip the -x/-yscrollcommand update until the canvas is mapped (test canvas-24.1). The text widget does the same.
      ⬜923174ffff2004Can’t catch loss of X server displaying slave toplevelREPRODUCED (main)

      Open ticket 923174ffff · 67. Unix Window Operations · filed 2004-03-25

      Full verdict

      REPRODUCED (main)

      • Platform X11
      • Tested main fb166f9f73, no WM. I killed the X server under a running wish: it died with "XIO: fatal IO error 25 ... X connection to :145 broken", and neither the WM_DELETE_WINDOW handler nor the <Destroy> binding ran. The second-display case needs a second X server, which I may not start. In main, generic/tkWindow.c:1747 still has the display close disabled (#if !defined(_WIN32) && defined(NOT_YET)), so serhiy's test of 2026-09-23 still applies.
      • Cause Xlib's fatal IO error handler must not return, so Xlib calls exit(). Tk also never closes a display before exit (NOT_YET block in Tk_DestroyWindow), and TkCloseDisplay does not release fonts, colors, cursors, borders or bitmaps.
      • Prototype none. This needs per-display resource cleanup in TkCloseDisplay, which is too large, plus XSetIOErrorExitHandler (libX11 >= 1.7) for an orderly shutdown. serhiy's 2026-09-23 comment already describes the deferred-close experiment.
      • Suggested ticket comment Still present in main as of 2026-10-22: killing the X server takes the application down with the XIO error, and no Tcl-level handler runs. See bf65b83d43 for the case where the window was already destroyed.
      💬919066ffff2004Slow processing irregular transparenciesREPRODUCED (main, 8.6)

      Open ticket 919066ffff · 41. Photo Images · filed 2004-03-18

      💬 Since triage (updated 2026-11-11)

      Status now: Open (Remind)

      • 💬 2026-10-02 chrstphrchvz: A quick comment in case others are about to look into this issue. The suggested alternative implementation I originally linked to in my comment from 2023-07-21 is currently better viewed from [https://github.com/chrstphrchvz/tk/commits/photo-919066/] . But it involved clipping to a mask pixmap using XSetClipMask(). And it turns out that, in general, neither of the Win32 and Aqua Tk ports support clipping to a pixm…
      Full verdict

      REPRODUCED (main, 8.6)

      • Platform generic
      • Tested main, 8.6 tip on Xvfb. Tiling a 20x20 image with scattered transparent pixels via copy -to: 300x200 takes 0.25 s, 450x300 takes 1.6 s. 2.25× the pixels costs 6.4× the time, so the cost grows much faster than the image size. Both versions give the same numbers.
      • Cause generic/tkImgPhoto.c:Tk_PhotoPutBlock / TkpBuildRegionFromAlphaData. The valid region is built from alpha data one rectangle at a time with XUnionRectWithRegion, and each union costs time in proportion to the region's current size.
      • Prototype none. Too large and still a design decision. chrstphrchvz lists three approaches in the ticket, with a draft of approach 3 (compute the valid region lazily, only when TkPhotoGetValidRegion() is called, then maintain it). It touches about 30 validRegion uses in tkImgPhoto.c and tkImgPhInstance.c.
      • Suggested ticket comment Still present in 9.1 and 8.6. The lazy valid-region approach looks like the way forward and needs a decision on TkPhotoGetValidRegion().
      ⬜917631ffff2004Grid information in wm may be incorrect (patch included)REPRODUCED (main, 8.6). This is a design limitation.

      Open ticket 917631ffff · 48. Geometry Management · filed 2004-03-16

      Full verdict

      REPRODUCED (main, 8.6). This is a design limitation.

      • Platform generic (X11 tested)
      • Tested main, 8.6 tip on Xvfb without a WM. A listbox -setgrid 1 -height 10 and a text widget are packed with -expand 1 -fill both. After wm geometry . 20x20, wm geometry reports 20x20 grid units, but the listbox only shows about 15 lines because the text widget takes part of the extra space.
      • Cause Tk_SetGrid (tkUnixWm.c UpdateSizeHints, and the same on Windows and Aqua) derives the base size from requested sizes. The code assumes the gridded widget absorbs all size changes, and wm.n does not say so.
      • Prototype none. The attached patch adds a new public API (Tk_UpdateGrid) and changes UpdateSizeHints on all three platforms, so it is too large and changes the API. fvogel reported it did not help on Windows. Related: 3466588fff, 7c51624a96. A cheap alternative is a wm.n note that the gridded widget should be the only one that expands.
      • Suggested ticket comment Still present: gridding assumes the gridded widget gets all extra space. The proposed Tk_UpdateGrid needs new API (a TIP). Meanwhile, documenting this assumption in wm.n (GRIDDED GEOMETRY MANAGEMENT) would help.
      ✅873655ffff2004Missing event when posting cascade menuspatchREPRODUCED (main, 8.6, 8.4)

      Open ticket 873655ffff · 12. Unix Menus · filed 2004-01-09

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-10-01 check-in dce65730e7 on cascade-grab-leave, in main, 9.0: Fix 873655ffff: missing <Leave> event when posting cascade menus.
      • 🔧 2026-10-15 check-in 44e2460267 on trunk, in main, 9.0, 8.6: Fix 873655ffff: missing <Leave> event when posting cascade menus
      • 🔧 2026-10-15 check-in a0aec1c373 on core-9-0-branch, in main, 9.0: Fix 873655ffff: missing <Leave> event when posting cascade menus
      • 🔧 2026-10-17 check-in 20e2b20246 on core-8-6-branch, in main, 9.0, 8.6: Fix 873655ffff: missing <Leave> event when posting cascade menus
      • 💬 2026-10-01 serhiy.storchaka: Still reproducible in 8.6 and 9.x on X11. When tk_popup sets the grab, Tk generates a event for the button, but it is processed only while postcascade waits for the cascade menu to be mapped. The grab is in effect by then, so TkPointerEvent drops the event as outside the grab. This is why "after idle .m postcascade 1" works around it. Proposed fix in branch [cascade-grab-leave]: events generated by a grab change a…
      • 💬 2026-10-17 nemethi: Merged by Jan into main and core-9-0-branch by commits 44e24602 and a0aec1c3; backported to core-8-6-branch by commit 20e2b202. Closing the ticket.
      Full verdict

      REPRODUCED (main, 8.6, 8.4)

      • Platform X11
      • Tested main, 8.6 tip, 8.4.20, no WM. Script work-a1/casc.tcl warps the pointer onto the button and synthesizes a click. Without postcascade the button gets <Leave> NotifyGrab and goes back to normal. With .m postcascade in the same command it gets no <Leave> and stays active.
      • Cause generic/tkGrab.c. Tk_Grab queues synthesized Leave events (MovePointer2) and then the grab-window-change event. The Leave events do reach Tk_HandleEvent (checked with instrumentation), but TkPointerEvent drops them because by then grabWinPtr is already .m and .b is outside the grab tree. Probably the grab-change event is serviced while mapping .m.c waits for its map event. The restrict proc defers window events during that wait but not the grab-change event. I did not confirm this last step.
      • Prototype 873655ffff.patch (37 lines, test menu-42.1). TkPointerEvent no longer applies grab-tree filtering to crossing events that carry GENERATED_GRAB_EVENT_MAGIC. The new test fails without the fix and passes with it. grab, bind, event, menu, menubut, menuDraw, button, entry and unixEmbed tests pass. focus-2.6/2.7/2.9/5.1 and window-2.4..2.9 also fail on unpatched main, so they are not caused by this.
      • Suggested ticket comment Still present on X11 (main, 8.6). The Leave events that Tk synthesizes on grab are dropped in TkPointerEvent when the grab window becomes current before they are processed. Fix prototype with test menu-42.1.
      ⬜842504ffff2003embedded windows (in text/canvas) are not croppedREPRODUCED (main, 8.6, 8.4)

      Open ticket 842504ffff · 48. Geometry Management · filed 2003-11-15

      Full verdict

      REPRODUCED (main, 8.6, 8.4)

      • Platform generic (tested on X11)
      • Tested main, 8.6, 8.4.20, pixel checks. A canvas with -bd 10 and a 200x200 window item: the red child covers the canvas border at the right, the bottom and the corner. A text widget with -bd 10 and a 300px-wide embedded frame: the frame covers the right border on main.
      • Cause tkCanvWind.c and tkTextWind.c (EmbWinDisplayProc/EmbWinBboxProc) place the child with its full requested size and never clip it to the widget's inner area (inside bd and highlightthickness). X child windows overlap the parent's border. Same class of issue as panedwindow bug 795869 and duplicate cb21f579dd.
      • Prototype none. A fix means clipping each embedded window to the inner area (shrinking or unmapping it, or clipping through an intermediate window). That is a behaviour change in two widgets and needs design.
      • Suggested ticket comment Still present in main and 8.6 for both canvas and text. Embedded windows would have to be clipped to the area inside borderwidth and highlightthickness. See also cb21f579dd.
      ⬜787402ffff2003No KP_Enter on WindowspatchREPRODUCED (9.1 main): real keypad Enter (extended VK_RETURN) gives %K Return. event generate <KP_Enter> gives ?? with keycode 0.

      Open ticket 787402ffff · 68. Win Window Operations · filed 2003-08-12

      Full verdict

      REPRODUCED (9.1 main): real keypad Enter (extended VK_RETURN) gives %K Return. event generate <KP_Enter> gives ?? with keycode 0.

      • Platform Windows
      • Tested code reading of win/tkWinKey.c. KeycodeToKeysym uses EXTENDED_MASK to tell Control_R from Alt_R, but not VK_RETURN, so keypad Enter maps to XK_Return. TkpSetKeycodeAndState cannot map KP_Enter either, so event generate <KP_Enter> gets keycode 0. On X11 the new test (Return then KP_Enter) passes.
      • Cause win/tkWinKey.c KeycodeToKeysym. There is no case VK_RETURN that checks EXTENDED_MASK.
      • Prototype 787402ffff.patch (27 lines, test bind-38.1). Returns XK_KP_Enter for extended VK_RETURN, and maps KP_Enter to VK_RETURN+EXTENDED_MASK in TkpSetKeycodeAndState. Not compiled or tested on Windows. Compatibility cost: <Return> bindings stop firing for keypad Enter on Windows, and Text/Entry insertion changes, because the library binds <KP_Enter> {# nothing}. That is the same behaviour X11 already has.
      • Repro win/787402ffff.tcl (checks event generate and asks for real key presses; prints a VERDICT line)
      • Suggested ticket comment Still present per code reading. Keypad Enter is reported as Return. A prototype returns KP_Enter for extended VK_RETURN, like X11 and macOS, but this changes behaviour for existing <Return> bindings.
      Windows VM, run 5
      • Platform Windows
      • Verdict REPRODUCED (9.1 main): real keypad Enter (extended VK_RETURN) gives %K Return. event generate <KP_Enter> gives ?? with keycode 0.
      • Tested 787402ffff_auto.tcl (act.exe key enter then ekey 0D into a focused text widget).
      • Cause win/tkWinKey.c KeycodeToKeysym ignores EXTENDED_MASK for VK_RETURN (see result-51).
      • Prototype 787402ffff.patch, unchanged. It compiles cleanly with MSVC. With it: real keys give Return KP_Enter, and event generate <KP_Enter> gives KP_Enter, keycode 13. As on X11, keypad Enter no longer inserts a newline in text/entry (class binding <KP_Enter> {# nothing}). bind, key, winKey, entry, text and spinbox tests: 1878 run, 0 failed (bind-38.1 passes; the ref build fails it).
      • Suggested ticket comment Confirmed on Windows 11 with 9.1: keypad Enter is reported as Return. The prototype (extended VK_RETURN → KP_Enter) works, and the tests pass. As on X11, <Return> bindings no longer fire for the keypad key.
      Prototype verified 2026-10-24, Windows VM
      • Status VERIFIED+TEST
      • Applies to main clean
      • Repro before: Return ?? / after: Return KP_Enter
      • Test bind-38.1 (existing; fails before: yes; passes after: yes)
      • Test file run bind.test 559/0, event.test 42/0 (pre-existing failures: none)
      • Patch changes none
      Prototype verified 2026-10-24, Windows VM, generic patches
      • 3119824fff: FAILS ON WINDOWS: text-40.1 errors before and after with no keycode for keysym "KP_Enter" (event generate -keysym KP_Enter); the repro also gives no newline. Windows has no KP_Enter keysym until 787402ffff (phase 1, bind-38.1) is applied. The patch is unchanged; commit it after 787402ffff, or give text-40.1 a constraint.
      ⬜770578ffff2003long text item won’t scrollpatchREPRODUCED (main fb166f9f73, 9.0.5, 8.6.18; both Xft and --disable-xft builds)

