fix: use path normalization to fix win32 issues

[pschiel]

Fixes #107

In reference to anomalyco/opencode#6763

Summary

This fix solves a whole series of bugs, resulting from filepath issues with backslashes occuring on Windows.

→ Solution: normalize all internal paths to forward slashes using centralized Filesystem wrappers

Why this works: all win32 shells work with forward slash paths, this is de-facto industry standard (git, vscode, cmake, ... all use this pattern)

Observed issues

• git rev-parse --show-toplevel returns E:/x/y forward slash format → worktree/directory mismatch
• path.resolve(), path.relative(), realpathSync.native(), realpath (bash) break with git bash paths (various issues)
• /d/x within C drive (/c/) results in things like C:\d\x
• relative paths not working like expected using cross-drive
• contains logic not working → issues with permission system and external_directory
• /tmp inside git bash is something else outside of it
• tools break using backslash paths (some backslashes get "eaten")
• escaping hell for agents → requires double/quadruple escaping
• observed in all win32 native shells (git bash, cmd, powershell/pwsh)
• few issues in app/desktop (wrong filepath splits/usage)
• bun hard crashing with segfault when spawn() is used with non-existing directory

Tested with fix

• local test suite succeeds now with 100% (bun test in packages/opencode)
• all possible path variants (C:\a\b, C:/a/b, /c/a/b, /cygdrive/c/a/b) are normalized into C:/a/b format
• full functionality of all tools (ENOENT errors gone, permissions, relative paths, external directory, LSP)
• worktree/sandbox handling correct
• tested TUI and desktop
• bun spawns via bash/pty not segfaulting

How to reproduce/test

• test suite
• running prompts with various problematic/mixed filepaths (see attached md example)

Notes

A few more places would require normalization - but they're unused/dead code (should be removed)
Desktop likely having more issues, especially bun.spawn() usage

[marius-kilocode]

@pschiel Can you confirm this was tested, how you tested it and how I can reproduce?

[pschiel]

@pschiel Can you confirm this was tested, how you tested it and how I can reproduce?

I fixed only issues that were not working, I reproduced all of the errors.

test suite (bun test in packages/opencode) in git bash, cmd and powershell:

your current dev branch:

with patch:

other than these, just plenty of usage since ~2 months. backslashes are the biggest problem, relative paths and cross-drive paths. i had a testing command that checked some of these (requires some structure/files setup).

when working in a windows env, it's inevitable that mixed paths come from anywhere, all of them should work:

test-windows-posix-paths.md

[pschiel]

if you require more info / steps to reproduce for a certain tool or so, let me know.

some others were
"run bash tool ls -l with cwd /foo" -> bun segfault (if dir not found, happens regularly)
"write invalid TS file to D:/bla.ts" (cross-drive) -> LSP shows no diagnostics (silent error)
external directory permissions ("accept all") are not stored

issue always same - string matches, contains() checks etc can only work if paths are internally normalized

[pschiel]

you might want to look packages\app\src\custom-elements.d.ts again.

this is a symlink in the repo, which doesn't work in windows, hence the change into a regular file with export. (I noticed a check failing due to this file in your CI)

[marius-kilocode]

Interesting that the windows runner tests are failing but passing locally for you

[pschiel]

Interesting that the windows runner tests are failing but passing locally for you

these are just e2e tests which do a few clicks and that's it - I don't think the unit+integration tests run in a win32 runner.

[pschiel]

Since last opencode merge, llm.test.ts fails due to BunProc install on Windows. Install takes much longer on Windows, and is not needed for the tests -> Solution: mock BunProc and Plugin 669fbf1

[pschiel]

e963daa fixes the cause of the test (linux) fails:

KILO_API_KEY and KILO_ORG_ID are not set in the branch pipeline, but required for tests:

-> no providers connect -> model list is empty -> model picker has no list items -> 4 model picker tests fail

added a check for these vars and skipping the 4 tests so the pipeline is green - if you fix the credentials issue they should run

[pschiel]

the other issue was that the two custom-elements.d.ts files still had type 120000 in git (symlink) but need to be 100644 (regular files)

reason for changing was: symlinks don't work reliably on windows, and vite will fail reading them. changing into regular files with an export works

[pschiel]

all unit/integration tests pass again

 1011 pass
 1 skip
 0 fail

[pschiel]

General Rebase/Conflict info

entire patch is 1 rule "normalize paths on win32 systems via Filesystem wrapper, where otherwise things will break"

• use Filesystem.normalize() around paths that are stored internally, used for matching, or passed to tools
• use Filesystem.join() instead if path.join() ONLY when segments are absolute paths and should be converted, otherwise use Filesystem.normalize(path.join(...))
• use Filesystem.relative() instead of path.relative()
• use Filesystem.resolve() instead of path.resolve()
• use Filesystem.dirname() instead of path.dirname()
• use Filesystem.realpath() instead of fs.realpath() or fs.realpathSync