Repository navigation
roadmap(windows): make Windows a supported platform #2142
Description
Activity
Windows follow-up update (2026-08-07):
- test(storage): close project catalogs before cleanup #2339 is merged. Its Windows baseline improved from 648 pass / 38 fail / 38 skip to 664 pass / 22 fail / 38 skip, removing the 16 project-catalog cleanup failures.
- Rebased and retriggered CI after the GitHub Actions incident for test(storage): close agent run stores before cleanup #2341, fix(runtime): classify expired probe budget as timeout #2344, and test(computer-use): synchronize session cancellation #2345. Required checks are passing or still running; their non-blocking Windows baselines are in progress.
- Opened fix(storage): close usage stores after lease revocation #2365 for the remaining usage-store SQLite handles: close now releases local telemetry/model-call/pricing resources even after root identity failure or lease revocation, while ordinary operations remain lease-bound. This targets the final 3 usage-store cleanup failures from the measured baseline.
Next: verify the fresh Windows artifacts, merge the isolated fixes as their checks settle, then recalculate the remaining #2142 failure inventory.
Windows Phase 1 update (2026-08-07, second pass):
- test(storage): close agent run stores before cleanup #2341, fix(runtime): classify expired probe budget as timeout #2344, and test(computer-use): synchronize session cancellation #2345 are merged.
- Refreshed fix(storage): close usage stores after lease revocation #2365 with
aa1fe4ac4after reproducing the remaining usage-store failure. The usage-store target now reports 7 pass / 0 fail / 1 narrowly explained skip on Windows; new CI is pending. - Opened fix(storage): keep portable path tests stable on Windows #2395 for two portable-path failures: Codex SQLite cwd filtering now preserves source-native and host-normalized path variants before
LIMIT, and the sandbox-boundary fixture now uses the POSIX path domain required by the contract. - On the combined local Windows branch (Node 22), storage moved from 677 pass / 18 fail / 44 skip to 678 pass / 14 fail / 45 skip. This is a development measurement, not the final Node 24 Actions baseline; the fresh PR artifacts will be authoritative.
The remaining local failures are now grouped into Windows symlink privilege coverage, Git fixture/process cleanup, Git worktree removal, root-authority file identity/locking semantics, session-bundle path/symlink behavior, and two Node/environment-sensitive import checks.
Next: settle #2365/#2395 CI, recalculate from their Windows artifacts, then take the Git process/worktree cleanup group without adding broad filesystem retries.
Windows Phase 1 update (2026-08-07, post-rebase baseline):
- fix(storage): close usage stores after lease revocation #2365, test(headless): remove trim residue #2410, test(desktop): stop exporting trimmed helpers #2414, and fix(desktop): drop dead css-test-helpers after #2406 #2415 are merged.
- fix(storage): keep portable path tests stable on Windows #2395 has been rebased onto current
main; all required checks and its non-blocking Windows baseline pass. The PR remains open pending merge. - The authoritative Node 24 Windows storage run on fix(storage): keep portable path tests stable on Windows #2395 reports 691 pass / 10 fail / 39 skip. The previous run was 692 / 9 / 39; the additional failure is
sqlite-recovery-concurrency, so that group is currently environment/timing-sensitive rather than a stable regression count. - The 10 failures are grouped as Git workspace/ref (1), Git worktree retirement/removal (2), root-authority identity/locking/races (5), session-bundle path/atomic hydration (1), and SQLite recovery concurrency (1).
- The Windows skip inventory is stale and needs regeneration.
Next: take the three Git workspace/worktree failures as one narrowly scoped Windows cleanup group, without adding broad filesystem retries; update the skip inventory separately.
Follow-up implementation: #2424 fixes the two deterministic Windows worktree retirement failures by running the final
git worktree removefrom the stable Git common directory rather than from inside the directory being removed. The affected tests pass in three consecutive local Windows runs (6/6).The managed head-ref conflict test passes on current
mainlocally; the next Node 24 baseline will determine whether that prior failure remains after recent main changes. A separate temp-root cleanupEBUSYwas observed and is intentionally not hidden with broad retries in #2424.Windows Phase 1 update after #2395 and #2424 merged:
- The combined
mainNode 24 Windows storage baseline is now 699 pass / 7 fail / 41 skip. - Both deterministic Git worktree retirement failures fixed by fix(storage): remove Windows worktrees outside their cwd #2424 are gone.
- Opened docs(windows): refresh test skip inventory #2434 to refresh the stale skip inventory (68 -> 70 declarations; both additions are pet-pack portable candidates).
- Opened test(storage): enable long paths for bare Git fixtures #2435 for the final Git workspace failure. The product runtime already enables
core.longpaths=true; the conflict fixture's direct bare-Git helper did not, so Actions failed while creating a deeply nested ref before exercising the CAS behavior. The target passes locally with a CI-depthTEMProot after aligning the helper.
Once #2435 lands, the remaining stable groups are root-authority identity/locking (5) and session-bundle path/atomic hydration (1).
- The combined
Windows Phase 2 update (2026-08-09):
- test(windows): run managed workspace crash gates #2509 merged and completes the managed-workspace real-process crash evidence.
- docs(windows): define crash durability boundary #2553 is open with green checks and documents the Windows durability boundary: file-content and SQLite WAL/FULL guarantees are distinguished from POSIX-equivalent parent-directory durability during sudden power loss.
- Opened ci(windows): require crash recovery evidence #2562 to make the already-landed recovery evidence release-blocking through a dedicated required windows_recovery job.
- The new clean Windows Node 24 job passed end to end in about three minutes: install, full test build, SQLite crash recovery, Runtime resume/continuation recovery, Runtime Host owner-death recovery, and managed-workspace crash convergence.
- The Windows installer/ZIP packaging and verification job also passes on ci(windows): require crash recovery evidence #2562.
- Runtime's two real-process crash harnesses retain the 60-second POSIX budget and use a 120-second Windows budget; a contract test pins that platform-specific boundary.
Current PR-level noise is unrelated to the recovery changes: Storybook repeatedly reports the existing Daily Review 480px overflow and Lucide icon story failure, while one E2E rerun missed the quote companion selection overlay. Standard workspace tests, typecheck, Runtime Host tests, Windows baseline, Windows recovery, and Windows packaging pass.
Next: merge #2553 and #2562 after maintainer review, then mark the Phase 2 durability/release-evidence items complete and move to the remaining process-tree evidence/distribution work.
Windows CI efficiency update (2026-08-09)
Opened #2599 to reduce the ordinary Windows baseline critical path without weakening the evidence owned by each affected surface.
Measured baseline from #2523: 6m31s total. The serialized Storage suite took 3m22s, managed-workspace crash recovery took 46s, and Windows smoke spent another 26s while rebuilding artifacts already produced by build:test.
#2599 now:
- reuses the existing repository impact planner for Windows scheduling;
- skips the Windows runner for documentation-only changes;
- keeps build and real CLI/Electron startup smoke for Desktop/UI changes;
- selects the Runtime PTY gate only for Runtime impact;
- selects the full Storage and managed-workspace crash gates only for Storage impact;
- reuses the existing test build for smoke instead of rebuilding;
- fails safe to the complete Windows suite for missing Git history, global changes, or unknown production paths.
Expected result: Desktop/UI pull requests avoid roughly 4 minutes of unrelated Storage/Crash work while retaining Windows startup evidence. Runtime and Storage changes keep their respective platform evidence. The dedicated release-blocking recovery gate remains tracked by #2562; #2599 does not replace it.
Verification: 12 tests pass, 1 Windows-only test is skipped locally; workflow YAML parsing, Biome, and diff checks pass.
中文进展
已提交 #2599，用于缩短日常 Windows baseline 的关键路径，同时不削弱各影响面负责的证据。
从 #2523 测得：总耗时 6 分 31 秒，其中串行 Storage 测试 3 分 22 秒，managed-workspace crash recovery 46 秒；Windows smoke 还用了约 26 秒重复构建 build:test 已经生成的产物。
#2599 的调整：
- Windows 调度复用仓库现有的影响面计算；
- 纯文档改动不再占用 Windows runner；
- Desktop/UI 改动仍保留构建和真实 CLI/Electron 启动 smoke；
- Runtime 受影响时才运行 Windows PTY gate；
- Storage 受影响时才运行完整 Storage 与 managed-workspace crash gate；
- smoke 复用已有测试构建，不再重复构建；
- Git 历史缺失、全局配置或未知生产路径变更时，安全回退为完整 Windows 套件。
预期 Desktop/UI PR 可减少约 4 分钟无关的 Storage/Crash 工作，同时保留 Windows 启动证据。Runtime 与 Storage 变更仍保留各自的 Windows 平台证据。#2562 的 release-blocking recovery gate 仍独立推进，#2599 不会替代它。
本地验证：12 项通过、1 项仅 Windows 运行而跳过；workflow YAML、Biome 与 diff 检查均通过。
Windows Phase 2 completion update (2026-08-10):
- test(windows): run managed workspace crash gates #2509 and docs(windows): define crash durability boundary #2553 are merged, completing managed-workspace crash evidence and the documented Windows durability boundary.
- ci(windows): require crash recovery evidence #2562 is merged. Its dedicated
windows_recoveryworkflow passed on the merge commit in 3m41s, covering SQLite crash recovery, Runtime continuation recovery, Runtime Host owner-death recovery, and managed-workspace crash convergence. - The merge commit Windows baseline also passed in 2m44s and retains the residual-process audit; process-tree behavior was previously completed by test(runtime): validate Windows process lifecycle #2465.
- Phase 2 engineering work is now complete. One repository-administration follow-up remains: configure
windows_recoveryas a required check in the effective GitHub ruleset.
Next: audit the existing Windows package output and start Phase 3 with the smallest complete distribution slice: installer production plus clean-run verification.
中文进展
Windows Phase 2 已完成：#2509、#2553、#2562 均已合并；合并提交上的 Windows recovery（3分41秒）和 baseline（2分44秒）均通过。恢复流水线覆盖 SQLite、Runtime continuation、Runtime Host owner-death 与 managed workspace 崩溃恢复，进程树证据由 #2465 和 baseline 残留进程审计覆盖。
剩余一项仓库治理配置：需要管理员在实际生效的 GitHub ruleset 中把
windows_recovery设为 required check。下一步进入 Phase 3，先审计现有 Windows 打包产物，再完成安装器生成与干净环境验证闭环。Phase 3 has started with #2650.
The installer-production checkbox is now marked complete based on #2182 and the existing Windows packaging workflow. #2650 adds the missing artifact-level lifecycle evidence: install the final NSIS
.exeinto an isolated directory, verify the installed app (resources, version, ConPTY, renderer), then run the shipped uninstaller and require complete removal. Upgrade/rollback remains explicitly out of scope for this PR.Windows Phase 3 lifecycle update (2026-08-11):
- test(windows): exercise installer lifecycle #2650 merged and completes clean NSIS install, installed-app verification, Runtime Host drain, and uninstall-directory removal evidence.
- Opened test(windows): verify installer upgrades #2658 for version transition evidence. It selects the newest stable release below the candidate, verifies the published SHA-256, installs that release, proves a malformed candidate cannot change the installed version, then upgrades to and fully verifies the current candidate before uninstalling.
- The failure contract is deliberately narrow: deterministic pre-install failure isolation, not arbitrary power-loss recovery during NSIS replacement.
Windows Phase 3 upgrade update (2026-08-12):
- test(windows): verify installer upgrades #2658 merged. The release workflow now pins the reviewed v0.1.9 NSIS artifact by exact tag, asset name, and repository-owned SHA-256.
- On the merge-ready Windows run, the verifier completed the real closed-app transition: v0.1.9 install and full packaged-app smoke -> installed-process drain -> current candidate upgrade in the same install/profile paths and full smoke -> process drain -> real NSIS uninstall and directory removal.
- The evidence boundary remains explicit: this proves a successful closed-app upgrade, not persisted business-state migration, running-app upgrade, automatic update, or rollback after a mid-install failure. Therefore the combined install/upgrade/rollback/failure checkbox remains open.
Next: design the smallest deterministic NSIS failure/rollback experiment. It must enter the real installer replacement path and then prove the previous installation tree is byte-identical and still launches; a loader-rejected or pre-install invalid executable is not sufficient evidence.
中文进展
#2658 已合并。发布流水线现在按精确 tag、资产名和仓库固定 SHA-256 锁定 v0.1.9 NSIS 安装包。Windows 实测完成了真实的关闭状态升级闭环：安装并完整验证 v0.1.9 -> 等待安装目录内进程退出 -> 在相同安装目录和 profile 路径覆盖升级当前候选并完整验证 -> 再次等待进程退出 -> 运行真实 NSIS 卸载器并确认目录移除。
证据边界保持明确：这证明成功的关闭状态升级，不证明业务状态迁移、运行中升级、自动更新或安装中途失败后的 rollback。因此 install/upgrade/rollback/failure 组合项暂不勾选。
下一步设计最小、确定性的 NSIS 失败/rollback 实验：失败必须真正进入安装器替换路径，随后验证旧安装目录树逐文件哈希完全一致且旧版本仍可启动；仅让 Windows loader 拒绝无效 exe 不构成 rollback 证据。
Phase 3 follow-up: opened #2926 for Windows preview publication and installation documentation. It documents checksum verification, the expected unsigned SmartScreen warning, setup/uninstall steps, and the exact upgrade evidence boundary without advertising Windows as fully supported.
