Caps Lock launcher: the Windows half (launch-game.ps1, game-launch.ahk, snapshot-game.ps1)

Delegated run·Sep 11, 2026  15:13·Prompted by Fable·Run by Opus 5Opus 5 xhigh
Ctrl + F
Show

Transcript

Fable·12:13:54·Sep 11, 2026Prompter

You are the implementer of an agreed design brief. The high-level decisions below are closed; the local details are yours. If you hit a high-level choice the brief does not make, find the brief wrong in a way that changes the design, or a step that would reach outside the stated frame, STOP and report instead of deciding. Do not commit anything to git; the directing session reviews and commits.

Task: port the Caps Lock game launcher to Windows

Frame

  • Machine: Windows 11, user a327e, home C:\Users\a327e. Git Bash exists at C:\Program Files\Git\bin\bash.exe. Python 3.13 (python, python3). PowerShell 5.1. AutoHotkey v2.0.27 at C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe.
  • Working tree: C:\Users\a327e\Desktop\a327ex (NOT a git repo itself). Anchor\ inside it IS a git repo (the engine + workflow + launcher). ABAgames\knightvspawns and ABAgames\windrang are game folders with their own repos.
  • Destination for everything you write: C:\Users\a327e\Desktop\a327ex\Anchor\launcher\. It currently holds the LINUX version of this launcher: README.md, launch-game.sh, snapshot-game.sh, roots.txt. Read all four first; the Windows port must mirror the bash script's semantics.
  • Source to port from (READ-ONLY archive drive, never write there): Z:\2025-2026\code\a327ex-hdd-2026-08\tools\game-launch\ holding the 2026-08 Windows version: README.md, game-launch.ahk, launch-game.ps1, snapshot-game.ps1, play.bat, roots.txt. Read all of them. They target a dead root E:\a327ex and an engine surface that has since changed (details below).
  • Engine: the shared build is C:\Users\a327e\Desktop\a327ex\Anchor\engine\engine\build\anchor.exe. A game may carry its own anchor.exe in its folder. Command line of the engine today: anchor.exe <game-folder-name> run from the PARENT directory of the game (the "run.bat way"), or anchor.exe <path\to\file.apr> to play a recording (positional; the old --replay= flag NO LONGER EXISTS). Recordings live in <game>\replays\<timestamp>.apr.
  • Web replay pieces: packaging script C:\Users\a327e\Desktop\a327ex\Anchor\engine\scripts\package-web-game.sh (bash; run it through Git Bash the way the archived launch-game.ps1 does, with the same argument shape: ./package-web-game.sh '<game posix path>' '<out posix path>' . <w> <h> pixelated executed from the scripts directory); the wasm engine artifacts anchor.js + anchor.wasm in C:\Users\a327e\Desktop\a327ex\Anchor\engine\engine\build-web-engine\.

Approach (agreed)

Bring launch-game.ps1, game-launch.ahk and snapshot-game.ps1 back from the archive into Anchor\launcher\, next to the Linux scripts, updated to today's surface. Same four chords: Caps Lock relaunches the last game; Shift+Caps Lock prompts for a name (AutoHotkey InputBox, partial names ok); Ctrl+Caps Lock replays the last game's newest recording on desktop; Alt+Caps Lock replays it in the browser. Discovery, matching and the replay loop stay as they are in the archived script; what changes is the engine surface underneath and where state lives.

Opus 5xhigh·12:13:56·Sep 11, 2026Worked for 10m 59s

I'll start by reading all the source material.

ReadC:\Users\a327e\Desktop\a327ex\Anchor\launcher\README.md12:13:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\README.md
ReadC:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.sh12:13:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.sh
ReadC:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.sh12:13:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.sh
ReadC:\Users\a327e\Desktop\a327ex\Anchor\launcher\roots.txt12:13:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\roots.txt
Thinking12:14:01

I'll switch to reading the archived Windows version next, using Bash rather than PowerShell as the instructions specify.

Now the archived Windows source.

Bashls -la "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" 2>/dev/null || ...12:14:03Output not recorded
Input
ls -la "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" 2>/dev/null || ...
Bashcd "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" && echo "=== README...12:14:03Output not recorded
Input
cd "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" && echo "=== README...
ReadZ:\2025-2026\code\a327ex-hdd-2026-08\tools\game-launch\launch-game.ps112:14:03Output not recorded
Input
Z:\2025-2026\code\a327ex-hdd-2026-08\tools\game-launch\launch-game.ps1
ReadZ:\2025-2026\code\a327ex-hdd-2026-08\tools\game-launch\snapshot-game.ps112:14:03Output not recorded
Input
Z:\2025-2026\code\a327ex-hdd-2026-08\tools\game-launch\snapshot-game.ps1
Thinking12:14:12

I should check the current state of the working tree—engine build, games, replays, etc. I want to confirm the verification step's expected commit message format matches exactly what verbatim copy would produce: "timeline: snapshot <hash> (<name> <stamp>)". Let me check the environment next.

Bashcd "/c/Users/a327e/Desktop/a327ex" && echo "=== top level ===" && ls -la && echo...12:14:13Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex" && echo "=== top level ===" && ls -la && echo...
Thinking12:14:17

I should check whether `~/a327ex` actually exists, since the path differs on Windows, and also scan newer folders like `character-lookdev` and `mmo-character` for potential games.

Bashls -d /c/Users/a327e/a327ex 2>&1; echo "--- discovery scan (2 levels, main.lua +...12:14:18Output not recorded
Input
ls -d /c/Users/a327e/a327ex 2>&1; echo "--- discovery scan (2 levels, main.lua +...
Thinking12:14:43

I notice playground and replay-test live three levels deep under Anchor/engine, not in the default root—yet the verification expects them to show up in the listing, so the root path must need adjusting to include that engine subdirectory.

