The workspace members, the path dependencies, the include_str! and test
fixture paths, the recorder's indexed source keys, and the Nix source
filter and path literals follow the new layout. Cargo.lock is unchanged.
Every package of the workspace now takes its version from [workspace.package]
(version.workspace = true) instead of its own number. The numbers before were
0.1.42 (Windows libraries and packages), 0.1.0 (KDE, the macOS version crate,
komorebi-shortcuts and the e2e guest crates) and 0.1.1 (macOS libraries); the
unified release is 0.2.0 for all of them.
Cargo.lock only changes the version of the workspace's own packages.
Not changed here: the version text in schema doc comments and the generated
schemas (next revision), and the shim packages, which are deleted later and
only pick up the number meanwhile.
Verified: cargo metadata --offline --locked succeeds.
Pure formatting by the same nightly rustfmt invocation the nix fmt check uses (tools/format-all, edition 2024, the tree's rustfmt.toml). No code change. Applied last on the candidate because it touches many files other agents were editing. Verified: tools/format-all --check --tree windows passes afterwards; light gates (fast-check) after all three.
Pins the cost rule the macOS fix depends on: Registry::insert of a new id runs no
handle's drop, a replaced handle is handed back and dropped only when the
caller drops it, get drops nothing, and remove hands the handle back. A probe
handle counts its drops.
Verified: native cargo test -p komorebi-backend: 40 unit, 7 doc, 1 compile_fail
passed, 0 failed.
Candidate for the Windows e2e gate: the Windows Backend switches and lock order (wip-u-kfw) with the macOS registry and Window methods (wip-u-kfm). Merge of two independent stacks that both touch crates/komorebi-backend; conflicts, if any, are resolved in this revision.
Windows call sites read the frame about 60 times and rely on the Win32
error (or unwrap_or_default). The shared window_rect is built on window(id),
which also asks title, class, exe and path and gives None for every failure.
Add the provided method Backend::frame(id) -> Result<Rect>: WindowGone for
an unknown window, Platform for a failed read. The default goes through
window() (a table lookup on KDE through Shared<B> and on the fake, so the
cached/requested-rect semantics of seam.md section 2 row 2 are kept);
WindowsBackend overrides it with Window::window_rect (one DWM call) and
MacosBackend with MacosApi::window_rect (the two Accessibility reads, no
metadata). The contract tests check frame agrees with window().rect and is
WindowGone for an unknown id; FakeBackend gets a test. A provided method, so
no adapter had to change to compile.
Not a behaviour change (nothing calls it yet). Verified: tools/fast-check
(native crate tests pass, Windows clippy -D warnings, KDE check + lib tests,
macOS fake-CC type-check) PASS; macOS cannot be run here.
Step 11(ii), the pure part. macOS cannot act on a window from its id: it
needs the Accessibility element, the pid and the observer. Registry<H> is
that table, generic over the handle so that its rules run on any host.
Rules, written at the top of registry.rs and tested: thread-safe; a
lookup clones the handle and releases the lock before anything is done to
it; a removed or replaced handle is handed back to the caller instead of
being dropped under the lock (dropping a macOS handle can make an AX round
trip, which must never happen with a lock held: seam.md 3.3); a stale or
unknown id is Error::WindowGone, never a panic; an entry ends on remove,
remove_pid (application quit) or a failed call; re-inserting replaces.
Nothing uses it yet; no daemon changes.
Verified: native cargo test -p komorebi-backend: 38 unit (9 new) + 7 doc
(1 new) + 1 compile_fail passed, 0 failed. Not run: the Windows, KDE and macOS
checks (the new module is std-only and no other file changed).
The doc of Backend::focus said Windows and macOS do focus and then center_cursor_in; writing the adapters showed that Window::focus on both takes mouse_follows_focus and centres the pointer itself (Windows also defers the focus while the window slides in). Doc comment only.
Verified: see the gates below; native cargo test -p komorebi-backend passes (doctests unchanged).
komorebi_backend::contract: a Harness trait (how a test puts a live window on the desktop of the backend under test) and 14 checks, run by check_all::<H>(): windows listed by id; window agrees with windows/is_window/window_rect; an unknown window is not a window; every window command on an unknown or on a destroyed window fails with WindowGone naming that window; a destroyed window is neither alive nor listed; set_rect on a live window succeeds and the frame stays readable (a request: the value read is not checked); hide/restore tolerate repeats; focus names a known window; capabilities and Unsupported agree (supported means never Unsupported, unsupported means always Unsupported, with a live or unknown window); monitors are sane and found by point; Batch scopes nest around writes; the backend works from another thread; it is usable as a trait object.
Only what the Backend documentation promises is checked, because a write is a request and platforms act at different speeds. The fake passes under three configurations (plain, with a frame offset, with no capability at all, as macOS would be) and a deliberately broken backend (set_rect answers Ok for a gone window) is shown to fail exactly the two checks that cover it.
Verified: native cargo test -p komorebi-backend: 29 unit tests passed, 6 doctests passed, 1 compile_fail doctest passed, 0 failed; clippy --all-targets -D warnings clean.
The in-memory desktop of komorebi-for-kde (backend/fake.rs) moved into the shared crate, rewritten against WindowId/MonitorId/Result and the generic Backend only (look, border colours, wallpaper and drag frames stay out: they are KDE's).
Same behaviour as the KDE fake (permissive mode, minimum sizes, user_* helpers, maximize, fullscreen, workspaces, the ordered Command log with 15 of KDE's 18 variants; close() leaves the foreground window alone, as KDE's did), plus:
- failure injection: fail_next(Op, Error) and fail_next_for(Op, WindowId, Error), oldest first, one per call; the failed call is still logged as a command and changes nothing else (none existed on KDE);
- frame offsets: set_frame_offset / set_window_frame_offset make window_rect differ from the request by a constant, as the Windows shadow border does;
- capabilities: all four on by default, set_supported switches one off and its operations return Unsupported without being logged;
- batches: begin/end are counted (open_batches) and logged as BatchBegin/BatchEnd;
- FakeBackend<X>: a platform extension logs its own commands in order with record_ext (Command::Ext(X)); the shared fake is FakeBackend<NoExt> (new/permissive/with_one_monitor), an extension uses empty/permissive_ext/one_monitor because default type parameters are not inferred.
Verified: native cargo test -p komorebi-backend: 24 unit tests and 6 doctests passed, 0 failed (the KDE fake tests were ported and 6 failure/offset/capability/batch/extension tests added); clippy --all-targets -D warnings clean; cargo fmt applied.
New crate crates/komorebi-backend (edition 2024, root workspace member, depends only on komorebi-layouts for Rect).
- WindowId and MonitorId: opaque u64 newtypes. KDE and macOS ids convert with from_u32/to_u32; a Windows HWND/HMONITOR goes through from_hwnd/to_hwnd, which keep the bits of the isize unchanged (isize as usize as u64), so no handle is lost whether or not it fits in 32 bits (the e2e samples stay below 2^32, which is an observation, not a guarantee).
- WindowInfo and MonitorInfo: plain data; only id, pid, rect, minimized and alive are required, the rest (title, class, exe, path, fullscreen, maximized, normal; name and serial id of a monitor) are Option (seam.md consequence 6).
- Error: WindowGone(id) (finding 9: macOS uses a failed write as its signal that a window died), Unsupported(Capability) and Platform(cause).
- Backend: only the operations of seam.md section 2.1 that at least two platforms have (5 queries, 9 commands), a Capability query for workspaces, maximize, opacity and focus-follows-mouse (their methods default to Unsupported), begin_batch/end_batch with a no-op default plus the Batch guard (finding 8), the documented set_rect contract (a request; reads after a write return what the platform defines, consequence 2 as corrected) and the thread contract (any non-main thread, no waiting for the platform's main thread: consequences 3 and 4, written as TARGET rules). Every doc comment says which platforms implement the method.
Verified: native cargo test -p komorebi-backend (2 unit tests, 3 doctests, 0 failed) and cargo clippy -p komorebi-backend --all-targets -- -D warnings clean. The Windows/KDE/macOS gates are not affected yet (new crate, nothing depends on it).