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.

31.0 KB
Raw Blame

Windows PTY host: end-to-end results and performance baseline

Date: 2026-10-15. Branch feat/windows-native-pty. Base f30915019; four fixes found by this run are committed on top of it (see Fixes made during the run). The final runs used build v0.4.9-dev.c71364a8e.

This is Task 23 of the Windows native PTY plan. The spec's testing section (§5, "End to end") is in the design.

How it was driven. Everything ran on the VM against the VM daemon's loopback WebSocket (ws://127.0.0.1:18473/api/local-ws). No account was signed in. The driver uses the same transport and frames as the desktop app: machine_select, agent_create, terminal_open with protocolVersion: 3, binary input, paste and keyframe frames, terminal_resize and terminal_resync. It lives in cli/scripts/windows-pty-e2e/. One check needs the user's account and was not run: the Mac desktop app as a remote client of the VM (see Not run).

Evidence. Each run's output is committed as cli/scripts/windows-pty-e2e/evidence/<run>.log, named after the run.cmd command or step that produced it: s1 (the first scenario 1 run), s1ps, s1b, s1c, codex-emoji, s2 and s2-rerun, s3-two-binary, s3-live and s3-live-close, s4, s4-resume, s4-setup and s4-rerun, s5, s6, s7, perf-vm, perf-stream-driver, perf-mac-tmux and bg-start. They hold no account credential (nothing was signed in) and no host token.

Summary

# Scenario Result
1 Claude Code, Codex and PowerShell agents: create, type, 30 KB paste, Vietnamese/CJK/emoji, resize Pass for the Harness side. Engine caveat: Codex drops emoji and splits large pastes (F13).
1b The same, from the Mac desktop as a remote client Not run: needs the user's sign-in
2 Daemon restart: ids kept, streams reattach Pass
3 Protocol bump drain: two-binary test, and live on the running daemon Pass (both). The live run used the production shutdown op path for the first time.
4 taskkill /F the host: agents stop, resume relaunches Pass after fix F4. Before the fix, resumed and new engines were archived within seconds. Resume starts a fresh conversation, because hooks are off (F6).
5 Child of a PowerShell agent is gone after the agent closes Pass
6 Set-Date +3h, then back: identities unchanged Pass. The clock was set back and checked.
7 Every other engine in cli/src/engines See the engine table
— Performance baseline (VM host vs Mac tmux) See the tables

Four bugs blocked the run and were fixed, each in its own commit with a test (F1–F4). The other findings are recorded below and were not fixed.

Setup recipe

VM. A Windows 11 ARM64 VM in UTM (build 10.0.26100.4349, 8 logical CPUs), reached as ssh [email protected]. The SSH session is an elevated admin running cmd.exe. Installed beforehand:

  • Node v24.20.0, Git for Windows, Rust (aarch64-msvc) and MSVC;
  • Claude Code 2.1.300 and Codex 0.162.1, both signed in;
  • OpenCode 1.19.0.
  1. Sync the tree. Copy the sources to C:\work\oh (it excludes the target folders and cli/node_modules): cli/scripts/windows-pty-e2e/vm-sync.sh --cli. VM_SSH and VM_DIR name another machine or folder (defaults [email protected], C:/work/oh).
  2. Build the host on the VM.
    • Run ssh [email protected] "cd /d C:\work\oh\pty && cargo build --release".
    • The result was harness-pty.exe at 1,299,456 bytes, sha256 6106543D…D696, reporting {"version":"0.1.0","protocol":2}.
    • The ConPTY payload is pty/vendor/conpty/arm64/ from pty/scripts/fetch-conpty.ps1, with ConPTY 1.25.260930003: conpty.dll (116,024 bytes) and OpenConsole.exe (1,149,752 bytes).
  3. Build the CLI bundle on the Mac.
    • Run cd cli && . scripts/lib/build-label.sh && ADAPTER_VERSION="$(harness_build_label "$PWD")" npm run bundle.
    • Copy dist/cli.js and dist/notify.mjs to C:\work\e2e\dist\, for example with tar -cf - -C dist . | ssh … "tar -xf - -C C:/work/e2e/dist".
    • Copy harness-pty.exe, conpty.dll and OpenConsole.exe to C:\work\e2e\dist\pty\.
  4. Copy the scripts. Copy the scripts in cli/scripts/windows-pty-e2e/ to C:\work\e2e\:
    • env.cmd, start.cmd, stop.cmd, launch.ps1, run.cmd and cli.cmd;
    • procs.ps1, psl.ps1, perf.ps1 and s6.ps1.
  5. Build the driver. Run cli/scripts/windows-pty-e2e/build.sh <dir> and copy e2e.mjs to C:\work\e2e\.
  6. Start the daemon outside the SSH session.
    • Run powershell -NoProfile -ExecutionPolicy Bypass -File C:\work\e2e\launch.ps1.
    • This uses WMI Win32_Process.Create to run start.cmd. That runs node dist\cli.js start -f in the foreground, with output to C:\work\e2e\daemon.log.
    • Why: OpenSSH for Windows kills every process of a session when the session ends (F16). A plain background harness start from SSH died with the session.
    • Stop it with C:\work\e2e\stop.cmd (harness stop).
  7. Isolate the daemon. env.cmd sets:
    • ADAPTER_DATA_DIR, HARNESS_PTY_DIR, ADAPTER_COMPUTER_ID_FILE, HARNESS_LOGS_DIR, HARNESS_HOOK_ROUTES_DIR, HARNESS_AGENT_MESSAGES_DIR, DSH_DIR and HARNESS_DEVICES_DIR, all under C:\work\e2e;
    • DISABLE_GRID_INSTALL=true. The value has to be true; 1 is read as off (F9). Engine hooks stay on (B4.15);
    • With hooks on, the daemon writes the running user's real engine homes (~\.claude\settings.json, ~\.codex\hooks.json and config.toml, the OpenCode, Kilo, Pi, Amp, Copilot and Command Code hook files). Back them up with their sha256 before the first start, and verify the restore afterwards: the procedure and the list are in evidence/hooks.log, section 10;
    • HARNESS_PTY_BIN=C:\work\e2e\dist\pty\harness-pty.exe. This one is required, because the "beside the bundle" lookup does not work under start -f (F8).
  8. Drive the scenarios.
    • Run C:\work\e2e\run.cmd <command> [args], for example run.cmd s1 or run.cmd list.
    • The commands are in scenarios.mts. Evidence was captured from their output.
    • The driver selects the machine id the daemon serves, as the CLI's own clients do: the sign-in's machine id when the daemon is signed in, otherwise the isolated computer-id file. A signed-in core serves under its account's machine id, so the computer id alone was refused with NO_PEER_LINK. The loopback API needs no account.

Scenario results

Scenario 1: create, type, paste, Unicode, resize

run.cmd s1 created three agents in C:\work\e2e\proj:

Agent Pane Ready after
PowerShell (terminal) %8 58 ms
Claude Code %9 1.1 s
Codex %10 1.7 s

Each was opened at 120×40. Some first-run checks failed because of the driver, not Harness, and were run again (s1ps, s1b, s1c):

  • a wrap-sensitive assertion;
  • Claude's one-time "Try the new fullscreen renderer?" dialog swallowing the typed prompt;
  • Codex's slow paste, given too short a timeout.

PowerShell 5.1. All 7 checks pass (s1ps, 7/7):

  • Prompt. It appears on the stream about 3 s after create.
  • Typed command. Write-Output ('E2E_' + (40+2)) printed E2E_42.
  • 30 KB paste. 300 lines of 100 characters (30,299 bytes) were pasted into a [Console]::In.ReadLine() loop. PowerShell counted TOTAL=30000 HASH=333713715, which matches the checksum the driver computed. Every byte arrived in order, in 1–3 s.
  • Unicode paste. Tiếng Việt: Xin chào thế giới — 中文:你好世界 — 日本語:こんにちは — 한국어: 안녕하세요 — 🎉🚀👍 went in through Write-Output '…'. The output line is exactly the pasted text.
  • Typed Unicode. gõ tiếng Việt 汉字 ✓, sent 3 characters at a time, came back intact.
  • Resize. At 120×40 → 100×30 the keyframe had 30 rows, and $Host.UI.RawUI.WindowSize said 100x30. Back to 120×40, it said 120x40.

Claude Code 2.1.300. All 6 checks pass (s1b, 6/6):

  • Folder trust prompt. It is answered. The selection starts on "No, exit" in this build.
  • Prompt. "What is 6 times 7? Reply with only the number." was answered ● 42.
  • 30 KB paste. It showed [Pasted text #2 +299 lines], settled after about 7.6 s, then was cleared.
  • Unicode paste. The whole string showed in the input box, emoji included.
  • Resize. At 100×30 the keyframe had 30 rows and the widest line was 100. Back to 40 rows.