I'll resolve this by adding the third line since it's literally what the existing roots.txt comment suggests, it's reversible and low-stakes, and the verification section is a specific statement of expected behavior I should honor. I'll clearly flag this as a decision I made beyond the brief's literal wording.

Rereading the stop condition, this isn't really a high-level design choice—it's a config line, and skipping it would be overly literal given the brief's own verification demands it. I need to double-check the cross-platform behavior too: on Windows, the existing `~/a327ex` line resolves to a path that doesn't exist, so I should verify that doesn't break anything.

I'll go with adding the third line to roots.txt for the Windows path while keeping the Linux one intact, treating it as a minimal, symmetric addition rather than restructuring anything. Let me double check the directory depth to make sure playground and replay-test would actually be discoverable from that root before finalizing.

Let me check several of these things at once.

Bashcd "/c/Users/a327e/Desktop/a327ex" && echo "=== knightvspawns anchor.exe? ===" &...12:14:46Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex" && echo "=== knightvspawns anchor.exe? ===" &...
Bashcd "/c/Users/a327e/Desktop/a327ex" && echo "=== playground ===" && ls -la Anchor...12:14:46Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex" && echo "=== playground ===" && ls -la Anchor...
Thinking12:14:55

Checking the depth-2 roots, I'm counting five games total: knightvspawns, windrang, a327ex-site/renderer, and the two under Anchor/engine (playground and replay-test). I'm also considering whether the engine root's depth-2 expansion would pull in additional entries.

Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor" && echo "=== check-ignore ===" && git ...12:14:58Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor" && echo "=== check-ignore ===" && git ...
Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor" && git status --short launcher ; echo ...12:14:58Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor" && git status --short launcher ; echo ...
Thinking12:15:04

There's no ignore rule for launcher state yet, so I'll add a .gitignore under Anchor/launcher to exclude the state folder. Next I want to check the Startup folder and the abnt2-remap.lnk pattern.

Bashls -la "/c/Users/a327e/AppData/Roaming/Microsoft/Windows/Start Menu/Programs/Sta...12:15:05Output not recorded
Input
ls -la "/c/Users/a327e/AppData/Roaming/Microsoft/Windows/Start Menu/Programs/Sta...
Thinking12:15:08

The Startup folder doesn't have abnt2-remap.lnk as expected — I need to search elsewhere for it since it's supposed to serve as the pattern for the exact shape.

Bashfind /c/Users/a327e -maxdepth 6 -iname 'abnt2*' 2>/dev/null | head -20; echo "--...12:15:11Output not recorded
Input
find /c/Users/a327e -maxdepth 6 -iname 'abnt2*' 2>/dev/null | head -20; echo "--...
Thinking12:15:16

The Startup folder is empty aside from desktop.ini, but there's a backed-up shortcut in Desktop/tools/startup-tasks/, suggesting the owner migrated away from Startup-folder shortcuts to another mechanism, maybe scheduled tasks. I want to check that startup-tasks folder to understand what replaced it.

This creates a discrepancy with the brief, which claims abnt2-remap.lnk exists "in that folder" when it's actually sitting in a backup subfolder instead — I need to dig into this before deciding how to proceed.

Bashcd /c/Users/a327e/Desktop/tools/startup-tasks && ls -laR 2>&1 | head -60 && echo...12:15:17Output not recorded
Input
cd /c/Users/a327e/Desktop/tools/startup-tasks && ls -laR 2>&1 | head -60 && echo...
Thinking12:15:30

This is significant — the abnt2-remap shortcut was moved out of Startup into a backup earlier today and replaced with an elevated scheduled task, specifically because AHK global hooks can't reach elevated windows if launched unelevated. The game launcher is also an AutoHotkey global-hotkey script, so the same problem likely applies: running from Startup, it stays unelevated too.

The brief explicitly asks for one shortcut in the Startup folder matching that abnt2-remap pattern, but that premise no longer holds since the folder now has no .lnk files and the established convention has shifted to scheduled tasks. This feels like a case where the brief's premise is wrong — but I don't think it changes the actual design decision: creating a Startup shortcut is still exactly what's authorized, reversible, and functional, whereas a scheduled task would require elevation I can't grant myself and would violate the scope restriction to the launcher folder and Startup shortcut only. So I'll proceed with creating the Startup .lnk as instructed, since it's the authorized and reversible path.

I want to flag the discrepancy clearly in my report and suggest adding game-launch to install-startup-tasks.ps1 as a follow-up, so nothing gets blocked but the owner still learns about it. I should also check abnt2-remap.ahk for any CapsLock handling that might conflict with my SetCapsLockState call, read the backup .lnk for its exact shape as instructed, and verify whether AutoHotkey64.exe supports a /validate flag — I'll batch these checks together.

Bashecho "=== abnt2-remap.ahk ===" && cat /c/Users/a327e/Desktop/tools/abnt2-remap/a...12:15:31Output not recorded
Input
echo "=== abnt2-remap.ahk ===" && cat /c/Users/a327e/Desktop/tools/abnt2-remap/a...
PowerShell$w = New-Object -ComObj WScript.Shell; $s = $w.CreateShortcut("C:\Users\a327e\De...12:15:31Output not recorded
Input
$w = New-Object -ComObj WScript.Shell; $s = $w.CreateShortcut("C:\Users\a327e\De...
Thinking12:15:42

No CapsLock conflict with abnt2-remap, so that's clear. I'm confirming the shortcut file structure with TargetPath, Arguments, WorkingDirectory, and Description, and checking that AutoHotkey v2 supports the `/validate` switch to check scripts for syntax errors without running them.

I'll test that later once the file is finished. Next I want to double-check the engine's replay command-line surface to confirm the positional argument form, glance at SURFACE.md, see how run.bat is structured, and verify package-web-game.sh's argument shape still lines up.

Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor/engine" && echo "=== playground/run.bat...12:15:43Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor/engine" && echo "=== playground/run.bat...
Thinking12:15:48

The argument shape stays unchanged. I notice playground/run.bat calls anchor.exe directly from its own directory rather than following the run.bat pattern described in the brief—working directory set to the game's parent folder with the folder name as argument—but since the Linux script matches that pattern and the engine accepts a relative folder name, this should still work fine.

Now I want to verify the engine's replay CLI uses positional arguments, so let me check SURFACE.md.

Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor/engine" && grep -n -i 'apr\|replay' doc...12:15:49Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor/engine" && grep -n -i 'apr\|replay' doc...
Thinking12:17:19

Confirmed the positional file.apr replaces --replay. Now I have what I need to write the files in the launcher folder: the ported resolver script with a -Print flag, the archived AHK launcher with an updated header, a verbatim snapshot script, and I'm figuring out the remaining files to include.

I'm laying out the design for launch-game.ps1: using script-relative paths, expanding roots.txt entries with tilde and forward-slash support, discovering game folders up to two levels deep that contain main.lua and an anchor directory while skipping dot-folders, and keeping the -Print mode side-effect free by only creating the state directory lazily when actually needed rather than during a dry-run print.

I realize PowerShell's recursive directory search doesn't prune descent into dot-folders the way the bash filter effectively does in practice, so I'm switching to a manual two-level directory enumeration instead of using -Recurse, which lets me properly skip hidden folders and also run faster.

I confirm the sort-by-path ordering matches a nested DFS traversal — level-1 dirs sort before their own children but after alphabetically earlier siblings, so my nested approach naturally reproduces the same order without needing to sort full paths.

On hidden dirs, Get-ChildItem without -Force already excludes hidden/system directories, matching the bash find's exclusion pattern, so no extra handling is needed. For matching game keys I'll preserve insertion order using an ordered hashtable, and for listing I'll just sort the keys directly. Now I'm turning to engine resolution.

I'll derive the engine path from $PSScriptRoot rather than hardcoding it, since deriving is more robust for the launcher sitting alongside the engine in the repo. For replay selection, I'll keep the same newest-by-name approach as the archived behavior, matching what the bash script does with sort/tail.

For launching the replay, I need to redirect the engine's stdout and stderr into last-run.log, but Start-Process won't let me redirect both streams to the same file directly—I'll need to use separate output and error files, or find another way to combine them into one log.

Since PowerShell can only redirect stdout and stderr to separate files natively, I realize the cleanest way to meet the single-combined-log requirement is to shell out through cmd.exe, building a quoted command string that redirects both streams into one file and launching it hidden via Start-Process.

I'm passing this as a single argument string to Start-Process so PowerShell forwards it mostly verbatim to CreateProcess. There's a minor tradeoff with the lingering hidden cmd.exe node in the process tree, but it's acceptable since it exits when the anchor does—alternatively I could use separate stdout/stderr redirection files instead of the cmd wrapper.

But I'm now second-guessing whether `-WindowStyle Hidden` on the cmd.exe wrapper could inadvertently hide the actual game window too—since STARTUPINFO's wShowWindow gets inherited by child processes, and if SDL doesn't explicitly call ShowWindow with SW_SHOW, the hidden flag might propagate down and suppress the game's own window.

I'm also considering redirecting stdout and stderr to the same file for one combined log, but PowerShell 5.1 actually rejects that with an error when both paths are identical, so that approach won't work directly.

Alternatively, I could just pipe the engine's output through `Out-File` synchronously within the script, but that blocks the launcher process until the game closes — unlike the bash version which detaches with setsid. I need to weigh whether keeping a lingering powershell.exe process around is actually a problem here.

Using .NET Process directly would keep PowerShell alive to pump streams, which isn't ideal for a fire-and-forget launcher. I'm weighing whether cmd.exe's hidden window actually propagates safely to the child process, but I'm genuinely uncertain about that STARTUPINFO inheritance behavior, so I'm leaning toward the option with the least window-state risk even if it deviates slightly from the brief's exact redirection approach.

I should check whether anchor.exe is actually a console or GUI subsystem binary, since that determines whether a window would flash on launch. I also recall the original Linux script only redirected output because setsid detaches from the terminal, whereas the archived Windows version didn't redirect at all and worked fine — that's a useful precedent.

I realize I can satisfy the "stdout+stderr" requirement by merging streams with `2>&1` inside cmd rather than juggling two files. But I'm wary of testing whether a hidden window actually stays hidden, since popping a test window would itself be a visible desktop action I'm supposed to avoid.

Decision: use `Start-Process -NoNewWindow` with separate `-RedirectStandardOutput` and `-RedirectStandardError` files instead of a cmd wrapper — this avoids spawning any console window at all, since there's nothing for one to inherit. I'll document this deviation clearly, producing `last-run.log` and `last-run.err.log` as two separate files.

I briefly consider whether I could merge stderr into the single log file afterward by appending its contents post-run, to stay truer to the original spec's single-file expectation.

I realize multiple relaunches could stack processes, but that's already an existing edge case, not something new I'm introducing. The bigger risk is the hidden console potentially leaking through to the game window depending on how SDL creates it, so rather than keep debating that, I'll settle on using Start-Process with redirected standard output.

PowerShell$p='C:\Users\a327e\Desktop\a327ex\Anchor\engine\engine\build\anchor.exe' $b=[IO....12:17:22Output not recorded
Input
$p='C:\Users\a327e\Desktop\a327ex\Anchor\engine\engine\build\anchor.exe' $b=[IO....
Thinking12:17:53

I'm realizing anchor.exe is a console subsystem app, so launching it via Start-Process without -NoNewWindow would spawn a new, visible console window — similar to how the archived script's run.bat approach also opened a visible cmd window despite @echo off, which the original owner apparently just accepted.

