Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

58.8 KB
Raw Blame

Windows desktop end-to-end results

Status: functional acceptance and local VM cleanup complete; one backend cleanup restriction, 20 October 2026. This is the C5.2 acceptance record for feat/windows-desktop. The seven scenarios are defined in the Windows desktop design, §4.4. Desktop 1.0.0, 1.0.1, 1.0.2, uninstall-fix 1.0.3, and first-key-fix 1.0.4 were published to staging for ARM64 and x64. An earlier ARM64 1.0.1-rc1 ZIP and local x64 Debug/Release builds were used for diagnosis; those were not the staging N1. The final native ARM bootstrap provisioned staging CLI 0.4.21-dev.win1 and PTY 0.4.19; its dedicated Windows account was removed after testing. The x64 VM retains the same staging CLI and PTY files with Desktop 1.0.4. Its test-only PTY 0.4.18 host was retired after the update checks. After identifying and retiring an obsolete test daemon, the published 0.4.19 host now serves the x64 GUI in the same Windows session. No PR, merge, or production release is part of this record.

Summary

# Scenario Result at this snapshot Driver
1 First install, sign-in, same-version launch Accepted across separate ARM/x64 runs: final ARM 1.0.4 provisioned missing Node/CLI/PTY and rendered Sign in in 160.404s; supported x64 test PC QR sign-in had already returned success and automatically signed in Desktop. Same-version launch converged to one installed window. Published x64 MOTW launch passed its observed Unknown Publisher warning. No fresh ARM sign-in or single uninterrupted journey is claimed. Script, visual automation; supported QR sign-in
2 This computer, terminal, Add Phone, remote Mac Passed: owned terminal creation, WIN_42 output, HTTP 200 handoffs on ARM and real x64 test PC x64, and remote row observed. Published x64 1.0.3 and PTY 0.4.19 passed GUI ping/Ctrl+C and synthetic text copy/paste. Duplicate cursor replies caused a new terminal to lose its first key. The source fix passed clean x64 Debug checks and the published 1.0.4 retest: the first physical a appeared in a fresh GUI-created terminal and was independently confirmed in passive PTY output. The owner completed Mac → Windows control; passive VM output confirmed two MAC_42 command/result pairs. Visual automation, Script, Person
3 Notifications, tray, links, paste and font Accepted with explicit combined evidence: tray/links/font observed; published x64 1.0.4 composer bitmap, raw PNG-only and file-drop paste passed. On published x64 1.0.3, a completed Kilo turn posted a native Finished notification; clicking its card restored the correct harness. The live tray menu showed the working Kilo row, Open Harness and Quit. Published 1.0.4 cold protocol launch rendered one installed window; both-architecture native tests prove exactly-once forwarding. Transient banner and live URL-specific delivery log/count remain unobserved; the release app has no such delivery log. Visual automation, Script, native integration
4 Staging updates and broken-build rollback Passed on real x64: real 1.0.0 → 1.0.1 GUI update passed on both VMs. The 1.0.1 → 1.0.2 update kept the ARM counter rising and repaired Settings DisplayVersion on both VMs; the x64 test PC then updated from 1.0.2 → 1.0.3 through the GUI and rendered signed in. Private fake-CDN truncated/no-executable packages were refused; a no-first-frame package rolled back to 1.0.3 after two failed starts. Counter and daemon/PTY processes survived. Script, visual automation
5 Uninstall through Windows Settings Passed on 1.0.3 after a 1.0.2 timing failure was fixed: holding the final dialog open over three minutes did not start cleanup early; after OK, the app folder and registrations disappeared while CLI, daemon and agents stayed. UI Automation
6 x64 desktop on ARM64 Passed across two runs of published x64 1.0.3: install, rendered window, tray and Settings uninstall passed; a later isolated x64 runtime returned HTTP 200 for Add Phone while the backend was connected. The installed ARM app and seven agents were restored afterward. Script, visual automation, UI Automation
7 Install desktop by one PowerShell command Passed on real x64: cold staging 1.0.3 install rendered immediately, console showed the full completion copy and still here, MOTW/temp counts were zero, and the subsequent Settings uninstall removed the app while preserving agents. A later cold ARM 1.0.4 install rendered at the first passive capture and showed the full completion copy; the earlier white 1.0.3 run is retained below. Script, visual automation, UI Automation
— Windows PC controlled from the Mac desktop Passed: owner confirmation plus passive VM MAC_42 output. Person, passive Script

The Windows Desktop workflow run 39298570517 passed its x64 and ARM64 test jobs at code commit 290ce5e83; that is build and test evidence, not a pass for the remaining live scenarios. The staging-trigger commit immediately after it had the same source tree. The first CLI/PTY staging attempt for newer versions stopped at a Windows-only test-fixture error before publication. The corrected CLI/PTY run 39309042273 published CLI 0.4.21-dev.win1 and PTY 0.4.19-dev.win1, with served bytes verified. The Desktop N1 run 39309318609 published x64 and ARM64 1.0.1 ZIPs and verified the served manifest and artifacts. The real N → N1 updates described below add live evidence; the broken-build rollback and the later one-command installer retests passed on real x64. The full Windows run 39313769196 passed both x64 and ARM64 jobs with the Settings-version launcher fix at 147ca03bd, including the one-command installer specs. The 1.0.2 staging run 39323499313 built both packages and passed the served-byte checks. The Windows run 39330487316 passed both architectures at the uninstall-dialog fix 826608097. The 1.0.3 staging run 39338310778 built both packages and verified served bytes; the CDN manifest names 1.0.3 on both architectures. The 1.0.3 ARM Settings retest below passed after the 1.0.2 timing failure. The staging app-installer run 39343243147 published desktop/install.ps1 and verified its served staging lines. The served file's SHA-256 matched the local --app --print output: a6309048b7987a1c0062b4a77fa364bf0bd1f25afa11a454475cb51e1921ae09. The full Windows run 39343243094 then passed its x64 and ARM64 build, Rust, full Flutter, trusted-pipe, single-instance and one-command installer jobs at 4ba0c9359. Subsequent commits through 50cefaace changed only the Desktop workflow trigger, its focused test and this E2E record. The trigger removal passed its focused Vitest test and actionlint. The later cursor-report fix f6bceb53d changes runtime Dart code; its Windows validation run passed both architectures, including their builds, native integrations and 46/46 installer specs. Each full Flutter result contained 6,359 passes, 170 recorded baseline failures and zero new or mandatory-case failures. The downloaded test artifacts matched their API SHA-256 digests. The 1.0.4 staging run 39375401781 then built both packages, published only staging, and passed its served size and SHA-256 checks. The x64 package used for the live retest additionally matched its manifest hash after download.

