Verify Cua Driver

Read the feature map first. This verifies the daemon's desktop-tool surface, not the bench, and does not certify every mapped entry point by exercising one convenient path. Use only a disposable desktop: the foreground cells intentionally move focus and input.

Launch

Run from the repository root. Use the CI toolchain (Rust 1.97.1, Node 22) and install the Linux dependencies from .github/workflows/e2e-rust-linux.yml:

sudo apt-get update
sudo apt-get install -y --no-install-recommends \
  xvfb dbus-x11 at-spi2-core openbox picom python3-gi gir1.2-gtk-3.0 \
  libgtk-3-dev clang pkg-config libdbus-1-dev libpipewire-0.3-dev \
  libspa-0.2-dev libei-dev libx11-dev libxi-dev libxtst-dev libxext-dev \
  libwebkit2gtk-4.1-dev libssl-dev libxdo-dev \
  libayatana-appindicator3-dev librsvg2-dev ffmpeg jq x11-utils

CUA_E2E_RETRY_ONCE=0 .cursor/skills/verify-cua-driver/scripts/verify-linux.sh x11

The helper prints an absolute Evidence: directory under artifacts/cua-driver/verify/run-XXXXXXXX. Each invocation creates a new directory; never reuse an older canonical artifact directory. The canonical runner builds the release driver once and stages only Electron using tests/fixtures/build/linux.sh --only electron. Reuse that build with:

timeout 1260 .cursor/skills/verify-cua-driver/scripts/verify-linux.sh x11 --no-build

The X11 recipe follows CI: xvfb-run -a --server-args='-screen 0 1920x1080x24' dbus-run-session, openbox, then picom with the xrender backend. at-spi2-core is D-Bus activated, not bootstrapped by another accessibility launcher. Bounded probes wait for the WM and org.a11y.Bus.GetAddress. Testkit creates private daemon sockets and empty isolated HOME/XDG state, then owns the daemon, MCP proxy, fixture, and sentinel processes. Do not attach to your installed daemon.

The runner first executes canonical_e2e_environment_is_ready, then only shared_web_action_matrix_is_state_verified. The default substring filter runs two cells: linux-electron-left-click-ax-background and its -typed-sdk sibling. Both must deliver, not skip or refuse. READY … daemon usable … AX and journal marker present means a successful daemon tool path, a discovered real window, and WEB_HARNESS_MARKER_v1 in both AX and the fixture's read-only event journal. Socket existence alone is not ready. The helper completes the action and tears down automatically; it is not an interactive server launcher.

Nested Wayland uses the existing CI wrapper, with its default runner and the same focused lane:

timeout 1260 .cursor/skills/verify-cua-driver/scripts/verify-linux.sh wayland --no-build

Install the compositor dependencies from .github/workflows/e2e-rust-linux-wayland.yml before this lane. scripts/ci/linux/run-rust-e2e-wayland.sh provisions its runtime, Sway and AT-SPI; do not duplicate its bootstrap. Compositor and AT-SPI logs honor the unique artifact directory. Wayland is unverified here; X11 results do not certify its capture or input routes.

Doctor

Read-only diagnostics run inside the same session before driving:

• gdbus call --session --dest org.a11y.Bus --object-path /org/a11y/bus --method org.a11y.Bus.GetAddress probes the accessibility bus; at-spi-address.txt retains its address.
• libs/cua-driver/rust/target/release/cua-driver doctor --json writes doctor.json; no --fix. Inspect warnings even if the exit status is zero. Linux health_report checks the display and AT-SPI availability.
• libs/cua-driver/rust/target/release/cua-driver --version writes version.txt.
• Preflight calls get_config, health_report, and check_permissions on its owned daemon. Its recording directory retains health_report.json, check_permissions.json, and daemon.json (PID, socket, owner UID, mode, version/build config). Preflight asserts source version and optional requested SHA, Unix socket type, same-user ownership and no other-user access. It also proves fixture AX, capture and video readiness.

A permissions report is not authorization coverage: canonical testkit explicitly enables unrestricted GUI operations and bypasses interactive approvals. Health warnings must be interpreted alongside preflight, not silently called healthy.

Drive

Reuse libs/cua-driver/rust/crates/cua-driver-e2e/tests/cross_platform_behavior_test.rs. It launches the real Electron fixture from libs/cua-driver/tests/fixtures/apps/cross-platform/electron, discovers its new PID/window, and obtains fresh snapshot-bound handles. The actual MCP tool sequence is:

let windows = fixture.driver.call("list_windows", serde_json::json!({}));
let pre = snapshot(fixture);
let args = action_target_args(fixture, &pre, "border-click-target", "ax", "background");
let response = fixture.driver.call("click", args);

action_target_args sends pid, window_id, delivery_mode: "background", and the discovered element_token from require_element. Do not paste numeric handles or reuse a token after another snapshot. The fixture journal must independently report last_action=left_click; an accepted response alone is insufficient. The typed SDK sibling additionally exercises stale/mismatched handle refusals before a valid click.

Use the commands in the feature map to select other user journeys. Prefer AX handles; use existing action_target_args pixel conversion only for pixel coverage, not guessed screen coordinates. Do not use DOM setters, fixture control endpoints, or direct journal writes to make target state pass. The fixture journal observes effects; the real driver performs the action.

Evidence

Keep the printed unique directory. It contains launch.log, doctor/version output, preflight logs, results.jsonl, cases.jsonl, environment.jsonl, and summary.md. Each existing per-cell recording directory contains:

• tool-calls.jsonl: actual MCP requests and full response envelopes, including transport errors/timeouts. Image/base64 payloads are replaced by encoded byte length and a hash; PNGs remain separate. Typed SDK traffic is not MCP transcript traffic.
• trajectory.json, recording.mp4, and action turn-* directories with action.json, before.png, after.png from the product recording mechanism.
• fixture-state.json: the independent fixture journal after the action, also saved during unwinding.
• sentinel-events.jsonl: foreground heartbeat, focus and input journal.
• desktop-observation.txt: native before/after focus, z-order, pointer, transient focus journal, unsupported oracles and violations. Result errors retain these diagnostics too.

The runner uses cua-e2e-report --require-video and ffprobe validation. A passing summary must include delivered click rows and playable owned videos. Inspect action, response, before/after screenshots and fixture side effect together. For background input, require a live heartbeat, no leaked input/blur, stable native focus and pointer, no target raise, and no unsupported oracle. A refusal is a valid refusal test, not proof of successful delivery. Failed setup must remain a failure, not be replaced by a screenshot from another run.

Cleanup

Testkit's child reaper owns daemon/fixture groups; the sentinel owns its own group. Scope exit reaps them. The X11 wrapper traps exit and kills/waits only the openbox and picom PIDs it captured with $!; xvfb-run and dbus-run-session own their sessions. The Wayland wrapper owns its compositor and bus PIDs. Never use process-name matching or kill another desktop.

After session teardown, the helper invokes scripts/check-evidence.sh on the printed directory, even if the desktop command failed. The checker requires complete current-attempt evidence and checks kill -0 against recorded daemon PIDs and discovered CuaTestHarness window PIDs, without killing them. Malformed extraction, zero recorded PIDs or any surviving PID fails verification, not permission to kill by name. Do not delete the evidence directory. A timeout or crash may leave incomplete evidence; inspect the launch log and recorded PIDs before retrying.

Helpers

The shipped executable is:

.cursor/skills/verify-cua-driver/scripts/verify-linux.sh x11
.cursor/skills/verify-cua-driver/scripts/verify-linux.sh x11 --no-build
.cursor/skills/verify-cua-driver/scripts/verify-linux.sh wayland --no-build

It reuses scripts/ci/linux/run-rust-e2e.sh through CUA_E2E_INTERNAL_LANE=verify and CUA_E2E_ARTIFACT_DIR; the runner is responsible for strict fixtures, preflight, recordings and report validation. Each desktop run is bounded to 1200 seconds internally. No bench code or second harness is involved.

macOS and Windows: unverified

macOS needs the installed TCC-authorized daemon (libs/cua-driver/scripts/install-local.sh, cua-driver permissions grant) and Accessibility/Screen Recording consent; its tests use the existing daemon proxy rather than Linux Xvfb. Windows needs an interactive desktop, native UIA and named-pipe transport; testkit uses Job Objects for ownership. Use their existing CI recipes and recording paths. Neither platform was run here; Linux's unrestricted testkit mode does not prove their production permission paths.

Execution status and maintenance

Linux X11 CI proof is pending in .github/workflows/e2e-rust-linux.yml, job verify-skill. It runs on relevant skill/driver pull requests or workflow_dispatch with lane=verify, with retries disabled and a 25-minute job bound. The helper owns Electron staging and desktop setup; check-evidence.sh gates success after cleanup. The verify-cua-driver-evidence artifact retains the printed evidence directory even on failure. No successful CI run is claimed. See execution instructions, historical attempts and remaining stand-ins, including the same command for a networked local host. Claim proof only after both click cells deliver and all required evidence survives cleanup; setup logs alone do not prove delivery or owned-process teardown.

Use /maintain-verification-skill when tools, fixture markers, CI lanes or supported input routes change.