To suppress this while still redirecting stdout/stderr to log files, I need Start-Process with -RedirectStandardOutput and -RedirectStandardError alongside -NoNewWindow, since that sets CreateNoWindow=true and hides the console without losing output — and importantly this only suppresses console creation, not the game's actual SDL/GL window.

Since PowerShell rejects identical stdout/stderr redirect paths, I'll just use two separate files: `last-run.log` for stdout and `last-run.err.log` for stderr, and document that split clearly.

Wait, actually there's a cleaner trick: piping through `cmd /c` with `-NoNewWindow` lets me merge both streams into one file before cmd launches anchor.exe as a hidden child process, inheriting the no-console setup.

This solution gives me exactly one combined log file with no visible console, which fits the requirement precisely. The main risk is command-line quoting, so I want to test it carefully first using a harmless placeholder command before trying it with an actual game executable.

I'm also weighing whether the child process's window will actually display when it inherits a hidden console from PowerShell run via AHK — since the game window is created independently by SDL with its own show call, it should appear regardless of the console's visibility state.

Let me just grep anchor.c directly for the window creation flags rather than reasoning about it further.

Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor/engine/engine/src" && grep -n 'SDL_Crea...12:17:54Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor/engine/engine/src" && grep -n 'SDL_Crea...
Thinking12:18:50

SDL3 explicitly calls SW_SHOW rather than SW_SHOWDEFAULT, so the STARTUPINFO show-state concern doesn't actually apply here — the window creation isn't hidden unless agent mode requests it. That means even a hidden window style would be safe, but I'll stick with the cmd wrapper using -NoNewWindow anyway.

Tracing through the console inheritance: when the launcher runs from AHK's hidden PowerShell, cmd inherits that hidden console; if PowerShell exits right after Start-Process, the console persists since cmd is still attached to it, and it only gets freed once anchor and cmd both exit. That confirms the approach is safe, so I'm settling on the cmd.exe wrapper with -NoNewWindow writing to a single combined log file, and now I want to test the quoting behavior with something harmless first.

Now moving to the web replay branch, I'm mapping out the same pattern: a synchronous git bash call to capture packaging's exit code, then copying artifacts, staging apr, and launching python's http.server hidden before opening the browser. I'm setting up the http.server Start-Process call with hidden window style and directory argument, plus a TCP port check matching the earlier archived pattern. For the bash invocation, I'm working out how to redirect both stdout and stderr into the packaging log file, since native command error streams need special handling with 2>&1 piped to Out-File.

But I catch a real hazard: in PowerShell 5.1, redirecting stderr with `2>&1` on a native executable wraps it in ErrorRecords, and combined with `$ErrorActionPreference='Stop'`, any stderr output could throw an exception even on success. I'm now considering having bash redirect its own output internally instead, to sidestep this PowerShell quirk entirely.

5.1 shouldn't throw on nonzero exit codes, which is fine here.

For the -Print flag on web, I need it to print "would web-replay: <key> (<apr>)" before packaging and exit 0. I'm thinking of a single Notify function that branches on $Print — writing to console when set, showing the auto-closing popup otherwise — so misses print instead of popping up, matching verification 6 (exit code 1, console message, no popup) and verification 7's no-last-game case.

I'll have the non-print path both show the popup and write to console (mirroring the bash notify's dual notify-send+echo behavior), while making sure -Print has zero side effects by creating the state directory lazily via a Get-StateDir helper only invoked on real paths, since -Last reads from state\last-game.txt.

I'm checking argument binding: with `-Print knight` the positional value correctly binds to $Name since -Print is a switch, and this holds even with additional switches like -Replay before the positional arg. Tracing through the prefix-matching logic confirms "play" resolves to playground and "knight" resolves to knightvspawns, based on the insertion order of games discovered across the root directories.

For `-Print -Replay knight`, the newest matching .apr file by timestamp needs to be selected. On the AHK launcher script, I'm keeping it archived and unchanged as instructed—no #SingleInstance Force addition—just updating the header comment to reflect its new location, since it already correctly references launch-game.ps1 relative to itself and passes -Name properly.

Now I'm thinking through the README consolidation into a single cross-platform document in plain prose, making sure to avoid em-dashes since the existing docs use them.

For roots.txt, I need a comment explaining it supports both platforms. I'm also verifying that a tilde-prefixed path like `~/Desktop/a327ex/Anchor/engine` expands correctly under Git Bash on Windows, since $HOME resolves properly there and the listing mode doesn't actually need the ENGINE path to exist.

Now I'm ready to write the actual files, starting with launch-game.ps1: I need to make sure $PSScriptRoot works when run with -File, handle tilde expansion by swapping it for $HOME and converting slashes, use Test-Path with -LiteralPath, and derive the engine path relative to the script root before resolving it. I'll model the header comment on the archived version's block comment style.

Writing directory discovery function...

Still writing directory traversal logic... I'm also double-checking that recursing into a game's subfolders like `anchor/` or `assets/` won't mistakenly pick up nested false positives, matching how the original bash script handled it.

Playground and replay-test cover it — my earlier 4-level find under Anchor/engine already confirmed no other games exist there. Now I'll move to writing launch-game.ps1 using the Write tool for cleaner handling of quotes and encoding.

I'll stick to pure ASCII in the script to avoid PowerShell 5.1's ANSI-without-BOM issues, skipping em-dashes and any fancy characters even though the archived version used them. Ready to write it now.

Now I have everything I need. Writing the files.

WriteC:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps112:19:42Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1
Thinking12:19:52

I realize I referenced Start-Engine without defining it, so I need to add that function right after Get-StateDir since PowerShell requires definitions before their call sites. I'm writing it to ensure the state directory exists, log to state\last-run.log, and launch the process through cmd.exe with combined output redirection.

