Skip to content

armgit / cua

Public Imported from GitHub
0

Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.

14.6 KB
Raw Blame

Discovery foundation validation

Note (18 Oct 2026): Cua Cloud (Fleets) ended on 18 Oct 2026. This is a dated record of validation run on it; the Fleet references below are historical.

The discovery-only plugin passed a native build and lifecycle checks on Cua Cloud Fleet on September 4, 2026. These results support review of the foundation; the full promotion gates in the test plan remain open.

The September 5 merge-readiness pass records a fresh candidate build and the remaining system-installation gate.

Candidate and environment

The tested source was based on Cua commit 4da67bb61c262c663fa88559d1d3f1a138b3a463. The 27-file plugin tree had SHA-256 32c4c2224bcf260921083797a4c5bfd72dbca0ced70cbd0d24738de2caf83007. This digest hashes sorted relative file paths and raw file contents, each followed by a NUL byte. It excludes this subsequently added validation report.

Component Tested value
Plugin 0.1.0
Omarchy 4.0.1-1
Hyprland 0.56.2-1
xdg-desktop-portal-hyprland 1.4.1-1
Installed Cua Driver 0.22.2
Compositor and plugin compiler GCC 16.1.1 20260728
CMake / Ninja 4.4.3 / 1.13.2
Shared C++ runtime libstdc++.so.6.0.36

This Fleet image matches the Hyprland baseline but has older Omarchy and Cua Driver versions than the complete acceptance target. Driver integration was not exercised by this plugin run.

The exact compiler package was selected from Hyprland's package build metadata and its Arch signature verified. The plugin used the shared C++ runtime already loaded by the compositor. The actual Arch recipe ran with makepkg --nodeps because CMake and Ninja were task-local Python wheels. Its independent checks for installed Hyprland and header versions still ran.

Artifact SHA-256
Arch package a9e401ca9ac1c1f0532519f74a650e84ff679ce2e29ab8751b85903e1e0a7e99
Packaged module 842a4aa768e95434bba7c9ac2c56c4846a518f0cff883f8c9fee7b2480736f43

Results

  • Native package build and check(): six of six CTests passed, including the production Linux SOCK_SEQPACKET transport.
  • Package inventory: shared module and license, plus package metadata; exact hyprland=0.56.2-1 dependency; no configuration edits or autoload.
  • Real Hyprland load: matching ABI fingerprints, disabled initial transport, and enabled discovery socket with mode 0600 after config reload.
  • Ten consecutive live unload/reload cycles: all six mutation types returned background_unavailable, measured compositor state stayed unchanged, old connections closed, and replacement transports published fresh epochs.
  • Disable/re-enable: socket and client closed, disabled epoch became zero, and re-enabling created a fresh epoch.
  • Synthetic ABI mismatch: altering the compiled fingerprint caused explicit load refusal before command/socket registration. The matching module loaded afterward. This was not a second real Hyprland version.
  • Supplemental native transport fixture: same-UID acceptance, different-UID rejection through SO_PEERCRED, and independent sockets/epochs. Only that fixture's permissions were relaxed to test peer rejection independently of filesystem access. This was not two simultaneous compositor instances.
  • macOS review-host suites: six of six tests passed in normal, ASan/UBSan, and TSan configurations. macOS uses a transport shim and mocked compositor APIs.

The reusable live check is live_discovery.py. It measures active-window identity, workspace, cursor coordinates, window geometry/mapping, and focus history. It does not prove application delivery, complete stacking order, physical input isolation, or behavior during an active foreground grab. The peer and synthetic ABI checks were supplemental fixtures; they are not claimed as automated by that script.

Runtime-linkage failure investigated

An earlier compiler staging directory omitted its separately packaged shared runtime. GCC selected libstdc++.a; a later reload of that module hung Hyprland. Debugger attachment was unavailable, so static linkage remains the leading suspected cause rather than a backtrace-confirmed diagnosis.

Correcting the runtime search path produced a module linked to the compositor's shared C++ runtime. A fresh Fleet passed the repeated lifecycle checks above. The build now rejects modules without a shared C++ runtime dependency.

