Background
PR #79 shipped two related changes to desktop/src/main/prerequisite-installer.ts:
- EINVAL-safe
runCommand wrapper that auto-flips shell: true for .cmd/.bat paths on Windows (works around Node 18.20.2+ CVE-2024-27980 mitigation)
installClaude rewrite: replaced npm install -g @anthropic-ai/claude-code with Anthropic's native installer (claude.ai/install.{ps1,sh})
Validated before merge: typecheck clean, 881 tests pass, build produces signed installer, bootstrap scripts pre-flight read confirmed non-interactive + no admin + clean exit codes.
Not validated: end-to-end first-run flow on a clean OS. We attempted a VirtualBox + Win11 ISO test rig but got blocked by VBox's Hyper-V Platform fallback on the dev machine (5–10x slowdown, install hung at 10% for hours). Documented in docs/superpowers/investigations/2026-04-29-vbox-hyperv-conflict.md.
Filing this so the verification gaps don't get forgotten.
What needs verifying
Critical (real-user repro path)
Important (other supported platforms)
Edge cases
Nice to have
How to actually run these
See docs/local-dev-vm.md in the workspace repo for the VirtualBox-based test rig (works on machines without an active hypervisor — read the pre-flight section first). For real hardware tests, just run a debug installer from a recent CI build of master against the target machine.
Closing criteria
Critical items checked. Important items: at least one of the three "clean" platforms verified per OS family. Edge cases tracked but not strictly required for closing.
Background
PR #79 shipped two related changes to
desktop/src/main/prerequisite-installer.ts:runCommandwrapper that auto-flipsshell: truefor.cmd/.batpaths on Windows (works around Node 18.20.2+ CVE-2024-27980 mitigation)installClauderewrite: replacednpm install -g @anthropic-ai/claude-codewith Anthropic's native installer (claude.ai/install.{ps1,sh})Validated before merge: typecheck clean, 881 tests pass, build produces signed installer, bootstrap scripts pre-flight read confirmed non-interactive + no admin + clean exit codes.
Not validated: end-to-end first-run flow on a clean OS. We attempted a VirtualBox + Win11 ISO test rig but got blocked by VBox's Hyper-V Platform fallback on the dev machine (5–10x slowdown, install hung at 10% for hours). Documented in
docs/superpowers/investigations/2026-04-29-vbox-hyperv-conflict.md.Filing this so the verification gaps don't get forgotten.
What needs verifying
Critical (real-user repro path)
spawn EINVAL. Native installer dropsclaude.exe,detectClaudefinds it post-install, OAuth flow completes. (This is the literal failure mode the PR was written to fix — a friend's machine hit the original bug here.)claude.cmdfrom before this change. Existing users upgrading shouldn't regress. TherunCommandwrapper should makedetectClaude/detectAuth/submitApiKeywork on the lingering.cmdshim too.Important (other supported platforms)
bash -c "curl -fsSL https://claude.ai/install.sh | bash"runs the bootstrap,claudeends up on PATH, OAuth completes.apt.Edge cases
claudedoesn't end up on the running Electron process's PATH, the user should see the "Quit and reopen YouCoded — the new PATH entry will be picked up on next launch" message rather than a generic failure. Easiest repro: run on a fresh Windows account where the user PATH isn't yet read into the Electron process.ExecutionPolicy AllSignedorRestrictedset at machine/group-policy level. The-ExecutionPolicy Bypassflag should override at process scope, but worth verifying.claude.aiordownloads.claude.ai. Should fail with a clear error, not silently hang.detectClaudereports installed) shouldn't break anything in the Detect → Install → Auth state machine.Nice to have
~/.claude/bin/claude.cmdfrom a prior npm install. What doeswhich clauderesolve to? Does the native installer add itself ahead of or behind the npm shim on PATH?How to actually run these
See
docs/local-dev-vm.mdin the workspace repo for the VirtualBox-based test rig (works on machines without an active hypervisor — read the pre-flight section first). For real hardware tests, just run a debug installer from a recent CI build of master against the target machine.Closing criteria
Critical items checked. Important items: at least one of the three "clean" platforms verified per OS family. Edge cases tracked but not strictly required for closing.