Rollback audit: electron-builder inserts
customInstallonly after application files, registry data, and shortcuts have already been replaced. Forcing an error there would create a partial install; NSIS/electron-builder does not restore the previous tree. Therefore a synthetic failpoint alone cannot honestly complete rollback evidence. A future rollback slice first needs a transactional backup/restore design, then a deterministic mid-install failure and byte-identical old-tree plus old-app smoke proof.中文进展
已提交 #2926，补充 Windows 预览版发布与安装文档：校验和、未签名 SmartScreen 提示、配置/卸载步骤，以及准确的升级证据边界，同时不提前宣称 Windows 已正式支持。
Rollback 审计确认：electron-builder 的
customInstall钩子发生在应用文件、注册表和快捷方式已经替换之后；此处强制失败只会制造半安装状态，NSIS/electron-builder 不会恢复旧安装树。因此不能用一个合成 failpoint 冒充 rollback 证明。后续必须先设计事务式备份/恢复，再加入确定性的安装中途失败，并验证旧目录逐文件哈希一致且旧版仍能完整启动。10 remaining items
Roadmap checklist sync (2026-08-21):
- Recorded fix(desktop): drain untracked Runtime Host before update #3382 merged / windows: auto-update silently fails when a Runtime Host process survives the app quit (installer cannot clear it) #3340 closed and the successful Runtime Host update-handoff evidence.
- Recorded feat(win): Abort-path installer rollback with backup retention #3265 exact head
68520adcas CI/audit/Windows L3 green but still pending independent human review; kept it unchecked. - Split the installer boundary honestly: Abort-path rollback/backup retention is the current PR, while hookless Quit/hard-kill/power-loss automatic recovery remains an explicit formal-support decision rather than being called transactional.
- Checked the Phase 4 RFC/shipped-slice alignment item as completed by feat(windows-sandbox): place sandboxed children on a private desktop #3174 (including production-identity readiness and private-desktop placement).
- Reworded the remaining Phase 4 lifecycle task as promotion to a required packaged/per-release gate plus Runtime-Host-mid-launch, explicit cancel, soak, and unsettled-state recovery evidence.
- Kept Authenticode signing, the
windows_recoveryruleset requirement, the adversarial matrix, independent human security review, Windows baseline cleanup, and Windows computer-use open.
Related tracker updates:
- ci(windows): track remaining Release Windows check product failures #3377 received a current-status correction because this account cannot edit its body.
- runtime-host: intermittent 'stopped responding during startup' after a Windows upgrade over an existing profile #3279 received the latest clean-run non-occurrence plus a concrete stress/capture plan; it remains open.
Windows remains preview-only.
中文：已同步 #2142 正文，勾选 #3174 完成的 Phase 4 对齐，记录 #3382/#3340 与 #3265 当前状态，并把 Abort/Quit/硬杀/断电边界拆开表述；签名、required ruleset、安全证据/人审、baseline 与 computer-use 仍保持开放。
Roadmap sync (2026-08-23):
- feat(win): Abort-path installer rollback with backup retention #3265 merged as
9de05e266; Phase 3 Abort-path rollback/backup-retention item is now checked. The formal-support decision for hookless Quit, hard kill, and power loss remains open. - ci(windows): gate packaged sandbox lifecycle evidence #3558 exact head
229d09c72is mergeable and Core/Release Windows green. It promotes the existing cancellation/parent-death/concurrency/Job-drain/ACL-recovery smokes into the shared packaged and formal-release verifier. It still requires independent human review. - test(windows): verify packaged sandbox client cancellation #3586 exact head
bbbb6b4f1is mergeable and Core/Release Windows green. It adds explicit post-dispatch client cancellation evidence (aborted,launch,dispatched:true), broker/AppContainer process-tree exit, a successful recovery launch, and no matching recovery ledger. The packaged stage passed three times in one run: initial package, pinned upgrade, and automatic update. It still requires independent human review. - The Phase 4 lifecycle checkbox remains open. Remaining technical slices are Runtime Host death mid-launch, sustained concurrency soak, and documented unsettled-state recovery/quarantine; the adversarial matrix and independent human security review remain separate gates.
Windows remains preview-only.
- feat(win): Abort-path installer rollback with backup retention #3265 merged as
Related context: sandbox approval pre-dispatch
Discussion #3629 is the design context for the Phase 4 permission-escalation flow tracked here.
The proposal is to let Runtime preflight the original tool invocation before dispatch:
allowed: dispatch immediately;approval_required: keep the original invocation pending, ask the user, revalidate the current boundary, then dispatch it for the first time;denied: return a typed denial;unsupported: return a typed capability error.