Windows shortcuts

Status: 8/8 delegated live checks passed using composed source-specific evidence; local repair not merged or published, 20 October 2026. This is the acceptance record for the Windows shortcut design for CI/live candidate 38dde786ec6aecc2b071390d068b07f3e8116513 and local image-paste repair 9fcc9f769fba61399430d412e75d4654b25649af on feat/windows-desktop. The earlier C5.2 results cover their recorded builds; they do not validate this shortcut change.

Mac evidence is composed across the reviewed source change:

  • At frozen c420c1bac63df3788d6239f04d4affd5f2ae748e, the full VM suite passed uninterrupted: 7,533 passed, 41 skipped, 0 failed, all 644 selected VM files, no loader failures or retries, in 855.811 seconds. Full analyze exited 1 with the same 46 baseline diagnostics (5 warnings, 41 infos, 0 errors), zero new normalized diagnostics and no warnings/errors in changed files. Analyze is a known failing baseline, not a passing check. The baseline at 87779178e had 7,463 passes and the same 41 skips and 46 diagnostics.
  • The reviewed correction from c420c1bac to 38dde786e has 336 passing cases across 17 affected files, composed of 250 passing cases in 14 unchanged files plus 86 passing cases in three corrected files. The first broad invocation had 334 passes and four invalid Ctrl+K test assumptions; after correcting those tests, the three amended files passed. Production source stayed unchanged between those invocations. This is completed-file coverage, not one uninterrupted passing invocation. Final analysis of all 11 changed files reported no issues.
  • Both Mac scopes used Flutter 3.47.5 / Dart 3.13.4 and the unchanged dependency lockfile. The integrated correction's file and log hashes were verified, and independent review approved all five Important fixes and the evidence composition. The full suite was not rerun on 38dde786e; its unchanged coverage is reused with the scoped correction checks. VM tests do not establish physical Windows keyboard, IME, clipboard, or live-agent acceptance.

Exactly one dedicated Windows Desktop run 39398347220 was dispatched at 38dde786e and completed with SUCCESS. Both x64 and arm64 build-test jobs succeeded, including their final installer steps. Both completed full VM test artifacts were downloaded and their SHA-256 digests verified. Each architecture reports 6,473 passed, 99 skipped, and 170 recorded known failures (115 errors and 55 failures), with zero new failures against desktop/test/windows_known_failures.txt and zero must-pass failures. All 150 cases in the six affected shortcut, keymap editing/focus, clipboard, composer, and Edit menu files passed on each architecture. Workflow success includes the recorded known-failure baseline; it does not mean every VM test passed. Release jobs remain SKIPPED, and no staging release is part of this shortcut record. This CI receipt covers 38dde786e only; it does not cover the later local image-paste repair.

Live screenshot paste failed on that CI build. Local correction 9fcc9f769 adds native Windows local panes to the existing guarded clipboard-PNG/session image route. The existing CLI fallback saves the PNG and pastes its file path when a Windows clipboard writer is unavailable; no CLI or engine-specific change was made. Scoped verification passed 105 cases across four files on macOS and 105 cases across the same four files on Windows, with no Windows skips or errors. The Windows scope used Flutter 3.47.2 / Dart 3.13.2 x64 under ARM64 emulation with verified source hashes and the unchanged lockfile. Fresh analysis of the two changed files reported no issues, and independent review approved the correction. The changed-source local Windows x64 release build also PASSED with flutter build windows --release --no-pub; its PE header confirms x64. This build used the same x64 toolchain under ARM64 emulation. The running repaired release bundle's executable, native DLL, and app.so hashes matched that completed build. Actual Ctrl+V then showed an image attachment, and genuine Claude identified the controlled screenshot's text and red rectangle. This live image retest PASSED on 9fcc9f769. The single full CI receipt remains limited to 38dde786e.

The owner delegated live checks in Windows 11 ARM64 under UTM using this run's exact x64 CI build under emulation. Initial checks used a separate standard-user profile: cases 3 and 8 passed. Cases 1, 2, 4, and 7 also have passing receipts from administrative-profile testing. The real remote-composer check also passed on 38dde786e. The image case passed on local repair 9fcc9f769, bringing live acceptance to 8/8. At the owner's request, live testing has resumed in an existing administrative profile; the standard test session was preserved. The exact x64 bundle was copied to a dedicated task folder and its hashes matched the CI artifact. The owner completed fresh Harness login in this profile; the exact candidate has a usable local workspace. Existing provider sessions are used in place through their normal clients; credentials are not copied. A genuine Claude Code session produced an actual text response. Screenshot Ctrl+V produced no attachment on the CI build; that failure remains recorded at 38dde786e. Its retest passed on the verified local 9fcc9f769 bundle. In the real remote Claude composer, Ctrl+W deleted a word and Ctrl+T opened a persistent new tab with a free slot. The earlier unsuccessful attempt matched the existing New Tab menu behavior; freeing an owned idle test tab allowed Ctrl+T to create the tab. This checks the terminal composer, without substituting New Harness. No native physical-x64 live acceptance is claimed; the occupied physical PC console and its existing session remain undisturbed. Keep machine addresses, accounts, private clipboard content, and screenshots with private content out of this record.

Live acceptance combines cases 1–4 and 6–8 from 38dde786e with the corrected image case from 9fcc9f769; it is not a full eight-case rerun on the later source. The local repair has not been pushed, merged, or released. Merge and publication remain outside this acceptance record.

# Delegated live check Result Driver
1 From the sidebar or New Harness composer, Ctrl+T opens a tab, Ctrl+W closes it, and Ctrl+N opens New Harness. PASS — standard-profile Ctrl+T/W plus administrative-profile find-field Ctrl+N opened the New Harness form Agent (owner delegated)
2 Inside the terminal view, Ctrl+R, Ctrl+W, and Ctrl+U reach the shell or agent. PASS — Ctrl+R/W/U reached a real local PTY program Agent (owner delegated)
3 Inside the terminal view, Ctrl+Shift+T opens a tab. PASS — standard profile: from the real Claude Code terminal at provider onboarding, Ctrl+Shift+T opened a tab Agent (owner delegated)
4 Ctrl+C with a terminal selection copies; without a selection it interrupts. PASS — selected text copied without interrupt; unselected Ctrl+C interrupted actual Claude generation Agent (owner delegated)
5 Ctrl+V pastes text, and a pasted screenshot reaches the agent. PASS at9fcc — earlier source38 text response retained; repaired image attached and genuine Claude correctly described its controlled text/red rectangle; source38 image failure retained Agent (owner delegated)
6 In the terminal composer, Ctrl+T opens a tab and Ctrl+W deletes a word. PASS at source38 — real remote composer Ctrl+W deleted a word; with a free slot, Ctrl+T opened a persistent new tab Agent (owner delegated)
7 The Edit menu's Cut, Copy, Paste, and Select All work in a text field; in a terminal, Cut is disabled and Copy, Paste, and Select All work. PASS — all four field actions worked; terminal Cut disabled and Select All/Copy/Paste worked Agent (owner delegated)
8 Holding Ctrl shows the tab numbers. PASS — standard profile: holding physical Ctrl showed tab-number hints in UTM Agent (owner delegated)

Two reviewed limits remain explicit: the spec's Ctrl+backslash sidebar example is excluded because no existing catalog command owns it, and removing an editor while its Edit menu stays open can briefly leave enabled rows visible. Invoking such a row revalidates the target and does nothing; reopening the menu updates availability. That display issue is an accepted deferred Minor, not an unvalidated stale edit.

Setup and evidence

The ARM64 Windows 11 VM was exercised in a dedicated standard-user desktop session, using the C5.1 one-shot interactive-session runner, Windows UI Automation where it exposed controls, screenshot-grounded visual automation for Flutter controls, and the owned CLI/PTY driver for read-only terminal observations. The driver was changed to a passive screen reader after its earlier screen command was found to take terminal control from Desktop. A separate Windows x64 VM was used for local Debug/Release checks. Both VMs had owned test terminals; existing unrelated panes were left untouched.

The detailed receipts and screenshots remain in the ignored C5.2 execution scratchpad, not in this public document. They include account or pane content in some captures. No screenshot is included here until it has been reviewed and masked. Test account names, machine identifiers, QR codes, credentials, private clipboard contents and unrelated terminal content are omitted here. Tooling acceptance combines separate positive receipts for PS5 parsing, native helpers, five real GUI checks including console detection, conditional text clipboard restoration and an actual owned-account lifecycle cleanup. Initial failed selftest/clipboard attempts remain recorded; no single all-green guarded transaction or arbitrary-format clipboard restoration is claimed. Shared Winlogon and unrelated sessions are preserved under the reviewed ownership rules. Final C5.2 cleanup is recorded separately below.

Scenario results

1. First install

The staging ARM64 N ZIP was extracted into a path containing Vietnamese letters, spaces and shell punctuation. Its launcher installed the desktop for the owned user and a signed-in session was observed. Launching the same N ZIP again resolved to one window running from the installed path. A later local reinstall of N after a diagnostic build opened one installed window in 14.7 seconds; this measures that reinstall only, not the first-install journey.

The Mark-of-the-Web copy carried a downloaded-file marker. Its first launch timed out after 45 seconds. A bounded retry returned after roughly 51 seconds, but the existing installed process and launcher log did not change, and no SmartScreen prompt was visible. Execution of that marked launcher is therefore unproved. The first cold start also showed a temporary computer-check timeout; the later signed-in state does not establish a clean first-run sign-in path.

A later published x64 1.0.3 ZIP matched the verified staging SHA-256. A separate copy of the ZIP and all nine extracted .exe/.dll files were marked ZoneId=3, in a folder containing Vietnamese letters, spaces, an apostrophe, &, % and !. Shell launch displayed Open File – Security Warning, Unknown Publisher, Run and Cancel. After a screenshot-grounded click on Run, one installed 1.0.3 process opened and rendered its existing workspace without a focus or resize workaround. The “More info / Run anyway” SmartScreen screen did not appear. No trust or SmartScreen settings were changed. The helper initially misread the Unicode path until its PowerShell 5 script was encoded with a UTF-8 BOM; its launch wait also timed out while the warning awaited input. These helper failures are retained separately from the subsequent successful marked launch.

The final clean-state retest used the existing dedicated standard Windows user, with its entire prior Harness home moved to an owned marked backup. The legacy .machine home was absent. Desktop, registrations, CLI, managed Node, PTY and auth/state paths were absent before launch; child environment overrides and PATH fallbacks could not reuse the old install. This was clean Harness state, rather than a newly created Windows profile or a reset OS.

The verified native ARM64 1.0.4 ZIP was freshly extracted into another Vietnamese/metacharacter path. First launch at 06:28:28 UTC provisioned the missing components; the passive 06:31:08 UTC capture showed the fully rendered Sign in screen and green ready footer after 160.404 seconds. There was no Retry, relaunch, focus/resize workaround, CLI prewarming or auth copy. The 300-second console watcher observed zero console windows. Two observation commands hit a private helper's startup heuristic; extending that helper's guard allowed capture without changing or relaunching the app.

Only afterward, managed Node answered 22.23.2, CLI 0.4.21-dev.win1 and PTY 0.4.19-dev.win1, protocol 2; Desktop, Node and PTY were all ARM64. Desktop was 1.0.4 with its first-frame marker, no staging directories and loggedIn=false; no auth session file existed. Same-version relaunch briefly created a windowless forwarding process, then settled to the original installed process and one visible window. The installer streaming transcript was not retained, so the actual rendered transition and resulting managed files/versions are the bootstrap evidence.

Sign-in acceptance reuses the earlier real x64 test PC supported QR flow: approval through the authorized Mac session returned status: success, restarted harnessd on that account and automatically moved the still-running Desktop from Sign in to its account and This Computer. This is separate x64 sign-in evidence, not a new ARM sign-in or one continuous first-install journey. The ordinary QR approval/confirmation flow remains supported by the final source; intervening optional sign-in features do not make the entire CLI auth implementation byte-identical.

2. This computer

The app's File → New Harness composer created an active Terminal harness in an owned folder on the ARM VM. In the installed RC1 after relaunch, visual typing of Write-Output WIN_42 and physical Enter produced WIN_42 plus a returned PowerShell prompt in both the GUI and a passive PTY read. A later single physical printable key was seen in the correct owned pane; this does not prove cold first-key reliability. Earlier first-launch attempts had missed printable input, and their lifecycle cause was unknown at that point; the later duplicate-cursor diagnosis below resolved the first-key loss.

The ARM RC1 GUI started ping -t 127.0.0.1, but three Ctrl+C input variants did not stop it. The exact owned process was stopped separately for cleanup. A Windows PTY host fix was verified with a failing-then-passing ARM regression and the full Windows PTY suite. A local x64 Release desktop using that patched test host then sent Ctrl+C through GUI → daemon → PTY, stopped its owned ping, and returned to the prompt. The published x64 N1 desktop later repeated the GUI Ctrl+C check on that same test-only patched PTY host: an owned ping -t child exited, Control-C and the prompt appeared, and the app and host PIDs stayed running. This proves the N1 GUI path with the patched host, while the published PTY 0.4.19 runtime had not yet been checked at that point; the later real x64 run below closes that gap. The local x64 Release also passed Ctrl+Shift+V paste and Ctrl+Shift+C copy of a synthetic marker. That local clipboard result does not validate staging N1.

Add Phone returned HTTP 200 before and after the owned daemon restarted. The app discovered a changed trusted pipe; the observed recovery took about 91 seconds, not the plan's sample 20-second wait. The remote Mac row was visible and untouched. ARM clipboard checks were held because the shared pasteboard could not be safely restored by the test helper. The x64 local Release clipboard result does not substitute for ARM staging acceptance.

A later real x64 check used installed staging 1.0.3 with the published 0.4.19 PTY host, both in the same interactive session. Before moving the empty host, an identity-guarded operation retired only an obsolete test CLI master and its two children; the current daemon stayed alive. A GUI-created Terminal printed WIN_42, then GUI Ctrl+C stopped its owned ping -t child, showed Control-C and returned the prompt; a process query confirmed the child exited. Ctrl+Shift+V executed a synthetic marker, and Ctrl+Shift+C copied it with the previous clipboard restored. This passes the published x64 Ctrl+C and text clipboard path.

The first command in that fresh terminal lost its leading character. Physical key injection reproduced the loss, while a temporary Debug trace showed the character reaching the Desktop input queue. In an isolated published-host experiment, a fresh PowerShell accepted its first a; injecting a duplicate cursor-position reply before it made the same a disappear. Both isolated panes were removed and that test host shut down. Commit f6bceb53d separates cursor reports from user input and consumes the renderer duplicate while preserving other replies. Its regression first failed with unwanted input frames; 153 related Mac Flutter tests, 129 Windows x64 tests, targeted analysis and independent review then passed. A GUI-created fresh terminal on the clean x64 Debug build then accepted its first physical a; a passive PTY screen independently confirmed the character. The full Windows CI subsequently passed both architectures. On published x64 1.0.4, the GUI created another fresh Terminal harness against published PTY 0.4.19. Its first single physical a appeared at the prompt without a warm-up key; passive PTY output independently confirmed PS C:\Users\<user>> a. This passes the published first-key retest.

3. Native Windows behavior

On ARM RC1, hide-to-tray and restore kept a single app process; the tray menu showed Open and Quit. Edge's native protocol prompt opened an owned terminal, a shell fork link reached the app's message flow, and a link into a closed app opened one window in roughly 2.5 seconds. Two shortcut launches 300 ms apart converged to one process and window. The notification link had no clear dispatch log, so exactly-once delivery was not proved in that live run. The font picker visibly listed Cascadia Code variants.

On local x64 Release, first hide showed the Windows “still running in the tray” toast, second hide did not repeat it, and the tray icon restored the same window. Bitmap, PNG and file-drop clipboard representations each attached one synthetic image in the New Harness composer. A test-helper clipboard restore error was diagnosed and corrected for the synthetic text baseline; arbitrary pre-existing clipboard formats were not proved restorable.

An owned Kilo turn on hidden staging N1 also did not complete: the backend was offline before the turn and then reset the connection. The counter pane continued to advance, and the N1 window restored from its tray icon without changing app, daemon or PTY PIDs. At that point, the product's completed-turn toast and its click target remained unverified. An interactive native probe showed that Windows could post a generic toast, but the product's permission check had returned HRESULT 0x80070490 earlier; the probe is not product-toast evidence. The working-agent tray row had also not been demonstrated; later x64 receipts below cover both paths.

With the backend connected, a later native ARM64 1.0.3 app had Desktop Notifications enabled and was hidden while one owned Kilo turn ran. Its pane returned the requested answer in 9.4 seconds, and the app log received a turn_summary notification with kind=done. A toast check about 46 seconds later and a masked screenshot about 74 seconds later found no visible toast. The toast may have expired before observation; its appearance and click target were unverified in that ARM run. The test Kilo agent was deleted, the original seven agents and rising counter remained available, and the native 1.0.3 window was restored. Desktop Notifications remained enabled.

On published x64 1.0.3, a subsequent owned Kilo turn completed with Desktop Notifications enabled and the app hidden. A native C++ WinRT history probe returned Setting=Enabled and the product's own C52 toast probe → Finished notification, including its protocol URL for the correct owned harness. At 2026-10-20 05:16 UTC, a same-session GUI worker opened Notification Center and captured that card while Harness remained hidden. After reviewing the screenshot, automation clicked the card with guards requiring unchanged foreground and input state. The sole installed app process restored with the correct Kilo harness focused; the app log independently recorded that harness opening and becoming terminal-ready. This verifies native notification delivery, visible Notification Center presentation and the product click path. The short-lived banner was not captured. UI Automation did not expose the notification card on this Windows build, so its earlier absence from an automation search was insufficient evidence of missing delivery.

At 05:22 UTC, a further real turn on that owned Kilo harness produced a native tray menu showing Working (1), C52 toast probe with 5s elapsed, Open Harness and Quit in the captured screenshot. The tray selector excluded taskbar app buttons before opening the actual notification area icon. This verifies the live working-row criterion.

On published x64 1.0.4, a final composer check separately attached a blue bitmap, an orange PNG copied as an Explorer-style file drop, and a green PNG-only clipboard payload. Each visible attachment was captured; none was submitted to an agent. The raw PNG case advertised only PNG, with a valid 190-byte PNG and its expected signature, so bitmap fallback could not account for the result. The helper initially offered a managed payload without getting an attachment, then prematurely disposed its raw clipboard owner window. Keeping that private hidden owner alive through the paste produced the green attachment. These failed helper attempts are retained; no product change was needed. Every completed probe restored the prior supported clipboard formats and text, and the synthetic attachments and fresh first-key terminal were then removed.

At 06:25 UTC, the installed published x64 1.0.4 app quit normally. With zero Desktop processes, one shell invocation of an owned notification URI started the installed app. After 16 seconds, the same-session probe saw one process, one visible window and a rendered workspace. Process argv contained the URI once, and the launcher recorded one start. The restored terminal tab cannot independently prove which URI selected it, and a passive agent-sync event is not a delivery log.

The acceptance ruling combines that actual release-app cold-launch evidence with final-source x64 and ARM64 native integrations that assert exactly [cold, warm] deliveries, including after settling and Quit. The C3.2 routing contract says a cold notification opens the app before its workspace handler listens. Production code has no URL-delivery log, so C5.2's sample log line cannot be captured without adding new observability. Live URL-specific multiplicity remains uninstrumented; it is not relabeled as an observed exactly-once result. The separately observed warm notification-card activation proves the real notification route. No duplicated or lost link was observed.

4. Update and rollback

On the x64 VM, the installed N binary and launcher matched the published N ZIP. Its signed-in UI offered 1.0.1. Clicking Update showed installation progress, then changed current to 1.0.1 and previous to 1.0.0. One N1 window opened; its executable and launcher matched the published N1 ZIP. The PTY host PID and all six owned Node PIDs were unchanged. About showed 1.0.1; the old and new app folders remained, with no staging folder. This was a real staging GUI update, but the bounded GUI attempts to start a long-running terminal command on N yielded only blank prompts, so a live counter was not established there. After the update, a bounded physical-key and GUI check on the same owned terminal executed a synthetic Write-Output marker and showed the expected output and a new prompt; the input failure did not recur in N1.

On ARM64, the owned user installed staging CLI 0.4.21-dev.win1 and PTY 0.4.19 first. This intentionally restarted that user's daemon; its new master/core PIDs were recorded before the app update. N offered 1.0.1 through the real staging manifest. After Update, current=1.0.1, previous=1.0.0, one N1 window rendered, and About showed 1.0.1. The N1 ZIP matched the manifest's size and SHA-256. The master, core and PTY host PIDs stayed unchanged across the app update. A corrected, credential-aware passive E2E driver then listed the same five owned agent IDs as before the CLI and app updates. A scoped quit and shortcut start left one rendered N1 window, only the N and N1 app folders, and no staging folder. The footer visibly separated words and counts; the terminal font was compact and unclipped in the normal pane. A new owned terminal counter rose from COUNT_14 to COUNT_31 before the next update and later to COUNT_2862; continuity across this first N → N1 update cannot be claimed.

The x64 VM then upgraded its owned CLI to staging 0.4.21-dev.win1 and installed PTY 0.4.19; the new PTY executable matched the published hash. Its previously running test-only 0.4.18 PTY host remained alive and served even a newly created GUI Terminal, so the x64 GUI Ctrl+C checks cannot be attributed to the published PTY. The old terminal reported offline ready after the daemon restarted, despite its PowerShell and host processes staying alive; the cause remains unproved. The new owned Terminal was active and passed GUI ping/Ctrl+C on that old host. The x64 N1 app stayed signed in.

After the x64 N1→N2 continuity checks, both owned test agents were stopped through the Harness API and their two pane processes ended. The idle patched 0.4.18 host was then stopped with an exact process-owner and image guard. A new host started from the published 0.4.19 executable, whose hash matched the served artifact, but another daemon in Windows session 0 won the shared-home startup race. The GUI and owned test daemon were in session 1. No GUI terminal was opened against the session-0 host, so the published-host GUI Ctrl+C path is unverified at that point; the later same-session x64 run below passed.

A later read-only ARM check found its rendered 1.0.3 app and serving PTY host both in Windows session 2, but the host process still ran from the 0.4.18 runtime path. The installed 0.4.19 file had a different hash and was not the serving process. No new terminal or Ctrl+C attempt was made on that host; its seven existing agents were left running.

Both VMs then offered the published 1.0.2 update in their installed N1 GUIs. On x64, Update changed current/previous to 1.0.2/1.0.1, wrote .started, removed temporary staging folders, and opened one signed-in 1.0.2 window. The app executable and launcher matched the verified x64 ZIP. About and the Windows uninstall entry both showed 1.0.2, including after one scoped quit and Start-shortcut relaunch. Daemon and PTY PIDs and both owned agent IDs and statuses survived. The old test-only PTY host still served the terminals.

On ARM64, Update also changed current/previous to 1.0.2/1.0.1, wrote .started, removed temporary staging folders, and rendered one signed-in 1.0.2 window. About and the uninstall entry showed 1.0.2. The same owned counter pane advanced from COUNT_4108 before Update to COUNT_4403 after it; daemon, core, worker and PTY host PIDs stayed unchanged. A GUI Quit later hid the window and removed the tray icon, but the app process remained alive for at least 47 seconds and exited by itself within 4 minutes 38 seconds. No process was killed. A subsequent installed-shortcut launch started one rendered, signed-in 1.0.2 window. The pointer and registry stayed at 1.0.2; the same seven agent IDs remained and the counter rose from COUNT_4940 to COUNT_5094 after that launch. The daemon, core, worker and PTY host kept their original PIDs and start times.

Both VMs still had DisplayVersion=1.0.0 in their Windows uninstall registry entry after N1 was running. The updater swaps the installed files and pointer without calling the install path that writes that value. A launcher fix in 147ca03bd now refreshes the existing entry from the version it opens. Its focused Rust desktop suite passed 97/97, and the published 1.0.2 GUI update confirmed the corrected registry value on both VMs.

On real x64, installed staging 1.0.3 then exercised three unpublished packages through its actual update banner and a private fake HTTPS CDN. A half-downloaded 1.0.6 ZIP and a complete, hash-valid 1.0.4 ZIP without the app executable were refused before changing current. A 1.0.5 package without data/app.so was staged, exited twice before its first frame, and the installed launcher logged the rollback to 1.0.3. The good app rendered again, current was 1.0.3, the consumed previous pointer and bad app folders were absent, and the owned counter advanced from COUNT_1 to over COUNT_450. Daemon and PTY PIDs and start times stayed unchanged. These versions were never published.

The fake-CDN transaction used a name-constrained CA, loopback-only listener and reverse SSH forward, with a restoration watchdog registered first. After testing, the hosts file was byte-identical to its original hash, the test CA and watchdog were absent, the reverse forward and local server were stopped, VM port 443 had no listener, and the real staging manifest again returned 1.0.3. No shared ARM hosts or trust settings changed.

5. Uninstall; 6. x64 on ARM; 7. One-command install

Windows Settings found the installed Harness 1.0.2 row, opened its menu, and invoked both Uninstall controls through UI Automation. The installed-app count fell from 43 to 42. Harness showed the final sentence that its command line and agents stay on the PC. The protocol and uninstall registry entries and Start shortcut disappeared; the same daemon PID, seven agent IDs and rising counter remained. But app/1.0.2 and the install folder still existed after the final dialog closed. The cleanup script's own log said it left the folder because its launcher was still running. Its failure timestamp preceded the final OK click: the script's roughly two-minute wait expired while the modal dialog kept the launcher alive. This is a failed full-uninstall check, not a successful Settings uninstall. The partial state was preserved for a fix and retest. Commit 826608097 moves script startup after the modal closes and adds a regression for that order. Windows CI passed.

The published 1.0.3 ARM64 ZIP matched the staging manifest's SHA-256 f754c3fd976647a83ef13db18095885ab18e5bf1dc7085358ea74f00c15082af and 103046548-byte size. Its launcher installed over the partial 1.0.2 state without manual cleanup: current=1.0.3, previous=1.0.2, Settings version, shortcut and protocol registration were restored, and one signed-in window rendered. The seven agent IDs and their daemon/PTY PIDs stayed intact.

Windows Settings then invoked Uninstall for 1.0.3. The final Harness dialog stayed open for over three minutes; before OK, no new cleanup script, task file or failure log existed. After the exact OK click at 23:21:48 UTC, the install root was absent by 23:22:15. Registry entries, shortcut and desktop app state were absent; no desktop process remained. The CLI daemon/core/worker and PTY host kept their original PIDs and start times, the same seven agent IDs remained, and the owned counter advanced from COUNT_11929 to COUNT_11966. The new cleanup script removed itself without leaving a failure log. This passes Scenario 5 on the fixed staging build.

The published x64 1.0.3 ZIP matched the staging manifest's 104011999-byte size and SHA-256 d22a129e40dcb43918f7449456181f54b9b81b79903480e326a24983bd34cf26. Its launcher installed the x64 app under ARM64 emulation. One signed-in window rendered with This Computer visible; the native tray icon existed. Add Phone opened its dialog, but /api/auth/handoff repeatedly returned HTTP 502 while the daemon reported backend offline and connection resets. No HTTP 200 was observed, so this does not prove the x64 FFI/pipe handoff criterion. A later read-only check still found 30-second ECONNRESET retries, while DNS, TCP 443 and unauthenticated TLS/HTTP to the backend succeeded; the authenticated WebSocket path remained unhealthy at that time. Settings then uninstalled the app through its final dialog. The install root, registry entries, shortcut and app state disappeared; no desktop process remained. The CLI stayed signed in with the same seven agents and unchanged daemon/core/worker and PTY host PIDs; the owned counter advanced to COUNT_12973.

After the backend recovered, the same published x64 1.0.3 runtime was copied to an isolated folder outside the launcher's self-install layout. Its executable hash matched the published ZIP. The installed ARM app quit, and the x64 app ran as the sole desktop process under emulation, rendered the signed-in UI, and opened Add Phone. One real Add Phone click produced POST /api/auth/handoff → 200 in its app log in 664 ms. No QR image was retained. The x64 app then quit, and the installed ARM 1.0.3 app was restored as the sole desktop process with its registration and version pointer intact. The same seven agent IDs, daemon/core/worker/PTY PIDs and process start times remained; the owned counter continued rising through COUNT_18687. Together, the two runs satisfy Scenario 6's x64 emulation, pipe handoff, tray and cleanup criteria. The earlier 502 was a backend outage, not an x64 FFI failure.

The real x64 test PC was also checked after the user prioritized x64. Its already signed-in, rendered 1.0.2 app ran in the active GUI session with the staging 1.0.3 update banner. Clicking Update installed 1.0.3 and restarted one app process in the same GUI session. Its home surface rendered signed in, the installed current pointer read 1.0.3, and harness status remained backend connected with zero available agents. On this x64 app, one actual Add Phone click produced POST /api/auth/handoff → 200 in 478 ms in the app log. The dialog was dismissed, the app remained at 1.0.3, and the daemon kept its PID and connected status. No QR screenshot was retained. This is live x64 GUI/update/FFI evidence; it does not establish published PTY 0.4.19 GUI Ctrl+C because that host then ran in a different Windows session.

A bounded attempt to create one Terminal from the x64 test PC's 1.0.3 GUI did not confirm a start; harness ps remained empty, so no input or Ctrl+C was sent. The x64 test PC's GUI daemon logged pty host is unavailable and repeated attempts to start a host despite an existing 0.4.19 host. Read-only probes found that host alive in Session 0, its discovery file naming the same PID, and its pipe connectable from both Session 0 and the limited GUI Session 1. However, the Windows process-table scan from Session 1 reported the host row with an empty creation marker; the Session 0 scan read the marker. The daemon drops rows without that marker, so it cannot prove that the discovery file still names the same process and will not trust the host. This bounds the failure to the cross-session process-identity check, rather than a terminal input or Ctrl+C result. A limited-session read-only follow-up could read the same creation time through CIM, while a .NET process-handle start-time read failed. An authenticated host inventory then confirmed zero panes. Its protocol accepted a guarded shutdown, but a Session 0 process started the replacement 0.4.19 host before the limited Session 1 task did. The test stopped after that single attempt. The x64 test PC remained signed in, backend connected, with the same daemon and zero agents; the replacement host also had zero panes. At that point, published 0.4.19 GUI Ctrl+C still needed a host started in a compatible user context. The later same-session x64 receipt in §2 passed that check.

The staging desktop/install.ps1 served the staging manifest base and enabled app installation. Its real irm …/staging/desktop/install.ps1 | iex command ran in the owned user's GUI PowerShell session. After 53.7 seconds, one installed ARM64 1.0.3 app process and window appeared. The version pointer, .started, Settings version, protocol and Start shortcut matched 1.0.3. No SmartScreen prompt was observed during the watched interval. Installed files had zero Zone.Identifier streams, and zero staging temp folders remained. A single same-version rerun showed part of the expected completion copy and the still here sentinel; the app obscured the full text before the console closed. The first-install screenshot showed a white client area, so a rendered signed-in UI was not established by the process and file checks.

Settings then uninstalled 1.0.3 through the native final dialog. The install root, registry entries and shortcut disappeared, with no desktop process left. The CLI remained signed in with the same seven agents and original daemon/core/ worker and PTY host PIDs; the owned counter advanced from COUNT_12973 to COUNT_14102. Scenario 7's install and cleanup checks passed in that run, but its first render and full completion-copy checks were unverified; that run was partial. The later x64 and ARM retests below supersede that status.

A later diagnostic install from the already verified ARM64 ZIP started one 1.0.3 app process. Its first screenshot had Windows Settings in front, so it could not establish whether the app client was white. Bringing that same app window forward without resizing showed the signed-in home UI. The Flutter child window had a normal nonzero client rectangle; the available app log had backend connection warnings but no renderer error. No resize was applied. This confirms that a later direct launch rendered, but does not establish why the one-command first-run capture was white. The diagnostic app was left installed at that point for the owner's Mac-to-VM check; later retests and cleanup below supersede that state.

On real x64, a later cold one-command test began with the install root and uninstall entry absent after an actual Settings uninstall. The visible PowerShell ran the real staging one-liner and completed in about 98 seconds, including the CDN download. The sole installed 1.0.3 process rendered its signed-in UI before any test focus or resize operation. The original console remained open, displayed the full installed/opening sentence followed by still here, and returned to the PowerShell prompt. Both sentences were captured with the rendered app behind it. Download markers and staging temp folders both counted zero; no SmartScreen prompt was observed.

The post-install Settings uninstall also passed: the root, uninstall/protocol registry entries and Start shortcut were absent, no app process remained, and the daemon/core/updater/PTY retained their exact PIDs and start times. The unrelated OpenCode harness was preserved. This passes Scenario 7 on real x64; it does not resolve the earlier ARM first-render observation.

A final cold ARM64 staging 1.0.4 one-command run started at 2026-10-20 05:42:54 UTC after the 1.0.3 root and registrations were removed. The real staging irm … | iex; Write-Host "still here" completed in about 49 seconds. At 05:43:57 UTC, the first passive screenshot showed the signed-in composer rendered, before any app focus or resize. The installed app was the sole native 1.0.4 process; installed MOTW and staging-folder counts were zero. Its daemon, core, worker and PTY PIDs and start times matched the pre-install state, and the seven original agents were preserved. A later full console screenshot showed the complete installed/opening sentence, Start-menu sentence, still here and the PowerShell prompt. Focusing that console initially returned false in the helper; hiding only the owned app exposed the console for the successful capture, and also displayed the native first-hide toast. The installed shortcut then restored the app, and only the completed installer shell was stopped. The prior white 1.0.3 run did not recur here; this is a successful retest, not a diagnosis or a claimed rendering fix. The earlier post-uninstall receipts still cover the unchanged Rust uninstaller inputs.

The Windows PC seen from the Mac

The owner confirmed Mac → Windows control was complete. A fresh passive VM terminal read showed two Write-Output MAC_42 commands, each followed by MAC_42 and a returned PowerShell prompt. The owner performed the Mac side; this continuation did not operate the Mac's real Harness app, CLI, daemon or ~/.harness.

Findings and fixes

ID Finding Current evidence and impact
G1 ARM RC1 GUI Ctrl+C did not interrupt an owned PowerShell child. Isolated PTY host inherited Ctrl+C ignore from its process group. A Windows regression failed before the host fix and passed after it; the full Windows PTY suite passed. Local x64 Release and staging N1 passed GUI Ctrl+C with the patched test host. A later real x64 1.0.3 GUI check passed Ctrl+C through published PTY 0.4.19 after a guarded same-session host setup. ARM preserved its live sessions on 0.4.18 until final cleanup; the final fresh bootstrap provisioned 0.4.19.
G2 Printable terminal input depended on launch/relaunch state. ARM RC1 failed early input attempts, then WIN_42 passed after relaunch. An isolated published-host experiment proved a duplicate cursor-position reply causes the lost first key. Fix f6bceb53d suppresses only the renderer duplicate, preserving user input and other queries; 153 related Mac tests, 129 x64 Windows tests and independent review passed. The clean x64 Debug GUI accepted its first physical a in a fresh terminal; a passive PTY screen confirmed the character. Published x64 1.0.4 also accepted the first physical a in a new GUI-created terminal; passive PTY output confirmed it.
G3 The ARM client area turned black/white during RC1, N and one-command 1.0.3 sessions. The process stayed responsive at the OS level. N did not repaint after one app restart, but one scoped window-size change restored its UI; the surface went black again after the CLI restart and another scoped size change restored it. The GPU accepted D3D feature level 12_1. N1, N2 and x64-on-ARM 1.0.3 rendered in other launches; the one-command ARM64 1.0.3 first-run capture was white. A later direct ARM64 ZIP launch rendered after focus without a resize. The final cold ARM64 1.0.4 one-command run rendered in its first passive capture, without focus or resize. That retest passed; the earlier intermittent white surface is retained without claiming a root-cause fix.
G4 Notification permission gate reported unavailable while a native probe could post a toast. The first-toast Setting() error is handled in N1 source. Published x64 1.0.3 subsequently posted its own Finished notification, displayed it in Notification Center, and restored the correct harness when clicked. Transient banner presentation remains unverified.
G5 Published N footer joined its words and counts. Staging N1 on both x64 and ARM64 visibly separates the words and counts.
G6 Windows Settings uninstall entry kept the old version after a successful in-app update. Both staging 1.0.0 → 1.0.1 updates left DisplayVersion=1.0.0 while About showed 1.0.1. Launcher fix 147ca03bd has a failing-then-passing regression; live 1.0.1 → 1.0.2 updates on both VMs now show DisplayVersion=1.0.2.
G7 An x64 owned terminal reported offline ready after the staging CLI upgrade. The old PowerShell and test-only PTY host stayed alive, while the restarted daemon saw the host and reopened its stream. A new owned terminal worked on the same host. Exact cause and recovery of the old agent remain unproved.
G8 ARM 1.0.2 Quit took more than 47 seconds. The menu Quit removed the tray icon and hid the window, but its process remained alive on two checks and exited by itself within 4 minutes 38 seconds. The native teardown boundary is suspected, not proved; no forced termination or source fix was made.
G9 Settings uninstall left the installed files. The hidden cleanup script started before the final modal dialog, exhausted its launcher-exit wait while that dialog remained open, and logged that the launcher was still running. Registry and shortcut were already removed. Commit 826608097 fixes the order; Windows CI passed and the 1.0.3 ARM Settings retest passed after a >3-minute dialog hold.

Functional acceptance uses combined real x64 and ARM64 evidence, following the owner's x64 priority. The original spec names an ARM VM; Ctrl+C, completion notification activation, final composer paste and bad-update rollback use the explicitly labeled x64 receipts rather than fictional ARM passes. The clean ARM bootstrap, separate supported x64 sign-in and combined cold-link ruling above close the corresponding functional gates. Historical render/quit/old-terminal findings and the live observability limits remain visible. Final Windows-account cleanup and backend-record disposition are separate from those product results.

C5.3 local regression evidence reuses the complete Mac base comparison at 866f9347f and full suite at 290ce5e83: 7,181 passes / 17 skips versus 7,437 passes / 41 skips, with zero regressions. The newer runs used Homebrew Flutter 3.47.5 / Dart 3.13.4. After that full run, only three Desktop Dart files changed for the cursor fix; the affected nine-file check passed 153 tests and scoped analysis found no issues. Its raw logs do not retain a structured historical toolchain/argv receipt, so no full Mac run at f6bceb53d is claimed. Pubspec/lockfile and native macOS/Linux directories were unchanged across that last change. The exact final Windows matrix supplies separate final-source evidence. Temporary Desktop triggers were removed and focused workflow checks passed. The standalone Linux Desktop CI run then passed all four shards and ci/required at docs-only head c0d39cbcd, with the same runtime as f6bceb53d, using scope=desktop. Its four official receipts cover 640/640 files, 7,431 passes and 50 skips, with zero failed tests or errors. Downloaded archives matched their API digests; source, identity and raw-log hashes were verified. These are the same Linux shards without opening a PR. Only the PR-specific external record remains deferred by the owner's instruction.

Timing, limits and safety

The live observations span 18–20 October 2026. Measured intervals above are per-operation only. Implementation, validation, CI waiting, VM waiting, and document-writing durations were not captured as a reliable common ledger, so no total is claimed. No merge or production publication occurred. Bootstrap, link launch and cleanup are timed separately from the completed CI runs.

On ARM64, only the dedicated standard user's app, daemon, agents, terminals and scratch folders were changed. The separate x64 checks used only their owned app, terminal and scratch. C:\work\e2e, other users' agents and terminals, and the Mac's Harness home were not accessed. The later x64 rollback test temporarily changed only its fake-CDN hosts routes and name-constrained trust; hosts were restored byte-for-byte, and the CA, watchdog and reverse forward were removed. Its private CA keys were deleted and the loopback server stopped. No ARM hosts or trust settings changed. QR codes, account details and private screenshots stay outside tracked evidence. The ARM user's earlier Settings uninstalls finished cleanly. After the final clean bootstrap, the owned app quit and normal 1.0.4 uninstall completed; final account/profile teardown completed successfully. The x64 VM retains installed Desktop 1.0.4, its account/machine and shared CLI/core/updater/PTY. All nine run-owned x64 agent rows are stopped with no panes; the unrelated OpenCode harness retains its active pane. Ninety-eight selected test files/folders were removed, followed by the cleanup helpers; final scratch count was zero. Preserved shared process PIDs and creation times matched before and after, including the unrelated terminal window.

The old run-owned ARM backend machine was identified through supported CLI metadata with a process-local override to its marked backup, without reading credentials. The supported machine list from the x64 test PC uniquely identified that remote record with zero agents and proved it was not the x64 test PC's current machine. Its exact-ID deletion returned HTTP 403 / NEEDS_AUTONOMOUS_SIGN_IN requiring a Google or Apple sign-in. No successful deletion is claimed. Reauthenticating the x64 test PC would disturb the preserved session, so it was left unchanged. Its backend cleanup restriction is retained separately from the completed local Windows-account removal.

Final ARM local cleanup completed with exit 0 in 400.628 seconds. The initial 300-second wrapper had timed out and the original process later exited; its retained mounted test-user hives prevented ordinary profile readiness. After successful strict process ownership and recorded-task checks, the approved C5.1 exact-test-SID recovery was independently reviewed, parsed and compiled, then executed once. Both exact owned hives disappeared and the non-special profile explicitly reported unloaded. Native privilege state was restored; no admin/shared hive, process, service or Winlogon setting was changed.

Ordinary cleanup then removed the owned profile, account, marked backup, helpers and admin scratch. The complete preflight comparison confirmed preexisting user, process, task and listener identities were preserved. A separate postcheck verifies account/profile/directory/hives/scratch absence and zero owned processes. The final local cleanup is complete; the failed backend record deletion remains an explicit external authentication limit.

Final review found a separate CLI at sign-in task left after the normal cleanup: the helper recognizes the main daemon task name, while this newer logon task was outside its recorded task list. Its full principal SID, logon trigger user, exact owned launcher path and arguments matched the removed test profile; it was absent from the run-creation baseline. A guarded follow-up removed that task. The final inventory also found three OneDrive and three Windows SoftLanding tasks created for this new SID. All six were absent from the creation baseline and were removed by exact task path/name and principal, preserving the 247 other task identities. A fresh successful Scheduler inventory confirmed zero tasks for the owned SID/user. These manual supplements and the helper's task-coverage limit are retained; the earlier preflight containing the orphan tasks is not used as proof that all owned tasks were absent.