I'm reasoning through the `-NoNewWindow` behavior — since output is fully redirected to the log file, nothing should print to the console whether run from AHK's hidden shell or an interactive terminal, and the console stays alive if the parent exits since cmd remains attached. Now I'm constructing the cmd.exe argument string that wraps the executable and arg with output redirected to the log file, then plan to invoke Start-Process with that command line, working directory, and -NoNewWindow, before inserting the function into the script.

EditC:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps112:19:55Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1
Thinking12:20:05

I'm moving on to the AHK script, snapshot-game.ps1, roots.txt, .gitignore, and README.md, planning to copy snapshot-game.ps1 verbatim from the archive—though I'm noting the em-dashes in its comments could cause encoding issues if the file isn't saved with proper BOM handling for PS 5.1.

I'll copy the file bytes directly rather than read-and-rewrite to avoid mangling any special characters, following past experience where similar roundtrips silently corrupted icon glyphs. For game-launch.ahk, I need to change just the header comment while keeping everything else unchanged, so I'm weighing whether to copy the bytes and edit just the header lines, or rewrite the file fresh in plain ASCII since the README calls for no em-dashes—first I should check the actual encoding of the archived files.

Bashcd "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" && for f in snapsho...12:20:06Output not recorded
Input
cd "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/" && for f in snapsho...
Thinking12:20:12

I'm noting that snapshot-game.ps1's em-dashes might render oddly under PowerShell 5.1's ANSI encoding, but since they're only in comments and the here-string content is pure ASCII, this won't cause functional issues—so I'll just copy the file as-is.

Bashcp "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/snapshot-game.ps1" "/...12:20:14Output not recorded
Input
cp "/z/2025-2026/code/a327ex-hdd-2026-08/tools/game-launch/snapshot-game.ps1" "/...

Now the AHK script, roots.txt and the ignore rule.

WriteC:\Users\a327e\Desktop\a327ex\Anchor\launcher\game-launch.ahk12:20:28Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\game-launch.ahk
WriteC:\Users\a327e\Desktop\a327ex\Anchor\launcher\roots.txt12:20:28Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\roots.txt
WriteC:\Users\a327e\Desktop\a327ex\Anchor\launcher\.gitignore12:20:28Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\.gitignore

Now let me verify the script runs before writing the README.

PowerShellpowershell -NoProfile -ExecutionPolicy Bypass -File "C:\Users\a327e\Desktop\a327...12:20:40Output not recorded
Input
powershell -NoProfile -ExecutionPolicy Bypass -File "C:\Users\a327e\Desktop\a327...
PowerShell$s = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1" "### -Print...12:20:40Output not recorded
Input
$s = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1" "### -Print...
Thinking12:21:10

That exit code 1 just reflects the last $LASTEXITCODE from the earlier run, nothing wrong there. And the "-Print zzzz" query returned no suggestions since no game name contains z-z-z-z as a subsequence, which makes sense.

I want to also test a case with genuine ambiguous subsequence suggestions rather than a unique match, so I'm trying a few candidate substrings like "wrng", "kn", "pln", "rn" to see which ones avoid exact/prefix/substring matches and instead hit multiple names only via subsequence.

I find that "rn" matches renderer, windrang, and playground purely as subsequences, giving three candidates — a good demonstration of the miss-with-suggestions path, which I'll run to confirm. After that I still need to work through the remaining verification steps: the scratch snapshot, AHK validation, the bash listing, git status, and the shortcut check.

I also want to test the Start-Engine command quoting safely, using a harmless exe in a scratch directory rather than actually launching a game, so I'll do that alongside the snapshot test.

PowerShell$s = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1" "### -Print...12:21:14Output not recorded
Input
$s = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.ps1" "### -Print...
Bashset -e SCRATCH="/c/Users/a327e/AppData/Local/Temp/claude/launcher-test" rm -rf "...12:21:14Output not recorded
Input
set -e SCRATCH="/c/Users/a327e/AppData/Local/Temp/claude/launcher-test" rm -rf "...
PowerShell$snap = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.ps1" $g = "...12:21:14Output not recorded
Input
$snap = "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.ps1" $g = "...
Thinking12:21:29

Snapshot works, so now I want to safely test the Start-Engine command-quoting mechanism using a harmless console exe like where.exe, checking that logs capture both stdout and stderr and that the working directory is respected. I'll set up a test with where.exe searching for "cmd" in a scratch directory, logging output to test-run.log.

I'll also verify the stderr path with a nonexistent target like "zzznotfound", which should print an error message and exit with status 1 -- running both cases to confirm the logging handles each correctly.

PowerShell$work = "C:\Users\a327e\AppData\Local\Temp\claude\launcher-test" function Start-...12:21:38Output not recorded
Input
$work = "C:\Users\a327e\AppData\Local\Temp\claude\launcher-test" function Start-...
Thinking12:21:46

The quoting handles the real-world case fine, with stdout and stderr both captured into one log. That third test failure was a false negative since I'd passed an invalid single-argument command to cmd, not representative of actual usage — I should verify the working directory a different way, maybe using powershell.exe to print it.

Actually, a cleaner approach: I'll drop a marker file into the scratch directory and run `where marker.txt`, since `where` searches the current directory first. That confirms the working-directory behavior without launching any actual game, matching the real invocation shape closely enough.

PowerShell$work = "C:\Users\a327e\AppData\Local\Temp\claude\launcher-test" Set-Content -Pa...12:21:53Output not recorded
Input
$work = "C:\Users\a327e\AppData\Local\Temp\claude\launcher-test" Set-Content -Pa...
Thinking12:21:56

With the working directory confirmed, I've got a list of remaining checks to run: AHK validation, a bash listing, git status, the shortcut check, and the README. I'll start with the AHK validation and bash listing now.