For Windows/AppContainer, user approval must not silently turn an unenforceable capability into unsandboxed execution. Unsupported capabilities should remain explicit and fail closed. A
bypassboundary should dispatch directly without creating an approval request. Bash may retainrequired_boundarybecause Runtime cannot safely infer the resources used by an arbitrary shell command; Runtime should own the approval state machine around that declaration.This is design context only; it does not claim that the Phase 4 sandbox-security gates are complete. The preview-only support boundary and the remaining lifecycle, adversarial, and independent-review gates remain unchanged.
Phase 3 boundary decision and Phase 4 continuation
The Phase 3 formal-support boundary is now recorded for the Windows 11 x64 preview:
- verified Abort-path failures recover automatically;
- hookless
Quituses retained backup evidence and exact-identity rerun recovery; - hard-kill and power-loss scenarios remain outside the automatic rollback guarantee and may require repair or reinstall;
- full automatic recovery is deferred to a future release-hardening slice.
This closes the boundary-decision item only. It does not declare Windows fully supported.
Phase 4 remains the active track. #3558 is at exact head
f65dc9d15with Core, dependency-audit, and Release Windows checks green; it remains open pending maintainer review and one review-thread resolution. #3586 has merged the explicit packaged client-cancellation evidence.The remaining Phase 4 gates are Runtime Host death during launch, sustained concurrency soak, unsettled-state recovery/quarantine, the adversarial escape matrix, and independent human security review.
Phase 4 automated evidence update
#3722 now has exact-head packaged Windows L3 green at
6970405e99a7e0aaba28c1f5eb30227aeb57b320:- Runtime Host parent-death cleanup;
- 64-launch packaged concurrency soak;
- quarantined ACL-ledger non-reuse;
- packaged filesystem/network/IPC/environment/registry/parent-token/descendant matrix;
- pinned upgrade/uninstall;
- automatic update;
- deterministic rollback.
The W0 native broker protocol lane is also green. The ordinary
testlane is independently red in the pre-existingpackages/eval/harbor/test_relay_lifecycle.pytest (late_writeremained present); #3722 does not modifypackages/eval. The Phase 4 lifecycle checkbox remains open until #3722 is merged and the required independent human security review is complete. UDP/DNS/SMB, inbound listener enforcement, Authenticode, no-Win32k/window-station isolation, power-loss automatic recovery, and direct Credential Manager/DPAPI probes remain explicitly deferred.The Computer Use portion of this roadmap now has a dedicated implementation issue: #3785.
#3785 reviews the current
maka-agent/maka-cuWindows prototype and proposes the concrete Phase 5 work: reuse the platform-neutralmaka.cu/2host contract, bind actions to process/window generations and opaque snapshot tokens, use a target-window capture API, make UIA/Win32 fallbacks capability-driven and verifiable, define DPI/session/UIPI/security behavior, and add fixture plus Windows E2E evidence.It is intentionally scoped as a child of this umbrella issue rather than a second Windows-support roadmap. The acceptance criteria in #3785 are meant to be the Computer Use-specific gates for Phase 5 here.
Roadmap sync (2026-08-25)
The issue body is now synchronized after #3558, #3586, and #3722 merged.
- The shipped Windows 11 x64 W1 filesystem-worker AppContainer surface has completed its Phase 4 automated lifecycle/adversarial evidence and independent maintainer/security review gates.
- ci(windows): gate packaged sandbox lifecycle evidence #3558 owns the shared packaged/formal-release lifecycle verifier and its trigger/source-closure authority; test(windows): verify packaged sandbox client cancellation #3586 owns explicit client cancellation; test(windows): close packaged sandbox lifecycle evidence #3722 owns Runtime Host parent death, validated owner binding, Job drain, the 64-launch soak, quarantined-state non-reuse, and the scoped adversarial matrix.
- Windows remains preview-only. Phase 3 now has one implementation blocker: Authenticode identity/signature verification in infra: obtain ASF code signing identities for Desktop artifacts #3414.
- The next execution order is: signing (infra: obtain ASF code signing identities for Desktop artifacts #3414) -> authoritative Windows baseline and required
windows_recoverygate (ci(windows): eliminate hidden failures in the non-blocking baseline #2624 plus repository administration) -> one support-declaration cutover PR. - Direct Credential Manager/DPAPI probes, broader network channels, no-Win32k, dedicated window-station/clipboard isolation, power-loss automatic recovery, W2 general-command sandboxing, and Windows computer-use remain explicit separate work. They are not claimed by the completed W1 preview surface.