      Open ticket 770578ffff · 05. Canvas Items · filed 2003-07-13

      Full verdict

      REPRODUCED (main fb166f9f73, 9.0.5, 8.6.18; both Xft and --disable-xft builds)

      • Platform X11 (generic canvas code)
      • Tested Xvfb, no WM. Test script with a non-periodic text ("00000|00001|..." in Courier 12) plus a short red reference item placed where the visible word should be. Scroll offsets 20000/40000/70000/140000. Up to about 32k the text is correct. At 40000 the long item shows 00547 where 00668 belongs, so it is shifted by 40000-32768 px. This happens on installed main, 9.0.5, 8.6.18 and the wt build, Xft and core fonts alike. The report's "bug!_" string repeats every 5 characters, which hides the shift; that is why the 2026-09-22 comment saw it as working with Xft.
      • Cause generic/tkCanvText.c:DisplayCanvText() gets the text origin from Tk_CanvasDrawableCoords(), which clamps it to a short. Once the item starts more than 32768 px left of the drawable, the whole layout is drawn from -32768. Separately, with core X fonts unix/tkUnixFont.c:Tk_DrawChars() passes x to XDrawString() as INT16, so a run starting below -32768 wraps around and could be drawn in the wrong place.
      • Prototype 770578ffff.patch (104 changed lines, mostly comments and two small static helpers). DisplayCanvText computes the origin as an unclamped int (DrawableCoord()), and the cursor/selection polygon points are clamped with ClampShort() instead of a wrapping (short) cast. Tk_DrawChars skips the XDrawString* call (but still advances x) for runs starting outside the 16-bit range. Checked by screenshot: the item is correct at all four offsets in both Xft and core-font builds. canvText.test, canvas.test, font.test and unixFont.test pass on the Xft build. The core-font build gives the same 3 unixfont-2.x failures with and without the patch (font availability). No new test: it would need pixel checks on a 40000-px item.
      • Suggested ticket comment The text is not only clipped but drawn shifted past 32k: DisplayCanvText() takes its origin from Tk_CanvasDrawableCoords(), which clamps to a short (the report's periodic string hides this). A prototype computes the origin without clamping and, for core X fonts, does not send runs starting below -32768 to XDrawString().
      Prototype verified 2026-10-24, Linux
      • Status VERIFIED
      • Applies to main clean
      • Repro new pixel-comparison script work-a3/p770.tcl (Courier 12 "00000|00001|..." canvas scrolled by S px, compared with a short reference item; count of differing pixels): before S=20000:0, 40000:1030, 70000:880, 140000:1051 / after all 0 (Xft build)
      • Test none (needs a pixel comparison on a canvas scrolled by more than 32k px; no portable read-back)
      • Test file run canvText.test 100/0, canvas.test 175/0, font.test 304/0, unixFont.test 50/0
      • Patch changes none. Review: DrawableCoord duplicates the static RoundCoord from tkCanvUtil.c, which is acceptable. Skipping core-font runs that start below SHRT_MIN is safe because Tk_DrawChars splits runs at about 200 chars, so a skipped run never reaches the visible area. Runs starting above SHRT_MAX are not skipped (asymmetric), but they wrap to positions far to the left, which are off-screen anyway. The core-font build was not rebuilt here (triage checked it).
      Prototype verified 2026-10-24, Windows VM, generic patches
      • 770578ffff: generic part builds; canvText 102/0 (the unix/tkUnixFont.c part does not apply to Windows).
      ⬜742773ffff200310,000 widgets leaves some undisplayedpatchREPRODUCED (9.1 main: 113 frames; 8.6 tip: 110 frames with winfo id 0)

      Open ticket 742773ffff · 48. Geometry Management · filed 2003-05-24

      Full verdict

      REPRODUCED (9.1 main: 113 frames; 8.6 tip: 110 frames with winfo id 0)

      • Platform Windows
      • Tested pack version, main + 8.6 on X11 (Xvfb, no WM): all 10000 frames mapped and placed correctly.
      • Cause (likely): Windows limits each process to 10000 USER objects by default (USERProcessHandleQuota). 80x125 frames plus their parents and the toplevel go over that. TkpMakeWindow (win/tkWinWindow.c) doesn't check whether CreateWindowExW() returned NULL, so the extra widgets get a NULL HWND and draw through the desktop DC. That's why they show up in the top-left corner of the screen. At most Tk could report an error; it can't create more windows than the OS allows.
      • Repro win/742773ffff.tcl (counts widgets whose winfo id is 0)
      • Prototype none (OS limit; only improvement would be an error from TkpMakeWindow on a NULL HWND)
      • Suggested ticket comment Not reproducible on X11. On Windows this is most likely the per-process limit of 10000 USER objects: the extra child windows can't be created. Tk doesn't check the result of CreateWindowExW(), so they end up drawing on the desktop.
      Windows VM, run 5
      • Platform Windows
      • Verdict REPRODUCED (9.1 main: 113 frames; 8.6 tip: 110 frames with winfo id 0)
      • Tested 742773ffff.tcl plus a full-screen capture: the frames without a window paint red at the desktop origin (0,0).
      • Cause the per-process USER object quota (10000). CreateWindowExW in TkpMakeWindow returns NULL, which is not checked. TkWinGetDrawableDC then calls GetDC(NULL), the DC of the whole screen.
      • Prototype 742773ffff.patch (7 lines, win/tkWinDraw.c). TkWinGetDrawableDC returns a NULL DC for a window without an HWND, so drawing into it does nothing. After the fix the desktop corner stays clean. frame, winDraw, canvas, button and winfo tests: 856 run, 0 failed. It fits with a58cda72fe.patch (same function, NULL twdPtr). A Tcl error is not possible without an error path from Tk_MakeWindowExist.
      • Suggested ticket comment Confirmed on Windows 11 (8.6.18, 9.1): past about 9900 windows, CreateWindowExW fails because of the USER object quota, and the windowless widgets draw on the screen DC. A small fix stops the drawing on the desktop; the OS limit remains.
      Prototype verified 2026-10-24, Windows VM
      • Status VERIFIED
      • Applies to main clean. With a58cda72fe it applies cleanly in either order (fuzz 2 if 742773 goes first).
      • Repro 10000 frames give 111 without an HWND. Before: 25 red pixels in the 40x40 desktop corner. After: 0.
      • Test none (exhausting the per-process USER quota with ~10000 windows would break the rest of the test process)
      • Test file run B2: frame 207/0, canvas 176/1 (canvas-31.1 is the test of the unapplied 1055974, expected), button 416/0, winfo 76/0
      • Patch changes none
      ✅680660ffff2003Menu items accidentally selectedpatchREPRODUCED (main, 8.6, 8.4), both menubutton and ttk::menubutton

      Open ticket 680660ffff · 01. Bindings · filed 2003-02-05

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-09-29 check-in c81290221a on menu-post-under-pointer, in main, 9.0: Fix [470331], [680660], 1596662fff, f9d316a5e0: a menu posted over its menubutton or menubar entry invoked the entry under the pointer on the release of the
      • 🔧 2026-09-30 check-in 0f86666370 on trunk, in main, 9.0: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
      • 🔧 2026-09-30 check-in 5c47ae3e9c on core-9-0-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
      • 🔧 2026-09-30 check-in 3aacabb15f on core-8-6-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
      • 🔧 2026-09-30 check-in a33031d2f8 on menu-post-under-pointer, in main, 9.0: Add a widget demo for popup (context) menus posted with tk_popup, with menus of different sizes.
      • 💬 2026-09-30 serhiy.storchaka: The same problem as [470331]. Proposed fix in branch menu-post-under-pointer: the release of the click which posted the menu over the menubutton is ignored while the pointer has not moved, for menubutton and ttk::menubutton. A menu which does not fit below the menubutton is posted above it, or beside the pointer. See [470331] for details.
      • 💬 2026-09-30 nemethi: Serhiy, many thanks! The fix is now in main, core-9-0-branch, and core-8-6-branch; see commits 0f866663, 5c47ae3e, and 3aacabb1.
      Full verdict

      REPRODUCED (main, 8.6, 8.4), both menubutton and ttk::menubutton

      • Platform X11
      • Tested Xvfb with no WM, real XTest press/release on a menubutton whose 60-item menu is taller than the screen, so it is posted over the button. A quick click invokes the item under the pointer (main, 8.6.18, 8.4.20; ttk too). With a small menu that doesn't overlap, click-move-click works.
      • Cause library/menu.tcl tk::MenuInvoke. The ButtonRelease of the click that posted the menu goes to the menu when the menu is under the pointer, and gets treated as a selection.
      • Prototype 680660ffff.patch (23 code lines in menu.tcl and ttk/menubutton.tcl, plus tests menubutton-11.1/11.2 in menubut.test). MbPost (when posting from a button press) and ttk Pulldown save the pointer position in Priv(postPointer). MenuInvoke ignores the first release if the pointer hasn't moved (less than 3 px) and keeps the menu posted; for ttk it calls TransferGrab. MenuUnpost and MbButtonUp clear the value. Checked: click, move, click works for big and small menus, for Tk and ttk; press-drag-release still selects. menu.test and menubut.test pass; the new tests fail without the fix. Side effect: a quick click on a tk_optionMenu now leaves the menu posted instead of re-selecting the current value.
      • Suggested ticket comment Still present in 8.6 and 9.x (also for ttk::menubutton) when the menu is posted over the menubutton, e.g. a menu taller than the screen. The release of the posting click now doesn't invoke an entry if the pointer hasn't moved.
      ⬜657656ffff2002-initialfile doesn't hilight filenamespatchREPRODUCED (main, 8.6)

      Open ticket 657656ffff · 37. [tk_get*File] · filed 2002-12-22

      Full verdict

      REPRODUCED (main, 8.6)

      • Platform X11 (Tk script file dialog)
      • Tested tk_getOpenFile -initialdir d -initialfile b.txt on main and 8.6.18: the entry shows b.txt but no icon is selected. The multi-file part of the request (-initialfile as a list) was rejected by dkf; see TIP 141.
      • Cause library/tkfbox.tcl ::tk::dialog::file::Update fills the icon list but never selects data(selectFile).
      • Prototype 657656ffff.patch (11 code lines, test filebox-7.3-0 in filebox.test). After filling the file list, Update selects the icon that matches data(selectFile), sets the anchor and scrolls it into view. filebox.test and choosedir.test pass; the new test fails without the fix.
      • Suggested ticket comment Still present. The Unix file dialog now highlights and scrolls to the -initialfile entry. Multiple initial files stay out of scope (-initialfile is a single name).
      Prototype verified 2026-10-24, Linux
      • Status VERIFIED+TEST
      • Applies to main clean
      • Repro tk_getOpenFile -initialdir/-initialfile: before no icon selected / after the -initialfile icon is selected, anchored and scrolled into view
      • Test filebox-7.3-0 (existing in the patch; fails before: yes; passes after: yes). It uses after 100 because the dialog is modal: the check has to run inside the dialog's event loop, the same way ToPressButton does, so pause cannot be used here.
      • Test file run filebox.test 117/0 (unpatched 117/1, only 7.3-0), choosedir.test 14/0
      • Patch changes none
      Prototype verified 2026-10-24, Windows VM, generic patches
      • 657656ffff: filebox-7.3 is x11-only (Tk dialog), skipped; filebox.test 63/0.
      ⬜601954ffff2002destroying menubuttons leaves them unpostablepatchREPRODUCED (main, 8.6, 8.5)

      Open ticket 601954ffff · 10. Generic Menus · filed 2002-08-29

      Full verdict

      REPRODUCED (main, 8.6, 8.5)

