fix(init): detect Git Bash on Windows instead of assuming PowerShell

[timvw]

Fixes #112.

Problem

wt co printed wt navigating to: ... but the shell stayed put, and wt init from Git Bash installed a PowerShell script.

Two places decided the target shell from runtime.GOOS alone, so Windows always resolved to PowerShell:

• detectShell (cmd/init.go) — short-circuited to powershell before the $SHELL checks, so wt init under Git Bash wrote to Documents\WindowsPowerShell\Microsoft.PowerShell_profile.ps1. The bash session never loads that, so the wt wrapper function was never defined and there was nothing to perform the cd.
• shellenvTargetShell (cmd/shellenv.go) — had the identical short-circuit. This is why passing the shell explicitly did not help: the installed bash block was eval "$(wt shellenv)", and with no argument shellenv re-ran the same GOOS-only detection at every shell startup and emitted PowerShell into bash.

Changes

Detection — a shared isPOSIXShellEnv() gates the Windows→PowerShell default. Git Bash and MSYS2 export MSYSTEM; Cygwin does not, but all of them set $SHELL to a Unix shell path, which neither PowerShell nor cmd does. $SHELL is matched on its basename against a known set (sh, bash, zsh, fish, dash, ash, ksh, with an optional .exe) — a substring match would have read powershell as a POSIX shell, since it ends in shell. Explicit arguments still win, so wt init powershell from Git Bash works.

Config block — getShellConfigContent now writes the resolved shell into the block: eval "$(wt shellenv bash)", Invoke-Expression (& wt shellenv powershell). The installed line no longer depends on startup-time detection. Existing blocks are rewritten in place on the next wt init (marker handling is unchanged, and removal of old-format blocks still works).

script(1) fallback — Git for Windows does not ship script, so the bash wrapper would fail there even with detection fixed. Added a command -v script guard; without it, stdout is redirected to the log file and replayed with cat. Deliberately not tee: a pipeline makes $? tee's status, and PIPESTATUS/pipestatus are clobbered by the very next command — including the test needed to choose between bash's and zsh's spelling of it. I hit exactly that while writing this (bash returned 1 on success, zsh returned 0 on failure); the redirect form has no such trap. stderr is left alone so errors still stream live. Interactive promptui menus need a TTY and won't render in this mode — documented, and wt co <branch> works fine.

Path translation — cygpath -u is applied to the navigation target when available, so native C:\... paths become the POSIX form cd understands. No-op everywhere else, since cygpath only exists on MSYS/Cygwin.

Also refactored detectShell to take an injected goos, matching the pattern shellenvTargetShell already used, so the Windows branches are testable off-Windows.

Testing

• New table cases in TestDetectShell and TestShellenvTargetShell cover Git Bash via MSYSTEM, via Unix $SHELL, the .exe suffix, explicit-argument override, and the powershell/bash-named-directory false-positive guards. Env vars are set unconditionally via t.Setenv so an ambient MSYSTEM on a Windows runner cannot leak in.
• Verified the generated script parses under both bash -n and zsh -n.
• Exercised the no-script(1) branch against a stub binary on a stripped PATH in both bash and zsh: auto-cd works and exit codes propagate (confirming the PIPESTATUS fix).
• go test ./... and all 292 e2e scenarios pass. Two TestBashTabCompletion* tests fail on this machine, but they fail identically on unmodified main (macOS bash 3.2) — pre-existing and unrelated.

Not verified

I have no Windows machine, so the Git Bash path is reasoned from the code and tested by proxy (stubbed PATH, injected goos) rather than run for real. The three assumptions worth a second pair of eyes: that Git for Windows genuinely lacks script(1), that cygpath is on PATH in a Git Bash session, and that $SHELL/MSYSTEM are exported to the wt.exe child process. @celluj34 — if you're able to try this branch, that would confirm it.

[codecov]

Codecov Report

❌ Patch coverage is 42.50000% with 23 lines in your changes missing coverage. Please review.
✅ Project coverage is 32.32%. Comparing base (22d83db) to head (cc218f5).

Files with missing lines	Patch %	Lines
cmd/shellenv.go	4.34%	22 Missing ⚠️
cmd/init.go	94.11%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #113      +/-   ##
==========================================
+ Coverage   32.11%   32.32%   +0.20%     
==========================================
  Files          27       27              
  Lines        3036     3066      +30     
==========================================
+ Hits          975      991      +16     
- Misses       1975     1989      +14     
  Partials       86       86              

Files with missing lines	Coverage Δ
cmd/examples.go	67.74% <ø> (ø)
cmd/init.go	49.55% <94.11%> (+3.54%)	⬆️
cmd/shellenv.go	31.69% <4.34%> (-1.97%)	⬇️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.