BUG: Make SyN and velocity field registration work with GlobalReleaseDataFlag by gdevenyi · Pull Request #7095 · InsightSoftwareConsortium/ITK

gdevenyi · 2026-11-02T23:10:56Z

Make SyN, BSplineSyN and time-varying velocity field registration run with DataObject::GlobalReleaseDataFlag on; on main they throw or segfault. Results are unchanged.

Root cause

With the global flag, every internal mini-pipeline calls ReleaseData() on its inputs after Update(). Several of those inputs are read again later:

SyNImageRegistrationMethod and BSplineSyN set the current-level virtual domain image as the ResampleImageFilter reference image, which is a pipeline input. After the first resample, the metric's virtual image had no buffer ("Virtual domain and moving transform displacement field must have the same size").
The resamplers released the smoothed level images (m_FixedSmoothImages, m_MovingSmoothImages) and the mask images, which every iteration reads again. ImageRegistrationMethodv4 released the user's fixed and moving images and the virtual domain through its per-level shrink and smoothing filters.
SyN's composers and inverter released the transform fields and the intermediate total field, which is inverted twice. InvertDisplacementFieldImageFilter released its own input and output through its internal composer.
TimeVaryingVelocityFieldTransform::IntegrateVelocityField released the transform's velocity field through the integrators.
ImageToImageMetricv4 released the metric's fixed and moving images through the gradient filters, and GradientRecursiveGaussianImageFilter released its own input after the first of its per-dimension passes.

Fix

These pipelines now get grafts, which share the buffers. Releasing a graft only drops the graft's own pointer. ImageRegistrationMethodv4 gets a protected MakeGraft() helper for the registration methods.
The resamplers set their geometry with SetOutputParametersFromImage(virtualDomainImage) instead of a reference image input, which gives the same output region and geometry.
GradientRecursiveGaussianImageFilter grafts its input once per pass, only when the input would be released. After the passes the filter drops the last graft.
First commit, BUG: Copy the pixel accessor in ImageAdaptor::Graft: Graft now grafts the adapted image, which carries a VectorImage's number of components, and copies the pixel accessor. Without it, a graft of an NthElementImageAdaptor or VectorImageToImageAdaptor read element 0, so the metric and gradient-filter grafts of an adaptor input used the wrong channel. The registration commit depends on it.
TimeVaryingVelocityFieldTransform::IntegrateVelocityField points the shared velocity field interpolator back at the velocity field after integrating, also on exceptions. The integrators had left it on their released input graft.
The metric grafts only when ShouldIReleaseData() is true, so an unchanged image still does not trigger a new gradient computation.
Values are unchanged; nothing is copied.

Test results

New GTest itkSyNImageRegistrationMethodReleaseDataGTest.cxx: a two-level 2-D registration with MeanSquaresImageToImageMetricv4 (default gradient filters) for SyN and BSplineSyN (each with and without fixed and moving masks) and for TimeVaryingVelocityFieldImageRegistrationMethodv4. It requires the displacement and inverse displacement fields with the flag on during Update() to equal exactly those of a run with the flag off. On main all 5 cases fail with exceptions. With this PR they pass at 1, 4, 8 and 16 threads. Each masked case fails when its mask grafts are removed. ITKGoogleTest is added to the module's TEST_DEPENDS.
New GTest itkGradientRecursiveGaussianImageFilterGTest.cxx: the gradient of an NthElementImageAdaptor input equals exactly that of a scalar copy of the selected element, with the flag off and on. The flag-off case passes on main and fails if the filter always grafts. The flag-on case fails on main and fails without the accessor copy in ImageAdaptor::Graft.
ImageAdaptor.GraftKeepsThePixelAccessor: grafts of an NthElementImageAdaptor and of a VectorImageToImageAdaptor read the selected element at every pixel, before and after UpdateOutputInformation(). It fails if only the pixel container and the accessor are copied, because the graft's vector length is then 1.
itkSyNImageRegistrationMethodReleaseDataGTest.cxx, two more tests:
- GlobalReleaseDataFlag.MetricKeepsTheAdaptorChannel: the Mean Squares value and derivative for channel 1 of two-channel adaptor inputs are equal with the flag off and on. It fails without the accessor copy.
- GlobalReleaseDataFlag.VelocityFieldIntegrationLeavesTheInterpolatorOnTheField: after integration with the flag on, the interpolator's input is the velocity field, and evaluating it returns the field's value. It fails without the restore.
All 575 tests of the registration build (RegistrationMethodsv4, Metricsv4, DisplacementField, ImageGradient, ImageGrid, Smoothing and dependencies) pass. The build adds no warnings.
SyN harness (MNI152 T1 at shrink 4; CC, Mattes and MSQ in float and double; with and without derivative downsampling and a mask) with cmp against main: 18/18 identical.
TimeVaryingBSplineVelocityFieldImageRegistrationMethod is unchanged here. Its transform needs the same treatment, so it will be a separate PR with its own test.