      • Platform generic (tested on X11)
      • Tested main, 8.6 tip, 8.5 on Xvfb with icewm. Post the menu via the menubutton, destroy it, recreate it: the new menubutton never posts again (tk::Priv(postedMb) still names the old .mb, so tk::MbPost returns early).
      • Cause library/menu.tcl: nothing clears tk::Priv(postedMb) when the posted menubutton is destroyed, and tk::MbPost returns early when $w eq $Priv(postedMb).
      • Prototype 601954ffff.patch (33 lines, test menubutton-12.1). Adds a bind Menubutton <Destroy> class binding that resets Priv(postedMb) and restores Priv(focus). menubut.test passes; the new test fails without the fix.
      • Suggested ticket comment Still happens on 9.1/8.6: tk::Priv(postedMb) keeps naming the destroyed menubutton, so a new one with the same name can never post. Branch fixes it with a Menubutton <Destroy> class binding, test menubutton-12.1.
      Prototype verified 2026-10-24, Linux
      • Status VERIFIED+TEST
      • Applies to main ported (the menubut.test hunk was rejected because menubutton-12.x through 14.x now exist on main)
      • Repro post via tk::MbPost, destroy .mb, recreate, post again: before the menu is not mapped (Priv(postedMb) is stale) / after it is mapped
      • Test menubutton-15.1 (from the patch, renumbered from 12.1 and moved after 14.2, title gets "- bug 601954ffff"; fails before: yes; passes after: yes)
      • Test file run menubut.test 118/0 (unpatched: 118/1, only 15.1), menu.test 574/0
      • Patch changes test renumbered and moved; the menu.tcl change is unchanged. Original kept as 601954ffff.patch.1024.
      Prototype verified 2026-10-24, Windows VM, generic patches
      • 601954ffff: CHANGED. menubutton-15.1 hangs on Windows before and after: tk::MbPost posts a native menu, which runs a modal loop. Added -constraints x11. With that, menubut.test is 118/0 (15.1 skipped). The <Destroy> binding was not checked on Windows.
      ✅600313ffff2002Several problems with tkfbox.tclpatchREPRODUCED (main, 8.6). Only the FocusOut item is left open; dkf fixed the others in 2005.

      Open ticket 600313ffff · 37. [tk_get*File] · filed 2002-08-26

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-10-23 check-in 1994866fa3 on trunk, in main, 9.0: Fix [600313]: the icon list of the file dialog no longer clears the selection when it loses focus
      • 🔧 2026-10-26 check-in b2ea87aa90 on core-9-0-branch, in main, 9.0: Fix [600313]: the icon list of the file dialog no longer clears the selection when it loses focus
      • 🔧 2026-09-28 check-in 37c17b63eb on iconlist-keep-selection, in main, 9.0: Shift-double-click in the icon list no longer invokes the file dialog. Add a test for mixed Control-click and Shift-click.
      • 💬 2026-09-28 serhiy.storchaka: Proposed fix in branch iconlist-keep-selection: * The icon list no longer clears the selection on FocusOut (added together with -multiple in 6d6e691ec6). The Motif dialog and other file dialogs keep the selection too. * Control-double-click and Shift-double-click no longer invoke the dialog, as in the KDE file dialog and Windows Explorer. Two quick Control-clicks on the same item toggled it twice and then opened…
      • 💬 2026-10-23 jan.nijtmans: Fixed on 1994866f If CI build is clean, can be merged further.
      • 💬 2026-10-26 oehhar: Merged to core-9-0-branch by b2ea87aa. Closing. Thanks, Harald
      Full verdict

      REPRODUCED (main, 8.6). Only the FocusOut item is left open; dkf fixed the others in 2005.

      • Platform X11
      • Tested main, 8.6 on Xvfb with icewm: select 2 files in tk_getOpenFile -multiple, focus -force another toplevel. The icon-list selection is cleared (the entry keeps the names). A later Ctrl-click then starts a new selection, so the earlier picks are lost. The same happens when a focus-follows-mouse WM moves focus away.
      • Cause library/iconlist.tcl method FocusOut always does $w selection clear 0 end, even when focus leaves the dialog's toplevel entirely.
      • Prototype 600313ffff.patch (48 lines, test fileDialog-3.1 in filebox.test). The selection is cleared only when the new focus is in the same toplevel, e.g. Tab to the entry, which keeps the Motif-like "entry is the selection" behaviour. filebox.test passes; the new test fails without the fix.
      • Suggested ticket comment The remaining FocusOut issue still exists: the icon list drops its selection whenever the dialog loses focus. Branch clears it only when focus moves within the dialog, test fileDialog-3.1.
      ✅546607ffff2002problem with cursor changeREPRODUCED (9.1 main); FIXED on branch win-foreign-cursor (manual check)

      Open ticket 546607ffff · 68. Win Window Operations · filed 2002-04-20

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Fixed)

      • 🔧 2026-10-23 check-in e836917d7d on trunk, in main, 9.0: Fix 546607ffff: Do not reset the cursor over a window not managed by Tk
      • 🔧 2026-10-29 check-in fef2f0d039 on core-9-0-branch, in main, 9.0, 8.6: Fix 546607ffff: Do not reset the cursor over a window not managed by Tk
      • 🔧 2026-10-29 check-in 3602edd9b5 on core-8-6-branch, in 8.6: Fix 546607ffff: Do not reset the cursor over a window not managed by Tk
      • 🔧 2026-09-23 check-in a283eb2a52 on win-foreign-cursor, in main, 9.0: Do not reset the cursor over a window not managed by Tk.
      • 💬 2026-10-22 serhiy.storchaka: Confirmed on Windows 11 with current main: over a window which Tk does not own, inside a frame with -cursor watch, the cursor turns into the arrow as soon as the mouse stops. With branch win-foreign-cursor the frame cursor stays.
      • 💬 2026-10-23 jan.nijtmans: Fixed on e836917d If CI build is clean, can be merged further.
      • 💬 2026-10-29 nemethi: Backported to core-9-0-branch and core-8-6-branch by commits fef2f0d0 and 3602edd9. Closing the ticket.
      Full verdict

      REPRODUCED (9.1 main); FIXED on branch win-foreign-cursor (manual check)

      • Platform Windows
      • Tested Linux not applicable (Windows mouse timer code path).
      • Cause win/tkWinPointer.c:Tk_PointerEvent → Tk_UpdatePointer(NULL) → UpdateCursor(NULL) → TkpSetCursor(None), which sets IDC_ARROW over a child HWND that Tk does not own. This repeats on every 250 ms mouse-timer tick.
      • Prototype none (already on branch win-foreign-cursor). Repro script: win/546607ffff.tcl. PowerShell creates an EDIT child (I-beam cursor) inside a Tk frame, keeps the mouse still over it, counts arrow and I-beam samples, and prints a verdict.
      • Suggested ticket comment The fix is on branch [/timeline?r=win-foreign-cursor|win-foreign-cursor], awaiting review and merge.
      Windows VM, run 6
      • Platform Windows
      • Verdict REPRODUCED (9.1 main); FIXED on branch win-foreign-cursor (manual check)
      • Tested The PowerShell EDIT child was created, but all 40 cursor samples were neither the I-beam nor the arrow, so the script could not decide. Same on branch win-foreign-cursor (40 of 40 "other"). winWm.test, wm.test and cursor.test pass on the branch (430 tests, 0 failed). Manual check by serhiy on the VM desktop (2026-10-23, win/546607ffff-manual.tcl: a frame with -cursor watch holding a foreign EDIT child): on main the busy cursor turns into the arrow while the mouse rests on the foreign window; on win-foreign-cursor it stays. (The first version of the check expected the EDIT I-beam, which can never show: TkWinChildProc answers every WM_SETCURSOR, also the one the child forwards to its parent.)
      • Prototype none
      • Suggested ticket comment Confirmed on Windows 11 with current main: over a window which Tk does not own, inside a frame with -cursor watch, the cursor turns into the arrow as soon as the mouse stops. With branch win-foreign-cursor the frame cursor stays.
      ⏳531658ffff2002ButtonRelease on Grab WindowREPRODUCED on X11 (main, 9.0, 8.6, 8.5, 8.4) with eserte's example (grab set in the <ButtonPress-1> binding). The original Windows case (grab set b…

      Open ticket 531658ffff · 69. Events · filed 2002-03-19

      ⏳ Since triage (updated 2026-11-11)

      Status now: Pending (Works for Me)

        Full verdict

        REPRODUCED on X11 (main, 9.0, 8.6, 8.5, 8.4) with eserte's example (grab set in the <ButtonPress-1> binding). The original Windows case (grab set before the press) is not reproduced: serhiy on Win11, and on X11 it goes through Note 6.

        • Platform generic (original report Windows; eserte's variant is X11)
        • Tested main (fb166f9f73), 9.0, 8.6 tip, 8.5, 8.4.20 on Xvfb+twm, with real XTest events: press on .b, drag to the root window, release. "release" is never printed, .b stays sunken and grab current is still "."; the ButtonRelease goes to "." (bind . <ButtonRelease-1> sees it at 598,571).
        • Cause generic/tkGrab.c. When a local grab is set while a button is down, Tk_Grab calls ReleaseButtonGrab (buttonWinPtr = NULL) and turns the grab into a temporary global XGrabPointer on the grab window, which replaces the X implicit grab of the pressed window. TkPointerEvent then delivers the release to the grab window by design ("Note 5": "forget about the auto-grab for the button press"), not to the window that got the press.
        • Prototype none. A fix would set dispPtr->buttonWinPtr to the pressed window in Tk_Grab when buttons are down. That needs the press window recorded outside grabs, and it changes where menus see the release (tk::MbPost/tk_popup set grabs with the button down), so it is a behaviour change to review carefully, not a quick patch.
        • Suggested ticket comment The Windows case from the report no longer reproduces. eserte's example (grab set from the ButtonPress binding) still loses the ButtonRelease on X11 in all versions: the release goes to the grab window, because Tk_Grab() deliberately forgets the button's implicit grab (tkGrab.c, Note 5). Workaround: also bind <ButtonRelease-1> on the grab window.
        ⬜528330ffff2002geometry X resource ignoredpatchREPRODUCED (main, 8.6, 8.4)

        Open ticket 528330ffff · 48. Geometry Management · filed 2002-03-11

        Full verdict

        REPRODUCED (main, 8.6, 8.4)

        • Platform X11
        • Tested with Wish.geometry: 400x300+100+50 loaded by xrdb on Xvfb and wish -name wish, option get . geometry Geometry returns 400x300+100+50 but wm geometry . is 200x200+0+0 on main, 8.6 and 8.4.20.
        • Cause generic/tkWindow.c Initialize() applies only the -geometry command line option and never looks up the geometry resource of the main window as Xt apps do.
        • Prototype 528330ffff.patch (17 lines, no test: the resource must be present at Tk startup). If -geometry was not given, Tk_GetOption(main, "geometry", "Geometry") is applied with wm geometry .; invalid values are ignored. Verified: the resource is honoured, -geometry still wins, and a bogus value is ignored. wm.test has the same 10 no-WM failures as without the patch. This is a behaviour change, since users with *geometry resources will see wish follow them, so it may need discussion or a TIP.
        • Suggested ticket comment Still true: wish ignores the Wish.geometry resource. A small change in Initialize() would apply it when -geometry isn't given. That changes behaviour for users who have *geometry resources set.
        Prototype verified 2026-10-24, Linux
        • Status VERIFIED
        • Applies to main clean (offset 34 lines)
        • Repro with Wish.geometry: 400x300+100+50 loaded by xrdb (an xprop client kept Xvfb from resetting) and -name wish: before 200x200+0+0 / after 400x300+100+50. With -geometry 300x200+10+10 the command line wins (300x200+10+10). A bogus resource is ignored (200x200+0+0).
        • Test none (the resource must be in RESOURCE_MANAGER on the root window before Tk starts; core Tk cannot set root properties, so it would need xrdb in the test suite)
        • Test file run wm.test 302/0 (same as unpatched)
        • Patch changes none. Behaviour change for users who have *geometry resources set, so it may need discussion.
        Prototype verified 2026-10-24, Windows VM, generic patches
        • 528330ffff: no test; window/wm/frame tests pass.
        ⬜509974ffff2002pkg_mkIndex libtk.so requires DISPLAYREPRODUCED (main): package require Tk without DISPLAY fails with "no display name and no $DISPLAY environment variable". This is by design, and T…

        Open ticket 509974ffff · 63. Tk_Win Functions · filed 2002-01-29

        Full verdict

        REPRODUCED (main): package require Tk without DISPLAY fails with "no display name and no $DISPLAY environment variable". This is by design, and Tk ships a hand-written pkgIndex.tcl.

