Codex CLI on Windows: neo-tree markers, path handling, and shell-create detection

[Cannon07]

Summary

Makes the Codex CLI backend work correctly on Windows (validated end-to-end with codex-cli 0.137, GPT 5.5 and GPT 5.3). The integration was already wired for Windows; this PR fixes the Windows-specific path/marker bugs that surfaced under real use, plus a couple of related quality/consistency fixes. Claude Code on Windows is unaffected, and Unix behaviour is preserved (see Unix safety below).

What was broken & fixed

1. neo-tree markers silently dropped (path separators)

Codex routes edits through apply_patch with relative paths; apply/patch.lua joined them as cwd .. "/" .. relpath, and on Windows cwd is backslashed → a mixed-separator key (D:\proj\sub/file.txt). neo-tree keys nodes by native backslash paths, so the lookup missed and no created/modified/deleted marker rendered (the diff still opened). Claude Code was immune (it sends native paths).

• changes.lua — normalize() folds separators to the OS-native one on Windows (central key canonicalization).
• neo_tree.lua — the ancestor-deletion check used a hardcoded "/"; now uses the native separator.

2. Existing file shown as "new file" (absolute-path doubling — open)

For some projects Codex emits absolute paths in apply_patch directives (*** Update File: D:\proj\sub\file). resolve_path only treated a leading / as absolute, so a Windows-absolute path was doubled onto cwd (D:\proj/D:\proj\sub\file). That bogus path fs_stats as missing → file mis-marked created, diff opened at the wrong position (rendered as a new file), and a junk neo-tree node appeared.

• apply/patch.lua — resolve_path now recognizes Windows drive-letter (X:\/X:/) and UNC (\) absolutes and passes them through.

3. PostToolUse didn't clear the diff/marker after accept (open/close parity)

post_tool.lua had its own copy of the path-join with the same Unix-only absolute check, so the close path was still doubled while the open path was fixed → close_for_file: no active diff … skipping → marker/diff never cleared.

• post_tool.lua — now reuses apply.patch.resolve_path (exported), so open and close resolve identically and can't diverge again.

4. Diff-tab label showed the absolute path

On Windows the cwd-prefix strip compared D:\proj\… against D:\proj/ (forward slash) and never matched, so the tab title fell back to the full absolute path. Additionally, apply_patch labelled the tab with file.rel_path — the literal directive path — which GPT 5.3 writes as absolute.

• pre_tool/init.lua — extracted display_path() with a separator-insensitive cwd-prefix test (relative result keeps native separators); used for Edit/Write and apply_patch, so the label is cwd-relative regardless of OS or what the model emits.

5. New-Item file creation not detected (shell-create)

Codex/GPT creates files on Windows via PowerShell New-Item -ItemType File, which the shell-detect grammar didn't recognize → no neo-tree mark.

• shell_detect.lua — added new-item/ni as a create category routed through the existing write path (a brand-new file becomes bash_created, like Out-File to a new file). -ItemType Directory is deliberately skipped (parity with mkdir and the locked New-Item -ItemType Directory … | Out-Null → {} spec row); POSIX touch is left untouched.

6. Healthcheck no longer steers Codex users away

• health.lua — removed the stale Windows-only warning "Codex CLI on Windows is not yet validated … use Claude Code on Windows" now that Codex is supported on Windows. (Copilot-on-Windows remains correctly deferred.)

7. Housekeeping

• .gitignore — ignore .codex/ (local hook install written by :CodePreviewInstallCodexCliHooks), matching the existing opencode/copilot entries.

Unix safety

Windows path transforms are either gated on package.config:sub(1,1) == "\\" (changes.normalize, display_path) or use the native separator package.config:sub(1,1) (neo_tree), so the Unix paths are byte-identical to before. The resolve_path absolute-detection and the New-Item grammar are not OS-gated but only match Windows-shaped input (X:\ paths / a PowerShell cmdlet) that doesn't occur in real Unix workflows. No Unix workflow changes.

Tests

Regression specs added/extended:

• changes_registry_spec — mixed-separator key canonicalization (Windows-gated)
• apply_patch_spec — relative / Unix-absolute / Windows-absolute / UNC resolution (cross-platform)
• post_tool_handle_spec — Windows-absolute patch path not doubled on close
• display_path_spec — separator-insensitive label + an integration test driving pre_tool.handle with an absolute-path codex patch
• pre_tool_shell_detect_spec — New-Item file create (incl. the exact ; … | Out-Null two-file command), ni alias, and directory-create still {}

Full plugin spec suite passes on Windows (per-file; PlenaryBustedDirectory hangs headless on Windows). diff_lifecycle_spec fails 13/14 headless on Windows on origin/main too — a pre-existing UI limitation, not introduced here.

Validated end-to-end on Windows 11 with codex-cli 0.137 (GPT 5.5 / 5.3) and Claude Code: apply_patch (relative & absolute) and PowerShell shell ops mark correctly, open the diff at the right position with a relative tab label, and clear on accept.

Relates to #46.