Codex 0.162.1. Harness passes all its checks; Codex's own handling of input has caveats (s1c, 5/6, plus the emoji probe):

  • Answer. • 42.
  • 30 KB paste. It showed as [Pasted Content 5952 chars][Pasted Content 1122 chars], settled after about 23 s. In the first run it showed as [Pasted Content 5125 chars][Pasted Content 25241 chars], and later lines leaked into the composer as typed text.
  • Unicode. CJK and Vietnamese arrive. Every emoji is dropped, whether typed or pasted: P1 🎉🚀 中文 Việt → P1 中文 Việt and a🎉b → ab.
  • Resize. At 100×30 the keyframe had 30 rows. Codex redraws its inline UI.

The same bytes reach PowerShell and Claude intact, so this is Codex's console input on Windows, not the host (F13).

Tile size. Resize is covered by terminal_resize. A "tile" is a desktop concept, and the desktop was not part of this run.

Not run: the Mac desktop as a remote client

This check needs the VM signed in to the user's account, which this run must not do. The VM has since been signed in (2026-10-15, after this run), so the check can now be run as below; its result is still not run. To run it:

  1. On the VM, start the daemon as in the recipe. Then sign it in with the same account as the Mac desktop:
    • C:\work\e2e\cli.cmd login --qr, which a signed-in phone scans; or
    • cli.cmd login --google in a browser in the VM's desktop session.
  2. Restart the daemon (stop.cmd, then launch.ps1). Check that C:\work\e2e\cli.cmd status shows it signed in.
  3. On the Mac, open the Harness desktop app. The VM appears as a remote machine. Create a Claude Code, a Codex and a terminal (PowerShell) tile on it. Type, paste about 30 KB, paste the Unicode line above, and resize the tile. Check that the tile's screen matches.
  4. Sign the VM out afterwards (C:\work\e2e\cli.cmd logout), if it should not stay linked.

Scenario 2: daemon restart

s2before, then stop.cmd, then launch.ps1, then s2after (evidence in s2-rerun, 7/7):

  • Timing. The daemon stopped at 12:22:51 and listened again by 12:22:59.
  • Agents. The three live agents kept their ids, panes (%8, %9, %12) and engines. No new rows appeared.
  • Streams. All three reattached:
    • PowerShell answered AFTER_42;
    • Claude's screen still showed the earlier ● 42;
    • Codex's TUI was back.
  • Processes. The process table before and after is identical for the host (pid 12232, started at 03:32 and alive through 7 daemon restarts), claude.exe, the Codex node.exe and powershell.exe.
  • Status. A terminal row's status reads offline after a restart (F11). That is cosmetic: its stream reattaches.

Scenario 3: drain on a protocol bump

Two-binary test on the VM. vitest run src/lib/ptyHost/drain.spec.ts, with:

  • RUN_TERMINAL_CONTRACT=1;
  • HARNESS_PTY_BIN_PREV = the protocol-1 debug build at C:\work\oh-p1;
  • HARNESS_PTY_BIN = the protocol-2 release host.

All 23 tests pass, including "streams the old host's pane, makes new panes on the new host, and lets the old one go once its pane closes" (3.7 s).

Live, against the running daemon. This bumps from 2 to 3. It is the first time the production shutdown path ran, because Task 21's live drain could only ever be 1 → 2.

  1. A protocol-3 build was made for the test only:
    • the host was built with PROTOCOL = 3 in a copy of the tree;
    • the bundle was built with PTY_PROTOCOL = 3, into a scratch folder;
    • the source edit was reverted at once;
    • both were installed at C:\work\e2e\dist-p3.
  2. With a Claude, a Codex and a PowerShell agent on the p2 host, the p2 daemon was stopped and the p3 daemon started. The log shows ptyhost: available (p2, draining) and ptyhost: available (p3).
  3. Old agents keep working (s3check, 6/6):
    • all three are still listed on instance p2;
    • Claude's and Codex's screens stream;
    • PowerShell on p2 answered DRAIN_42.
  4. New agents go to the new host. A new terminal landed on p3 (pane %1) and answered NEW_42.
  5. The old host exits once its panes close:
    • the old agents were closed (engines exited, terminal rows deleted);
    • the log shows draining p2 has gone (shut down with no panes left (pid 2936)), then ptyhost:p2 drained and removed;
    • pid 2936 exited, pty-host-2.json was removed, and the p3 host was untouched.
  6. Afterwards the p3 agent was deleted, the p3 host ended and the p2 daemon restored.

Scenario 4: kill the host, then resume

s4setup, then taskkill /F /PID <host>, then s4after (evidence in s4-rerun, 9/9, after fix F4):

  • Job kill. Within 3 s of the taskkill, every pane process was gone: claude.exe, Codex's node.exe, powershell.exe and each pane's OpenConsole.exe. That is the Job's kill-on-close.
  • Reconnect. The daemon logged ptyhost: lost the connection to the host; reconnecting. About 1 s later it logged reconnected (p2, a new host: its panes are gone).
  • Rows. The three rows went stopped with no pane, removed · terminal runtime absent after 2 confirmed scans.
  • Resume. agent_resume on each:
    • the terminal came back with resumed: true on %2001;
    • Claude (%2002) and Codex (%2003) came back with resumed: false. The log says [resume] … hook=no · fresh, because no conversation was bound with hooks off (F6).
  • 15 s later. All three still have their engine, and their screens are the engine's UI.

Before fix F4. The first run (s4, s4-resume) failed. The resumed Claude and Codex were archived 8 s after resume (retained · engine process absent after 2 confirmed scans) while claude.exe and Codex ran in their panes. Their panes then turned into terminal rows. A brand-new Claude agent made then failed the same way within 1 s. See F4.

Scenario 5: no leftover after closing a PowerShell agent

In a PowerShell agent: Start-Process powershell -ArgumentList '-c','Start-Sleep 600'; … Get-CimInstance … 'Sleep 600'. It listed the child 5932 with parent 13104, the pane's powershell.exe.

After agent_delete (deleted: true), 3 s later, neither 13104 nor 5932 exists. The six powershell.exe the person had open in the desktop session beforehand (parent 9628) were untouched (evidence in s5).

Scenario 6: clock change

s6.ps1 does the following:

  1. It records the live rows and the registry's process identities (pid and w: start marker).
  2. It runs Set-Date (Get-Date).AddHours(3) and waits 20 s, which is four reconcile passes.
  3. It checks again.
  4. In a finally, it runs Set-Date (Get-Date).AddHours(-3) and waits 15 s.
  5. It checks a third time.

Results (evidence in s6):

  • At +3h and after the clock was set back, there are the same 3 ids, panes and engines, and no new rows (4/4 and 4/4).
  • The identities are unchanged, for example 20d8966b 9456 w:134365409702318334.
  • After the run the VM clock read 12:31:49Z against the Mac's 12:31:55Z (the end of the SSH call). That is the same offset of a few seconds as before the scenario. The VM has no time service running (w32tm: the service has not been started).

Scenario 7: other engines

run.cmd s7 ran agent_create for each engine in cli/src/engines. Four npm-installable engines were installed for this test with their recipe's command, and uninstalled afterwards. Nothing was signed in.

Engine Version Result
Claude Code 2.1.300 Works (scenarios 1–6). Stop is refused (F5). Resume is fresh (F6).
Codex 0.162.1 Works with caveat: emoji dropped and large pastes split (F13). Stop is refused (F5). Resume is fresh (F6).
PowerShell (terminal) 5.1 Works
OpenCode 1.19.0 Does not run on Windows ARM64: the TUI exits (bun:ffi unavailable in the ARM64 build, known). The pane hands on to PowerShell and the row stops.
Pi 1.2.0, npm install -g --ignore-scripts @earendil-works/pi-coding-agent Works with caveat: the TUI is up (it downloads fd and rg), then says "No models available. Use /login". The launch stayed starting (no readiness without hooks).
Kilo 7.8.12, npm install -g @kilocode/cli Works: the TUI is up and the launch is ready. npm 11 blocked the postinstall until --allow-scripts=@kilocode/cli was given.
GitHub Copilot CLI 1.0.98, npm install -g @github/copilot Works with caveat: the TUI is up at its folder-trust prompt. It needs a GitHub sign-in to be used, which was not done.
Command Code 1.82.3, npm i -g command-code Does not run on Windows: its binary is cmd, which resolves to C:\Windows\System32\cmd.exe. The pane runs Windows cmd and the launch stays starting (F7). Fixed (B4.13): 1.84.0, installed with npm i --prefix, starts as command-code and shows its sign-in screen; the launch is ready.
Cursor, Hermes, Devin, Muse, Amp, Grok, Antigravity (agy) — Not installed. Their recipes are curl … | bash. Create is refused with NATIVE_LAUNCH_HELD, "<engine> is not installed. Installing engines is not supported on Windows yet: install it, then try again" (the F2 fix).

Performance baseline

All CPU figures are percent of one core, as % Processor Time and ps %cpu both report. Each figure is an average over 60 s.

The idle agents are plain shells with nothing running: PowerShell on the VM, zsh -f on the Mac. "Streaming" is one agent printing continuously with a client reading its output:

  • VM: while ($true) { Get-ChildItem -Recurse C:\Windows }, with the driver's WebSocket stream open. About 24,000 characters per second reached the client.
  • Mac: while true; do find /System/Library -print; done, with a tmux control-mode client attached. It prints much faster.

VM: Windows 11 ARM64 in UTM, 8 vCPUs (perf.ps1)

Get-Counter '\Process(harness-pty)\% Processor Time' -SampleInterval 1 -MaxSamples 60, with the working set read at the end. Each pane also has its own OpenConsole.exe, the bundled ConPTY, shown separately. The daemon's core is shown for context.

Agents Host CPU avg (max) Host working set Host private OpenConsole CPU / working set (all panes) Daemon core CPU / WS
0 idle 0.67 % (6.2) 9.8 MB 1.7 MB — 1.07 % / 97.7 MB
4 idle 0.67 % (4.7) 13.8 MB 21.0 MB 0 % / 54.3 MB 1.04 % / 97.4 MB
8 idle 0.83 % (6.2) 18.3 MB 41.4 MB 0 % / 108.6 MB 1.38 % / 98.6 MB
1 streaming 7.36 % (34.2) 18.6 MB 12.7 MB 40.5 % / 13.6 MB 8.93 % / 99.2 MB

Per idle pane, that is about 1.1 MB of host working set, about 5 MB of host private memory (the grid and its scrollback) and about 13.5 MB for its OpenConsole. The idle host CPU (about 0.7 %) is the daemon's periodic inventory, process table and capture calls, and it hardly grows with panes.

Mac: tmux 3.8, M2 Max, 12 cores (tmux-perf.sh)

This uses a private server (-L e2eperf -f /dev/null, its own TMUX_TMPDIR), never the user's. CPU is the server's ps -o %cpu sampled each second, and its CPU-time delta over the 60 s. RSS is from ps -o rss.

Agents tmux CPU (ps avg / CPU time) tmux RSS
0 idle (exit-empty off) 0.00 % / 0.00 % 1.5 MB
4 idle 0.00 % / 0.00 % 1.6 MB
8 idle 0.00 % / 0.00 % 2.3 MB
1 streaming 42.6 % / 43.5 % 4.6 MB

Not directly comparable.

  • The VM is ARM64 under virtualization, while the Mac is bare metal, and it was loaded by other work (load average 7–12).
  • The streaming workloads print at very different rates.
  • tmux has no per-pane ConPTY process. On Windows the console host is a separate process per pane: tmux's grid work is split between harness-pty.exe and OpenConsole.exe.
  • The idle Windows host is not at zero, because the daemon polls it. The tmux server is polled through tmux commands, whose CPU lands in short-lived clients and not in the server.

Fixes made during the run

Each fix is a separate commit with a test that fails without it. Each was verified on the VM.

  • F1 · 10a40058d — the hook credential is reused on Windows.
    • readHookCredential required file mode 0o600, but NTFS reports 0o666.
    • So the second start of a daemon on the same data folder failed EEXIST: file already exists, open '…\hook-credential' and stayed in safe mode ("Nothing else on this machine works").
    • Now the mode is checked only off Windows. The folder's location check (secureState) protects it on Windows.
    • Test: hookAuth.spec.ts.
  • F2 · 3fd2c4d8a — the native install history is not consulted on Windows.
    • NativePaths walks POSIX paths only, so every install-history read threw ("Unconfirmed missing installation route").
    • So every agent_create, a terminal's included, answered NATIVE_LAUNCH_HELD "Waiting to confirm durable engine installation ownership".
    • On win32, status now reports no obligation, and ensure holds with "is not installed. Installing engines is not supported on Windows yet". Install-first is sub-project B.
    • Test: nativeInstall.spec.ts.
  • F3 · 33fd3c11c — the profile is HOME for native launches.
    • Windows sets no HOME. OpenSSH does, which is why earlier VM runs, all over SSH, had one.
    • A daemon started any other way (WMI here; a service, a scheduled task or the desktop later) held every Claude and Codex launch at "Waiting to read the engine's launch environment".
    • Now os.homedir() is used when HOME is unset on win32.
    • Test: launchExecutable.spec.ts.
  • F4 · c71364a8e — engines in panes of a host the daemon started are found.
    • Discovery excludes the daemon's own descendants. Windows does not reparent, so a harness-pty.exe the current daemon started keeps it as parent. Every engine in every pane then read as the daemon's own process.
    • When it happens: after a host crash, after a reboot, or on a machine's first start, every new or resumed engine agent was archived within seconds and its pane turned into a terminal row, until the daemon restarted.
    • A pane's root now ends the exclusion. A tmux server daemonizes, so Unix is unchanged.
    • Test: terminalAgentDiscovery.spec.ts.
    • Scenario 4 shows it fixed on the VM.

Checks run.

  • tsc --noEmit is clean.
  • vitest on src/lib/terminalAgent*, hookAuth, src/engines/kit, prepareNativeLaunch, core/agents/discovery and registry.spec: 877 passed and 8 failed. The 8 failures are in exactTranscript.spec.ts and storePool.spec.ts, and they fail the same way with the four source files reverted to f30915019 (pre-existing on this Mac).
  • nativeInstall.spec.ts has existing tests that fail intermittently under the Mac's load: a different 2–5 of 38 on each run, with or without the fix. The new test passes.

Findings not fixed

  • F5 — Stop is refused for Claude Code and Codex on Windows.
    • agent_delete answers IDENTITY_UNAVAILABLE "Conversation identity is held: the native ps probe did not complete" (held, retryable).
    • The path: stopAgentService calls captureResumeIdentity, which calls readProcessEvidence, which runs nativeProbe('ps', …). There is no ps on Windows. It is reached because no sessionId is bound with hooks off.
    • This follows the spec's rule that "we could not look" is not "nothing is there", so it is left for B: hooks, or a Windows process-evidence reader.
    • Workaround used here: end the engine inside its pane (/exit or Ctrl+C). The row is archived and the pane becomes a terminal row, which can then be deleted.
    • With hooks on (B4.15, evidence/hooks.log): stop is accepted for Claude Code, Codex and Kilo once the pane's conversation is bound (deleted: true). A pane with no bound conversation is still refused this way.
  • F6 — Resume reopens no conversation for Claude Code or Codex (resumed: false, "fresh"). Hooks are off, so no conversation is ever bound (sub-project B).
    • With hooks on (B4.15, evidence/hooks.log): resume answers resumed: true for Claude Code and Codex, keeps the sessionId, and the new pane shows the earlier prompt and answer.
  • F7 — Command Code did not start on Windows. Fixed (B4.13).
    • ENGINE_CLI_COMMANDS.commandcode is cmd, which resolved to C:\Windows\System32\cmd.exe before npm's cmd.cmd. Every Command Code pane opened the command processor instead.
    • Fix: on win32 Command Code starts as command-code, from the same package (engineBin), and a launch never searches PATH for a program named cmd there (launchSearchNames). macOS and Linux still start cmd.
    • Verified on the VM with Command Code 1.84.0 (B4 install, hooks off): the pane shows Command Code's own sign-in screen and launch is ready. The engine is node.exe …\command-code\dist\index.mjs, a direct child of the pty host, with no cmd.exe (cli/scripts/windows-pty-e2e/evidence/commandcode.log).
    • Hooks not verified: needs a Command Code sign-in. Hooks were off in this run; they are verified in B4.15.
    • Command Code 1.84.0 does not read COMMANDCODE_HOME: it puts its files under HOME, or USERPROFILE when HOME is unset, plus .commandcode. It also always writes telemetry-install-id and updates.json to %USERPROFILE%\.commandcode, even when HOME is set.
  • F8 — The "beside the bundle" host lookup misses under harness start -f.
    • The master runs the core from data\lean\<hash>\, so ptyHostBinary() looks for data\lean\<hash>\pty\ and logs ptyhost: unavailable (harness-pty is not installed), although dist\pty\ has it.
    • This only affects local testing. Releases use runtime\pty\current-pty (B). HARNESS_PTY_BIN was used here.
  • F9 — DISABLE_HOOK_INSTALL=1 does not disable hooks, because flag() accepts only true.
    • With hooks not disabled on Windows, every engine's hook install fails with "the configuration directory exceeds its path limit" (B).
  • F10 — The master's startup log on Windows is misleading (cosmetic).
    • It logs that process.execve is unavailable, and "this data folder is too deep for the daemon's local socket".
    • The second reason is wrong on win32, which has no local socket at all (localSocketPath returns null).
  • F11 — Terminal rows read status: offline after a daemon restart until something touches them.
    • By code reading this is not Windows-specific. The reconciler reactivates only rows with an observed engine process, and a shell has none.
    • Streams reattach normally.
  • F12 — Lock inspection is a risk on Windows (not observed).
    • registry.ts:445 (registry lock) and daemonSpawnLock.ts:142 (spawn lock) check POSIX modes 0700/0600 when they meet an existing lock.
    • On Windows that check would throw "unsafe owner, mode, or type" for a stale or contended lock. A daemon killed while it holds the registry lock would then block registry writes until someone removes the lock by hand.
  • F13 — Codex on Windows (engine, not Harness):
    • non-BMP characters (emoji) are dropped, typed or pasted;
    • a 30 KB paste arrives as two or more [Pasted Content …] chunks (some lines as typed text) and takes 15–23 s to settle;
    • a ? was once lost in fast typing, and was not reproduced in a later probe.
    • PowerShell and Claude get the same bytes intact.
  • F14 — OpenCode's ARM64 build has no bun:ffi, so its TUI exits (known since Task 20). Measured again in Task B4.12 (cli/scripts/windows-pty-e2e/evidence/opencode-arm64.log), outcome C:
    • OpenCode 1.19.3, the latest, still stops: Failed to initialize OpenTUI render library: bun:ffi dlopen() is not available in this build.
    • Its x64 build crashes (access violation) under emulation before OpenCode's screen shows. npm i opencode-ai --cpu=x64 fails in OpenCode's install script and leaves no program. Only the x64 baseline build drew, unpacked by hand.
    • Harness now refuses the Arm build before a pane opens. create opencode answers ENGINE_LAUNCH_REFUSED with "OpenCode does not run on Windows on Arm yet: its Arm build stops before its screen opens. Use another engine on this computer."
  • F15 — Side effect on the VM's Claude Code settings.
    • Claude's one-time "Try the new fullscreen renderer?" dialog was answered by typed text in the first run, which turned the fullscreen renderer on.
    • It was set back with /tui default, and the feedback form was skipped. ~\.claude\settings.json now contains "tui": "default".
  • F16 — Background start from SSH. harness start (background) from an SSH session reported "started", but the daemon died when the session ended, because OpenSSH kills the session's job. The host survived, because it breaks away. The service and autostart story on Windows is B.
  • F17 — A product-root write outside the isolated folders. ~\.harness\connections\projections.json was written even though the data folders were redirected. It was removed afterwards.
  • F18 — Pane ids jump by 1000 at each new host (%1001, %2001, …). This is how the counter keeps ids unique across hosts. It is noted because it is visible in tmuxPane.

Known limits (by design, sub-project B or later)

  • Hooks are on (sub-project B, Phase B4); see evidence/hooks.log.
  • Engine install-first is not carried on Windows (F2 refuses with a reason).
  • There is no named-pipe trust for the daemon socket and no local socket, so services run in the core's process.
  • The clients (desktop, mobile, hn) do not yet word PTY_HOST_UNAVAILABLE or ENGINE_LAUNCH_REFUSED.
  • The host version is still 0.1.0 across protocol builds. Version folders and the current-pty pointer are the installer's job (B).
  • OpenCode does not run on Windows on Arm (F14; measured with 1.19.3). Harness says so when an OpenCode agent is created, and no pane opens.

Clean-up done on the VM

  • The daemons and every harness-pty.exe this run started were stopped, and no process from C:\work\e2e is left.
  • The four npm engines were uninstalled, and the state they created was removed:
    • ~\.pi, ~\.copilot, ~\.commandcode;
    • ~\.config\kilo, ~\.local\share\kilo;
    • AppData\Local\copilot.
  • Also removed: ~\.harness, the test project's Claude history (~\.claude\projects\C--work-e2e-proj), and the p3 build folders.
  • Left in place: C:\work\e2e, with the bundles, scripts, logs and isolated state, for a rerun.
  • Claude Code (loggedIn: true) and Codex ("Logged in using ChatGPT") are still signed in.