Follow-up at 3789f0e6e

A second Fleet run on September 4, 2026 (September 5 UTC), tested commit 3789f0e6eeb350587fe48c4da35b327975a5bd9c through the Cua Sandbox SDK. The image had the same component versions listed above. The source archive contained the committed plugin directory and repository license, with SHA-256 27835f9192a35d9b721396642e91a81eb18088075d15c5b4ed2f1f912e159aa4.

The matching signed compiler package and shared runtime were used again. The Arch recipe built successfully with makepkg --nodeps; all six native CTests passed. The downloaded artifacts matched the guest hashes:

Artifact SHA-256
Arch package 2c4881b4a17aef554edcaa1181f73ae2453e39b05c7bb5169db8ae9be34adebf
Packaged module 693d6470a011956e58320de67fae8ed8d4628476e605f74cea018eb2eb5d5705

Additional native checks passed:

  • Two simultaneous Hyprland 0.56.2 processes, nested under the Fleet desktop, exposed distinct discovery sockets and epochs. The live discovery harness passed all six mutation refusals and its compositor-state comparison in each.
  • Restarting one nested compositor closed its old connection and removed its socket. Its replacement used a fresh instance socket and epoch. The other compositor's existing discovery connection remained responsive through both restart and shutdown of its sibling.
  • A genuine Hyprland 0.56.1-3 Arch package, verified against its signature, ran as another nested compositor. Loading the unmodified 0.56.2 plugin explicitly failed with an ABI fingerprint mismatch. No plugin, status command, or discovery socket remained registered, and the compositor still answered version queries. This closes the earlier synthetic-only ABI gap for this specific version pair, not arbitrary upgrades.

The lifecycle orchestration was task-local; these additional checks are not automated by live_discovery.py alone. Initial headless startup failed before loading the plugin; nested Wayland startup succeeded. The nested tests do not certify direct DRM startup, physical input, application delivery, or isolation during a foreground grab. All input mutations remained disabled.

The SDK shell ran as an unprivileged desktop user. sudo -n -l required a password, and the SDK shell interface provided no user-selection parameter. System installation and upgrade-transaction checks therefore remain open; package construction is not evidence of pacman -U installation.

The hosted Linux CI run also passed all six tests at this SHA, using Ubuntu 24.04, Clang 18.1.3, and CMake 3.31.6. It covers the native Unix transport and mocked compositor APIs, not a loaded compositor module. A fresh macOS Debug build passed six tests.

All test compositor units were stopped, the plugin was unloaded, and the original desktop remained responsive with its configuration hash unchanged. No system package was installed and no autoload setting was added. The SDK accepted deletion of the follow-up pool. Immediate lookup still returned it; later pool and namespace lookups returned 403, so final resource absence could not be independently confirmed with that credential. A subsequent successful account-level SDK namespace listing returned no matching namespace. Direct lookup still returned 403; that response alone is not treated as proof of deletion.

Reproducible lifecycle follow-up at 0a67bf5c8

The checked-in nested lifecycle runner passed on Cua Cloud Fleet at 0a67bf5c844f5b358651737fb3545aaed5522007 on September 5, 2026 (UTC), through the Cua Sandbox SDK. The committed plugin, workflow, and license archive had SHA-256 960beaebeba6b4b10e9b08cdeb56e3843313b5e41a4154ca95693615e2a06ce0.

The guest had Omarchy 4.0.1-1, Hyprland 0.56.2-1, portal 1.4.1-1, and installed driver 0.22.2. This run reused the matching-toolchain module from the 3789f0e6e build, SHA-256 693d6470a011956e58320de67fae8ed8d4628476e605f74cea018eb2eb5d5705. Plugin C++, headers, CMake, and packaging were unchanged between that build and this candidate; this run tests the new orchestration, not a new package build.

Both exact-PID-owned nested compositors passed the live six-mutation refusal and measured-state check. One restarted with a fresh socket and epoch; its old connection closed and socket disappeared. The sibling's connection remained responsive through restart and shutdown. Both owned processes stopped, only the original parent remained in the instance inventory, and the parent's plugin list was empty. Seven runner control-flow/cleanup tests also passed on the guest and on the macOS review host. Hosted CI now runs those tests alongside the existing six CTests; CI itself does not launch nested Hyprland.