        • Platform X11
        • Tested main tclsh9.1 with DISPLAY unset.
        • Cause Tk_Init creates the main window, which needs a display connection. dgp (2018) wants the ticket kept open as a reminder to make Tk a package that doesn't create windows until asked.
        • Prototype none (design change, far over 60 lines)
        • Suggested ticket comment (none; dgp asked to keep this open as a reminder.)
        ✅483387ffff2001BadFont error using multiple X serverspatchREPRODUCED (main, by instrumentation; the code path is the same in 8.6/8.5/8.4). No X error is visible here because only one X server is available:…

        Open ticket 483387ffff · 46. Unix Fonts · filed 2001-11-19

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-09-28 check-in be529645fb on font-multiple-displays, in main, 9.0: Fix [483387]: BadFont error when a named font used on another display was changed or recreated. The fonts which depend on a named font were recreated for the di
        • 🔧 2026-10-07 check-in 41f156bc91 on trunk, in main, 9.0: Fix [483387]: BadFont error when a named font used on another display was changed or recreated
        • 🔧 2026-10-07 check-in c7586ac9f9 on core-9-0-branch, in main, 9.0, 8.6: Fix [483387]: BadFont error when a named font used on another display was changed or recreated
        • 🔧 2026-10-07 check-in d42c762a70 on core-8-6-branch, in main, 9.0, 8.6: Fix [483387]: BadFont error when a named font used on another display was changed or recreated
        • 💬 2026-09-28 serhiy.storchaka: Still reproducible with the X11 core fonts (configured with --disable-xft --disable-bidi); Xft is not affected. When a named font was changed or recreated, the fonts which depend on it were recreated for the display of the main window, including the one used on the other display. Proposed fix in branch font-multiple-displays, with test font-47.4 (needs TK_ALT_DISPLAY).
        • 💬 2026-10-13 nemethi: Merged by Jan into main, core-9-0-branch, and core-8-6-branch by commits 41f156bc, c7586ac9, and d42c762a. Closing the ticket.
        Full verdict

        REPRODUCED (main, by instrumentation; the code path is the same in 8.6/8.5/8.4). No X error is visible here because only one X server is available: a second connection to the same server via -screen unix/:122.0 shares the XIDs, which hides the error.

        • Platform X11
        • Tested main, 8.6, 8.5, 8.4 with the ticket script on -screen :122.1 and unix/:122.0; no X error. With a temporary check in Tk_DrawChars on main, fontPtr->display != display fires twice after font delete; font create, and does not fire with the fix.
        • Cause generic/tkFont.c UpdateDependentFonts() reloads every cached instance of the named font through TkpGetFontFromAttributes(fontPtr, tkwin, ...) with the window that ran font create/configure. Instances allocated for another screen/display are therefore reopened on the wrong display (wrong font XID/XftFont and wrong DPI), which gives BadFont on a real second server. font configure of a named font takes the same path.
        • Prototype 483387ffff.patch (34 lines, no test: needs a second X server). Adds a FindWindowOnScreen() helper that finds a window on fontPtr->screen (falling back to tkwin) and passes it to TkpGetFontFromAttributes. font.test passes.
        • Suggested ticket comment Still there: UpdateDependentFonts() reloads font instances for other displays through the window that ran font create/configure, so they end up on the wrong display. Branch reloads each instance with a window on its own screen.
        ✅481255ffff2001palette 2 doesn't work for windowspatchREPRODUCED (8.6 tip): the -palette 2 row is inverted and scrambled (wrong line stride). 9.1 main: not visible, because since the main merge 3bbe54b…

        Open ticket 481255ffff · 41. Photo Images · filed 2001-11-13

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-09-28 check-in b87aa58e0b on win-mono-image, in main, 9.0: Fix [481255]: photo images with -palette 2 were corrupted on Windows. TkAlignImageData() computed the line length of the output bitmap from the line length of t
        • 🔧 2026-09-29 check-in 9a4cbaace7 on win-mono-image, in main, 9.0: Add test imgPhInstance-7.3: -palette 2 draws black and white pixels in place [481255].
        • 🔧 2026-10-23 check-in 710caca0d3 on trunk, in main, 9.0: Fix [481255]: photo images with -palette 2 were corrupted on Windows
        • 🔧 2026-10-29 check-in 9572dd82e1 on core-9-0-branch, in main, 9.0, 8.6: Fix [481255]: photo images with -palette 2 were corrupted on Windows
        • 🔧 2026-10-29 check-in 697d9c3699 on core-8-6-branch, in 8.6: Fix [481255]: photo images with -palette 2 were corrupted on Windows
        • 🔧 2026-10-15 check-in 7e263169f7 on win-mono-image, in main, 9.0: Rebase to main
        • 🔧 2026-09-24 check-in b7178d60f1 on win-photo-gamma-palette, in main, 9.0: Add tests imgPhInstance-7.1 and 7.2 for -gamma and -palette 32f2e19de8.
        • 💬 2026-09-28 serhiy.storchaka: Still reproducible on Windows with 8.6 and 9.0, for widths 1-16, 33-48, 65-80, etc. There are two bugs: * TkAlignImageData() computed the line length of the word-aligned bitmap passed to CreateBitmap() from the line length of the photo data, which is padded to 4 bytes, instead of from the width, so the lines were shifted. * Windows draws the 1 bits of a monochrome bitmap with the background color, while X draws th…
        • 💬 2026-09-29 serhiy.storchaka: Added test imgPhInstance-7.3 to the branch win-mono-image. It checks the pixels of -palette 2 images with testpixel. It fails on Windows without the fix and passes with it.
        • 💬 2026-10-26 serhiy.storchaka: Fixed in 710caca0d3 (main), with test imgPhInstance-7.3.
        • 💬 2026-10-29 nemethi: Backported to core-9-0-branch and core-8-6-branch by commits 9572dd82 and 697d9c36. Closing the ticket.
        Full verdict

        REPRODUCED (8.6 tip): the -palette 2 row is inverted and scrambled (wrong line stride). 9.1 main: not visible, because since the main merge 3bbe54bc74 (2026-06-15) win/tkWinPort.h defines TK_CAN_RENDER_RGBA, and photos are drawn from pix32. -palette therefore has no visible effect on Windows in 9.1: p2 shows the smooth grey ramp.

        • Platform Windows
        • Tested X11 Xvfb, 9.1 main / 8.6 / 8.4: a -palette 2 photo with stripes, width 101, read back from the screen matches the source pixel for pixel. Windows repro: win/481255ffff.tcl (visual check; the window text explains PASS/FAIL).
        • Cause The 2002 bit-order fix (bitmap_bit_order = LSBFirst in win/tkWinImage.c XCreateImage) is still in place. There is a likely second problem in win/tkWinDraw.c TkPutImage(), in the 1-bit branch. It sets SetTextColor(fg)/SetBkColor(bg), but when Windows blits a monochrome bitmap to a colour DC, 0 bits get the text colour and 1 bits the background colour. That is the opposite of X, so the photo would be drawn as a negative. XCopyPlane in the same file already swaps the two colours ("seems backwards"). Only photos use the 1-bit path on a colour display: palette 2, or the fallback when the colormap is full.
        • Prototype 481255ffff.patch (swaps the two colours plus a comment; 25-line diff). NOT compiled or tested (Windows only). The Windows run should also check that bitmaps drawn via XCreateBitmapFromData (mono → mono pixmap) still look right.
        • Suggested ticket comment The bit-order fix is in. The 1-bit TkPutImage path on Windows still sets text and background colours the X way round, which probably draws -palette 2 photos inverted. Candidate fix: swap them, as XCopyPlane already does. Needs checking on Windows.
        Windows VM, run 5
        • Platform Windows
        • Verdict REPRODUCED (8.6 tip): the -palette 2 row is inverted and scrambled (wrong line stride). 9.1 main: not visible, because since the main merge 3bbe54bc74 (2026-06-15) win/tkWinPort.h defines TK_CAN_RENDER_RGBA, and photos are drawn from pix32. -palette therefore has no visible effect on Windows in 9.1: p2 shows the smooth grey ramp.
        • Tested 481255ffff.tcl (captures of the source, p2, p16, a -bitmap label, a bitmap image and a canvas bitmap). To exercise the 1-bit path on 9.1, I built tk-wt91 with a temporary #undef TK_CAN_RENDER_RGBA in tkImgPhInstance.c. The colour swap alone fixes the inversion, but the stripes stay scrambled. With the second fix, p2 is correct: black block, stripes, dithered ramp.
        • Cause (1) win/tkWinDraw.c TkPutImage 1-bit branch: the text and background colours are the X way round (mono→colour blits map 0 bits to the text colour). (2) generic/tkImgUtil.c TkAlignImageData keeps the source's 32-bit line padding (bytes_per_line). CreateBitmap expects lines of minimal width padded to WORD, so every line after the first is shifted.
        • Prototype 481255ffff.patch, updated (14 changed lines): the colour swap plus dataWidth = (image->width + 7) / 8 in TkAlignImageData. Bitmaps (-bitmap, bitmap image, canvas bitmap) are pixel-identical to ref91. imgPhoto, imgBmap, bitmap, winDraw, canvImg, image, button, canvas and menu tests: 1574 run, 0 failed (without the test hack). Relevant for 8.6/9.0; on 9.1 the path is only used without TK_CAN_RENDER_RGBA.
        • Suggested ticket comment Still broken in 8.6.18 on Windows 11: -palette 2 photos are drawn inverted and with shifted lines. Two causes: TkPutImage sets fg/bg the X way round, and TkAlignImageData keeps the 32-bit line padding that CreateBitmap doesn't expect. Prototype fixes both. In 9.1, Windows draws photos from RGBA, so -palette no longer affects the display there.
        ✅470331ffff2001Displaced Menus do not stay open.patchREPRODUCED (main, 8.6, 8.5)

        Open ticket 470331ffff · 12. Unix Menus · filed 2001-10-11

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-09-29 check-in c81290221a on menu-post-under-pointer, in main, 9.0: Fix [470331], [680660], 1596662fff, f9d316a5e0: a menu posted over its menubutton or menubar entry invoked the entry under the pointer on the release of the
        • 🔧 2026-09-30 check-in 0f86666370 on trunk, in main, 9.0: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
        • 🔧 2026-09-30 check-in 5c47ae3e9c on core-9-0-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
        • 🔧 2026-09-30 check-in 3aacabb15f on core-8-6-branch, in main, 9.0, 8.6: Fix for 470331ffff: Displaced menus do not stay open; 680660ffff: Menu items accidentally selected; 1596662fff: Menu behaviour at screen bottom; [f9d316a5
        • 💬 2026-09-30 serhiy.storchaka: Still present in 8.6 and 9.x on X11. The same problem is reported in [680660], [1596662] and f9d316a5e0. If a menu does not fit below its menubar entry or menubutton, it is moved up over the entry. The release of the click which posted it then invokes the entry under the pointer. The same happens with tk_popup near the bottom or right edge of the screen. I implemented a fix in two parts, similar to other toolkit…
        • 💬 2026-09-30 nemethi: Serhiy, many thanks! The fix is now in main, core-9-0-branch, and core-8-6-branch; see commits 0f866663, 5c47ae3e, and 3aacabb1.
        Full verdict

        REPRODUCED (main, 8.6, 8.5)

        • Platform X11
        • Tested menubar window at the bottom of a 1280x1024 Xvfb, a single click on "File"; no WM; event generate on main/8.6/8.5, and a real XTEST click confirmed it on main. The menu is shifted up under the pointer, the release invokes "Item 8" and the menu unposts.
        • Cause library/menu.tcl: bind Menu <ButtonRelease> calls tk::MenuInvoke without checking whether the pointer moved since the press that posted the menu.
        • Prototype 470331ffff.patch (+26/-1 lines). MenuButtonDown and the Menubutton <Button-1> binding record the press window and root position. A new tk::MenuButtonUp ignores a release in a different menu within 2 px of the press, so the menu stays posted. After the patch the menu stays open; a second click on an item invokes it; a press-drag-release still invokes. menu/menubut/menuDraw/unixMenu tests pass. No test added because it needs screen-edge placement.
        • Suggested ticket comment Still happens on 8.6 and 9.1 (X11). Fix: ignore the button release when the pointer has not moved since the click that posted the menu (a shifted menu under the pointer).
        💬220849ffff2000bug in font actualREPRODUCED (main, 8.6, 8.4)