- The eval descendant-cleanup timing issue test(eval): harden relay descendant cleanup against process-group timing #3770 remains separate and is not a Windows support blocker.
Support should not be declared until the remaining clean-run, fail-closed, signing, and required-gate criteria in the body are satisfied.
Windows plan item 2 progress: #3796 merged the current-main bootstrap repairs; #3789 is now rebased and exact-head green for both required \ est\ and automatic \windows_recovery. The recovery run executed every native crash/owner-death group, not just workflow setup. #3789 also declares \windows_recovery\ in the ASF-managed required contexts; the roadmap administration checkbox remains open until #3789 merges and ASF infrastructure applies the updated protection. #2624 still tracks PTY/Git cleanup, symlink-capability classification, and the required three clean scheduled baseline runs.
Windows baseline progress update (2026-08-26)
The post-#3789 full baseline was started on byte-identical
apache/main@32a1db0ebin the maintainer fork because this account cannot dispatch the Apache workflow (GitHub HTTP 403, Actions-admin permission required).The first run exposed one deterministic root-authority fixture failure in both focused and full Storage. That was fixed separately in #3872, which adds the exact root race to
windows_recoverywith strict1 test / 1 pass / 0 skipevidence. #3872's hostedtestandwindows_recoveryare green.Three subsequent full-baseline runs on the fixed exact head are clean:
Each run reports focused Storage 88 pass / 0 fail / 12 skip, full Storage 940 pass / 0 fail / 38 skip, successful Runtime PTY and PowerShell UTF-8 gates, clean CLI/Electron smoke, and an empty residual-process audit. These are fork diagnostics on the Apache main tree; they are not a replacement for Apache-owned scheduled/admin-triggered evidence.
Remaining Phase 1 administration work:
- trigger or authorize three equivalent full-storage runs on
apache/main; - apply and verify
windows_recoveryas an effective required check in the GitHub ruleset.
After those organization-owned checks are clean, the Windows baseline item in #2624 can be closed. Authenticode #3414 remains the separate Phase 3 blocker before the roadmap can change Windows from Preview to Supported.
- trigger or authorize three equivalent full-storage runs on
#3872 has merged as 235a12d. On byte-identical �pache/main@235a12d, three full fork baseline runs are clean: 950/0/38 full Storage, 88/0/12 focused Storage, and empty residual-process evidence in each. The technical baseline work is complete; remaining Phase 1 actions are ASF-owned three-run confirmation and effective windows_recovery ruleset enforcement. Phase 3's remaining product blocker is Authenticode #3414.
Status refresh, 2026-10-09. This roadmap is still current; Windows stays a preview target.
- Phase 3 signing: infra: obtain ASF code signing identities for Desktop artifacts #3414 closed as not planned for 0.2.0 when the Apple and Windows signing requests were withdrawn to focus on the source release. The Authenticode item, and the "signed install and update artifacts are published" exit criterion, are deferred, not done. The support-declaration cutover (step 3 in the plan) stays blocked until signed convenience artifacts return to a release plan.
- Baseline: ci(windows): eliminate hidden failures in the non-blocking baseline #2624, eliminating hidden failures in the non-blocking baseline, is still open, and so is making
windows_recoverya required check in the effective ruleset. - Phase 5, computer use: feat(windows): harden and productionize the maka-cu Computer Use executor #3785 and its implementation PR feat(computer-use): add guarded Windows maka.cu/2 integration #4668 are in progress.
- Since the last update: the Windows pre-commit hook is fixed (bug(ci): pre-commit hook cannot spawn biome.cmd on Windows (EINVAL) #5668 via fix(ci): run Biome native binary on Windows #5677). It had blocked every commit on Windows with Node's
.cmdspawnEINVAL.
No phase checkboxes change.
Problem
Maka can already build and launch its CLI and Electron desktop app on Windows, and the codebase includes Windows-specific paths for named pipes, PowerShell/cmd shell detection, ConPTY, process-tree termination, and filesystem behavior. However, Windows is not currently a supported platform because these paths are not continuously tested or released, and several product guarantees remain platform-specific.
A local Windows baseline on Node 22 found:
core.longpaths=true);514 pass / 100 fail / 40 skip;EBUSYcleanup failures caused by SQLite handles still owningruntime.sqliteorruntime.sqlite-shm, which POSIX permits unlinking but Windows does not;Being able to open the app is therefore not the same as having a tested, secure, releasable Windows product.
Desired outcome
Define and deliver an explicit Windows support tier for Maka. At minimum, supported CLI and desktop workflows should build, test, install, update, recover from crashes, and fail safely on every release. Features that cannot initially be supported, especially sandboxing and computer-use, must be clearly surfaced rather than silently degraded.