The guest still lacked renderD128 and passwordless package-install access. No driver was installed, no system package transaction ran, and no input capability was enabled. Nested startup can affect the disposable parent's focus, so parent foreground preservation is not claimed by this lifecycle run. Application delivery, foreground-grab isolation, and direct-DRM behavior remain outside its evidence. The report-only follow-up does not change the tested runner or executable sources.

Outstanding acceptance gates

System installation with pacman -U, dependency-upgrade refusal, the complete Fleet baseline, and physical-host certification remain unverified. Native two-instance and clean-restart evidence is limited to nested compositors. Second-seat delivery, target tokens, operator authorization, driver integration, and the application/foreground-grab matrix remain future work.

The initial source differed from the first Fleet-tested tree only in documentation and formatting: shutdown wording describes the worker join accurately, this report records the evidence, and Markdown/Python formatting follows repository tools. The Python harness's parsed syntax tree is checked against the tested copy. Plugin C++ source, CMake files, and packaging are unchanged.

The guest configuration was restored, the plugin unloaded, and both initial disposable Fleet pools deleted with fresh lookups confirming their removal. The follow-up report changes only this Markdown file after testing 3789f0e6e; it does not change the plugin, packaging, CI, or test harness.

Merge-readiness revalidation (September 5, 2026)

Tested source: e254619817e75019d025e2714a395378ea533a28, after merging 1336a5e419c933eb750bfefe2d9db6aad63235de from main. The merge resolved one installation-document conflict without changing the plugin implementation. The test-helper follow-up refuses optimized Python before validation or compositor access; four regression tests cover that guard and packet exchange. The plugin remains version 0.1.0, discovery-only, and outside the Driver installation/release path. There are no Rust delivery changes in this PR.

A fresh disposable Cua Cloud Fleet claim used Omarchy 4.0.1-1, Hyprland 0.56.2-1, and portal 1.4.1-1. The installed Driver remained 0.22.2 and was not replaced or used as input-delivery evidence. Hyprland's ELF compiler record and the selected plugin compiler both reported GCC 16.1.1 20260728; the guest's default compiler was 16.2.1 and was not used for the plugin. The module linked the guest's shared libstdc++.so.6.0.36.

Validation results:

  • The unchanged Arch recipe built with CMake 4.4.3 and Ninja 1.13.2. makepkg --nodeps skipped dependency-database checks because CMake/Ninja were task-local wheels; the recipe's exact installed Hyprland and header checks ran. All six native Release CTests passed.
  • All 11 Python helper tests passed on the guest and macOS review host. Six portable Release CTests also passed on macOS; those are not Linux proof.
  • Package metadata pins hyprland=0.56.2-1. Its payload contains only the module and license plus their parent directories, with no autoload config.
  • nested_lifecycle.py passed against the freshly built module: two actual owned compositors, six mutation refusals per start, a clean restart, old connection/socket closure, fresh socket and epoch, and sibling liveness. Both children stopped; the original compositor remained responsive and its plugin list was empty. This does not claim unchanged parent focus during nested-window startup.
  • A real sudo -n pacman -U attempt refused because the guest requires a sudo password. A subsequent package query confirmed that the plugin was not installed. System installation and dependency-upgrade refusal remain unverified; package creation and metadata inspection do not replace them.
Artifact SHA-256
Exported plugin source and license archive 4feaafc406f6557ea21b358a5f883cd77053d70a31623483d2ac06b82b4d90e9
Arch package 300cb04ba8c001daeb84f37232fedc135920e5c424ea651d3db1b3cbb3b6d23c
Native module e179eaa165a20d2f0365e0062c1cca8ac4c4caa9532124417636a603ee4ba73d

The remaining merge gates are a recorded discovery-only RFC disposition and the system package-transaction evidence above. Direct-DRM restart, the exact physical-host baseline, and input/client certification remain separate promotion gates. The nonshipping input experiment in #3572 is not included in this foundation's scope or results.

The claim and pool were deleted after evidence collection. A fresh successful SDK namespace inventory found zero matching namespaces. The local evidence archive is retained; no Fleet image or package was published.