        Open ticket 220849ffff · 46. Unix Fonts · filed 2000-11-01

        💬 Since triage (updated 2026-11-11)

        Status now: Open

        • 💬 2026-09-29 serhiy.storchaka: Still present in 8.6 and 9.x, on all platforms; also reported in python/cpython#143990. font actual returns the size rounded to an integer number of points. If the size was specified in pixels, different sizes can give the same result, and font create {*}[font actual $font] creates a different font. For example, -12 and -13 pixels both give 12 at 75 DPI. At 96 DPI (100%), -10/-11, -14/-15, -18/-19 and -22/-23 coll…
        Full verdict

        REPRODUCED (main, 8.6, 8.4)

        • Platform X11 (generic rounding)
        • Tested Xvfb at tk scaling 1.389: font actual {times -12} and {times -13} both give -size 9, while reqwidth (8.6: 40 vs 43) and linespace (13 vs 14) differ.
        • Cause generic/tkFont.c (around lines 3631/3696): font actual turns the actual pixel size into points and rounds to an integer ((int)(size + 0.5)). 12 px = 8.64 pt and 13 px = 9.36 pt both round to 9. fa.size is a double internally.
        • Prototype none. The fix changes the API: report a fractional point size, or report negative pixels when pixels were requested. That needs a design decision, maybe a TIP.
        • Suggested ticket comment Still present. font actual rounds the size to whole points, so -12 and -13 px both show as 9 pt at 100 dpi. Fixing it means reporting fractional points or pixels.
        ⬜220776ffff2000update idletasks does not update repacked widgets on WindowsREPRODUCED (9.1 main, 8.6 tip)

        Open ticket 220776ffff · 69. Events · filed 2000-11-01

        Full verdict

        REPRODUCED (9.1 main, 8.6 tip)

        • Platform Windows
        • Tested X11/Xvfb without a WM, main (9.1b1) and 8.6.18. An automated pixel check (PIL screen grab after each pack -anchor ...; update idletasks) found no stale button pixels; a sanity run confirmed that the check detects the button.
        • Cause win/tkWinWindow.c XMoveResizeWindow uses MoveWindow(..., TRUE), which only invalidates the uncovered area of the container. Windows then sends WM_PAINT, which becomes an Expose event, which schedules the idle redraw. [update idletasks] never processes window events, and Tk ignores WM_ERASEBKGND (tkWinX.c), so nothing repaints the container until the event loop runs. X11 is not affected because the X server fills uncovered areas with the window background itself.
        • Prototype none. A fix would need Windows paint messages handled synchronously inside [update idletasks], or Tk to erase backgrounds itself. That is a redesign and cannot be tested here. dkf (2004) considered the current behaviour acceptable; the workaround is [update] or timer-driven code.
        • Windows repro win/220776ffff.tcl (interactive: animates the repack, asks yes/no, prints the verdict and appends it to 220776-result.txt)
        • Suggested ticket comment Still Windows-only. On X11 the server repaints the uncovered area, so it does not occur there (checked with 9.1b1 and 8.6.18). On Windows the uncovered area is repainted only through WM_PAINT → Expose, which [update idletasks] does not process.
        Windows VM, run 4
        • Platform Windows
        • Verdict REPRODUCED (9.1 main, 8.6 tip)
        • Tested 220776_auto.tcl, the ticket script with screen shots while Tcl is blocked in after 700. The red area grows from 1.7 to 7.4 button areas. The final shot shows 9 red "Hello World!" copies (one at each anchor) on the blue background.
        • Cause moving the child window invalidates the parent area it uncovered. On Windows this only reaches Tk as WM_PAINT, which becomes a queued Expose event (win/tkWinX.c GenerateXEvent). update idletasks processes neither the message nor the window event, and WM_ERASEBKGND does nothing (tkWinX.c: result 0). On X11 the server repaints the window background by itself, so the problem does not show there. It is not a flushing problem (the 3017005 check shows GDI output appears at once), so a GdiFlush in XSync would not help.
        • Prototype none. A fix would mean either erasing with the Tk background pixel in WM_ERASEBKGND plus RedrawWindow(RDW_ERASENOW) in the XSync stub, or processing paint/Expose in update idletasks. That is a behaviour change, and there is a flicker risk.
        • Suggested ticket comment Still present on Windows 11 (8.6.18, 9.1): the areas a repacked widget uncovers are only repainted from WM_PAINT/Expose, which update idletasks does not process. X11 is unaffected because the X server paints window backgrounds.
        ⏳220787ffff2000Stipple patterns don’t work in arc/oval items on Windows/MacREPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple

        Open ticket 220787ffff · 05. Canvas Items · filed 2000-11-01

        ⏳ Since triage (updated 2026-11-11)

        Status now: Pending (Fixed)

        • 🔧 2026-10-23 check-in f2d5d1a9d2 on win-canvas-arc-stipple, in main, 9.0: Fix [220787]: Stipple patterns don't work in arc/oval items on Windows
        • 🔧 2026-10-23 check-in f34f83d4e1 on trunk, in main, 9.0: Fix [220787]: Stipple patterns don't work in arc/oval items on Windows
        • 🔧 2026-10-24 check-in e277e5e56a on trunk, in main, 9.0: Stipple tests canvas-27.1, canvas-27.2 and canvRect-13.1 to 13.3 (from [220787]) fail on Aqua, which does not support stippling: add the notAqua constraint.
        • 🔧 2026-10-28 check-in a8e033419d on core-9-0-branch, in main, 9.0: Fix [220787]: Stipple patterns don't work in arc/oval items on Windows
        • 🔧 2026-10-23 check-in 5b6f94abcd on win-canvas-arc-stipple, in main, 9.0: Update changes.md
        • 💬 2026-10-22 serhiy.storchaka: Re-tested on Windows 11 with today's main: stippled ovals, arc pie slices and rectangle outlines are still drawn solid. Branch win-canvas-arc-stipple still fixes all of them, and canvas.test, canvRect.test and canvArc.test pass with it. The Windows part is ready for review; the macOS part is still open.
        • 💬 2026-10-23 jan.nijtmans: Fixed in f34f83d4e1. Tested on Windows: just works fine! If CI is OK, can be backported further.
        • 💬 2026-10-23 jan.nijtmans: Tested on MacOS as well... I see a black oval, no fill pattern. Conclusion: a similar fix will be needed for MacOS too.
        • 💬 2026-10-23 kevin_walzer: Stippling is not, to my knowledge, supported on CoreGraphics/macOS.
        • 💬 2026-10-23 chrstphrchvz: I have a stale draft implementation of stippling for Aqua from a few years ago: see 3bfa5eb4 if interested.
        • 💬 2026-10-24 kevin_walzer: Does the macos-stipple branch accomplish what is needed here? I took Christopher's suggested API and screen shot as an example - see my attached screen shot.
        • 💬 2026-10-24 serhiy.storchaka: The new stipple tests failed on Aqua in CI, since Aqua does not support stippling yet. I added the notAqua constraint to canvas-27.1, canvas-27.2 and canvRect-13.1 to 13.3 in e277e5e56a; it can be removed when the macos-stipple branch is merged.
        • 💬 2026-10-24 jan.nijtmans: > it can be removed when the macos-stipple branch is merged Unfortunately, the notAqua constraints cannot be removed yet. The macOS changes itself are now merged to trunk and core-9-0-branch too. I'm seeing: canvRect.test ==== canvRect-13.1 stippled oval fill - bug 220787 FAILED ==== Contents of test case: .c create oval 10 10 110 110 -fill red -stipple gray50 -outline {} update stippleColors .c 60 60 ---- Result …
        • 💬 2026-10-24 serhiy.storchaka: Thank you, Jan. The differences are probably caused by the color space conversion on macOS. stippleColors in canvas.test and canvRect.test could compare the colors with a small tolerance, as pixelNear in imgPhInstance.test does.
        Full verdict

        REPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple

        • Platform Windows (and Aqua)
        • Tested X11 only, where stipples work. win/220787ffff.tcl prints VERDICT: FIXED with tktest on X11, which checks that the script works. Aqua not tested.
        • Cause win/tkWinDraw.c: the GC stipple is ignored for Arc/Chord/Pie and rectangle outlines (see serhiy's comment).
        • Prototype none (exists on branch win-canvas-arc-stipple)
        • Win repro win/220787ffff.tcl. Needs tktest or Img (img::window). Checks oval, arc and rectangle stipples with testpixel and prints PASS/FAIL lines and a VERDICT line.
        • Suggested ticket comment Fix is ready on [/timeline?r=win-canvas-arc-stipple|win-canvas-arc-stipple], with tests canvRect-13.x and canvas-27.x. It needs review and a merge. The macOS part remains open.
        Windows VM, run 6
        • Platform Windows
        • Verdict REPRODUCED (9.1 main); FIXED on branch win-canvas-arc-stipple
        • Tested testpixel: oval -stipple, oval -outlinestipple, arc pieslice -stipple and rectangle -outlinestipple are drawn solid; rectangle -stipple (control) is stippled. Branch win-canvas-arc-stipple (built on the VM 2026-10-23): all 5 cases PASS, VERDICT FIXED; canvas.test, canvRect.test, canvArc.test pass (227 tests, 0 failed).
        • Prototype none
        • Suggested ticket comment Still present on main on Windows 11 (4 of 5 cases drawn solid). Branch win-canvas-arc-stipple is waiting for review.
        ✅220796ffff2000-stipple option for canvas text item disrupts positioning, bREPRODUCED (9.1 main); FIXED on branch win-text-stipple

        Open ticket 220796ffff · 05. Canvas Items · filed 2000-11-01

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-10-27 check-in fde93bd248 on trunk, in main, 9.0: Fix [220796]: -stipple option for canvas text item disrupts positioning. Testcase now complete yet
        • 🔧 2026-10-27 check-in 5b8a105362 on trunk, in main, 9.0: canvText-21.1 (bug [220796]): the text, given in points, did not fit in the 200x60 canvas at a high screen resolution, so countPixels read outside of the canvas
        • 🔧 2026-11-02 check-in 4fd4398349 on win-text-stipple-9.0, in main, 9.0: Fix 220796ffff for 9.0: -stipple option for canvas text item disrupts positioning on Windows. Stippled text was not drawn, and angled stippled text was mispla
        • 🔧 2026-11-02 check-in 29e1638bb3 on core-9-0-branch, in main, 9.0, 8.6: Fix [220796]: -stipple option for canvas text item disrupts positioning
        • 🔧 2026-11-02 check-in 89594d4584 on core-8-6-branch, in 8.6: Fix [220796]: -stipple option for canvas text item disrupts positioning
        • 🔧 2026-10-19 check-in 221311c835 on win-text-stipple, in main, 9.0: Merge main.
        • 🔧 2026-11-07 check-in 834172a0bd on tests-min-width, in main: Test textDisp-33.3 failed on Windows when the 480 pixel wide frame .f and the 80-column text widget were wider than the screen: the text widget was laid out a s
        • 💬 2026-10-22 serhiy.storchaka: Re-tested on Windows 11 with today's main: stippled text is still drawn solid, and angled stippled text is not drawn at all. With branch win-text-stipple both are stippled and the bounding box does not change; canvText.test, font.test and winFont.test pass with it.
        • 💬 2026-10-27 jan.nijtmans: Running the testcase on Windows 11, I see: canvText.test ==== canvText-21.1 stippled text - bug 220796 FAILED ==== Contents of test case: .c create text 10 10 -anchor nw -text MMMM -font {Courier 30 bold} -fill black update set bbox [.c bbox 1] set plain [countPixels .c {*}$bbox] .c itemconfigure 1 -stipple gray50 update set stippled [countPixels .c {*}$bbox] # About half of the pixels are drawn with a gray50 stip…
        • 💬 2026-10-27 serhiy.storchaka: The failure of canvText-21.1 was in the test: the text is given in points, and from 200% display scaling it no longer fits in the 200×60 canvas, so countPixels read outside of it. Fixed in 5b8a105362: a 200×200 canvas and a single "M", and the knownBug constraint removed. On Windows 11 with tk scaling 1.0 to 4.67 (350%), the test passes with the fix (half of the pixels are drawn, as expected) and fails without i…
        • 💬 2026-10-27 jan.nijtmans: Thanks! Confirming the testcase works for me now. I also fixed frame-12.3 from the "tests-high-dpi" branch, which was also failing for me and now passed. Let's see tomorrows CI run. If it passes, it can be backported to 9.0 and 8.6.
        • 💬 2026-11-02 nemethi: I have tried to cherrypick Jan's commit from main into core-9-0-branch, but after several hours I had to give up because of the file win/tkWinFont.c, which, due to the bidi/shaping stuff, in Tk 9.1 is much more complicated than in Tk 9.0. I see two possible ways for backporting to core-9-0-branch: 1. Someone more proficient in fonts and bidi/shaping on Windows masters the cherrypick; 2. Serhiy kindly provides a ve…
        • 💬 2026-11-02 jan.nijtmans: I tried the backport to 9.0 as well. Same result :-(
        • 💬 2026-11-02 kevin_walzer: Or, maybe we don’t backport?
        • 💬 2026-11-02 serhiy.storchaka: The 9.0 version is on branch win-text-stipple-9.0, 4fd4398349. In 9.0, Tk_DrawChars() and TkDrawAngledChars() drew stippled text into the destination instead of the memory DC, so it was not drawn at all; they now use the same BeginStippledText()/EndStippledText() as 9.1.
        • 💬 2026-11-02 nemethi: Serhiy, many, many thanks for providing the 9.0 version of your fix! This worked immediately, without any code change. The fix is now merged into main, core-9-0-branch, and core-8-6-branch by commits fde93bd2 (Jan), 29e1638b, and 89594d45. Closing the ticket.
        Full verdict