PowerShell& "C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe" /ErrorS...12:22:01Output not recorded
Input
& "C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe" /ErrorS...
Bashbash "C:/Users/a327e/Desktop/a327ex/Anchor/launcher/launch-game.sh"; echo "exit=...12:22:01Output not recorded
Input
bash "C:/Users/a327e/Desktop/a327ex/Anchor/launcher/launch-game.sh"; echo "exit=...
Thinking12:22:07

`/validate` works as expected—the good file gives no output while the broken one throws a syntax error. I realize `$LASTEXITCODE` is empty because AutoHotkey64.exe is a GUI-subsystem executable, so PowerShell doesn't wait on it; I need to use Start-Process -Wait -PassThru to capture the exit code properly.

The bash script lists the same 5 games too. Let me redo the AHK validation correctly this time.

PowerShell$ahk = "C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe" $o...12:22:16Output not recorded
Input
$ahk = "C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe" $o...
Thinking12:22:21

I'm setting exit codes for the script (0 for success, 2 for broken control) and confirming `/validate` is supported. Next I need to handle the startup shortcut, git status verification, and write the README covering what it is, the chords, and how the shared resolver and discovery rules work across both platforms.

Let me draft the README structure, covering Windows, Linux, timeline snapshots, replays, and adding a new root, then start writing it out.

WriteC:\Users\a327e\Desktop\a327ex\Anchor\launcher\README.md12:22:49Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\README.md
Thinking12:22:54

The Linux launch-game.sh header still points to a dead Windows path, so I'll update that comment to reference the sibling script correctly — the same fix applies to snapshot-game.sh, both comment-only edits that fit within the allowed scope.

The stale engine path under Git Bash is a behavioral concern the brief says not to touch, so I'll leave it as-is and flag it as a follow-up rather than change it now. Moving on to the Startup shortcut next.

EditC:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.sh12:22:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\launch-game.sh
EditC:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.sh12:22:56Output not recorded
Input
C:\Users\a327e\Desktop\a327ex\Anchor\launcher\snapshot-game.sh

Now the Startup shortcut and the remaining git checks.

PowerShell$startup = Join-Path $env:APPDATA 'Microsoft\Windows\Start Menu\Programs\Startup...12:23:08Output not recorded
Input
$startup = Join-Path $env:APPDATA 'Microsoft\Windows\Start Menu\Programs\Startup...
Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor" && echo "### git status --short launch...12:23:08Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor" && echo "### git status --short launch...
Thinking12:23:17

Everything looks clean, and state/ was never created by any run. I notice the launch-game.ps1 timestamp shows 12:19, matching when I edited it to insert Start-Engine, but the verification run that worked used the -Print paths which bypass Start-Engine entirely, so that function's behavior is still unverified.

I should run a syntax check to confirm the script parses correctly and Start-Engine is properly defined, then verify the file contents on disk reflect the edit — checking whether the file size matches what I'd expect after adding roughly 500 characters to the original.

Bashcd "/c/Users/a327e/Desktop/a327ex/Anchor/launcher" && grep -n 'Start-Engine' lau...12:23:18Output not recorded
Input
cd "/c/Users/a327e/Desktop/a327ex/Anchor/launcher" && grep -n 'Start-Engine' lau...
PowerShell$errs = $null; $toks = $null [System.Management.Automation.Language.Parser]::Par...12:23:18Output not recorded
Input
$errs = $null; $toks = $null [System.Management.Automation.Language.Parser]::Par...
Thinking12:23:33

Both parse cleanly. I'm sanity-checking the web packaging bash command construction without actually launching it, verifying the scripts path resolves and bash would find package-web-game.sh. Running the packaging for real would create files in the state/web directory as a side effect, which is documented behavior, but it also requires a real game to read from.

Given that full packaging is slow, may need emscripten tooling, and its downstream path eventually opens a browser—which I need to avoid—I'll instead verify just the shape of the bash command by echoing it and doing a harmless `ls` to confirm bash can cd into the scripts directory and see the script. I also want to confirm python exists for the http.server fallback, and test the .apr header parsing in isolation to make sure it produces sane width/height for the knightvspawns recording. Let me batch these final checks together.

PowerShell$Engine = [IO.Path]::GetFullPath('C:\Users\a327e\Desktop\a327ex\Anchor\launcher\...12:23:41Output not recorded
Input
$Engine = [IO.Path]::GetFullPath('C:\Users\a327e\Desktop\a327ex\Anchor\launcher\...
Thinking12:23:53

All prerequisites check out. I'm deciding whether to clean up the scratch test folder — since the verification output is already captured, cleaning it up is fine and keeps things tidy. I'm also considering the `-Print` flag's no-side-effect behavior.

I confirm from the code order that `-Print -Replay -Web` prints and exits before `Get-StateDir` runs, and `state/` stayed absent through all the print-only runs, which is correct. I still want to flag the Startup folder mechanism versus the elevated scheduled task convention, note the third roots.txt line decision, confirm the `last-run.log` cmd wrapper matches the brief's single-file requirement, and verify the engine path is derived dynamically rather than hardcoded. I also want to mention that `-Print zzzz` correctly returns no closest match since nothing subsequence-matches it, that `-Print rn` works as the suggestion example, and that discovery picked up `renderer` as a fifth game. The Linux state path in `~/.local/state/game-launch` stays as documented, and the ignore rule only needs to cover `state/`. Now I'll clean up the scratch folder and run a final git status.

Bashrm -rf "/c/Users/a327e/AppData/Local/Temp/claude/launcher-test" && echo "scratch...12:23:54Output not recorded
Input
rm -rf "/c/Users/a327e/AppData/Local/Temp/claude/launcher-test" && echo "scratch...
Thinking12:23:56

I'm ready to write the report now, making sure to flag the Startup-folder finding prominently as a key uncertainty since it affects whether the launcher works in elevated windows.

