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.
Windows CLI: end-to-end results
Date: 2026-10-18. Branch feat/windows-native-cli (run from 235d90535; fixes e3c28f981, 6617af96c and
dbeca9ff8 made during the run). Build 0.4.18-dev.win1 from staging
(https://cdn.autonomous.ai/harness/staging/cli/install.ps1, published 2026-10-18; CLI built at 808483702, terminal
host from Actions run 39082479868). Because a
-dev build never updates itself (selfUpdate.ts isLocalDevBuild), scenarios 3 and 4 ran on a non-dev N,
0.4.19-e2e.1, installed with the same irm line from the fake HTTPS CDN (B5.13), with updates 0.4.20-e2e.1,
0.4.21-e2e.1 (canary-failing) and 0.4.22-e2e.1 (failing on probation) served from the Mac. This is Task B5.14 of the
plan; the spec's testing section is
§6 of the design.
How it was driven. The UTM VM: Windows 11 Home ARM64, build 10.0.26100.9711, 8 logical CPUs, computer name
WIN-TESTVM. Two new standard users, hxe2e and hxe2e2 (in Users, not in Administrators), were reached
over SSH by key. The PTY driver (cli/scripts/windows-pty-e2e, through run.cmd) used hxe2e's loopback on port
18474. The Mac's update server (serve-updates.sh, 192.168.65.1:8799) served the updates, and the fake HTTPS CDN
(fake-cdn.sh + vm-cdn.ps1) was enabled for scenarios 3 and 4 only. The Mac's CLI (signed in to the same account)
read the VM as a remote machine for scenario 5; the desktop app was not driven (no computer use in this session; see
Known limits). Checks that need WMI as a standard user (check-install.ps1, the uninstall's last step) ran in a
one-shot scheduled task with the user's password (a batch logon), as B5.12 did: an SSH key logon is a network logon,
which WMI refuses. C:\work\e2e's daemon was not running before or after the run (nothing listened on 1847x).
Evidence. cli/scripts/windows-cli-e2e/evidence/, one log per step:
preflight, users, s1-*, s2-*, s3-*, s4-*, s5-*, s6-*, rf4-reinstall, s7-*, mac-*, linux-*,
cleanup. No password, token or credential is in them; the Mac's paths are written ~ and $SCR.
Time (Mac clock, +07). Preflight and users 07:44–07:47. Validation on the VM 07:47–09:13, of which about 21 minutes (08:00–08:21) were the VM hung in UTM's firmware after a warm reboot (waiting), and about 15 minutes were the reboots and the 5-minute trigger windows (waiting). Implementation of the three fixes, inside that window, about 15 minutes. macOS and Linux 09:13–09:25. Clean-up 09:25–09:27. Writing this document and the report about 25 minutes. No merge or publication.
Summary
| # | Scenario | Result | Time |
|---|---|---|---|
| 1 | Install from staging with irm, as a new standard user |
Pass (after fix e3c28f981 to the check) |
install 82 s; 07:47–07:56 |
| 2 | Reboot: back at logon; a dead master back within the 5-minute trigger; with nobody signed in | Pass at logon (84 s from boot) and by the trigger (1 min 38 s); stop honoured past a 5-minute mark; not up with nobody signed in (G1) | 07:56–08:37 (21 min of it the VM hung) |
| 3 | Update to N+1 with a terminal and Claude Code running: both survive | Pass (Claude Code not signed in, carries) | update 23 s after the offer; 08:40–08:45 |
| 4 | Canary-failing N+2 never used; N+3 failing on probation rolled back | Pass | rollback 32 s after the offer; 08:45–08:50 |
| 5 | Claude Code, Codex, OpenCode and Kilo hooks report to the Mac desktop (remote) | Partial: Kilo bound and seen from the Mac; Claude Code and Codex not verified (not signed in); OpenCode refused on Arm; desktop tiles not driven | 08:50–08:53 |
| 6 | Another local user: no file readable, no trusted call accepted | Files and named pipes: Pass (after fix 6617af96c to the check). Loopback TCP: see the note under Scenario 6, out of scope for this run |
08:53–08:57 |
| 7 | Uninstall: no task, no PATH entry, no files, no process, no hook; --keep-data keeps the data; a half install repaired, then uninstalled | Pass, with G2 (engine harnesses refused to stop over SSH) | 08:59–09:13 |
| RF4 | Reinstall over a live install: one restart onto the new build, agents keep running; the same version restarts nothing | Pass | run 1 24 s; 08:57–08:59 |
| — | Uninstall on macOS (launchd, or no service when a Harness job is loaded) and Linux | Pass after fix dbeca9ff8 (macOS: no service; Linux: container with an init). The launchd half ran once and is unsafe in a throwaway home (G7) |
09:13–09:25 |
| — | Windows workflow run (windows.yml, install job) |
run 39086071140 at 7bdeef21a: both install jobs success (the run before, 39083518767, failed on the check bug e3c28f981 fixes) |
— |
Setup recipe
- Preflight: the staging
install.ps1names…/harness/staging/; nothing listens on 18473/18475; the VM's build. - Users:
New-LocalUser -PasswordNeverExpires(as admin;net user … /addasks a Y/N question for a password over 14 characters and fails over SSH),Usersgroup;first-login.shfor each (profile and key). hxe2e:setx PORT 18474,setx ADAPTER_UPDATE_URL http://192.168.65.1:8799/cli/metadata.json; hxe2e2:PORT=18476,ADAPTER_UPDATE_DISABLE=true. Passwords stayed in the session scratchpad. - Batch logon for WMI:
SeBatchLogonRightfor both users (LsaAddAccountRights, removed at the end), and a one-shotschtasks /Create /RU <user> /RP …running the script with its output to a file. - Autologon (scenario 2 only): Winlogon values saved, set for hxe2e, restored exactly by an EXIT trap.
- Fake CDN (scenarios 3–4 only):
fake-cdn.sh make-ca(in the scratchpad,FAKE_CDN_DIR),stage 0.4.19-e2e.1with the staging publish's own pty payload (PTY_VERSION=0.4.18, the run 39082479868 artifacts; the arm64 zip is byte-identical to staging's),serve;vm-cdn.ps1 -Enableas admin. The stagedinstall.ps1is byte-identical to the published one (sha256042b2a3f…).NODE_EXTRA_CA_CERTSwas not needed: the updates came over plain HTTP. - Updates:
serve-updates.sh build 0.4.20-e2e.1,break-canary 0.4.21-e2e.1,break-core 0.4.22-e2e.1; eachnotify.mjsgot a// release <version>line so its hash names its build. - Engines for hxe2e:
%APPDATA%\npmand%USERPROFILE%\.local\binadded raw to the user PATH (REG_EXPAND_SZ), Claude Code 2.1.302 (irm https://claude.ai/install.ps1 | iex), Codex 0.163.0, OpenCode 1.19.3, Kilo 7.8.12 (npm). No engine was signed in. - Sign-in for scenario 5:
approve-qr.mts [email protected](prod), machine7e92c45a…, deleted at the end. - Warm reboots hang this VM in UTM's firmware ("Start boot option"); every reboot after the first was
shutdown /sandutmctl start Windows. - macOS uninstall. Do not run the launchd/service half (
harness service install) on a developer's real Mac account, even withHOMEpointed at a throwaway folder: launchd gives the job the account's realHOME, and that daemon works on the person's real~/.harnessand engine settings (G7, what run 1 did). Safe recipe: in the throwawayHOME, runharness start(no service), as run 2 did. To test launchd itself, use a separate macOS user or a macOS VM, or load a hand-made plist whoseEnvironmentVariablessetHOMEto the throwaway home.
Scenario results
Scenario 1: install from staging
irm https://cdn.autonomous.ai/harness/staging/cli/install.ps1 | iex as hxe2e over SSH, exit 0 in 82.3 s
(s1-install.log): Node v22.23.2 (13 s), Harness
0.4.18-dev.win1, terminal host 0.4.18, PATH, then about 62 s in harness service install waiting for a task that
cannot start over SSH, then Harness starts when you sign in to Windows; it is not running now: run 'harness start'.
and the guide. The line before it still says ✗ Task Scheduler has harnessd, but it did not come up… (G4).
harness start took 15.2 s (Task Scheduler did not start harnessd … — starting it directly). A new SSH connection
read ● running, same pid 7500: the daemon survived the session.
check-install.ps1 -Version 0.4.18-dev.win1
(s1-check.log):
- over SSH: version, task and PATH
ok, thenFAIL could not list processes through WMI (Access denied ), as B5.12 recorded; - in a batch logon, twice:
FAIL no harness-pty.exe runs after creating a terminalwith harness-pty 6932 running. A bug in the check, fixed ine3c28f981(below); - after the fix:
ok version,ok task Harness harnessd (hxe2e),ok PATH has c:\users\hxe2e\.harness\bin,ok terminal 297854ec… on harness-pty pid 6932 (4 shell(s) under it), exit 0.
Process tree (procs.ps1): master node.exe 7500 whose parent 9744 is gone (harness-launch.exe exited), so not under
sshd; core 9652 and updater 1004 under it; harness-pty.exe 6932 under the core, with OpenConsole and a PowerShell per
terminal. Defender: no 1116/1117 event in the last 30; no SmartScreen event. Layout
(s1-layout.log): bin\harness.exe,
bin\harness-launch.exe, cli\0.4.18-dev.win1\{cli.js,notify.mjs,package.json}, cli\current = 0.4.18-dev.win1,
cli\notify.mjs, cli\data\, runtime\node-v22.23.2-win32-arm64, runtime\pty\0.4.18.
Scenario 2: reboot, the 5-minute trigger, the stop marker, nobody signed in
The task is \Harness harnessd (hxe2e), Logon Mode Interactive only (InteractiveToken), logon trigger plus a
5-minute repetition, action harness-launch.exe --root … -- start --if-down. Winlogon before: AutoAdminLogon '0',
DefaultUserName 'admin', DefaultDomainName '.', no DefaultPassword; restored exactly (restored exactly: True,
s2-winlogon.log).
- Autologon reboot (
s2-reboot.log). The first warm reboot hung in UTM's firmware for 21 minutes; the VM was power-cycled. Boot 18:21:47 (VM clock), task ran 18:22:14 (result 0), master 18:22:35, core ready 18:23:03,● runningread at 18:23:11: 84 s from boot. Master 7356, parent gone. - The 5-minute trigger (
s2-trigger.log). Admintaskkill /PID 7356 /Fat 18:23:26; status○ stopped; the task ran at 18:25:01, core ready 18:25:04, new master 7892: 1 min 38 s. The terminal host 9672 lived through it. - The stop marker (
s2-stopped.log).harness stopat 18:25:14 wrote{"boot":1792286507337}. The task ran at 18:30:01 (result 0) and started nothing: at 18:31:56 still○ stopped. Reboot with the marker in place: boot 18:32:23, core ready 18:33:23 (60 s): a marker from an earlier boot is not honoured. B3.21's deferred steps 5 and 6 are covered. - Nobody signed in (
s2-nologon.log). Autologon restored, then a reboot: boot 18:33:57; three minutes later no hxe2e process, the task's last run still 18:32:51 (before the shutdown), status○ stopped. WithInteractive onlythe task never runs without a session: the remote machine is not back after a reboot with nobody signed in (G1).
Scenario 3: update to N+1 while harnesses run
First the brief's flow on the staging build
(s3-dev-check.log): with 0.4.20-e2e.1 offered,
0.4.18-dev.win1 fetched the Mac's metadata.json at 08:40:45 and 08:41:45 and staged nothing; its log says only
self-update on · v0.4.18-dev.win1 · every 60s (G5). So the carries' path was used: the fake CDN on, and the same
irm line installed 0.4.19-e2e.1 over the live daemon in 19.6 s, printing Moving the running Harness from 0.4.18-dev.win1 to 0.4.19-e2e.1 (your agents keep running)...
(s3-install-n.log,
s3-cdn.log).
Before (s3-before.log): terminal c66db6da… answered
BEFORE_42; Claude Code 47324568… showed its first-run theme picker (not signed in). Master 2312, core and updater
3344/5656, harness-pty.exe 5028 started 18:40:04.
0.4.20-e2e.1 offered at 08:43:30 (s3-offer.log);
harness version said it 23 s later. The log: staged 0.4.20-e2e.1 18:43:47, handed over to the new master (pid 2952) 18:43:49, the update stayed up — keeping it 18:44:20. After
(s3-after.log): harness version and harness status
0.4.20-e2e.1, cli\current = 0.4.20-e2e.1, cli\ holds 0.4.19-e2e.1 and 0.4.20-e2e.1; the stable
cli\notify.mjs hash a177f39f… is N+1's; the same two ids and panes (%3001, %3002); AFTER_42; Claude's screen
unchanged; master 2952, core/updater 3276/4368; harness-pty.exe 5028, start 18:40:04, unchanged. **Pass.** The
terminal's row read "status":"offline" after the update while its shell answered (G3).
Scenario 4: a failing build is rolled back
- Canary (
s4-canary.log):0.4.21-e2e.1offered;18:45:45 newer build availableand18:45:47 0.4.21-e2e.1 failed its canary (exit 3) — not trying it again for 120 s. 180 s latercli\current=0.4.20-e2e.1, no0.4.21-e2e.1folder, master 2952 unchanged. - Probation (
s4-probation.log):0.4.22-e2e.1offered; staged 18:48:47,took over from the master before it (pid 2952)18:48:51,the updated core failed (code 1) — rolled back to the previous bundle; restarting18:48:52. Seen 32 s after the offer. Thenharness status0.4.20-e2e.1(master 6704),cli\current=0.4.20-e2e.1,update-rejected.json=["0.4.22-e2e.1"],cli\=0.4.19-e2e.1,0.4.20-e2e.1(the two newest good builds; the rolled-back folder gone), both harnesses listed,ROLLBACK_42,harness-pty.exe5028 unchanged. **Pass.** After the rollback the updater loggedself-update on · v0.4.22-e2e.1(G5).
The fake CDN was then disabled: hosts sha256 back to 2d6bdfb3… (byte-identical) and the CA gone from Root.
Scenario 5: engine hooks, and the Mac reading the VM
Sign-in: approve-qr.mts → result success, machine registered: 7e92c45a…
(s5-login.log). After harness stop / harness start:
● started · connecting to the backend, then backend connected. The Mac's harness machines --json lists
WIN-TESTVM running; the Mac's harness ps --machine 7e92c45a… --all --json
(s5-remote.log) lists the VM's harnesses with Kilo
"state": "idle", "asks": ["Reply with only the number: what is 6 times 7?"], "about": "42".
| Engine | Result | Hook runner (B4), from s5-hookfiles.log |
|---|---|---|
| Claude Code 2.1.302 | not verified: needs Claude Code sign-in (it stays on its first-run screen) | command = the managed node.exe, args = notify.mjs … (no shell) |
| Codex 0.163.0 | not verified: needs Codex sign-in (its sign-in screen) | a PowerShell line, & '…\node.exe' '…\notify.mjs' --port 18474 … --engine codex |
| OpenCode 1.19.3 | refused: ENGINE_LAUNCH_REFUSED OpenCode does not run on Windows on Arm yet: its Arm build stops before its screen opens. Use another engine on this computer. |
plugins C:\Users\hxe2e\.config\opencode\plugin\launcher-register.js, …\plugins\launcher-register\tui.js |
| Kilo 7.8.12 | bound: answered 42 (free model), row sessionId ses_eb34…, log [hooks] ses_eb34 session-start · engine=kilo · isNew=true |
plugin C:\Users\hxe2e\.config\kilo\plugin\launcher-register.js |
Command Code's hooks went to C:\Users\hxe2e\.commandcode\settings.json (COMMANDCODE_HOME unset), as B4.13/B4.15
expect. Copilot, Pi and Amp files were written too. The Claude Code resume check (A's F6) was not run: there is no
Claude session without a sign-in. The desktop tiles were not driven (Known limits).
Scenario 6: a second user
s6-second-user.log: hxe2e2 installed
0.4.18-dev.win1 and started on 18476. check-install first passed while naming hxe2e's host (pid 5028): it
matched every user's harness-pty.exe by name. Fixed in 6617af96c; again: ok terminal 2541c62a… on harness-pty pid 7076, exit 0. schtasks lists \Harness harnessd (hxe2e) and \Harness harnessd (hxe2e2).
Trust (s6-other-user.log), as hxe2e2 against hxe2e:
daemon-18474.json,daemon-token,hook-credential: eachrefused EPERM;- pipes hxe2e2 could list:
harness-41b22efa…(hxe2e's daemon),harness-pty-bddfac6e…,harness-claim-6ae5b773…: eacherror EPERM(cannot open), with and without a forged proof; hxe2e2's own pipes answered401 PIPE_AUTH_REQUIREDor reset; - hxe2e's own
harness statusthrough its pipe:● running · backend connected.
hxe2e2's uninstall over SSH: every step, then ✗ Harness is only partly uninstalled: · Windows did not start the step that removes Harness's files once this command has ended. (Access denied) Windows refused to start it, as it does for a standard user signed in over SSH; run harness uninstall at the desktop, or delete C:\Users\hxe2e2\.harness yourself.
Terminal hosts by pid and start: before 5028 18:40:04.889 hxe2e, 7076 18:54:38.625 hxe2e2; after, only
5028 18:40:04.889 hxe2e. Run again from a batch logon: Harness is uninstalled., exit 0 in 2.1 s;
check-uninstall: clean. hxe2e's daemon (pid 6100) still ran and its task was still registered.
Note: a security observation from S6 was reported to the maintainers; details are not kept here.
Reinstall over a live install
rf4-reinstall.log. With hxe2e on 0.4.20-e2e.1
(signed in) and the update server stopped, a COUNTER terminal b5c7e44b… printed COUNT_n each second.
- Run 1, staging's
0.4.18-dev.win1over the live0.4.20-e2e.1, 24.4 s:Moving the running Harness from 0.4.20-e2e.1 to 0.4.18-dev.win1 (your agents keep running)...,machine stopped (pid 6100), started directly,harnessd is registered with Task Scheduler already … nothing to change. After:harness versionand status0.4.18-dev.win1; master 6100 → 7156, core/updater 7128/5356 → 2840/2848 (once);harness-pty.exe5028, started 18:40:04, unchanged;cli\holds0.4.18-dev.win1and0.4.20-e2e.1(N+1 kept as the build to go back to,0.4.19-e2e.1retired); COUNTERCOUNT_43/COUNT_44, thenCOUNT_51/COUNT_52five seconds later; same id;check-installallok(one PATH entry). No "being used by another process". - Run 2, the same version: no
Movingline, master 7156 before and after.
Pass.
Scenario 7: uninstall
Round 1 (s7-uninstall.log,
s7-check.log). Before: user PATH [ExpandString] %USERPROFILE%\AppData\Local\Microsoft\WindowsApps;C:\Users\hxe2e\.harness\bin;%APPDATA%\npm;%USERPROFILE%\.local\bin.
A Claude harness made 3 s before (not bound, B5.10), and cli\data\harnessd-task.xml deleted first (the task still
registered). Over SSH, harness uninstall --yes: ✓ stopped 4 harnesses, ✓ removed the Harness daemon from Task Scheduler, ✓ stopped the Harness daemon (pid 7156), ✓ ended 3 terminals in the Harness terminal host, the hooks of
Claude Code, Codex (and its config.toml record), Copilot, Command Code, the Kilo, Pi, Amp and both OpenCode plugin
files, ✓ removed ~\.harness\bin from your PATH; then exit 1 with three harnesses that did not stop (IDENTITY_UNAVAILABLE: Conversation identity is held: the native ps probe did not complete.) (both Claude harnesses,
including the unbound one, and Codex; G2) and the WMI sentence above. No hxe2e process was left (the host's end took
them). From a batch logon: ~\.harness goes once this command has exited (…harness-uninstall-1464-….log),
Harness is uninstalled., exit 0 in 1.7 s.
Checks: check-uninstall -HookFiles <the six plugin and hook files>: clean: nothing of Harness is left … and 6 hook file(s) named by -HookFiles. procs.ps1: nothing. PATH after: [ExpandString] %USERPROFILE%\AppData\Local\Microsoft\WindowsApps;%APPDATA%\npm;%USERPROFILE%\.local\bin: only Harness's entry gone,
kind and %VAR% kept (B5.3). Claude, Codex and Command Code settings hold "hooks": {}; rules\harness.rules
absent. No ~\.harness, no harness.exe~RF*.TMP, no harness-uninstall-* in %TEMP% (B5.9). The task, its file
deleted first, was removed (B5.9). The test variables were then deleted.
Round 2, --keep-data (s7-keepdata.log): install,
harness start (pid 5240), a reboot (nobody signed in, so adapter.pid names 5240, no process), then
harness uninstall --yes --keep-data from a batch logon: exit 0 in 21.2 s, Harness is uninstalled. Your sign-in and history are kept for the next install., and no problem about the stale pid (B5.9). Kept: cli\data (registry, logs,
daemon-18474.json …), computer-id, agent-messages, connections, dsh, hook-routes. Gone: bin, runtime,
cli\<version>, cli\current. auth did not exist: round 1 had removed it and this install was not signed in, so
the sign-in's survival was not shown here. With no harness.exe left, the full uninstall of that data is round 3's.
Round 3, a half install (s7-half.log): install over the
kept data; del cli\current; harness version → Harness is not installed completely; to install it again, open PowerShell and run: irm https://cdn.autonomous.ai/harness/cli/install.ps1 | iex, exit 3 (B1.8's code; the brief's
pass line says 1). The installer again: cli\current = 0.4.18-dev.win1. From a batch logon harness uninstall --yes: exit 0 in 12 s. check-uninstall: clean (its first run in the log is the run's own mistake, a list mangled
by zsh's echo). Pass.
Uninstall on macOS and Linux
macOS (mac-launchd.log). No Harness job was loaded,
so run 1 took the launchd half (mac-run.log). harness service install in the throwaway home printed ✗ launchd has harnessd, but it did not come up, and the launchd-started
daemon ran with the person's real home: launchd gives the job the account's HOME, which the plist does not set,
so it used the real ~/.harness/cli/data and rewrote the real engine hook files (G7; repaired, see Clean-up). The
uninstall then removed the launchd job (launchctl print afterwards: not found), both launchers and the ~/.zshrc
line, but ended could not remove ~/.harness/runtime, exit 1: the managed grid's bin folder is read-only (0555).
Fixed in dbeca9ff8. Run 2 (mac-run-2.log), with no
service (harness start inside the throwaway home) and the fix: ✓ stopped 1 harness, ✓ stopped the Harness daemon (pid 75875), every hook file, ~/.local/bin/harness and hn, the ~/.zshrc line, ✓ removed ~/.harness,
Harness is uninstalled., exit 0 in 1 s. No ~/.harness, ~/.local/bin empty, no installer line, no LaunchAgents,
no process. The real daemon was not stopped or restarted by either run. Do not repeat run 1 on a developer's
Mac: use the no-service recipe (Setup recipe, step 10), a separate macOS user or a VM.
Linux (Ubuntu 24.04 container, user hx). install.sh needs tmux and the user has no sudo, so tmux was installed
as root first (linux-install.log). Run 1, a container
with no init (linux-run-1-noinit.log): ✗ The Harness daemon (pid 2857) did not stop. None of Harness's files were deleted., exit 1; ps shows 2857 as a zombie of
PID 1 (G6). Run 2 with docker run --init
(linux-run.log): ✓ stopped 1 harness, ✓ stopped the Harness daemon (pid 2818), every hook file, both launchers, the ~/.bashrc and ~/.profile lines, ✓ removed ~/.harness (the managed grid included), Harness is uninstalled., exit 0 in 1 s; nothing left, no process.
Fixes made during the run
Each is its own commit with a test that fails without it.
e3c28f981fix(windows): check-install counts a lone terminal host. With oneharness-pty.exeon the machine,$hosts = Processes …held one CimInstance (a function's array is unrolled), and Windows PowerShell 5.1 gives a CimInstance no.Count: the check waited 30 s and failed with the host running. It is also why run 39083518767 failed. Test:windowsCiChecks.spec.ts(everyProcessesresult kept an array; on Windows, one process counts as one). Scenario 1 ran again: pass. Run 39086071140, with it: both install jobs pass.6617af96cfix(windows): check-install counts only this install's terminal hosts. WMI lists every user'sharness-pty.exe; a second user's check named the first user's host and counted its shell, so it would pass with its own host missing. Hosts now must run from$root\. Test:windowsCiChecks.spec.ts. Scenario 6's check ran again: pass, naming 7076.dbeca9ff8fix(cli): harness uninstall removes a read-only folder, as the managed grid installs one.removeNowmakes the target's folders writable when the first removal fails (links never followed) and removes it again. Test:removal.spec.ts(a 0555runtime/grid-…/binwith a 0555 file and a link to a 0555 folder outside: the tree goes, the outside folder keeps its mode). macOS and Linux ran again: pass.
Findings not fixed
After the final review, G4, G5, G6 and the closing-guide half of G8 are fixed; G1, G2, G3 and G7 are in the
spec's Known limits. That review also split the task in two (the five-minute task, and a logon task that forgets
a harness stop) and set the tasks to normal priority. This run predates those changes and did not exercise them.
- G1. With nobody signed in, the daemon does not come back after a reboot. The task's logon type is
InteractiveToken (
Logon Mode: Interactive only,harnessd/platform.tstask XML), so Task Scheduler runs it only in a signed-in session; the spec wants a remote machine back after a reboot with nobody signed in. (Running without a session needs a stored password or S4U, which B3 chose not to.) - G2. Over SSH, the uninstall cannot stop engine harnesses. For Claude Code and Codex harnesses (bound or not)
agent_deleteansweredIDENTITY_UNAVAILABLE … the native ps probe did not completein a daemon started over SSH, where WMI refuses the user (lib/uninstall/agents.ts→ the core's identity hold). The terminal host's end then closed them, so nothing was left, but the summary said theymay still be runningand exited 1. From a batch logon (as at the desktop) it was not seen. - G3. Terminal rows read
offlineafter a daemon restart (update, reinstall) while their shell keeps answering (lib/agentFrame.ts,status: s.active ? 'active' : 'offline'). Claude, Codex and Kilo rows stayedactive. - G4. install.ps1 prints
✗ Task Scheduler has harnessd, but it did not come up…on every install over SSH, then the corrected sentence, and waits about 62 s for the task first (harness service install). B5.12 fixed the exit code; the ✗ line still prints. - G5. The updater's log misleads. A
-devinstall logsself-update on · v… · every 60sand polls every minute, but never takes an update (isLocalDevBuild), and says nothing about why. After a rollback it logsself-update on · v<the rolled-back version>. - G6. A zombie daemon counts as alive. In a container with no init, the stopped master stays a zombie of PID 1;
isAlive(kill -0) says alive, andharness uninstallremoves nothing (The Harness daemon … did not stop). The help text names containers as a supported place to run. - G7.
harness service installunder another HOME starts a daemon on the real home (macOS). The plist carriesCARRIED_ENVbut notHOME, and launchd gives the job the account's home. A person's HOME is the account's, so this matters for tests (the brief's Step 10 launchd half), not for users. That half must not run on a developer's Mac; Setup recipe step 10 gives the safe way. - G8. Minor copy. The installer's closing guide says
harness login # 1. sign inon a machine already signed in;net user … /addwith a password over 14 characters fails over SSH withNo valid response was provided.
Known limits
- The Mac desktop app was not driven: this session had no computer use. The desktop tiles (working, then idle with
42) and their screenshots are not recorded; the Mac's CLI read the same remote rows instead. - No engine was signed in on hxe2e (carries): Claude Code and Codex bindings, the Claude resume check, and scenario 3's "Claude answering" are not verified. OpenCode is refused on Arm; its binding is the x64 CI job's.
- Scenarios 3 and 4 used a non-dev N from the fake CDN; the published staging build itself never updates.
--keep-datawas not shown keepingauth(none existed in that round).- The VM's warm reboot hangs in UTM's firmware; reboots were a shutdown and a power-on.
- WMI-dependent checks over SSH need a batch logon (B5.12).
Clean-up done
- VM (
cleanup.log):SeBatchLogonRightremoved from both users; both users deleted; hxe2e2's profile removed; hxe2e's after a forced unload of its two hives (as in B3.21): noC:\Users\hxe2e*, no ProfileList key. NoHarness harnessd (…)task and noHxE2ERun-*task left. User variablesPORT,ADAPTER_UPDATE_URL,ADAPTER_UPDATE_DISABLEdeleted before the users went;NODE_EXTRA_CA_CERTSwas never set. hosts sha2562d6bdfb341be3a6234b24742377f93aa7c7cfb0d9fd64efa9282c87852e57085, identical to before; no fake CDN CA in Root;C:\ProgramData\harness-fake-cdngone. Winlogon restored exactly, noDefaultPassword. The admin helper folders (C:\Users\admin\e2e,e2e-b514) removed.C:\work\e2euntouched (its daemon not running before or after; nothing on 1847x). - Account: machine
7e92c45a…(the only one this run registered) deleted and gone fromharness machines. An olderWIN-TESTVMrow (5d96588a…, 2026-10-15) is not this run's and was left. - Mac: the update server and the fake CDN stopped (8799 and 443 free); both throwaway homes removed; the scratchpad's
password file, CA and keys, staged CDN and update builds removed. The macOS run 1 repair (G7): in the real
~/.harness/cli/data,adapter.pidwritten back with the running master's pid (19304) and the straydaemon-18491.jsonremoved; in 16 real engine files (Claude, Codex, Copilot, Command Code, Gemini, Cursor, Grok, Hermes, Devin, the OpenCode, Kilo, Pi and Amp plugins, and~/.harness/hook-routes) the throwawaynotify.mjspath and port 18491 replaced with the real ones, and Codex'sconfig.tomlreview records for Harness's hooks recomputed.harness statusreads● running(pid 19304). Originals kept in the session scratchpad. - Linux: container
hx-uninstallremoved.
B3 lifecycle
Date: 2026-10-16. Branch feat/windows-native-cli at ff0705990, with the two real specs of Task B3.21 on top.
Machine: the UTM VM (Windows 11 ARM64, build 26100), ssh [email protected]. The times in the log excerpts are
the VM's clock.
Build under test.
- The pty binaries were built on the VM from this branch:
cargo +stable build --locked --releasewithHARNESS_PTY_VERSION=0.4.91. This includes theharness-launch.exethat inherits no handles (5a9c0c56c). - Four CLI bundles were built from this tree with
ADAPTER_VERSION: N0.4.91-b3, N+10.4.92-b3, N+20.4.93-b3and N+30.4.94-b3. Each patch number is distinct becausesemverGtcompares only major.minor.patch. None is a-devbuild, which an automatic update would leave alone. - N+2 is the same build with
if(process.argv[2]==="__run")process.exit(3);put first incli.jsand in its lean entries (withFault,e2e/harness/release.ts). It passes the updater's canary (cli.js version) and its core exits at start. - Each
notify.mjsends with// release <version>, so the copy in place shows which build it came from. - The install manifest came from
cli/scripts/windows-install-manifest.mjs, using N, the payload above and Node v22.23.2's official zip (sha256fec025a6…5ef3).
Who ran it. A fresh local user, hb3, ran everything. It was made as admin with New-LocalUser (the parameter
is -PasswordNeverExpires) and added to Users. Its profile came from userenv's CreateProfile, with no sign-in.
The key in C:\ProgramData\ssh\administrators_authorized_keys was copied to C:\Users\hb3\.ssh\authorized_keys
(the admin profile has no .ssh folder). hb3 set PORT=18621 with setx, so its daemon never takes 18473,
the port of the daemon in C:\work\e2e. The agent was driven over ws://127.0.0.1:18621/api/local-ws with a
throwaway script built on cli/scripts/windows-pty-e2e/driver.mts. Nothing was signed in. C:\work\e2e and the
admin profile were not touched.
Install: install.ps1 -Manifest C:\work\b3\manifest\install-manifest.json -NoService, run as hb3, took 14 s.
OK Harness 0.4.91-b3 -> C:\Users\hb3\.harness\cli\0.4.91-b3
OK terminal host 0.4.91 -> C:\Users\hb3\.harness\runtime\pty\0.4.91
OK harness 0.4.91-b3 installed.
Real specs
terminate.real.spec.ts and taskScheduler.real.spec.ts were run on the VM as admin, beside A's
launch.real.spec.ts: 4 passed (3 files). The first run found three things the specs had to say differently:
- The child in the TerminateProcess spec is detached. A plain Node child sits in libuv's kill-on-close job
object, so it ended with its parent. The first run failed at
expect(processExists(childPid)).toBe(true). TerminateProcess did not end that child; the job did. The core starts the pty host withdetached: true(lib/ptyHost/launch.ts), and the spec now does the same. - The task spec passes the user's SID and writes the file as UTF-16LE with a BOM, as
PlatformService.writedoes.taskSchedulerXmltakessidsince B3.12. - "Enabled" is the absence of
<Enabled>false</Enabled>.schtasks /Query /XMLleaves out every setting that has its default value, so<Enabled>true</Enabled>never appears.PlatformService.inspectreads it by the same rule. The registered XML also showed that schtasks adds<UserId>S-1-5-21-…-1001</UserId>to the principal, and shows the logon trigger's SID as the account name.
Steps
| # | Step | Result |
|---|---|---|
| 1 | harness start, launcher, survives the SSH session |
Pass |
| 2 | harness stop by message, agent survives |
Pass |
| 3 | The core alone ended, master restarts it, agent reattached | Pass (with Stop-Process, not taskkill: see F1) |
| 4 | harness service install, /Run refused over SSH, "starting it directly" |
Pass |
| 5 | Console sign-in, the task brings an ended master back | Deferred to B5.14 (needs a console sign-in and a reboot). The parts that need neither passed. |
| 6 | Stop honoured past the 5-minute mark; back at logon after a reboot | Deferred to B5.14 (needs a console sign-in and a reboot). The part that needs neither passed. |
| 7 | Self-update, handover, rollback, retirement | Pass |
| 8 | Clean up | Done |
1. harness start. It returned in 6.0 s over SSH (2.9 s on later starts). The master's parent is gone, because
the launcher started it and exited. The core and the updater are the master's children, and the pty host is the
core's child. The command ran in one SSH session and the processes were listed from the next one, after that
session had closed: all of them were still running.
1220 node.exe parent 7684 <gone> ~h\runtime\node-v22.23.2-win32-arm64\node.exe ~h\cli\0.4.91-b3\cli.js __harnessd
8656 node.exe parent 1220 node.exe ... ~h\cli\0.4.91-b3\cli.js __run
11368 node.exe parent 1220 node.exe ... ~h\cli\0.4.91-b3\cli.js __service updater
11984 harness-pty.exe parent 8656 node.exe ~h\runtime\pty\0.4.91\harness-pty.exe serve --home ~h\pty
2. harness stop. A PowerShell agent (terminal, pid 1888) answered B3_42. harness stop printed
machine stopped (pid 1220). in 1.4 s. After the stop, harness-pty (11984), its OpenConsole.exe and
powershell.exe 1888 were still running.
2026-10-16 03:35:58.795 [harnessd] the core asked to stop (harness stop) — stopping
2026-10-16 03:35:58.800 [cli] the harnessd master asked this core to stop — shutting down
2026-10-16 03:35:58.849 [harnessd] service updater stopped (code 0)
2026-10-16 03:35:58.886 [harnessd] core stopped (code 0)
3. The core ended alone. After harness start, harness status showed agents 1 available, and the agent
answered AFTER_STOP_1888. Run as hb3 over SSH, taskkill /PID <core> /F failed with ERROR: Access denied
(F1). The core was then ended with Stop-Process -Id <core> -Force, which is TerminateProcess on that one pid, as
taskkill /F without /T would be. The master restarted it, and the same agent answered AFTER_CRASH_1888.
2026-10-16 03:38:28.362 [harnessd] core exited (code 4294967295, crashed) — restarting in 500 ms
2026-10-16 03:38:29.024 [harnessd] core started (pid 10204) · restart 1
2026-10-16 03:38:29.848 [harnessd] core ready (pid 10204)
4. harness service install. It registered \Harness harnessd (hb3): Run As User: hb3,
Logon Mode: Interactive only, Scheduled Task State: Enabled, both triggers.
✓ Task Scheduler starts the Harness daemon when you sign in to Windows, and again within 5 minutes if it stops (pid 5496 runs now).
task Harness harnessd (hb3)
Then harness stop wrote harnessd-stopped.json ({"boot":1792115627718}). harness start took 12.8 s and
printed:
Task Scheduler did not start harnessd (it runs the task only in a signed-in Windows session) — starting it directly
The task's Last Run Time stayed 11/30/1999, so it never ran. The start removed the stop marker.
5 and 6, the parts with no console and no reboot.
The registered task, as
schtasks /Query /XMLshows it, has everything the steps rely on: logon trigger userWIN-TESTVM\hb3, which is hb3's SID in the file written;<Interval>PT5M</Interval>;IgnoreNew;InteractiveToken; and the actionC:\Users\hb3\.harness\bin\harness-launch.exe --root "C:\Users\hb3\.harness" -- start --if-down.The task's action was run by hand while the daemon was up. It exited 0 at once and the pid file did not change.
The master was ended with
Stop-Process. Its core and updater ended with it, through Node's job object. The pty host and the agent lived on, andharness statussaidstopped. The task's action was run by hand again, and the daemon was back 2 s later, with the agent reattached (AFTER_IFDOWN_1888):2026-10-16 03:50:11.838 [harnessd] core started (pid 9148) 2026-10-16 03:50:13.254 [harnessd] core ready (pid 9148)After
harness stop, the task's action run by hand started nothing: no new log line, and statusstopped.harnessd-stopped.jsonwas honoured.
7. Self-update. A Node server on 127.0.0.1:8765 served C:\work\b3\updates, holding the builds and
metadata.json ({ "cli": … } with absolute http URLs). hb3 ran setx ADAPTER_UPDATE_URL http://127.0.0.1:8765/metadata.json, then harness stop and harness start. The updater's next check staged
N+1 and handed over:
2026-10-16 03:41:45.057 [update] newer build available: 0.4.91-b3 → 0.4.92-b3
2026-10-16 03:41:49.939 [harnessd] the updater staged 0.4.92-b3 — asking the core to hand over
2026-10-16 03:41:51.785 [harnessd] the bundle on disk is not this master's code — re-executing on it
2026-10-16 03:41:53.520 [harnessd] took over from the master before it (pid 5948)
2026-10-16 03:41:53.758 [harnessd] handed over to the new master (pid 8908) — leaving
2026-10-16 03:41:55.562 [harnessd] core ready (pid 7036)
2026-10-16 03:42:25.654 [harnessd] the update stayed up — keeping it
After the handover:
cli\0.4.92-b3\was there andcli\currentread0.4.92-b3.- The sha256 of
cli\notify.mjswas N+1's (33724a3d…). adapter.pidheld8908, the new master, which runs fromcli\0.4.92-b3\cli.js.- The agent answered
AFTER_UPDATE_1888.
N+2's core exits at start. The master rolled it back and went back with it:
2026-10-16 03:44:04.753 [harnessd] took over from the master before it (pid 8908)
2026-10-16 03:44:05.918 [harnessd] the updated core failed (code 3) — rolled back to the previous bundle; restarting
2026-10-16 03:44:09.106 [harnessd] took over from the master before it (pid 3412)
2026-10-16 03:44:10.618 [harnessd] core ready (pid 6964)
2026-10-16 03:44:45.086 [update] 0.4.93-b3 was rolled back on this machine — waiting for a newer build (`harness update` installs it anyway)
After the rollback:
cli\currentread0.4.92-b3andupdate-rejected.jsonread["0.4.93-b3"].cli\0.4.93-b3\had been removed.adapter.pidheld9740, the master on 0.4.92-b3.- The agent answered
AFTER_ROLLBACK_1888.
Retirement. After a master restart, cli\0.4.91-b3\ stayed. That is by design: the retirement keeps the
build in use and the newest other build that was not rolled back, which is the build a rollback would go back to
(buildsToKeep, B3.17). So N was removed only once N+3 (0.4.94-b3) had been staged and kept. Nothing ran from
N by then, and its folder was gone when the updater of N+3's master started: cli\ held 0.4.92-b3 and
0.4.94-b3. The agent answered AFTER_N3_1888.
8. Clean up.
- As
hb3:harness service uninstall(✓ removed ~\.harness\cli\data\harnessd-task.xml: Task Scheduler no longer runs harnessd.),harness stop,PORTandADAPTER_UPDATE_URLdeleted fromHKCU\Environment, and the pty host ended. The agent had been deleted first (deleted: true), and itspowershell.exeandOpenConsole.exewere gone 4 s later. - As admin: the update server stopped, the empty
\HarnessTesttask folder removed, thehb3profile and user removed, andC:\work\b3removed. - The profile at first failed to go, because a service still had hb3's registry hive loaded after every hb3
process had ended. It went after a forced unload of the hive (
NtUnloadKey2,REG_FORCE_UNLOAD). Afterwards nohb3user, noC:\Users\hb3, no ProfileList key, no Harness task and no listener on 8765 were left.
Watched
- One log for the whole daemon. The master's own
[harnessd]lines and its children's ([cli],[update],[terminal],[hooks]) are all incli\data\harness.log. No master crash by exception was forced: that needs a build that throws. A master ended by TerminateProcess leaves no line, as expected. - "this data folder is too deep for the daemon's local socket — the services run in the core's process" is logged at every Windows master start that was looked at (03:34:09, 03:36:15, 03:40:51 and 03:50:11). See F2.
- Log trimming while the file is held open. The master, the core and the updater all held the file open.
Filler was appended to make it 11.5 MB. At the next trim tick, within 20 s, it logged
[log] trimmed — dropped the oldest 6 MB (cap 10 MB)and the file was 5.2 MB. There was no EBUSY or EPERM, no NUL gap, and the next master and core lines (a core restart 30 s later) landed after the header. harness stopduring a handover. Not seen. No stop was run in a handover window, and each handover took about 4 s from staged to handed over.ssh hb3@vm "harness start"returns at once: 6.0 s on the first start, 2.9 s later, and 12.8 s with the task installed (it waits for the task first).- Retirement and decision files. An old version folder was removed once nothing ran from it: N+2 after its
rollback, and N after N+3 was kept. The
<marker>.<token>.handoverfiles were swept. After the N+2 rollback only the newestharnessd-reexec.json.<token>.handoverwas left, and after a fresh master start there were none.
Findings
- F1.
taskkillis refused to a standard user in an SSH session, even for that user's own processes.taskkill /PID <pid> /Fand/T /FgaveERROR: Access denied(exit 1) ashb3, even on apingit had just started.Stop-Process, which is TerminateProcess, worked.tasklist /VandGet-CimInstance Win32_Processare refused the same way: an SSH key logon is a network logon, and WMI turns down a network-logon token that is not an administrator's.lib/windows/taskkill.tstaskkillTree(the agent signaller inptyHostBackend.ts) will therefore fail in a daemon a standard user started over SSH. Deleting a PowerShell agent still worked here, because its processes ended when its console closed. Fixed in B3.22: the signaller ends the tree with TerminateProcess over the host's process table (lib/windows/treeKill.ts); see "B3.22" below. - F2. On Windows every service runs in the core's process, and the log gives the wrong reason.
masterEnvgets no socket path on win32, setsHARNESSD_SERVICES=none, and logs the Unix-socket line at every start. Only the updater, which belongs to the master, ran in a process of its own. Fixed in B3.22: the line sayson Windows the services run in the core's process. - F3. Engine hooks are never installed on Windows. At every core start,
[hooks] failed to write …: HookPreparationHeld: Hook preparation is waiting: the configuration directory exceeds its path limitis logged with a stack trace, for 12 engines' files (over 100 lines per start).nativeHookFiles.tsdirectoryPathandsyncDirectoryChainwalk up to/. This is outside B3. - F4. Minor.
- A core ended by TerminateProcess is logged as
code 4294967295, which is -1 read as unsigned. Fixed in B3.22. harness service installwarns thatPORTis "set only here" even when it is set for the Windows user withsetx(the code says it cannot tell the two apart). Fixed in B3.22 forPORT.- B1.9's concern did not come back: no
harness service install exited 1under-NoService.
- A core ended by TerminateProcess is logged as
B3.22: an agent's tree ended without taskkill
Date: 2026-10-17. Branch feat/windows-native-cli on a5339f101 with B3.22's changes. Same VM. The pty host is the
release build B3.21 made (0.4.91); the CLI bundle 0.4.95-b22 was built on the Mac from this tree and installed from
a local manifest (windows-install-manifest.mjs, the B1 payload's conpty.dll and OpenConsole.exe).
Who ran it. A fresh standard user hb22, made as in B3 (New-LocalUser -PasswordNeverExpires, Users,
userenv CreateProfile, the admin key copied). Over SSH its token is a network logon (NT AUTHORITY\NETWORK), not an
administrator's. It set PORT=18622 and ADAPTER_UPDATE_DISABLE=1 with setx. Install: install.ps1 -Manifest … -NoService.
F1 reproduced, then gone.
- As
hb22,taskkill /PID <its own ping> /T /FgaveERROR: Access denied(exit 1);Stop-Processended it. - As
hb22, over SSH, with the real pty host:treeKill.real.spec.ts,processControl.real.spec.ts(a pane's engine ended throughsignalAgentProcess, and a restart that respawns in the same pane) andterminate.real.spec.ts: 4 passed (3 files). - The control: the same
processControl.real.spec.tsfroma5339f101(the taskkill signaller), run ashb22: 2 failed, each with[terminal] ptyhost: taskkill could not end pid … (SIGKILL|SIGTERM)and the engine still running after 30 s. - As admin:
treeKill.real.spec.ts(node → cmd → ping: ping, then cmd, then node; a bystander lives),terminate.real.spec.ts,taskScheduler.real.spec.ts(with its new disabled control) andprocessControl.real.spec.ts: all passed.
The daemon, as hb22. harness start came up on 18622. Its master's first line was
[harnessd] on Windows the services run in the core's process. A PowerShell agent was made over the local
WebSocket and ran ping -t. agent_restart ended the shell and its ping ([restart] … terminal · fresh shell) and
agent_delete left no ping. A terminal agent's restart and delete go through the host's respawn and pane kill (the
pane's Job), not the agent signaller: the signaller is for engine agents, which the real specs above cover on the
same host. An engine agent through the daemon was not reached: a stand-in opencode.exe was held at create with
NATIVE_LAUNCH_HELD ("Waiting for the selected engine executable to stop changing"), not looked into further.
F4. Stop-Process on the core, as hb22: [harnessd] core exited (code -1, crashed) — restarting in 500 ms,
then core started … restart 1 and core ready. harness service install with this shell's PORT the same as the
user's: no PORT warning; with PORT=18623 in the shell: PORT, ADAPTER_UPDATE_DISABLE set only here do not reach it. ADAPTER_UPDATE_DISABLE, also set with setx, is still warned of: only PORT is read back.
Clean-up. As hb22: harness service uninstall, harness stop, its pty host ended, PORT and
ADAPTER_UPDATE_DISABLE deleted. As admin: the user removed; its profile removed after a forced unload of its two
hives (a service held them, as in B3); C:\work\b22 removed. Afterwards: no hb22, no C:\Users\hb22, no
ProfileList key, no Harness or HarnessTest task, no Harness process. C:\work\e2e and the admin profile were not
touched.