AI assistance

Tool: Claude Code (claude-opus-5-5)
Role: reproduced the failures, traced each release to its mini-pipeline, wrote the fix and the GTests, addressed the review, and ran the local build, tests, mutation checks and harness
All code was reviewed and tested locally before committing

gdevenyi · 2026-11-02T23:11:11Z

@ntustison, could you review this when you have time? It makes SyN, BSplineSyN and time-varying velocity field registration work with GlobalReleaseDataFlag, which came up in the review of #7078. The results are unchanged.

hjmjohnson · 2026-11-03T01:49:07Z

@greptileai review

greptile-apps · 2026-11-03T01:55:30Z

Confidence Score: 3/5

[Medium impact] The changes reach registration workflows and a shared gradient filter when global data release is enabled.

Fix the metric's lost channel selection and the velocity interpolator's released input before merging.

Findings

Metric gradients use wrong channel ▶
Velocity interpolator reads released pixels ▶

Summary

The PR keeps images and transform fields buffered while internal filters run with GlobalReleaseDataFlag enabled.

Registration keeps needed images and fields when filters release their inputs.
Recursive Gaussian gradients keep their input available across passes.

Previous-thread context: gdevenyi accepted the loss of incidental upstream updates because transform consumers already use the current buffer and registration supplies buffered fields; Greptile withdrew that finding. gdevenyi deferred TimeVaryingBSplineVelocityFieldImageRegistrationMethod to a separate PR because its transform needs its own fix and test.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  Original["Original image"] -->|shares pixels| Graft["Temporary graft"]
  Graft --> Filter["Internal filter runs"]
  Filter --> Release["Graft releases its pixels"]
  Original --> Reuse["Original remains usable"]
  Interpolator["Owned velocity interpolator"] --> Graft
  Release --> Broken["Interpolator still points at released graft"]

_{Reviews (2) · Last reviewed commit: "BUG: Make SyN and velocity field registr..." · Reviewed by Greptile}

hjmjohnson

Thanks. For plain Image inputs the approach holds: a graft shares the buffer, so results do not change, and the test fails on main with the "must have the same size" error from the #7078 thread (reasoned from the code, not run).

Requesting changes. One item is blocking because it changes results even with the flag off, and Greptile flagged it too:

Blocking: GradientRecursiveGaussianImageFilter now always grafts its input, and ImageAdaptor::Graft (itkImageAdaptor.hxx:67-76) does not copy the pixel accessor. An NthElementImageAdaptor or VectorImageToImageAdaptor input then reads component 0. I confirmed this in the code. Details inline.
Should-fix: masked BSplineSyN still fails with the flag on. Its fixed and moving mask resamplers (around BSplineSyN.hxx:336) are not grafted, unlike SyN's four, and neither is the one at itkTimeVaryingBSplineVelocityFieldImageRegistrationMethod.hxx:738. The test has no masked case, so it cannot catch this.
Should-fix: after GenerateData, m_DerivativeFilter keeps the last graft, and so keeps the input's buffer alive after the outer pipeline would have released it (inline).
Should-fix: the TimeVaryingBSplineVelocityField changes have no test, and the body says they may be incomplete. Please cover them or move them to a separate PR.
Greptile's P2 on GetBufferPointer() in the test holds (conventions.md: "Use iterators, not raw buffer access").

Non-blocking:

Greptile's P1 on the velocity-field graft (TimeVaryingVelocityField.hxx:36-43) looks moot: SetVelocityField wraps the field's buffer as the transform's parameters, so a field whose producer has not run is not a supported state. A one-line reply on that thread would close it.
"New + Graft" is written five ways, and MakeGraft is a protected static that other modules cannot reuse.
EXPECT_EQ on 4,096-element vectors prints both vectors on failure, and GTest.cxx:54-55 hardcodes 16.0 and 15.0 next to the named imageSize.
The PR body still ends with the "Generated with Claude Code" footer; the AI disclosure in <details> is enough (non-blocking).

This review ran as /code-review max (Standards and Spec axes) on Opus 5.5. Once these are addressed I will re-check, then request a Greptile review and mark the PR ready.

gdevenyi · 2026-11-03T03:39:38Z

Thanks @hjmjohnson. I amended the single commit; the new head is 3bf7382.

Adaptor input (blocking): GradientRecursiveGaussianImageFilter grafts only when ShouldIReleaseData() is true. With the flag off it behaves as on main. The graft also copies an adaptor's pixel accessor, so the flag-on path keeps the selected element too. The new NthElementImageAdaptor GTest:
- flag off: passes on main, fails with the always-graft version;
- flag on: fails on main and fails without the accessor copy.
Masked BSplineSyN: its mask resampler now gets a graft. It has only the fixed-mask resampler; the moving mask only reaches SyN's four resamplers, which already get grafts. The release-data GTest now has masked SyN and BSplineSyN cases. Each fails when its mask grafts are removed.
Derivative filter input: after the loop, m_DerivativeFilter->SetInput(nullptr) drops the last graft.
TimeVaryingBSplineVelocityField: moved out. itkTimeVaryingBSplineVelocityFieldImageRegistrationMethod.hxx is now the same as on main, including the mask resampler at line 738. Making that method work with the flag also needs changes in TimeVaryingBSplineVelocityFieldTransform:
- its B-spline reconstruction takes the control point lattice, which is the transform's parameters, as a releasable input;
- both integrators share the reconstructed field;
- UpdateTransformParameters feeds the field to an adder.
That will be a separate PR with its own test.
Raw buffer access: the test uses iterators now.

Non-blocking:

Replied on the velocity-field thread.
The GTest derives the image center from imageSize. It reports the first difference instead of comparing whole vectors with EXPECT_EQ.
I left MakeGraft as it is. One helper for all five places would need a public home in Core/Common. I can do that in a follow-up if you want it, but I kept this PR's API unchanged.
The "Generated with Claude Code" footer is no longer in the body.

Tests (local, every test binary under ulimit -v 8 GiB):

The release-data GTest (5 cases) and the gradient GTests pass at 1, 4, 8 and 16 threads. With every changed header swapped back to main, all 5 release-data cases fail.
All 575 tests of the registration build pass (RegistrationMethodsv4, Metricsv4, DisplacementField, ImageGradient, ImageGrid, Smoothing and dependencies). The build adds no warnings.
SyN harness (MNI152 T1 at shrink 4; CC, Mattes and MSQ in float and double; with and without derivative downsampling and a mask) compared with cmp against main: 18/18 identical.

gdevenyi · 2026-11-03T04:14:42Z

The time-varying B-spline velocity field part of the GlobalReleaseDataFlag fix is now #7096, stacked on this PR. It also needed changes in TimeVaryingBSplineVelocityFieldTransform and its parameters adaptor.

Copilot

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

gdevenyi · 2026-11-03T13:12:34Z

@hjmjohnson, a heads-up on one change in 9b7034c that reaches outside the registration code. To fix Greptile's P1 on the metric grafts, ImageAdaptor::Graft now also copies the pixel accessor, instead of each graft site copying it. Before this, a graft of an NthElementImageAdaptor or VectorImageToImageAdaptor read element 0, whatever the source selected, and I can't think of a caller that relies on that.

With this change, the gradient filter's CopyPixelAccessor helper is removed. ImageAdaptor.GraftKeepsThePixelAccessor covers the change directly. The metric, gradient-filter and adaptor tests all fail without the copy.

The velocity field interpolator P1 is fixed too: it points back at the velocity field after integration.

These pass: the ImageAdaptors, ImageGradient, DisplacementField, Metricsv4, Smoothing and RegistrationMethodsv4 tests in my build (61/61), and the 7 release-flag GTests. #7096 is rebased on this as c968cc5, and its 9 tests pass.

hjmjohnson · 2026-11-06T00:34:54Z

Force-push 4209b16..ccb7da4 is a plain rebase on main, with no content change. The only conflict was in the RegistrationMethodsv4 GTest list, where main had added itkImageRegistrationMethodv4VirtualDomainGTest.cxx on the same line; both entries are kept.

hjmjohnson

Re-reviewed at ccb7da4 (the rebase on main). Every concern from my earlier review is addressed, and the Greptile threads check out against the code: the accessor copy for NthElementImageAdaptor, the mask grafts, the derivative filter releasing its last graft, the interpolator restore on both paths, and the iterator-based test. Thank you for the thorough rework.

Three new should-fix points are inline; one of them is a wrong-result case for VectorImageToImageAdaptor. Behavior claims are reasoned from the code, not run.

Nits that have no changed line to anchor on:

itkBSplineSyNImageRegistrationMethod.hxx lines 161 to 174 still pass ungrafted fields to the composers, where SyN's composers are grafted. Nothing reads the released field before SetDisplacementField replaces it, so this is consistency only.
The tests for the ImageToImageMetricv4 and TimeVaryingVelocityFieldTransform fixes live in RegistrationMethodsv4, so a build without that module does not run them.
The commit body says "Results are unchanged"; that holds for registration, while ImageAdaptor::Graft results change on purpose.
The graft idiom is still written several ways (MakeGraft, a lambda, and inline copies). Optional.

hjmjohnson · 2026-11-06T00:55:41Z

    // Now copy anything remaining that is needed
    this->SetPixelContainer(const_cast<Self *>(imgData)->GetPixelContainer());
+    // The accessor selects what the adaptor reads, for example the element of an NthElementImageAdaptor.
+    this->m_PixelAccessor = imgData->m_PixelAccessor;


should-fix. Copying only the accessor leaves a VectorImageToImageAdaptor graft inconsistent. Superclass::Graft calls ImageAdaptor::CopyInformation, which calls m_Image->CopyInformation(data) with the adaptor as source. The adaptor reports one component per pixel, so the graft's internal VectorImage gets vector length 1 while the copied accessor has the source's length. GetPixel at any index other than zero then computes the wrong buffer offset, and once the graft enters a pipeline UpdateOutputInformation calls UpdateAccessor, which resets the accessor's vector length to 1 as well. The gradient filter and metric grafts would then read wrong values for VectorImageToImageAdaptor inputs whenever the input is released.

Suggested fix: graft the internal image (m_Image->Graft(imgData->m_Image), which copies the vector length) before copying the accessor, or set the internal image's component count from the source's internal image.

Thanks, confirmed and fixed in 1a28b6f, as you suggested. Graft now calls m_Image->Graft(imgData->m_Image) in place of SetPixelContainer, so the internal VectorImage gets the source's vector length, then copies the accessor. With the old code, the strengthened test reads 1, 101 and 102 where 201, 202 and 212 are expected.

hjmjohnson · 2026-11-06T00:55:41Z

+  componentAdaptor->SetExtractComponentIndex(2);
+  auto componentGraft = ComponentAdaptorType::New();
+  componentGraft->Graft(componentAdaptor);
+  EXPECT_EQ(componentGraft->GetPixel({}), 3.0F);


should-fix (test). Index {} has buffer offset 0 for any vector length, so this passes even when the graft's vector length is wrong. Please also read a nonzero index, for example componentGraft->GetPixel({ { 1, 1 } }), and read one value through an iterator after componentGraft->UpdateOutputInformation(). That test settles the comment on itkImageAdaptor.hxx.