Done. Report follows.

1. Files written or changed

All under C:\Users\a327e\Desktop\a327ex\Anchor\launcher\:

  • launch-game.ps1 (new) — the Windows resolver. Reads roots.txt, discovers games (main.lua + anchor\, ≤2 levels, dot-folders skipped), matches exact→prefix→substring→unique subsequence, resolves the engine (game-local anchor.exe else shared build), launches the run.bat way, replays positionally, packages+serves web replays. Params: positional $Name, -Last, -Replay, -Web, -NoSnapshot, -Print.
  • game-launch.ahk (new) — the four Caps Lock chords, archived script with the header rewritten for the new home; SetCapsLockState("AlwaysOff") kept, A_ScriptDir lookup kept, rebinding/v1 notes kept.
  • snapshot-game.ps1 (new) — byte-identical copy of the archived file (cmp verified).
  • .gitignore (new) — one line, state/.
  • roots.txt (changed) — comment now describes both platforms; lines are ~/a327ex, ~/Desktop/a327ex, ~/Desktop/a327ex/Anchor/engine.
  • README.md (rewritten) — one document for both platforms: chords, how it works (roots, discovery, matching, engine, launching, replaying), Windows section (AutoHotkey, the Startup shortcut by full path, state/, -Print testing), Linux section (keyd, Hyprland, Omarchy pick, ~/.local/state/game-launch), timeline snapshots, replays, adding a root.
  • launch-game.sh / snapshot-game.sh (changed) — one comment line each; the stale tools/game-launch/...ps1 pointer now names the sibling file. No behaviour change.

Outside the launcher folder: %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\game-launch.lnk (new).

2. Decisions below the brief

  • A third roots line. The brief said add ~/Desktop/a327ex, but its own verification expects playground and replay-test, which sit at Anchor/engine/<name> — three levels below that root, so unreachable. I added ~/Desktop/a327ex/Anchor/engine as well (exactly the addition the old Linux roots.txt comment already suggested). Drop that line if you'd rather those two not be chord-launchable.
  • Engine path derived, not hardcoded. $Engine = <script dir>\..\engine\engine rather than the literal C:\Users\a327e\Desktop\..., so the launcher travels with the repo. Resolves to the path the brief named.
  • last-run.log is written through a cmd.exe wrapper. PowerShell 5.1's Start-Process refuses to point -RedirectStandardOutput and -RedirectStandardError at the same file, and the brief asked for one combined log. So the launch is cmd /c ""<exe>" "<arg>" > "<log>" 2>&1" with -NoNewWindow, which gives one file and no console window. -NoNewWindow sets CREATE_NO_WINDOW only, not SW_HIDE, and the engine calls SDL_ShowWindow explicitly (anchor.c:16784), so the game window is unaffected. I verified the quoting, the combined redirection and the working directory with where.exe (transcript below); I did not run the engine.
  • Show-Note replaces Show-Popup. Same auto-closing WScript popup on the real path, but it also echoes to the console (matching the bash notify), and under -Print it only writes to the console.
  • state\ is created lazily, only on a real launch/replay, so -Print leaves nothing on disk.
  • The bash packaging command redirects its own log. > '<log>' 2>&1 inside the bash -lc string rather than capturing in PowerShell, because & native 2>&1 under $ErrorActionPreference='Stop' in 5.1 turns stderr into a terminating error.
  • The .apr header read is 12 bytes, not the whole file (the archived script did ReadAllBytes).

3. Verification transcript

1. Listing

> powershell -NoProfile -ExecutionPolicy Bypass -File "...\launch-game.ps1"
Games found (5):
  knightvspawns
  playground
  renderer
  replay-test
  windrang
exit=0

renderer is the extra one: a327ex-site\renderer has main.lua + anchor\, so it is a game by the rule. A 4-level sweep also finds a327ex-site\renderer\games\{anchor3-playground,kimi-k3-playground,knightvspawns,skyland}, but those are at depth 3+ and correctly out of range.

2–7. -Print

> ... -Print knight
would launch: knightvspawns (C:\...\ABAgames\knightvspawns) via C:\...\ABAgames\knightvspawns\anchor.exe
exit=0

> ... -Print play
would launch: playground (C:\...\Anchor\engine\playground) via C:\...\Anchor\engine\engine\build\anchor.exe
exit=0

> ... -Print -Replay knight
would replay: knightvspawns (20260911-114233.apr) via C:\...\ABAgames\knightvspawns\anchor.exe
exit=0

> ... -Print -Replay -Web knight
would web-replay: knightvspawns (20260911-114233.apr) at http://localhost:8326/knightvspawns/index.html?replay=replays/20260911-114233.apr
exit=0

> ... -Print zzzz
Launch game: No game matching 'zzzz'.

Run the launcher with no name to see the full list.
exit=1

> ... -Print -Last
Launch game: No last game yet.

Shift+Caps Lock to pick one.
exit=1

> Test-Path "...\launcher\state"
False

zzzz has no subsequence match against any name, so it takes the fallback line rather than a Closest: list. To show the suggestion path actually fires:

> ... -Print rn
Launch game: No game matching 'rn'.

Closest:  playground,  renderer,  windrang
exit=1

No popup appeared in any of these, and state\ still did not exist afterwards.

8. Snapshot on a scratch copy (%TEMP%\claude\launcher-test\playground, a copy of Anchor\engine\playground with replays\ removed)

### run 1
timeline: snapshot 6e6fa21 (playground 2026-09-11 12:21:23)
exit=0
### run 2 (no changes)
timeline: no changes since last snapshot
exit=0
### git --git-dir=<scratch>\replays\.timeline log --oneline
6e6fa21 playground 2026-09-11 12:21:23
### tracked files
34

Scratch folder deleted afterwards. No real game's replays\.timeline was touched.

Launch mechanism (engine substituted with where.exe; no game was run)

cmdline: /c ""C:\Windows\System32\where.exe" "cmd" > "...\test-run.log" 2>&1"
--- log ---  C:\Windows\System32\cmd.exe                              (stdout captured)
cmdline: /c ""C:\Windows\System32\where.exe" "zzznotfound" > "...\test-run.log" 2>&1"
--- log ---  INFO: Could not find files for the given pattern(s).     (stderr captured, same file)
workdir = scratch     -> C:\...\launcher-test\marker.txt              (cwd honoured)
workdir = C:\Windows  -> INFO: Could not find files for the given pattern(s).

9. AutoHotkey /validate — supported by this build (2.0.27).

> Start-Process AutoHotkey64.exe /ErrorStdOut /validate "...\game-launch.ahk" -Wait -PassThru
exitcode = 0
stdout   = ''
> (control: a deliberately broken .ahk)
exitcode = 2

Parse check of both .ps1 files

launch-game.ps1 parse errors: 0
snapshot-game.ps1 parse errors: 0

10. Linux script after the roots.txt edit

> bash "C:/Users/a327e/Desktop/a327ex/Anchor/launcher/launch-game.sh"
Games found (5):
  knightvspawns
  playground
  renderer
  replay-test
  windrang
exit=0

Identical list to the PowerShell side.

11. Git

> cd Anchor && git status --short launcher
 M launcher/README.md
 M launcher/launch-game.sh
 M launcher/roots.txt
 M launcher/snapshot-game.sh
?? launcher/.gitignore
?? launcher/game-launch.ahk
?? launcher/launch-game.ps1
?? launcher/snapshot-game.ps1

> git check-ignore -v launcher/state/x
launcher/.gitignore:1:state/	launcher/state/x
exit=0

No pre-existing rule covered launcher/state/, and the repo has no root .gitignore, so I created launcher/.gitignore. Nothing committed, nothing staged.

12. The shortcut

Path       : C:\Users\a327e\AppData\Roaming\Microsoft\Windows\Start Menu\Programs\Startup\game-launch.lnk
TargetPath : C:\Users\a327e\AppData\Local\Programs\AutoHotkey\v2\AutoHotkey64.exe
Arguments  : "C:\Users\a327e\Desktop\a327ex\Anchor\launcher\game-launch.ahk"
WorkingDir : C:\Users\a327e\Desktop\a327ex\Anchor\launcher
Description: Caps Lock game launcher (Anchor/launcher/game-launch.ahk)

The AHK script was not started.

Web-replay prerequisites (checked, not executed)

shared exe        : True
scripts posix     : C:/Users/a327e/Desktop/a327ex/Anchor/engine/scripts
bash reaches it   : package-web-game.sh / OK / exit=0
.apr header parse : 480x270  (knightvspawns 20260911-114233.apr)
python            : C:\Users\a327e\AppData\Local\Programs\Python\Python313\python.exe
anchor.js/.wasm   : True / True

4. Uncertain or unverified

  • The Startup-folder premise in the brief is stale, and it matters. That folder currently holds only desktop.ini. abnt2-remap.lnk is not in it; it is in C:\Users\a327e\Desktop\tools\startup-tasks\startup-lnk-backup\ (moved there today, 12:08). C:\Users\a327e\Desktop\tools\startup-tasks\install-startup-tasks.ps1 moved both ShareX and abnt2-remap out of Startup into elevated logon scheduled tasks, for exactly this reason, quoting its own header: an unelevated process cannot hook or send input to an elevated window, so a Startup-launched AHK script silently does nothing while an admin terminal, Task Manager or regedit has focus. game-launch.ahk is the same kind of script, so Caps Lock will not fire over elevated windows. I created the Startup shortcut as the brief specified (it is authorized, reversible, and works everywhere else) rather than registering a scheduled task, which needs elevation and is outside the frame. I read the backed-up abnt2-remap.lnk for the shape and matched it.
  • No Caps Lock conflict with abnt2-remap.ahk (it only touches SC027, AltGr+C).
  • Nothing on the real launch path was executed. Start-Engine, the packaging call, the http.server start and the browser open were verified structurally (quoting, redirection, cwd, prerequisites) but never fired at the engine. Two things you will find out on first use: whether the cmd-wrapped launch feels right, and whether bash -lc picks up the profile the packaging script expects.
  • snapshot-game.ps1 is UTF-8 without a BOM and contains em-dashes in its comment header; PowerShell 5.1 reads it as Windows-1252, so those three characters render as mojibake in the comments. Functionally irrelevant (it parses clean, and the here-string content is ASCII), and the brief asked for a verbatim copy, so I left it.
  • launch-game.sh still hardcodes $HOME/a327ex/Anchor/engine/engine as its engine root, which does not exist on this machine. Listing works because that path is only used for the binary; running a game from Git Bash would fail. The brief said not to change its behaviour, so I did not.

5. Suggested follow-ups

  • Add a third job to install-startup-tasks.ps1 for game-launch and re-run it elevated, replacing the Startup shortcut with a RunLevel Highest logon task. Without it Caps Lock is dead over admin windows.
  • launch-game.sh's ENGINE could derive from SCRIPT_DIR the way the PowerShell one now does, which would make the bash half usable under Git Bash too.
  • renderer showing up in the game list is correct by the rule but probably not something you want to chord-launch. If it bothers you, the options are an exclude list in roots.txt or narrowing the roots.
  • The web/ folder left in the archived launcher directory was not carried over; the new one is created under state/ on first use.
Back to Horse Game 20

Settings