Current support status (2026-08-25)
Windows 11 x64 remains an active preview target, not a fully supported Maka platform. Phase 0-2 build, baseline, crash-recovery, process-cleanup, and durability evidence are in place. Phase 3 preview packaging, checksum/install documentation, closed-app upgrade/uninstall, automatic update, Runtime Host update handoff, and the accepted Abort/Quit recovery boundary are complete; Authenticode signing is the only remaining Phase 3 implementation gate and is tracked by #3414.
The shipped W1 filesystem-worker AppContainer surface has now completed its Phase 4 automated and independent-review gates. #3558 promotes the lifecycle checks into the shared pull-request/formal-release verifier and protects its trigger/source closure; #3586 verifies explicit post-dispatch client cancellation; #3722 closes Runtime Host parent-death, owner binding, Job drain, a 64-launch soak, quarantined ACL-ledger non-reuse, and the scoped adversarial matrix. All three are merged, and #3722 received exact-head independent maintainer approval with no P0-P3 findings.
This completes the current W1 preview sandbox gate, not general Windows support and not the wider W2 command tier. Full support still requires signed artifacts, a current authoritative Windows baseline with required gates, and a final support-declaration cutover. Windows computer-use and the explicitly deferred sandbox-hardening rows remain separate future work.
Phase 0: establish the baseline
Tracking PR: #2156
windows-latestand publish pass/fail/skip results.process.platform === 'win32'skip and classify it as POSIX-only, portable, or missing Windows implementation.Phase 1: continuous Windows correctness
Tracking PR: #2173
Current status (2026-08-08): the combined Node 24
mainWindows storage baseline is 703 pass / 6 fail / 47 skip, improved from the original 514 / 100 / 40. The non-blocking baseline lane retains complete diagnostics as artifacts.Completed Phase 1 failure groups include project-catalog and usage-store SQLite shutdown, portable Codex/sandbox path handling, and deterministic Git worktree retirement on Windows. Current implementation focus is grouped by root cause rather than individual failing tests:
root-authority file identity, replacement, lock, and initialization races: completed by fix(storage): serialize root marker repair on Windows #2438;
session-bundle deterministic metadata and atomic hydration semantics: completed by test(storage): cover Windows session bundle metadata #2449;
managed Git head-ref long-path fixture: completed by test(storage): enable long paths for bare Git fixtures #2435;
Windows skip inventory refresh: completed by docs(windows): refresh test skip inventory #2434.
Add a required or explicitly non-blocking
windows-latestCI lane while the backlog is being burned down. Completed by ci(windows): add non-blocking baseline lane #2173.Enable Git for Windows long-path handling in managed workspace operations. Completed by test(windows): establish support baseline #2156.
Close SQLite stores, owners, and leases deterministically before temporary-root cleanup; do not mask open-handle bugs with broad retry loops. Completed across test(storage): close project catalogs before cleanup #2339, test(storage): close agent run stores before cleanup #2341, fix(runtime): classify expired probe budget as timeout #2344, test(computer-use): synchronize session cancellation #2345, and fix(storage): close usage stores after lease revocation #2365.
Validate named-pipe endpoint lifecycle and two-client runtime-host flows on Windows. Completed by test(runtime-host): validate Windows named-pipe clients #2464.
Validate ConPTY, cancellation,
taskkill /T, and descendant cleanup behavior. Completed by test(runtime): validate Windows process lifecycle #2465.Unskip portable tests and retain narrowly explained skips only for genuinely POSIX-specific contracts. Completed by test(windows): resolve portable skip inventory #2470.
Phase 2: crash and recovery guarantees
windows_recoverylane and validates SQLite, Runtime continuation, Runtime Host owner-death, and managed-workspace crash recovery on every pull request andmainpush.Repository administration follow-up:
windows_recoveryas a required check in the effective GitHub ruleset. The workflow is green on the ci(windows): require crash recovery evidence #2562 merge commit, but the current token cannot verify or modify organization-level rulesets.Phase 3: distribution
Current status (2026-08-25): Windows x64 preview packaging, checksum verification, installation documentation, closed-app upgrade/uninstall, CI-verified automatic updates (#3240), safe Runtime Host update handoff (#3382), and the accepted Abort/Quit recovery boundary (#3265) are complete. Authenticode identity acquisition and release-workflow signature verification remain the only Phase 3 implementation gate and are tracked by #3414.