Done: the images now hold a different value at every pixel and component (100 c + 10 y + x). The test reads index (2, 1) of both grafts, then calls UpdateOutputInformation() on both and checks every pixel through an iterator. It fails with the previous Graft.

hjmjohnson · 2026-11-06T00:55:41Z

+}
+} // namespace
+
+TEST(SyNImageRegistrationMethod, GlobalReleaseDataFlagDoesNotChangeTheResult)


should-fix. The suite name SyNImageRegistrationMethod is already used by itkSyNImageRegistrationMethodGTest.cxx in the same GTest driver, and Documentation/AI/code-review-lessons.md asks for one unique suite name per .cxx file. That file also already has a GlobalReleaseDataFlagGuard class and a GlobalReleaseDataFlag test, so this file adds a second copy of the guard (line 118). Either move these SyN cases into the existing file and reuse its guard, or rename the suites here, for example SyNImageRegistrationMethodReleaseData.

Done: every test in itkSyNImageRegistrationMethodReleaseDataGTest.cxx is now in one suite, GlobalReleaseDataFlag (for example GlobalReleaseDataFlag.SyNResultIsUnchanged), which no other file uses; #7096's two tests follow it. I kept the file's own guard, which is in its anonymous namespace, rather than merge the two files.

hjmjohnson · 2026-11-06T00:55:41Z

  {
    // Now copy anything remaining that is needed
    this->SetPixelContainer(const_cast<Self *>(imgData)->GetPixelContainer());
+    // The accessor selects what the adaptor reads, for example the element of an NthElementImageAdaptor.


should-fix (history). This changes what Graft means for every ImageAdaptor subclass, in Core, inside a registration commit. It has its own test and the metric fix depends on it, so please split it into a first commit, for example BUG: Copy the pixel accessor in ImageAdaptor::Graft, with ImageAdaptor.GraftKeepsThePixelAccessor, and keep the registration commit second. Two small points for that commit: Graft changes the accessor without calling Modified(), and the doc block in itkImageAdaptor.h still ends with "then copies over the pixel container" without the accessor.

Done: BUG: Copy the pixel accessor in ImageAdaptor::Graft (1a28b6f) is now the first commit, with ImageAdaptor.GraftKeepsThePixelAccessor, and the registration commit (494697b) is second. Graft calls Modified() after changing the accessor, and the doc block now says it grafts the adapted image and copies the accessor. The first commit passes its tests on its own: 24/24 for ImageAdaptors, ImageGradient and Smoothing.

ImageAdaptor::Graft copied the pixel container but not the pixel accessor, so a graft of an NthElementImageAdaptor or of a VectorImageToImageAdaptor read element or component 0, whatever the source selected. Graft the adapted image, which also carries a VectorImage's number of components, then copy the accessor. A GoogleTest reads both kinds of graft at every pixel, before and after UpdateOutputInformation.

…DataFlag With DataObject::GlobalReleaseDataFlag on, internal mini-pipelines released inputs that are read again: the virtual domain (a resampler reference image), the level images, the masks, the transform fields fed to the composers, inverters and integrators, and the metric images fed to the gradient filters. Feed those pipelines grafts, which share the buffers, and set resampler geometry with SetOutputParametersFromImage. The velocity field interpolator points back at the velocity field after integration. Registration results are unchanged. GoogleTests cover each case with the flag on, including adaptor inputs to the metric.

kwrobot-v1 · 2026-11-06T17:24:51Z

Errors:

Failed to reserve ref 494697b for the merge request: invalid git ref: 'no such commit'.
Failed to run the checks: mr utilities error: failed to list commits of 494697b7684eac6408e48bf6bf82bc099eee5416 for https://github.com/InsightSoftwareConsortium/ITK/pull/7095: fatal: bad object 494697b7684eac6408e48bf6bf82bc099eee5416 .

gdevenyi · 2026-11-06T17:25:36Z

Thanks @hjmjohnson. Pushed 1a28b6f and 494697b on your rebase; #7096 is rebased on them (61a2cab). The three should-fixes are inline. The review-body nits:

BSplineSyN composers: now grafted, as SyN's are.
Commit body: says "Registration results are unchanged"; the ImageAdaptor::Graft change is described in its own commit.
Metric and transform tests in RegistrationMethodsv4: left there for now, because they share the release-flag guard and fixtures with the registration cases. I can move them into Metricsv4 and DisplacementField GTests if you'd rather they run without RegistrationMethodsv4.
Graft idiom: not unified yet. It is still MakeGraft in the registration methods, a lambda in the velocity field transform, and inline in the metric and the gradient filter, because those modules don't depend on each other. A shared helper would belong in Core; I left that out of this PR.

Test results:

ctest -R "Adaptor|ImageGradient|RegistrationMethodsv4|DisplacementField|Metricsv4|Smoothing": 67/67.
GTests: ImageAdaptors 5/5, ImageGradient 10/10, RegistrationMethodsv4 17/17.
With BUG: Make TimeVaryingBSplineVelocityField registration work with release flag #7096 on top: 42/42.
pre-commit passes.

This was referenced Nov 2, 2026

PERF: Coordinate ITK upstream performance fixes (N4 B-spline fit, N4 point set, AdaptiveDenoising box mean and thread race) ANTsX/ANTs#2037

Open

PERF: Avoid displacement field copies in SyN Gaussian smoothing #7078

Merged

greptile-apps Bot reviewed Nov 3, 2026

View reviewed changes

hjmjohnson requested changes Nov 3, 2026

View reviewed changes

Comment thread Modules/Filtering/ImageGradient/include/itkGradientRecursiveGaussianImageFilter.hxx Resolved

gdevenyi force-pushed the bug-syn-global-release-data-flag branch from d701177 to 3bf7382 Compare November 3, 2026 03:38

gdevenyi mentioned this pull request Nov 3, 2026

BUG: Make TimeVaryingBSplineVelocityField registration work with release flag #7096

Draft

gdevenyi marked this pull request as ready for review November 3, 2026 12:00

Copilot AI balanced review requested due to automatic review settings November 3, 2026 12:00

Copilot AI reviewed Nov 3, 2026

greptile-apps Bot reviewed Nov 3, 2026

View reviewed changes

Comment thread Modules/Registration/Metricsv4/include/itkImageToImageMetricv4.hxx Resolved

Comment thread Modules/Filtering/DisplacementField/include/itkTimeVaryingVelocityFieldTransform.hxx Resolved

gdevenyi force-pushed the bug-syn-global-release-data-flag branch from 3bf7382 to 9b7034c Compare November 3, 2026 13:10

github-actions Bot added the area:Core Issues affecting the Core module label Nov 3, 2026

gdevenyi force-pushed the bug-syn-global-release-data-flag branch 2 times, most recently from bf0d867 to 4209b16 Compare November 4, 2026 00:39

hjmjohnson force-pushed the bug-syn-global-release-data-flag branch from 4209b16 to ccb7da4 Compare November 6, 2026 00:34

hjmjohnson requested changes Nov 6, 2026

View reviewed changes

hjmjohnson marked this pull request as draft November 6, 2026 00:55

gdevenyi force-pushed the bug-syn-global-release-data-flag branch from ccb7da4 to 494697b Compare November 6, 2026 17:24

Conversation

gdevenyi commented Nov 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

gdevenyi commented Nov 2, 2026

Uh oh!

hjmjohnson commented Nov 3, 2026

Uh oh!

greptile-apps Bot commented Nov 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 3/5

Findings

Summary

Diagram

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

hjmjohnson left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

gdevenyi commented Nov 3, 2026

Uh oh!

gdevenyi commented Nov 3, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

gdevenyi commented Nov 3, 2026

Uh oh!

hjmjohnson commented Nov 6, 2026

Uh oh!

hjmjohnson left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kwrobot-v1 Bot commented Nov 6, 2026

Uh oh!

gdevenyi commented Nov 6, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

gdevenyi commented Nov 2, 2026 •

edited

Loading

greptile-apps Bot commented Nov 3, 2026 •

edited

Loading