feat(server): verified Cua Driver 0.37.0 lifecycle with telemetry off - #665

GOAT has no computer use. Upstream T3 Code has an open, unmerged Cua Driver stack by Julius Marminge (rebuilt from his #16527) that adds verified distribution, an opt‑in server lifecycle, an Electron host, and packaging, and attaches nothing to agents. This PR ports its server half, pingdotgg/t3code#21694 at frozen head b8569ac92 and #21695 at dc307c8a7, pins Cua Driver 0.37.0, and forces the driver's telemetry and update check off. Computer use stays off by default (enableCua: false), and no provider or UI uses it yet.

The first two commits are the upstream PRs applied as diffs, authored by Julius with the upstream commit list in each message. Desktop hosting and packaging follow in the stacked PR feat/cua-driver-desktop.

GOAT deltas from upstream

• Pin 0.37.0 (upstream pins 0.34.1). @trycua/cua-driver on npm stops at 0.37.0, and the SDK and executable must share a version. Archive names, byte sizes and SHA‑256 come from the official cua-driver-rs-v0.37.0 SHA256SUMS (re‑downloaded and compared). All six platform/arch entries match the downloaded archives and contain every required companion file. The 0.37.0 embedded host API is identical to 0.34.1's, so no ported logic changed.
• Telemetry and update check off at the daemon. The host is built with EmbeddedCuaDriverHost.withOptions(cuaDriverHostOptions(...)), passing CUA_DRIVER_RS_TELEMETRY_ENABLED=false, CUA_TELEMETRY_ENABLED=false and CUA_DRIVER_RS_UPDATE_CHECK=false. The 0.37.0 host clears the daemon environment and rejects names outside its allowlist (allowed_environment_name in cua-driver-sdk/src/embedded.rs), so DO_NOT_TRACK and CUA_TELEMETRY cannot be passed there; trying makes host construction fail. This differs from plan D4's four‑variable list on purpose.
• Opt‑outs on the agent path. The MCP descriptor that providers will spawn (cua-driver mcp --embedded …) is a separate driver process outside that allowlist, and upstream returns it with an empty environment. CuaDriver.acquire now adds all five opt‑outs (withCuaDriverPrivacyEnvironment), overriding any value the native descriptor carries.
• Renames via node scripts/goat-rename.ts --apply: @goat/* imports, GOAT_CUA_DRIVER_PATH, goat/cua/* service keys, temp prefix. The standalone host identity is si.goat.server (upstream com.t3tools.t3code.server; os-identities keeps com.t3tools only for installed OS services).
• server.ts: the runtime layer pipe already had 20 steps, its maximum, so layerCuaDriver joins the existing ProjectStore/ThreadSearch merge. Upstream's separate step fails typecheck in GOAT (TS2554).
• Merged with dev 881757889 (after the upstream chase chase(t3): merge upstream main 4f651cb10 #658). server.ts, packages/shared/package.json and third-party-licenses.config.json take dev's versions plus this PR's additions. pnpm-lock.yaml differs from dev only in @trycua/@ubjs entries, and vp i --frozen-lockfile passes.
• Docs: ADR‑0020 (structure/decisions/ADR-0020-cua-driver-computer-use.md), a "Computer use" section in structure/server.md, the shared export row, manifest and INDEX rows, and fork‑delta area cua-driver (devlog/_chase/scan.sh --delta ok).

Verification

Real macOS 26 (arm64), official 0.37.0 universal binary:

• Daemon env. I started the host through the SDK using the shared options, with a hostile parent env (CUA_DRIVER_RS_TELEMETRY_ENABLED=true, CUA_DRIVER_RS_UPDATE_CHECK=true) and an isolated HOME. ps -E on the live daemon showed all three switches =false and --host-bundle-id si.goat.server. Under that env, cua-driver telemetry status --json reports enabled: false, source: environment (baseline without overrides: enabled: true, source: default). The isolated HOME held no installation ID or telemetry markers afterwards. 0.37.0's get_config does not report telemetry state, so this check relies on the process env plus telemetry status.
• Agent path. I spawned the MCP proxy from the published descriptor: 62 tools listed over stdio (same count as upstream's Linux run), proxy env holds all five opt‑outs, and telemetry status under them reports source: do_not_track.
• Production standalone path. makeCuaDriverInstallation + FetchHttpClient downloaded the 0.37.0 darwin archive from GitHub, verified size and SHA‑256, and published tools/cua-driver/0.37.0 with exactly the five expected files in 1.6 s; cache reuse took 1 ms. makeStandaloneHostFactory with the real SDK then started and stopped the host. Upstream only qualified this path on Linux. This was a temporary test file, not committed.
• vp test run src/cua src/resourceTelemetry (apps/server): 80/81. The one failure, HostResources publishes load and disk, is darwin‑only on dev: diskPathForPlatform("darwin") returns $HOME while the test expects /; neither file differs from dev, and CI runs Linux. vp test run src/telemetry (apps/desktop): 5/5. vp test run src/settings src/resourceTelemetry (packages/contracts): 198/198.
• Red→green: emptying the host option environment fails starts an explicit driver path with telemetry and update checks off; restoring it passes. The existing lifecycle tests now assert the exact five‑variable descriptor, with a hostile CUA_TELEMETRY=1 in the native fixture.
• Typecheck exits 0 for apps/server, apps/desktop, packages/shared and packages/contracts (only pre‑existing Effect suggestions remain). Targeted vp lint and vp fmt --check pass on touched files; server.ts:944 has a pre‑existing warning.

Platform notes (re‑verified from 0.37.0)

• Windows UAC. cua-driver.exe and cua-driver-uia.exe embed requestedExecutionLevel level="asInvoker", so neither prompts on launch. install.ps1 installs under %LOCALAPPDATA% without elevation and prompts only for its optional -AutoStart scheduled task, which GOAT never uses. Unelevated, the driver refuses elevated targets and may see truncated UWP/AppContainer UIA trees. Not run on a real Windows host (win was unreachable).

• Linux arm64. Ships native aarch64 builds of the driver and SDK library (glibc ≥ 2.30, same floor as x64). The installer maps Linux-aarch64, and the release notes label all Linux builds "preview". A read‑only probe of an arm64 Ubuntu 24.04 host (glibc 2.39) found it compatible, but it is headless; I did not run the driver there.

• goat-ci on head 5ef6fbc (after merging dev 10 commits past 8817578, including chore: drop leftover T3 tooling names and unused root folders #661): pull_request run 39591463869 and full dispatch run 39591455467, both 18/18 jobs with goat-ci-result success. The previous head a1340bf was green on both runs (39574338635, 39574407950).

Not run / open

• Windows and Linux runtime with this pin, macOS TCC attribution, packaged desktop (next PR plus lane 3). No GUI capture or input was exercised.
• The driver's check_for_update MCP tool still contacts GitHub when an agent calls it. Removing it belongs to the capability ceiling in the safety lane; ADR‑0020 records it as pending.
• Overlaps: chore: drop leftover T3 tooling names and unused root folders #661 edits devlog/_chase/010_fork_delta.md and goat‑rename rules (this adds one block in its own alphabetical slot; re‑run goat-rename --check after chore: drop leftover T3 tooling names and unused root folders #661). feat(web): Boat cloud machines show the Boat logo and state in Fleet #659 edits third-party-licenses.config.json (this keeps upstream's two Cua package overrides). chase(t3): merge upstream main 4f651cb10 #658 is the upstream chase; nothing here comes from upstream main. structure/INDEX.md already drifts from its manifest on dev (iMessage rows, two scopes); this PR adds only its own three rows.

Review

Independent read-only review by Codex gpt-6.1-sol. Round 1 returned MUST-FIX with three findings: unrelated lockfile peer re-resolution, docs describing the desktop host as current at this layer, and the ADR commit citation. All three are fixed in 8573ccbfb. Round 2 returned PASS. Neither round found a runtime defect.

Implementation: anthropic/claude-opus-5-5 via the Codex harness in GOAT (lane 1 of the Cua Driver program). Upstream implementation credit: Julius Marminge (pingdotgg/t3code #16527, #21694, #21695).