Repository navigation
LCU 0.8.5
LCU 0.8.5
This release shrinks and hardens the Linux window-targeted input translation after a second review of 0.8.4. It adds no new harness verification claims and no new tested app pairs; Windows remains deferred.
Changes
- Held keys are no longer translated.
key_downandkey_upnow always go to the original service unchanged, and the hold bookkeeping is gone. 0.8.4's claim that translated holds were "always released" was wrong (see the correction in the 0.8.4 notes): partial chord releases, key aliases such asShiftandShift_L, a failed release and turn cleanup could all leave a modifier held. On a GTK 4 window a held key now behaves as in the original engine, which delivers it withXSendEventand so is ignored. To hold a key for a GTK 4 window, activate it and use the desktop-levelkey_down/key_up(nowindow).press_key(atomic chords, with the original engine owning the whole chord),click(including double),scroll,dragandmovestay translated for GTK 4, andscrollfor Qt. - Everything that can change focus or input runs through one queue.
activate_window, desktop-level input, window-targeted input of every kind,typeText, pass-through fallbacks and untranslated calls are serialized with translated requests, so nothing interleaves between a translated request's final focus check and its input (anactivate_windowof another window could previously move a key to that window). Read-only calls (list_windows,list_apps,get_screenshot,get_window_state) bypass the queue. The focus is still verified immediately before the desktop-level call. - Window ownership comes from the X server.
_NET_WM_PIDis the client's own claim in its own PID namespace, so a client sharing the X socket from another namespace (Flatpak, containers) could advertise an id that collided with an unrelated local process. LCU now asks the X server (X-ResourceXResQueryClientIdswith the PID mask, theSO_PEERCREDid; nothing for a remote client) and translates only when it equals_NET_WM_PID, the machine is this host and the process shares LCU's PID namespace. Anything else, or an unavailable query, fails closed to the original behavior. - New system package:
libxres1(X.org X-Resource client library, about 7 kB, Ubuntumain) was added to the installer'sSYSTEM_PACKAGES. Silo's image needs it as well; without it, or withoutpython3, GTK 4 and Qt windows behave as in the original engine. LCU_LINUX_INPUT_TRANSLATION=offand the per-pairnative_inputswitch are unchanged. See Linux window-targeted input and the verification note.
Source and packages
The archives were built from commit 55d8a00.
| Archive | SHA-256 |
|---|---|
lcu-0.8.5-darwin-arm64.tar.gz |
518187dcf66eb1134cb199957b81012a22fb213f3f8759df783305a22f29a963 |
lcu-0.8.5-linux-arm64.tar.gz |
3a0856210656207a1d80b08077888701a0278407343d5276b1f6607c637c3722 |
lcu-0.8.5-linux-x64.tar.gz |
0caa6e8fbbbfef8c6c8b9b7de9b2b5c123869c76023f9f75f2b7e5cbd3410f30 |
Each archive has a matching .sha256 sidecar. Archives contain no OpenAI application binaries or copied original instructions. The Linux archives are release builds of that commit; each tests/run.sh gate run tested its own build of it, and the native x86-64 run tested the release build of the x64 archive. The macOS archive was built from /Applications/ChatGPT.app (26.930.21537) as for earlier releases. Archives are not byte-reproducible across builds.
Gates
tests/run.shpassed in disposable Linux ARM64 (native) containers against the official ChatGPT 26.915.31945 (the pinned package) and 26.928.31416 packages, and in AMD64 (emulated) containers against the same two: 325 Python tests (3 environment skips), offline in-place installation, agent registration, the exported MCP contract, the original GTK desktop suite with its independent save oracle, the adapter-path suite, the GTK 4 window-targeted suite (now withkey_down/key_uppassed through unchanged, anactivate_windowracing a translated key in both orders and a desktop-level key issued together with a translated one), the controls suite, and the read-only mount gate. The ARM64 read-only gate ran with bubblewrap allowed and required the original sandbox to be active, with the negative control and the stricter-host-profile check; under emulation bubblewrap cannot start, so the sandbox could not be required on AMD64. 26.915.31945 has nokey_down/key_up, so the pass-through checks run only against 26.928.31416. The first ARM64 run against 26.928.31416 failed one unrelated host-harness unit test (test_macos_acceptance, a pseudo-terminal drain count of 1044223 against 1048576, a timing assertion); the unchanged rerun passed.- Native x86-64: on an Ubuntu host with native Docker, the release build of the x64 archive ran the read-only mount gate with the original sandbox required (the desktop suites as the unprivileged account) against both packages; both passed. As for 0.8.3 and 0.8.4, that host's full
tests/run.shwas not used. The run used a temporary directory and a throwaway image and volumes, all removed; no installed application, VM or user data on that host was touched. - The wrapper logic is covered by
adapters/test/linux-sky-service.test.mjs(23 in-memory tests against a fake desktop that follows the original engine's chord semantics; 6 fail against the 0.8.4 logic) andnode --test adapters/test/*.test.mjspasses (59 passed, 8 environment skips, 0 failed). The X-Resource query was also exercised directly against a real Xvfb in an Ubuntu 24.04 container: it returned the connecting process's id, and failed closed withoutlibxres1. - The macOS archive was built; no desktop action ran on macOS and no model-driven run was made for this release. Fractional scaling, other window managers, real Chrome or Electron applications, and a window from a genuinely remote or other-PID-namespace X client were not exercised; those rules are covered only by the in-memory tests. A Silo-like guest (Xvfb run as the same user by the desktop service) is the same shape as the gate's Xvfb fixtures, where the translated GTK 4 suite passed.