9de05e266): the final exact head passed CI/audit/Windows L3 with 234 entries restored with 0 diffs, registry-mismatch 103 retention/recovery, stale and incomplete backup 101 refusals, hookless Quit retention/adoption, registration-less 101 refusal, and fixture-scoped uninstall-registry ownership.Quit, hard-kill, and power-loss failures. Accepted for the Windows 11 x64 preview: verified Abort-path failures recover automatically; hooklessQuitsupports exact-identity rerun recovery with retained backup evidence; hard-kill and power-loss scenarios remain outside the automatic rollback guarantee and may require repair or reinstall. Full automatic recovery is deferred to a future release-hardening slice.Phase 4: sandbox security (shipped W1 preview surface)
Current status (2026-08-25): the Maka-owned packaged AppContainer backend (#2961), production-identity readiness probe, per-launch private desktop, and EN/zh RFC alignment (#3174) are on
main. The supported claim remains limited to the Windows 11 x64 W1 filesystem-worker preview surface with packaged fail-closed enforcement; it does not claim the wider W2 general-command tier or escape-proof GUI/window-station isolation.The remaining W1 lifecycle and adversarial evidence is now merged:
ci(windows): gate packaged sandbox lifecycle evidence #3558 merged as
7edc971c1: the pull-request and formal-release lanes share the packaged lifecycle verifier; esbuild-derived source closure, release-workflow triggers, and the gate workflow's own release-contract route are protected by tests.test(windows): verify packaged sandbox client cancellation #3586 merged as
1ecd9bd6d: explicit post-dispatchFilesystemWorkerClientcancellation drains the packaged broker/AppContainer process tree and permits a clean recovery launch.test(windows): close packaged sandbox lifecycle evidence #3722 merged as
7235069ad: Runtime Host parent death, validated owner-handle binding, kill-on-close Job drain, an 8-wave x 8-way (64-launch) soak, quarantined ACL-ledger non-reuse, and the scoped packaged adversarial matrix are verified. Exact-headtest,package, and W0 checks passed, and independent maintainer/security review approved the head with no P0-P3 findings.Complete the comparative threat-model/RFC research. The initial docs(windows): define sandbox backend security contract #2940 proposal is retained only as history and is superseded by the updated RFC in feat(windows): add brokered AppContainer sandbox support #2961.
Merge the packaged AppContainer backend, Runtime/filesystem-worker integration, release resource, and updated RFC in feat(windows): add brokered AppContainer sandbox support #2961. Completed by feat(windows): add brokered AppContainer sandbox support #2961 (merged 2026-08-17, squash commit
d3a21f9).Align the RFC's implemented guarantees with the shipped slice. Completed by feat(windows-sandbox): place sandboxed children on a private desktop #3174: production-identity readiness and initial private-desktop placement are enforced.
Promote cancellation, parent-death, concurrency, process-drain, and residual ACL/state evidence to the packaged/per-release gate. Completed by ci(windows): gate packaged sandbox lifecycle evidence #3558, test(windows): verify packaged sandbox client cancellation #3586, and test(windows): close packaged sandbox lifecycle evidence #3722.
Complete the automated adversarial matrix for the shipped W1 preview surface. test(windows): close packaged sandbox lifecycle evidence #3722 covers filesystem aliases/reparse points, restricted TCP, host named pipes, descendants, ambient environment, credential-file reachability, host registry values, parent-token access, and lifecycle failures.
Complete independent maintainer/security review for the shipped W1 preview surface. ci(windows): gate packaged sandbox lifecycle evidence #3558's review findings were closed; test(windows): close packaged sandbox lifecycle evidence #3722 was approved at exact head with no P0-P3 findings.
Explicitly deferred hardening, not claimed by the current W1 preview:
These deferred rows require their own scoped issues and threat-model decisions before implementation. They do not reopen the completed W1 preview evidence gate unless maintainers expand the advertised W1 contract.
Next execution plan
windows_recoveryas a required check in the effective ruleset. Acceptance: the clean Windows runner has an explicit pass/fail/skip inventory and no non-blocking failure can be mistaken for support evidence.The unrelated eval process-group timing failure observed while validating #3722 is tracked separately in #3770 and is not a Windows support blocker.
Phase 5: computer-use
Support criteria
Windows should be advertised as supported only when:
Alternatives or workarounds
Today, developers can run the CLI directly and launch the Electron desktop app in development mode on Windows. WSL2 can provide a Linux execution environment for some workflows. Neither workaround supplies a native Windows release, Windows sandbox enforcement, Windows computer-use, or release-level regression coverage.
This umbrella issue is intentionally phased. CI, storage lifecycle correctness, and distribution can proceed before the sandbox and computer-use projects, while the support criteria prevent partial availability from being mistaken for full platform support.