        REPRODUCED (9.1 main); FIXED on branch win-text-stipple

        • Platform Windows
        • Tested X11. With Xft, stippled text is not supported, so the script fails there as expected and prints a NOTE saying so. The bbox stays the same with -stipple.
        • Cause win/tkWinFont.c: the text-layout path ignores the stipple. TkDrawAngledChars() draws into the wrong DC, measures without the font selected, and passes UTF-8 to GetTextExtentPointA().
        • Prototype none (exists on branch win-text-stipple)
        • Win repro win/220796ffff.tcl. Counts the drawn pixels of plain and stippled text (horizontal and angled) and checks the bbox; prints PASS/FAIL and a VERDICT line.
        • Suggested ticket comment Fix is ready on [/timeline?r=win-text-stipple|win-text-stipple], with test canvText-21.1. It needs review and a merge.
        Windows VM, run 6
        • Platform Windows
        • Verdict REPRODUCED (9.1 main); FIXED on branch win-text-stipple
        • Tested tktest91 with testpixel: horizontal stippled text draws all 1368 pixels like plain text instead of about half; angled stippled text draws 0 of 572 pixels. The bbox is unchanged. Branch win-text-stipple: stippled text draws 684 of 1368 pixels and angled 286 of 572, VERDICT FIXED; canvText.test, font.test, winFont.test pass (457 tests, 0 failed).
        • Prototype none
        • Suggested ticket comment Still present on main on Windows 11: stippled text is drawn solid, and angled stippled text is not drawn at all. Branch win-text-stipple is waiting for review.
        ✅220895ffff2000Canvas inconsistantly draws lines on Linux/UNIXREPRODUCED (9.1 main); FIXED on branch win-thin-line-endpoints

        Open ticket 220895ffff · 05. Canvas Items · filed 2000-11-01

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-10-27 check-in 53abc36388 on trunk, in main, 9.0, 8.6: Fix [220895]: Canvas inconsistently draws lines on Windows
        • 🔧 2026-10-28 check-in 1152f5a2b3 on core-9-0-branch, in main, 9.0: Fix [220895]: Canvas inconsistently draws lines on Windows
        • 🔧 2026-10-30 check-in 2b9c2189ac on core-8-6-branch, in 8.6: Fix [220895]: Canvas inconsistently draws lines on Windows
        • 🔧 2026-10-24 check-in 0fbe298ab6 on win-thin-line-endpoints, in main, 9.0: Merge main. Main moved stippled drawing into RenderStippled(): the last point of a thin polyline is now also drawn in DrawPathProc().
        • 💬 2026-10-22 serhiy.storchaka: Re-tested on Windows 11 with today's main: the last point of a thin line is still not drawn, so each of the eight lines of my test differs from the same line with its points reversed. Branch win-thin-line-endpoints fixes all eight, and canvas.test and canvLine.test pass with it.
        • 💬 2026-10-30 nemethi: Merged by Jan into main and core-9-0-branch by commits 53abc363 and 1152f5a2. Backported to core-8-6-branch by commit 2b9c2189. Closing the ticket.
        Full verdict

        REPRODUCED (9.1 main); FIXED on branch win-thin-line-endpoints

        • Platform Windows (X11 part no longer reproduces)
        • Tested main tktest on X11. All 8 coordinate sets in win/220895ffff.tcl give the same pixels when the points are reversed (VERDICT: FIXED).
        • Cause GDI does not draw the last point of a cosmetic-pen polyline (win/tkWinDraw.c XDrawLines).
        • Prototype none (exists on branch win-thin-line-endpoints)
        • Win repro win/220895ffff.tcl. Compares the pixels of each line with those of the same line with its points reversed; prints PASS/FAIL and a VERDICT line.
        • Suggested ticket comment On X11 this no longer happens. The Windows variant is fixed on [/timeline?r=win-thin-line-endpoints|win-thin-line-endpoints] (test canvas-28.1), which is waiting for a merge.
        Windows VM, run 6
        • Platform Windows
        • Verdict REPRODUCED (9.1 main); FIXED on branch win-thin-line-endpoints
        • Tested All 8 coordinate sets draw different pixels when the points are reversed (the last point of a thin polyline is not drawn). Branch win-thin-line-endpoints: all 8 lines PASS, VERDICT FIXED; canvas.test and canvLine.test pass (169 tests, 0 failed).
        • Prototype none
        • Suggested ticket comment Still present on main on Windows 11 for all 8 test lines. Branch win-thin-line-endpoints is waiting for review.
        🟢220856ffff2000[winfo containing] doesn’t work on non-root E desktopspatchREPRODUCED (main, 9.1, 8.6 tip, 8.6.10+, and partly 8.6.9, 8.6.8, 8.4.20). This contradicts the 2026-09-23 comment, which could not test a virtual …

        Open ticket 220856ffff · 67. Unix Window Operations · filed 2000-11-01

        🟢 Since triage (updated 2026-11-11)

        Status now: Open

        • 🔧 2026-10-24 check-in 8a0667c7fc on x11-coords-to-window, in main: Fix 586f722cdd and 220856ffff in Tk_CoordsToWindow() on X11. A toplevel that was just withdrawn was still found while the window manager had not yet unmappe
        • 💬 2026-10-16 erikleunissen: First off: I don't understand the ticket's original post fully: * Does "the root winow" (sic) mean the Tk root window ? * I don't know what the term "non-root" desktop means. So, sorry in advance if I got it wrong because of this limited understanding. But barring that, the observed behaviour by the OP seems correct to me since the reproducible script uses ".t" as the binding tag. Therefore, no action is to be exp…
        • 💬 2026-10-16 serhiy.storchaka: "The root window" is the root window of the X screen, and a "non-root desktop" is an Enlightenment virtual desktop other than desktop 0. Enlightenment showed these through a virtual root window, a large child of the real root that holds the toplevels of that desktop. The binding on .t is fine: while button 1 is held, X keeps sending the pointer events to the window where it was pressed, so <B1-Motion> of .t fires …
        • 💬 2026-10-16 erikleunissen: OK, thanks. Then my understanding of the subject matter is not enough to express any judgment.
        • 💬 2026-10-24 serhiy.storchaka: I was wrong earlier: with a virtual root window this fails since 8.6.10 (6942f5af38, the fix for 18a4ba19bd) when each toplevel is in a window manager frame inside the virtual root, as with Enlightenment: Tk_CoordsToWindow() checked the point with coordinates relative to the real root instead of the virtual root. If the window manager puts several toplevels into one frame, only the first one was checked, in al…
        Full verdict

        REPRODUCED (main, 9.1, 8.6 tip, 8.6.10+, and partly 8.6.9, 8.6.8, 8.4.20). This contradicts the 2026-09-23 comment, which could not test a virtual root.

        • Platform X11 (virtual-root / reparenting window managers)
        • Tested I emulated a virtual-root WM on a WM-less Xvfb with a small Xlib helper (work-b4/vr/vroot.c). It creates a 3000x3000 window at -400,-300 with __SWM_VROOT, reparents Tk's wrappers into it, sets __SWM_ROOT on the wrappers and sends ICCCM synthetic ConfigureNotify. Three modes, probing points in ., in .t and on the desktop:
          • direct (wrappers are children of the vroot): all versions correct.
          • frames (one decoration frame per toplevel inside the vroot, as tvtwm/E do): OK in 8.4.20, 8.6.8, 8.6.9; broken in 8.6.10, 8.6.13–8.6.15, 8.6 tip and main ('' for both toplevels).
          • shared (several toplevels in one frame, the case raster describes): 8.4.20 and 8.6.8 find only .t; 8.6 tip and main find neither.
          • Normal WMs (twm, fvwm, icewm, no WM) work in every version.
        • Cause unix/tkUnixWm.c Tk_CoordsToWindow(), the child == wmPtr->reparent case, has two problems:
          • It calls PointInWindow(x, y, …) with coordinates relative to the previous level (the real root) rather than childX/childY (relative to the vroot, like winPtr->changes). This regression came with 359434f369 / 6942f5af38 (2019-01-04, fix for 18a4ba19bd), first released in 8.6.10.
          • It returns NULL as soon as the point is not in the first toplevel whose reparent window is hit. That is the old "one client per frame" assumption.
        • Prototype 220856ffff.patch (39 lines, ~10 changed). It uses childX/childY, and when the point is not in that toplevel it goes on with the other toplevels and then descends into the frame; title-bar points still give ''. Checked in all three modes, including a title-bar point, and with twm/fvwm/icewm; unixWm.test and winfo.test pass. No automated test: one needs an external reparenting client. The emulator could become a testvroot command if wanted.
        • Suggested ticket comment Reproduced by emulating a virtual-root WM. Since 8.6.10, winfo containing returns "" for any toplevel inside a decoration frame inside a virtual root, because Tk_CoordsToWindow() checks the point with coordinates relative to the wrong window. In all versions it also fails when the WM puts several toplevels into one frame. A small fix is ready.

        Partly fixed 7

        ✅22d2341c9d2024Building extensions for Tk 8 and Tcl 8 in a Tcl 9 environmentpatchPARTLY (9.0 branch; main N/A). The original request is fixed: e754018763 + 055d75f68c (trunk, before 9.0.1, 2024-12-21) and fa41fdc007 (core-8-bran…

        Open ticket 22d2341c9d · (unused) · filed 2024-11-12

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-10-04 check-in 1f0f247550 on tk8-mode-tcl-size, in main, 9.0: Fix 22d2341c9d: Building extensions for Tk 8 and Tcl 8 in a Tcl 9 environment. Since the fix for Tcl ticket ad08bca6b1 (Tcl 9.0.4), tcl.h undefines Tcl_Size a
        • 🔧 2026-10-04 check-in 9c5a5ddf6c on core-9-0-branch, in main, 9.0: Fix 22d2341c9d: Building extensions for Tk 8 and Tcl 8 in a Tcl 9 environment
        • 💬 2026-10-04 serhiy.storchaka: Since the fix for Tcl ad08bca6b1 (Tcl 9.0.4), tk.h of Tk 9.0 no longer compiles in Tcl 8 mode: tcl.h undefines Tcl_Size at its end, but tk.h and tkDecls.h use it. Proposed fix in branch tk8-mode-tcl-size: tk.h defines and undefines Tcl_Size as tclOODecls.h does. With it, the structures have the same sizes as in Tk 8.6. Tk 9.1 headers do not support Tcl 8 mode, so this is only for 9.0.
        • 💬 2026-10-04 jan.nijtmans: For Tk, "building with Tcl 8" actually meant "building with Tcl 8.7". Reason: Tk 9.0 uses Tcl_GetIntForIndex(), Tcl_CreateObjCommand2(), Tcl_DStringToObj() and Tcl_UtfToChar16DString(). Tk 9.0 was never meant to work with Tcl 8.6. Also, Tk 9.0 can only be compiled with TCL_UTF_MAX=4, in order to be able to use unicode codepoints > 0xFFFF. Tcl 8.6 doesn't support that at all. Tcl 8.7 does, but will never be release…
        • 💬 2026-10-04 jan.nijtmans: Hm, I was a little bit too quick. Something which is supported with Tk 9.0 is used its headers to compile actually with Tcl 8.6/Tk 8.6. In this situation, the tk8-mode-tcl-size makes sense. Thanks!
        Full verdict

        PARTLY (9.0 branch; main N/A). The original request is fixed: e754018763 + 055d75f68c (trunk, before 9.0.1, 2024-12-21) and fa41fdc007 (core-8-branch); TEA tclconfig c34cded0a3, then 18e57a1c95 (2025-04-08: --with-tcl8 also defines TK_MAJOR_VERSION=8). The rules.vc OPTS=tk8 was later dropped again (effef6c657), and trunk tk.h now has #error for TK_MAJOR_VERSION != 9. However, the 9.0 headers have regressed since Tcl 9.0.4.

        • Platform generic (build)
        • Tested compiled a file that includes only tk.h with -DTCL_MAJOR_VERSION=8 -DTK_MAJOR_VERSION=8 -DUSE_TK_STUBS (what TEA --with-tcl8 passes). Against Tk 9.0 branch tk.h:
          • Tcl 9.0.3 headers: compiles; sizeof Tk_CanvasTextInfo = 88 and Tk_OptionSpec = 72, the same as the real Tk 8.6 headers.
          • Tcl 9.0.4 headers (and core-9-0-branch): 109 errors, starting with tk.h:197: unknown type name 'Tcl_Size'.
        • Cause Tcl fix ad08bca6b1 / 7c1522f896 (Apr 2026, in Tcl 9.0.4). In Tcl 8 mode tcl.h now does #define Tcl_Size int and then #undef Tcl_Size at its end (_TCLSIZEHANDLED), but Tk 9.0 generic/tk.h, tkDecls.h and tkPlatDecls.h use Tcl_Size after #include <tcl.h>.
        • Prototype 22d2341c9d.patch (30 lines, against core-9-0-branch generic/tk.h, not trunk). In Tcl 8 mode it defines Tcl_Size after tcl.h and undefines it at the end of tk.h, the same way tclOODecls.h does. Verified: compiles with Tcl 9.0.3 and 9.0.4 headers in 8 mode (same struct sizes as 8.6), and the normal 9.0 compile is unchanged.
        • Suggested ticket comment The --with-tk8 part works since 9.0.1 / tclconfig 18e57a1c95. However, since Tcl 9.0.4 (ad08bca6b1), Tk 9.0's tk.h no longer compiles in Tcl 8 mode: tcl.h undefines Tcl_Size, which tk.h still uses. Doing the same _TCLSIZEHANDLED dance in tk.h fixes it; after that the ticket can be closed (9.1 headers deliberately no longer support Tk 8).
        ⏳026e2bb6852024text tag sets bad relief, gets error, but bad relief is saved anywaypatchPARTLY (main 9.1b1, 9.0.5, 8.6.18): -relief, -justify and 13 other tag options now reject a bad value and keep the old one. -tabs is still broken i…

        Open ticket 026e2bb685 · 18. [text] · filed 2024-01-14

        ⏳ Since triage (updated 2026-11-11)

        Status now: Pending (Fixed)

          Full verdict

          PARTLY (main 9.1b1, 9.0.5, 8.6.18): -relief, -justify and 13 other tag options now reject a bad value and keep the old one. -tabs is still broken in all three versions.

          • Platform generic
          • Tested for each of 15 tag options I set a good value, then a bad one, then ran cget (B-work/t026.tcl). -tabs: .t tag configure t -tabs {1c 2c}; catch {.t tag configure t -tabs {1c bogus}} leaves cget -tabs = {1c bogus}. Worse, every later tag configure of that tag, even -foreground blue, then fails with "bad tab alignment" (the foreground is still applied). This is the "-tags" (really -tabs) issue Jan mentioned on 2024-06-04. His branch bug-026e2bb685 (tip e34d9e15a8, 2024-06-04) was never merged and no longer applies to trunk.
          • Cause generic/tkTextTag.c:TkTextTagCmd (TAG_CONFIGURE). -tabs is a TK_OPTION_STRING. Tk_SetOptions stores it first, and TkTextGetTabs only parses it afterwards. When parsing fails it returns TCL_ERROR with the bad string already stored. Because the parse step runs again on every configure, the tag stays broken.
          • Prototype 026e2bb685.patch (135 lines). It makes -tabs a TK_OPTION_CUSTOM (SetTabArray/RestoreTabArray/FreeTabArray) that parses on set, like Jan's branch, and adds a restoreProc and empty→NULL handling. Tests textTag-1.40/1.41 added; 1.40 fails on unpatched trunk. textTag.test (179), text.test (695) and textDisp.test (402) all pass.
          • Suggested ticket comment Only -tabs is still affected: a bad value is kept, and after that every tag configure of the tag fails, even for unrelated options (8.6.18, 9.0.5, 9.1). Rebasing bug-026e2bb685 (parse -tabs in a custom option setProc) fixes it; a refreshed version with a test is ready. Not closable until that is merged.
          Prototype verified 2026-10-24, Linux
          • Status VERIFIED+TEST
          • Applies to main clean
          • Repro before tag cget -tabs = {1c fork} and a later tag configure x -foreground blue fails with "bad tab alignment" / after {1c 2c}, and -foreground works
          • Test textTag-1.40, textTag-1.41 (in patch; 1.40 fails before: yes; passes after: yes. 1.41 is a regression check for the empty value, passes both)
          • Test file run textTag.test 184/0, text.test 703/0, textDisp.test 409/0
          • Patch changes none. Reviewed: tags do not go through Tk_InitOptions (tabArrayPtr is NULL-initialised in TkTextCreateTag). Tk_FreeConfigOptions now frees the array through FreeTabArray, so removing the manual free is correct.
          Prototype verified 2026-10-24, Windows VM, generic patches
          • 026e2bb685: OK. textTag-1.40 fails before, passes after (1.41 passes both); textTag.test 184/0.
          ⏳c2483bfe4b2020tk fontchooser on macOS can automatically open on startup, can lead to crashespatchPARTLY (read trunk, core-9-0-branch, core-8-6-branch; the macOS behaviour itself was not tested)

          Open ticket c2483bfe4b · 33. Generic Dialog Support · filed 2020-09-19

          ⏳ Since triage (updated 2026-11-11)

          Status now: Pending (Fixed)

            Full verdict

            PARTLY (read trunk, core-9-0-branch, core-8-6-branch; the macOS behaviour itself was not tested)

            • Platform Aqua
            • Tested Read macosx/tkMacOSXInit.c and macosx/tkMacOSXDialog.c on trunk, 9.0 and 8.6.
              • The reopening is fixed by 489304eaa9 (2020-10-20, merged to 8.6 and 8.7). closePanels() runs from TkpExitProc, which unix/tkAppInit.c installs through TkpWantsExitProc, and SIGINT/SIGHUP/SIGTERM handlers call Tcl_Exit. All of this is still on trunk.
              • The crash guard Marc described (a check for a NULL Tk_Window in TkSendVirtualEvent) is not in the code anymore. In FontchooserParseConfigure (-font case), if ([fp isVisible]) Tk_SendVirtualEvent(fcdPtr->parent, ...) still runs with fcdPtr->parent == NULL when this interpreter never ran tk fontchooser show. Tk_SendVirtualEvent then dereferences the NULL window (Tk_Display(target)).
              • This can still happen. The exit cleanup only covers wish started from a tty, and the signal handlers are skipped when the host (Python) already has its own. Another interpreter in the same process showing the shared NSFontPanel triggers it as well, e.g. interp create x; load {} Tk x; x eval {tk fontchooser show}; tk fontchooser configure -font Courier.
              • The FontchooserSelection branch has the same unguarded send. It is only reached with fontchooserInterp set, but fcdPtr->parent can be NULL after the parent window is destroyed.
            • Cause macosx/tkMacOSXDialog.c:FontchooserParseConfigure and FontchooserEvent send TkFontchooserFontChanged to fcdPtr->parent without checking that it is set. The parent is only set by show or -parent, but the NSFontPanel is shared and can be visible without that.
            • Prototype c2483bfe4b.patch (30 lines, against trunk: skips both sends when fcdPtr->parent is NULL; not compiled, no Mac)
            • Suggested ticket comment The startup reopening is fixed by 489304eaa9, but the reported segfault path is still there. tk fontchooser configure -font sends <<TkFontchooserFontChanged>> to a NULL parent whenever the shared font panel is visible without this interp having shown it (e.g. a second interp, or a non-wish host that skipped the exit cleanup). A two-line NULL check (prototype attached) would let this ticket be closed.
            ⏳1f46cac0802020::tk::fontchooser contains a couple of issuespatchPARTLY (main, 9.0, 8.6)

            Open ticket 1f46cac080 · 46. Unix Fonts · filed 2020-07-12

            ⏳ Since triage (updated 2026-11-11)

            Status now: Pending (Fixed)

              Full verdict

              PARTLY (main, 9.0, 8.6)

              • Platform X11 (script fontchooser; also Windows/Aqua only if the script implementation is used)
              • Tested B-work/t1f4*.tcl, fcw.tcl. Already fixed: the initial style and Done use msgcat::mc (styles are translated and translated back correctly when the locale is set before first use: {.. 12 bold} round-trips), and the German "Schriftgrad:" label is not truncated (reqwidth == width). Still reproduced in all three versions: if msgcat::mclocale changes after the first tk fontchooser call, the style listbox stays English ("Bold"), while S(style) and Done use the new locale ("Fett"). Picking "Bold" then returns a font without bold. The proposed branch bug-1f46cac080 (2020) was never merged.
              • Cause library/fontchooser.tcl. S(styles) is built with msgcat::mc in the namespace eval at source time, while Canonical (called from Show) and Done/Init translate each time. Canonical also appends to S(styles,lcase) in a redundant first loop.
              • Prototype 1f46cac080.patch (63 lines). It moves the S(styles) list into Canonical and drops the duplicate loop. Test fontchooser-6.1 added (custom locale "xx"); it fails on trunk (Bold Xbold {}) and passes patched; fontchooser.test 20/20 passes.
              • Suggested ticket comment Most points are fixed by now. One remains: if the locale changes after the font chooser was first used, the style list stays in the old language, and choosing a style there drops it from the result. Fix: build the style list in Canonical. A small patch with a test is ready; after it the ticket can be closed and branch bug-1f46cac080 closed.
              Prototype verified 2026-10-24, Linux
              • Status VERIFIED+TEST
              • Applies to main clean
              • Repro before, after a locale change the style list stays "Bold" and the result loses bold / after "Xbold" and bold is kept
              • Test fontchooser-6.1 (in patch; fails before: yes; passes after: yes)
              • Test file run fontchooser.test 20/0
              • Patch changes none. Reviewed: S(styles) is used only after Canonical, which runs from Setup at source time and from Show/Init. The -listvariable binding still works.
              Prototype verified 2026-10-24, Windows VM, generic patches
              • 1f46cac080: fontchooser-6.1 (scriptImpl) skipped: Windows uses the native font dialog, library/fontchooser.tcl is not used. fontchooser.test 20/0.
              ✅3020550fff2010grid, pack, place on half-dead windows resurrects hash entrypatchPARTLY FIXED. pack and grid were fixed on main by ddfb9848d8 (2024-10-09, fvogel, Fix eb332846f3: GetGrid/GetPacker return NULL for TK_ALREADY_DE…

              Open ticket 3020550fff · 48. Geometry Management · filed 2010-06-24

              ✅ Since triage (updated 2026-11-11)

              Status now: Closed (Fixed)

              • 🔧 2026-10-09 check-in d5f1c2d5b7 on trunk, in main, 9.0, 8.6: Fix 3020550fff: Do not crash or leave dangling entries when grid or place is used on a half-dead window
              • 🔧 2026-10-16 check-in 3b9d5c7aa7 on core-9-0-branch, in main, 9.0: Fix 3020550fff: Do not crash or leave dangling entries when grid or place is used on a half-dead window
              • 🔧 2026-10-16 check-in 758007f6cc on core-8-6-branch, in main, 9.0, 8.6: Fix 3020550fff: Do not crash or leave dangling entries when grid or place is used on a half-dead window
              • 🔧 2026-09-24 check-in 8c00370c35 on halfdead-geometry-crash, in main, 9.0: Do not crash or leave dangling entries when grid or place is used on a half-dead window.
              • 💬 2026-09-24 serhiy.storchaka: Still crashes in 8.6 and 9.x (and since at least 8.4): grid .x -in .f in a <Destroy> binding of .f segfaults, and place on a half-dead window leaves a dangling entry (use-after-free). ddfb9848d8 fixed the stale entries for grid and pack, but not this path. Branch halfdead-geometry-crash is a safe, compatible fix for all branches: a half-dead window is ignored, and content put -in a half-dead container is left un…
              • 💬 2026-10-16 nemethi: Merged into main, core-9-0-branch, and core-8-6-branch by commits d5f1c2d5 (Jan), 3b9d5c7a, and 758007f6. Closing the ticket.
              Full verdict

              PARTLY FIXED. pack and grid were fixed on main by ddfb9848d8 (2024-10-09, fvogel, Fix eb332846f3: GetGrid/GetPacker return NULL for TK_ALREADY_DEAD). 8.6.15 still has the original stale-entry bug for grid/pack (a new window's content is not mapped). REPRODUCED for the rest, see below.

