Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.
Rust agentai-agentapplecomputer-usecomputer-use-agentcontainerizationcuadesktop-automationhacktoberfestlumemacosmanusoperatorswiftvirtualizationvirtualization-frameworkwindowswindows-sandbox
ci(cua-driver): build exact MCP candidates for guest qualification (#3694)
Francesco Bonacci87bd0eced019
diff --git a/.github/workflows/ci-driver-mcp-candidate.yml b/.github/workflows/ci-driver-mcp-candidate.ymlnew file mode 100644index 0000000000..f41175cdf5--- /dev/null+++ b/.github/workflows/ci-driver-mcp-candidate.yml@@ -0,0 +1,120 @@+name: "CI: Driver MCP candidate"++on:+ pull_request:+ paths:+ - ".github/workflows/ci-driver-mcp-candidate.yml"+ - "libs/cua-driver/rust/crates/cua-driver/src/mcp_envelope.rs"+ - "libs/cua-driver/rust/crates/cua-driver/src/driver_service_http.rs"+ - "libs/cua-driver/rust/crates/cua-driver/src/proxy.rs"+ - "libs/cua-driver/rust/crates/cua-driver/src/serve.rs"+ - "libs/python/cua-sandbox/cua_sandbox/interfaces/_driver_mcp.py"+ - "libs/python/cua-sandbox/cua_sandbox/interfaces/driver.py"+ workflow_dispatch:++permissions:+ contents: read++concurrency:+ group: driver-mcp-candidate-${{ github.ref }}+ cancel-in-progress: true++jobs:+ candidate:+ name: MCP candidate (${{ matrix.os }})+ runs-on: ${{ matrix.os }}+ timeout-minutes: 40+ strategy:+ fail-fast: false+ matrix:+ os: [ubuntu-24.04, windows-latest]+ defaults:+ run:+ shell: bash+ env:+ EXPECTED_SOURCE_SHA: ${{ github.event.pull_request.head.sha || github.sha }}+ CARGO_PROFILE_DEV_DEBUG: "0"+ CUA_TELEMETRY_ENABLED: "false"+ steps:+ - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4+ with:+ ref: ${{ github.event.pull_request.head.sha || github.sha }}+ persist-credentials: false+ - name: Verify exact source+ run: |+ set -euo pipefail+ test "$(git rev-parse HEAD)" = "$EXPECTED_SOURCE_SHA"+ test -z "$(git status --porcelain)"+ - uses: dtolnay/rust-toolchain@stable+ - uses: Swatinem/rust-cache@9d47c6ad4b02e050fd481d890b2ea34778fd09d6 # v2+ with:+ workspaces: "libs/cua-driver/rust -> target"+ key: mcp-candidate+ - name: Install Linux build dependencies+ if: runner.os == 'Linux'+ run: |+ sudo apt-get update+ sudo apt-get install -y --no-install-recommends \+ clang pkg-config libdbus-1-dev libpipewire-0.3-dev libspa-0.2-dev \+ libei-dev libxkbcommon-dev libx11-dev libxi-dev libxtst-dev libxext-dev+ - name: Run envelope receiver and service tests+ working-directory: libs/cua-driver/rust+ run: |+ set -euo pipefail+ cargo test -p cua-driver --bin cua-driver mcp_envelope --locked+ cargo test -p cua-driver --bin cua-driver driver_service_http --locked+ - name: Build disposable-test binaries+ working-directory: libs/cua-driver/rust+ run: |+ set -euo pipefail+ if [[ "$RUNNER_OS" == Linux ]]; then+ cargo build --locked -p cua-driver -p cursor-theme-cli --features cua-driver/portal-input+ else+ cargo build --locked -p cua-driver -p cursor-theme-cli -p cua-driver-uia+ fi+ - name: Bundle exact binaries and provenance+ run: |+ python - <<'PY'+ import hashlib+ import json+ import os+ from pathlib import Path+ import subprocess+ import tarfile++ source = subprocess.check_output(["git", "rev-parse", "HEAD"], text=True).strip()+ assert source == os.environ["EXPECTED_SOURCE_SHA"]+ assert not subprocess.check_output(["git", "status", "--porcelain"], text=True).strip()+ windows = os.environ["RUNNER_OS"] == "Windows"+ names = ["cua-driver", "cua-cursor-theme"] + (["cua-driver-uia"] if windows else [])+ binaries = [Path("libs/cua-driver/rust/target/debug") / (name + (".exe" if windows else "")) for name in names]+ assert all(path.is_file() and path.stat().st_size > 0 for path in binaries)+ bundle = Path("mcp-candidate")+ bundle.mkdir()+ metadata = {+ "source_sha": source,+ "runner_os": os.environ["RUNNER_OS"],+ "profile": "dev, debug=0",+ "features": ["cua-driver/portal-input"] if not windows else [],+ "rustc": subprocess.check_output(["rustc", "--version"], text=True).strip(),+ "version": subprocess.check_output([str(binaries[0].resolve()), "--version"], text=True).strip(),+ "files": {path.name: hashlib.file_digest(path.open("rb"), "sha256").hexdigest() for path in binaries},+ "qualification": "Test-only unsigned candidate; not a release or image qualification.",+ }+ provenance = bundle / "provenance.json"+ provenance.write_text(json.dumps(metadata, indent=2) + "\n")+ with tarfile.open(bundle / "candidate.tar.gz", "w:gz") as archive:+ for path in binaries:+ archive.add(path, arcname=path.name)+ archive.add(provenance, arcname=provenance.name)+ print(json.dumps(metadata, indent=2))+ PY+ - name: Upload candidate for separately authorized tests+ uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4+ with:+ name: driver-mcp-candidate-${{ runner.os }}-${{ github.event.pull_request.head.sha || github.sha }}+ if-no-files-found: error+ retention-days: 7+ path: |+ mcp-candidate/candidate.tar.gz+ mcp-candidate/provenance.jsondiff --git a/libs/cua-driver/docs/mcp-candidate-qualification.md b/libs/cua-driver/docs/mcp-candidate-qualification.mdnew file mode 100644index 0000000000..f570ca68c4--- /dev/null+++ b/libs/cua-driver/docs/mcp-candidate-qualification.md@@ -0,0 +1,42 @@+# MCP candidate qualification++The `CI: Driver MCP candidate` workflow builds the exact pull-request head+on hosted Linux and Windows runners. It runs the envelope receiver and service+tests, then uploads unsigned development binaries with their source SHA,+toolchain, build profile, version, and SHA256 hashes. It has read-only repository+permissions and no Fleet credentials. It does not publish a release, install a+product, change an image, or connect to a Fleet computer.++These artifacts support separately authorized disposable-guest tests. Check the+workflow repository, source SHA, successful jobs, and every binary hash before+using an artifact. Do not select an artifact by version alone: an unreleased+source build can report the same package version as a released binary.++## Live test gate++For each selected Linux and Windows Fleet image:++1. Record the immutable image reference and installed Driver, computer-server,+ and MCP-wrapper versions. Use an owned disposable claim and namespace with+ explicit resource limits and verified cleanup ownership.+2. Stage the exact candidate separately. Preserve computer-server. Start Driver+ in the guest's interactive desktop session through its existing socket or+ named pipe. Enable the envelope opt-in in both the daemon and MCP proxy.+3. Disable raw request/response logging before action-bearing tests. Reuse the+ existing named MCP service and access boundary; do not introduce a new+ network endpoint or substitute the operator's local desktop.+4. Use matching generated bindings with+ `sb.driver.connect(service="mcp", transport="mcp")`. Prove a useful guest+ effect with fresh before/after Driver state and an independent guest+ postcondition. A successful call alone is not the oracle.+5. Verify independent sessions, cancellation without claiming rollback, stale+ session rejection after replacement, access rejection, and continued+ computer-server functionality. Verify receiver and MCP-session cleanup+ independently from Fleet resource cleanup.+6. Delete only the task-owned resources and confirm their absence. Preserve+ failure and cleanup evidence, including any incomplete result.++A staged development-binary result is not packaged-image certification or the+full canonical desktop matrix. A release/image enablement decision still needs+the applicable exact-candidate desktop harness evidence, packaged-image replay,+and separate rollout approval. Keep existing defaults and images unchanged.