Skip to content

feat(computer-use): integrate native Windows maka.cu/2 path - #4595

Open
sunheyi6 wants to merge 12 commits into
apache:mainfrom
sunheyi6:codex/maka-cu-rust-comparison-pr
Open

feat(computer-use): integrate native Windows maka.cu/2 path#4595
sunheyi6 wants to merge 12 commits into
apache:mainfrom
sunheyi6:codex/maka-cu-rust-comparison-pr

Conversation

@sunheyi6

@sunheyi6 sunheyi6 commented Sep 2, 2026

Copy link
Copy Markdown
Contributor

Summary

This PR integrates the native Windows maka.cu/2 Computer Use path into Maka and records the current Windows x64 validation evidence.

  • Keeps the shared Computer Use surface semantic-only; model actions do not expose raw coordinate mutation.
  • Uses the shared supervised lifecycle and a thin Windows native-helper adapter instead of retaining the former private Windows protocol.
  • Enforces fresh target identity using HWND, PID, title, process start time, and windowGeneration; stale observations are invalidated and unknown outcomes are not replayed or upgraded.
  • Adds the helper selection, packaging/provenance seam, Windows host tests, protocol drivers, and bilingual cross-machine evidence.
  • The companion native executor implementation is in maka-agent/maka-cu#8.

Validation

The following evidence is recorded in experiments/maka-cu-windows/CROSS_MACHINE_RUST_NATIVE_SUMMARY.md:

  • Formal Rust native executor: 8 cargo tests passed; shared-protocol lifecycle: 29/29 passed.
  • WPF direct run: 17 pass, 0 fail, 1 honest unknown; six-run WPF matrix: 102 pass, 0 fail, 6 unknown.
  • Chromium direct run: 9 pass, 0 fail, 3 blocked, 1 honest unknown; six-run matrix: 54 pass, 0 fail, 18 blocked, 6 unknown.
  • Computer Use host integration: 115 pass, 0 fail, 2 skipped (the two skipped cases require explicit published-helper environment variables).
  • Desktop host targeted checks: 17 pass plus one Windows symlink EPERM environment block; native capability contract: 13/13 pass.
  • Windows x64 packaging completed with exit code 0; the earlier Spectre/MSB8040 failure did not recur. The installer SHA256 is AD79C2362FF05B5392D79655B85E2B3C20E341AE8ECC2C5E5BDAA387CA37DA12; the ZIP SHA256 is 825213FE198A3140C60B75FD83D80B365C0CD67776DE840D26AD6364DF81B7A2.
  • Real Electron Maka smoke: 7/7 programmatic checks passed.
  • LibreOffice real-app probe: 9 pass, 1 typed dispatch_refused; the mutation was not retried.
  • Manual real Maka Computer Use smoke: user-confirmed pass; an isolated Maka session opened a new Chrome test instance and navigated to https://www.google.com/ without using an existing Chrome profile.
  • The Chromium Enter case remains unknown; a page oracle is never rewritten as helper verification.

Known limits

  • The repository-wide npm run build on the clean PR branch is currently blocked by pre-existing @maka/ui type errors involving settledText, autoScroll, and trailingAction, outside this PR. The CU-scoped build, @maka/computer-use tests, and Desktop main-process build pass.
  • The legacy comparison harness still speaks the former private protocol and is not valid shared-protocol conformance evidence; the new maka-cu2 drivers are authoritative.
  • The dedicated automated real-model runner was not run because its provider/API-key/local-bridge environment was not configured.
  • Full WinUI/UWP mutation coverage was not accepted after Notepad restored an existing user session.
  • Clean-machine validation was intentionally skipped by the current scope. This PR therefore provides local-machine evidence, not clean-machine or production qualification; distributionReady remains false.

Refs #4318

中文翻译

摘要

本 PR 将 native Windows maka.cu/2 Computer Use 路径接入 Maka,并记录当前 Windows x64 验证证据。

  • 保持共享 Computer Use action surface 为 semantic-only;模型 action 不暴露原始坐标 mutation。
  • 使用共享 supervised lifecycle 和薄的 Windows native-helper adapter,不再保留原来的私有 Windows protocol。
  • 使用 HWND、PID、标题、进程启动时间和 windowGeneration 强制执行新鲜目标身份;过期 observation 会失效,unknown 结果不会被重放或升级。
  • 增加 helper 选择、打包/来源追踪 seam、Windows host 测试、protocol driver 和中英双语跨机器证据。
  • 配套的 native executor 实现位于 maka-agent/maka-cu#8

验证情况

以下证据记录在 experiments/maka-cu-windows/CROSS_MACHINE_RUST_NATIVE_SUMMARY.md 中:

  • Formal Rust native executor:8 个 cargo 测试通过;shared-protocol lifecycle:29/29 通过。
  • WPF 直接运行:17 pass、0 fail、1 个诚实 unknown;六次 WPF 矩阵:102 pass、0 fail、6 个 unknown
  • Chromium 直接运行:9 pass、0 fail、3 个 blocked、1 个诚实 unknown;六次矩阵:54 pass、0 fail、18 个 blocked、6 个 unknown
  • Computer Use host integration:115 pass、0 fail、2 个 skipped(两个 skipped 用例需要显式的已发布 helper 环境变量)。
  • Desktop host 定向检查:17 pass,另有 1 个 Windows symlink EPERM 环境阻塞;native capability contract:13/13 通过。
  • Windows x64 打包以退出码 0 完成;之前的 Spectre/MSB8040 失败没有重现。安装包 SHA256 为 AD79C2362FF05B5392D79655B85E2B3C20E341AE8ECC2C5E5BDAA387CA37DA12;ZIP SHA256 为 825213FE198A3140C60B75FD83D80B365C0CD67776DE840D26AD6364DF81B7A2
  • 真实 Electron Maka smoke:7/7 程序化检查通过。
  • LibreOffice 真实应用探测:9 pass、1 个类型化 dispatch_refused;没有重试该 mutation。
  • 手工真实 Maka Computer Use smoke:用户确认通过;隔离的 Maka 会话打开新的 Chrome 测试实例并访问 https://www.google.com/,没有使用已有 Chrome profile。
  • Chromium Enter 用例保持为 unknown;页面 oracle 永远不会被改写为 helper verified。

已知限制

  • 干净 PR 分支上的仓库全量 npm run build 当前被本 PR 范围之外既有的 @maka/ui 类型错误阻塞,涉及 settledTextautoScrolltrailingAction。CU 范围构建、@maka/computer-use 测试和 Desktop main-process 构建通过。
  • 旧 comparison harness 仍使用原来的私有 protocol,因此不是有效的 shared-protocol conformance 证据;新的 maka-cu2 driver 才是权威结果。
  • 专用自动真实模型 runner 未运行,因为没有配置其 provider/API key/本地 bridge 环境。
  • Notepad 恢复了已有用户会话后,没有接受完整的 WinUI/UWP mutation 覆盖结果。
  • 当前范围明确跳过 clean-machine 验证。因此本 PR 提供的是本机证据,不是 clean-machine 或生产资格证明;distributionReady 仍为 false

Refs #4318

@github-actions github-actions Bot added the effort/XL Over 1000 readable lines label Sep 2, 2026
@sunheyi6

sunheyi6 commented Sep 3, 2026

Copy link
Copy Markdown
Contributor Author

@M4n5ter @hqhq1025 cc

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

effort/XL Over 1000 readable lines

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant