tests-windows never executes an adoption: git-filter-repo is not installed there, so the eight adopt suites skip and v1.0.0's "adopt works on Windows" is unverified

Lane: openreposhape-3 (openRepoShape-3)

(lane openRepoShape-3, 2026-10-12T21:40Z) Filed by lane 3's v1.0.0 triage. Proposed holder: lane openRepoShape-4, on Brett Heap's word at 21:21Z, "try to assign to lane 4 some tasks". Lane 4 claims it under Rule 1 as usual. Not claimed by lane 3.

What is wrong. v1.0.0 is "core flows, all OSes" (Brett Heap, 2026-10-12): every AGENTS.md procedure, adopt included, must work end to end for ordinary repositories on Linux, macOS and Windows. But no adoption is ever executed on Windows, not in CI and not anywhere else.

.github/workflows/tests.yml's tests-windows job does not install git-filter-repo, on purpose. Its comment says the adoption suites "SKIP without it — which is the honest outcome here: what this job is for is the entry point and the materializer on Windows, not a second platform's worth of history rewriting." That was written before the release aim was set.
Eight suites skip when shutil.which("git-filter-repo") is None, so all of them skip on Windows:
- tests/test_adopt_e2e.py;
- tests/test_adopt_execute_refusals.py;
- tests/test_adopt_mount_ignored.py;
- tests/test_adopt_leg_paths.py;
- tests/test_adopt_submodules.py;
- tests/test_adopt_verify.py (the part that executes);
- tests/test_adopt_leg_path_collisions.py;
- tests/test_adopt_ignored_shape_files.py.
Between them they hold about 180 test functions before parametrisation.
A green PR run, 38303387703 on setup-project.py: the root clone of --local-remote-dir is --no-local and its two leg clones are rewritten to file://, and the two session fixtures that clone a scaffold do the same, so a maintenance run packing the just-pushed objects mid-walk no longer dies failed to copy file to ... No such file or directory (#440) #467, shows the gap in its counts:

job passed skipped

tests (ubuntu) 5042 10

tests-macos 5030 22

tests-windows 4566 486

The rest of Windows' skips are the WINDOWS_SKIP entry-point tests (setup.sh and shebang stubs). Those are covered case for case by tests/test_setup_project_py.py.

What v1.0.0 needs.

Install git-filter-repo in tests-windows at the version the ubuntu job pins (2.47.0), and let the adoption suites run there.
Fix what an ORDINARY repository hits on Windows in adopt-project.py: paths, line endings, how git filter-repo is found and launched, and read-only object files on cleanup. Each fix is a PR on this issue, or on a sub-issue if it is big.
A failure only an exotic input triggers (non-UTF-8 names, case-only twins, reserved names) goes on Tracking: adopt-project.py, converting an existing repository in place #457's checklist as a known limit, not into the fix.
Measure what the suites add to the Windows job's ~48 minutes. If it is too much to pay on every pull request, say so on this issue so Brett can choose: the suites on every PR, or a workflow_dispatch run before the tag.

Related, for the release gate. Both tests-macos and tests-windows are skipped on a push to main (if: github.event_name != 'push'). So "a full run green on all three OSes on the release commit" needs a workflow_dispatch run on that commit before the tag.

Parents: #461 (tests, CI and docs) and #457 (adopt).

CLAIMED — lane openRepoShape-4, session bf1c72df-1c56-4ac9-aace-3ddd139d7ea3@Eagle, 2026-10-12T21:41:40Z, for #471

Logged in lanes/log/openRepoShape-4.md at 72e34cd625b92f04edba86d112c99a61e26f0e77.

The three reads (lane-collision-protocol Rule 1):

1. existing `CLAIMED —` comments on opensoft/openRepoShape#471:
   none
2. `gh pr list --repo opensoft/openRepoShape --state all --search "#471"`, kept where the row or its body names `#471`:
   none
3. `git ls-remote --heads git@github.com:opensoft/openRepoShape` branches naming `471` as a whole token:
   none

(lane openRepoShape-4, 2026-10-12T22:51Z) Probe results, Brett Heap's ruling and the plan.

Probe 1. Run 38312924968 used tests-windows only, with git-filter-repo 2.47.0 installed, over all 23 files that skip without it. #471's body lists 8; two more skip under other spellings.

Result: 1 failed, 1696 passed, 34 skipped in 19m30s.
Adoption works on Windows apart from one ordinary case.

The failure is ordinary, not exotic. The test is test_adopt_ignored_shape_files.py::test_the_assembly_bootstraps_where_it_failed_with_shape_copy_missing[its-own-attributes].

When a source tracks its own root .gitattributes, the shape's * text=auto eol=lf goes beside it, at shape/.gitattributes, and governs only shape/**.
The 9 pinned copies written at root paths get only the source's attributes: Makefile, AGENTS-shape.md, the scripts/ copies, contracts/repository-naming.yaml and the workflow.
Under Git for Windows' default core.autocrlf=true, ANY root .gitattributes without the shape's rule checks them out with CRLF. Even one that only says *.png binary does.
validate-pins.py then reports shape-copy-drift for all 9, and make bootstrap exits 1.
It reproduces on Linux under core.eol=crlf and under core.autocrlf=true, with byte-identical digests.