              • Platform generic (tested on X11)
                • REGRESSION from that fix: grid .x -in .f in a <Destroy> binding of .f segfaults (main 9.1b1, 9.0.5, 8.6.18; 8.6.15 did not crash). -in sets positionGiven=1 before GetGrid() returns NULL. containerPtr stays NULL, and so does contentPtr->containerPtr, so the code jumps to scheduleLayout and dereferences NULL (tkGrid.c ConfigureContent, containerPtr->abortPtr).
                • place was never fixed: place .f -x 0 -y 0 in .f's <Destroy> binding leaves freed .f linked into its container, causing a use-after-free and a segfault at update / place content (main, 8.6.18). A reuse test gives bad=1 on 8.6.18.
              • Tested main, 9.0, 8.6.18, 8.6.15, and the wt build (a script that creates and destroys 200 windows, plus single-op scripts under gdb), no WM
              • Cause tkPlace.c:ConfigureContent has no TK_ALREADY_DEAD check. In tkGrid.c:ConfigureContent, CONF_IN sets positionGiven before checking the result of GetGrid.
              • Prototype 3020550fff.patch (+39/-1 lines). Grid: set positionGiven only after GetGrid succeeds. Like pack, the content then goes to its parent. Place: ignore a dead content window, and a dead -in container. New tests grid-25.1 and place-13.2; both crash without the patch. grid/place/pack tests pass, except place-8.1/8.2 and pack-18.x, which also fail without the patch (no WM on Xvfb).
              • Suggested ticket comment pack and grid were fixed by ddfb9848d8 for eb332846f3, but that fix makes grid .x -in .f crash when run in a <Destroy> binding of .f (8.6.16+ and 9.0). place still has the original problem, which leads to a use-after-free. A prototype fix with tests is ready.
              💬2408101fff2008font sizes limited to integer valuesPartly FIXED. The internal TkFontAttributes.size became double in 6c1859a0ce (2017-05-19, 8.6.7), which is what the ticket asked for. At script…

              Open ticket 2408101fff · 46. Unix Fonts · filed 2008-12-09

              💬 Since triage (updated 2026-11-11)

              Status now: Open

              • 💬 2026-09-30 serhiy.storchaka: The size is a double internally now (TkFontAttributes.size), but the Tcl interface still requires an integer: font create -size 11.5 fails with "expected integer but got "11.5"", and font actual rounds the size to an integer number of points. This is part of [220849]. I implemented fractional font sizes there: -size accepts a fractional number, and font actual reports the actual size in points without rounding. Wi…
              Full verdict

              Partly FIXED. The internal TkFontAttributes.size became double in 6c1859a0ce (2017-05-19, 8.6.7), which is what the ticket asked for. At script level, sizes are still integers: font create f -size 11.5 fails with "expected integer" (main, 8.6, 8.5), and font actual/configure return rounded integers "for compatibility" (09eec5d40f).

              • Platform generic
              • Tested main, 8.6, 8.5 on X11.
              • Cause generic/tkFont.c ConfigAttributesObj uses Tcl_GetIntFromObj for FONT_SIZE, and GetAttributeInfoObj rounds the size.
              • Prototype none. Accepting and returning fractional sizes changes script-visible behaviour, so it needs a TIP.
              • Suggested ticket comment The internal size has been a double since 6c1859a0ce (8.6.7). Fractional sizes at script level (-size 11.5) would change the API and need a TIP, so we propose closing this as fixed or converting it to an RFE.
              🟢1185793fff2005extra event called in windowsPARTLY (9.1 main): the crossing events are FIXED; the dismissing click still leaks

              Open ticket 1185793fff · 69. Events · filed 2005-04-19

              🟢 Since triage (updated 2026-11-11)

              Status now: Open

              • 🔧 2026-10-15 check-in 8d61ec7f00 on win-popup-phantom-events, in main, 9.0: Rebase to main
              Full verdict

              PARTLY (9.1 main): the crossing events are FIXED; the dismissing click still leaks

              • Platform Windows
              • Tested not on Linux (the reporter says Linux is unaffected). Checked with fossil that the branch is merged into main, 9.0 and 8.6.
              • Cause win/tkWinMenu.c: after TrackPopupMenu, the click that dismissed the popup is still processed by Tk.
              • Prototype none. Repro: win/1185793fff.tcl (phase A counts Enter/Leave and checks whether the click leaks to another widget; phase B checks for a re-post when dismissing on the tag)
              • Suggested ticket comment The spurious Enter/Leave events are fixed by 0800b18ca8 (also in 9.0 and 8.6). Still open: the click that dismisses the popup reaches the widget under the pointer, so clicking the tagged text posts the menu again. win/1185793fff.tcl checks both.
              Windows VM, run 6
              • Platform Windows
              • Verdict PARTLY (9.1 main): the crossing events are FIXED; the dismissing click still leaks
              • Tested Phase A: one Enter and one Leave (PASS), but the click that dismisses the popup reaches .input2 (FAIL). Phase B: dismissing on the tag does not re-post the menu (PASS).
              • Prototype none
              • Suggested ticket comment The spurious Enter/Leave events are fixed in main. The click that dismisses the popup menu is still delivered to the widget under the pointer.

              Other 2

              ⬜3001403fff2010Wish crash on Windows XP when trying to free a colorpatchNEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.

              Open ticket 3001403fff · 68. Win Window Operations · filed 2010-05-14

              Full verdict

              NEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.

              • Platform Windows
              • Tested code inspection only. win/tkWinColor.c:XFreeColors still calls Tcl_Panic("Tried to free a color that isn't allocated").
              • Cause XFreeColors decides by the *current* RASTERCAPS & RC_PALETTE. A color allocated while the display was true-color has no refCounts entry. If the display switches to a palette mode (a full-screen game), freeing that color panics.
              • Prototype 3001403fff.patch (9 lines): skip pixels without a refcount entry and drop the second panic. Not compiled (no mingw here), no test.
              • Suggested ticket comment XFreeColors panics when the display switches to palette mode after colors were allocated in true-color mode. Freeing an unknown pixel can simply be ignored instead of panicking.
              Windows VM, run 4
              • Platform Windows
              • Verdict NEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.
              • Tested each patch built in tk-wt91 (together with 2100553fff.patch). color.test, winfo.test, imgPhoto.test, frame.test, button.test, canvas.test, winDraw.test with 1708823: 1120 tests, 0 failed.
              • Cause XFreeColors decides by the current RC_PALETTE, while XAllocColor decided at allocation time
              • Prototype keep 1708823fff.patch (89 lines, mostly re-indentation). It drops the per-call GetDC/RC_PALETTE check and relies only on the refCounts table. That handles both directions: true-color → palette (no panic) and palette → true-color (entries are still freed, no leak). 3001403fff.patch only fixes the first direction and keeps the per-call GetDC. 3001403fff can be closed as a duplicate of 1708823.
              • Suggested ticket comment XFreeColors uses the current display mode to decide whether a pixel came from the palette. Using only the refcount table (as XAllocColor records it) fixes the panic after a mode switch. Fix in the 1708823 prototype.
              Prototype verified 2026-10-24, Windows VM
              • Status CANNOT-TEST (palette mode needed, see 1708823fff)
              • Applies to main clean, but conflicts with 1708823fff (alternative fix)
              • Repro not reproducible / not reproducible
              • Test none (needs palette mode)
              • Test file run B2: color 21/0, imgPhoto 248/0, frame 207/0, button 416/0 (pre-existing failures: none). It compiles.
              • Patch changes none. Close the ticket as a duplicate of 1708823 and use that patch.
              ⬜1708823fff2007Wish crash on Windows XP when trying to free a colorpatchNEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.

              Open ticket 1708823fff · 29. Colormaps and Visuals · filed 2007-04-27

              Full verdict

              NEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.

              • Platform Windows
              • Tested source review of main win/tkWinColor.c. XAllocColor and XFreeColors each check GetDeviceCaps(dc, RASTERCAPS) & RC_PALETTE separately, at the time they are called. So a color allocated in true-color mode and freed after switching to palette mode is not in cmap->refCounts, and Tcl_Panic("Tried to free a color that isn't allocated") fires. The opposite switch leaks palette entries. Repro script: win/1708823fff.tcl (needs a real depth switch while it waits).
              • Cause win/tkWinColor.c:XFreeColors decides by the current device mode instead of by how the pixel was allocated.
              • Prototype 1708823fff.patch (about 35 lines changed, mostly re-indentation; no test because it cannot be exercised automatically). XFreeColors no longer checks RC_PALETTE. It frees through the refCounts table whenever the pixel is found there. Palette pixels carry the PALETTERGB flag, so true-color pixels never match and are skipped instead of causing a panic. Not compiled (no mingw here).
              • Suggested ticket comment Still present in the code: XFreeColors checks the current palette mode rather than the mode used at allocation. The fix is to free through the refCounts table whatever the current mode is, and to ignore pixels not in the table.
              Windows VM, run 4
              • Platform Windows
              • Verdict NEEDS palette mode (not testable: Windows 11 has no 8-bit desktop mode). Both prototypes compile cleanly with MSVC.
              • Tested each patch built in tk-wt91 (together with 2100553fff.patch). color.test, winfo.test, imgPhoto.test, frame.test, button.test, canvas.test, winDraw.test with 1708823: 1120 tests, 0 failed.
              • Cause XFreeColors decides by the current RC_PALETTE, while XAllocColor decided at allocation time
              • Prototype keep 1708823fff.patch (89 lines, mostly re-indentation). It drops the per-call GetDC/RC_PALETTE check and relies only on the refCounts table. That handles both directions: true-color → palette (no panic) and palette → true-color (entries are still freed, no leak). 3001403fff.patch only fixes the first direction and keeps the per-call GetDC. 3001403fff can be closed as a duplicate of 1708823.
              • Suggested ticket comment XFreeColors uses the current display mode to decide whether a pixel came from the palette. Using only the refcount table (as XAllocColor records it) fixes the panic after a mode switch. Fix in the 1708823 prototype.
              Prototype verified 2026-10-24, Windows VM
              • Status CANNOT-TEST (needs an 8-bit palette display mode, which Windows 11 does not have)
              • Applies to main clean
              • Repro not reproducible / not reproducible
              • Test none (needs a display mode switch to palette)
              • Test file run color.test 21/0, imgPhoto.test 248/0 (B1 and B6; pre-existing failures: none). It compiles cleanly.
              • Patch changes none. Preferred over 3001403fff (it handles both directions of the mode switch). The two patches conflict.