Ruling (Brett Heap, 2026-10-12T22:48Z): option A. The split appends a marked block to the source's own root .gitattributes, one /<copy> text eol=lf line per pinned copy the root file governs, so a Windows clone works first time with no manual step.

It overrides "never written over" for that one file, and writes only the block.
Verification gets an exception for it.
update-shape.py --add extends the block.

Rejected:

(B) hashing with normalized line endings, which would reverse Templates ship a .gitattributes so a Windows clone of a scaffolded project passes its own gate (follow-up of #49) #51;
(C) refusing a kept root .gitattributes;
(D) a follow-up only, which leaves a Windows clone made before the follow-up broken.

Plan: two PRs on this issue.

PR 1, ci/windows-adoption-suites:
- tests-windows installs git-filter-repo 2.47.0 and runs the whole suite, as now;
- the comments that said Windows lacks it on purpose are corrected;
- the one case above is xfail(strict=True) on Windows, naming this issue.
- It is Part of #471.
PR 2, fix/adopt-gitattributes-lf-block:
- implements A, pinned on Linux under core.eol=crlf;
- removes the xfail;
- if it is small, makes the 11 tests whose gh or git-filter-repo stub is a #! script runnable on Windows. Today these ordinary failure paths never run there: gh refusing after the push, and git filter-repo failing.
- Closes #471.

The 34 skips, apart from those 11:

20 are exotic or platform limits, and go on Tracking: adopt-project.py, converting an existing repository in place #457's checklist:
- 11 use names that cannot exist on NTFS or are not UTF-8: 3 TAB/control, 4 non-UTF-8, 4 newline/TAB/backslash;
- 9 use tracked symlinks without core.symlinks=true.
3 are test-helper only.

Timing, for Brett's choice (the suites on every PR, or a workflow_dispatch run before the tag):

Only 178 tests are new to Windows. The other 1553 in these files already run there. They add about 8 to 11 minutes: 2.8 s per gated test, which Linux's with/without difference (599.6 s against 349.1 s) confirms.
For scale, the release dry run on main d6df36b (run 38309655530, all green) took this much pytest time per job:

job pytest time passed skipped

tests 11m52s 5171 10

tests-macos 36m51s 5159 22

tests-windows 1h07m56s 4692 489

One note for the release gate. tests.yml's concurrency group is the workflow plus the ref, with cancel-in-progress: true. A workflow_dispatch on main therefore cancels the last landing's Linux push run, and the next landing cancels it. Run the release commit's three-OS check on a throwaway branch at that commit, as the dry run did: ci/v1.0.0-dry-run-d6df36b, since deleted.

docs/release-notes/v1.0.0.md: brought up to main 437a360, so the note…

Merge origin/main (47b1e06) into fix/adopt-gitattributes-lf-block, so #…

README.md, AGENTS.md and the handbook: a root `.gitattributes` the so…

adopt-project.py: a source `.gitattributes` git tracks as a symlink g…

README.md: `update-shape.py apply --add` names the file it adds in a …

Merge #497 (ci/windows-adoption-suites at 6824181) into fix/adopt-git…

tests: the two index-only-link helpers #471's review round added hand…

Merge origin/main (f391e58) into fix/adopt-gitattributes-lf-block, so #…

tests/test_adopt_pins.py: the module fixture deletes its copied repos…

Merge origin/main (d41cc29) into fix/adopt-gitattributes-lf-block: th…

docs/release-notes/v1.0.0.md: brought up to main d41cc29, so section …

adopt-project.py: #471's four imports, `ATTRIBUTES`, `eol_block_paths…

Merge origin/main (3121945, #494) into fix/adopt-gitattributes-lf-blo…

tests/test_repo_hygiene.py and tests/test_readme_sections.py: a handb…

Merge main (5ed3e5d, #511) into l3-519-handbook-section-hashes: #511'…

Merge main 5ed3e5d (#511's squash for #471) into fix/adopt-plan-quote…

docs/handbook.html: `apply --add` names an added file in an adopted r…

docs/release-notes/v1.0.0.md: brought up to main 8beb3d9, so section …

docs/release-notes/v1.0.0.md: brought up to main 3121945, so section …

docs/release-notes/v1.0.0.md: brought up to main 5ed3e5d, so section …

update-shape.py: `apply` refuses to rewrite a file that has a second,…

adopt-project.py: `check` finds, and `execute` refuses before any leg…

adopt-project.py: #654's question writes each file of the split in pl…

adopt-project.py and shape_materialize.py: #471's line-ending block n…

Navigation Menu

Repository navigation

tests-windows never executes an adoption: git-filter-repo is not installed there, so the eight adopt suites skip and v1.0.0's "adopt works on Windows" is unverified #471

5 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

job	pytest time	passed	skipped
`tests`	11m52s	5171	10
`tests-macos`	36m51s	5159	22
`tests-windows`	1h07m56s	4692	489

job	passed	skipped
tests (ubuntu)	5042	10
tests-macos	5030	22
tests-windows	4566	486

tests-windows never executes an adoption: git-filter-repo is not installed there, so the eight adopt suites skip and v1.0.0's "adopt works on Windows" is unverified #471

Description

Activity

brettheap commented on Oct 12, 2026

brettheap commented on Oct 12, 2026

5 remaining items

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions