Bracket/brace filenames break all LSP features

[andriytyurnikov]

Description

Files with [, ], {, or } in their names (e.g., [id].rb, {slug}.rb) break multiple ruby-lsp features. These filename patterns are common in Rails/Next.js routing conventions.

When such a file exists in the workspace, all LSP features fail for it — indexing, go-to-definition, references, hover, completion, diagnostics, code lens, etc.

Root cause

Two independent issues compound:

1. URI encoding mismatch: URI::Generic.from_path includes [] in its safe character set, so brackets are left unescaped. However, VS Code's vscode-uri always encodes them ([ → %5B, ] → %5D). This mismatch means the server's URI for a file doesn't match the client's URI, so the document store can't find it.

2. Glob metacharacters in paths: Dir.glob treats [] as character classes and {} as alternation groups. When file or workspace paths containing these characters are interpolated into glob patterns (e.g., Dir.glob(File.join(workspace_path, "**/*.rb"))), the glob interprets them as pattern syntax rather than literal characters.

Reproduction

1. Create a file named [id].rb in a Ruby project
2. Open the project in VS Code with ruby-lsp
3. Open [id].rb — observe that no LSP features work (no syntax highlighting from semantic tokens, no go-to-definition, no hover, etc.)

Fix

PR #4022

[github-actions]

This issue is being marked as stale because there was no activity in the last 2 months

[andriytyurnikov]

Bookkeeping: test coverage, a harness bug, and related issues

Notes from rebasing #4022 onto current main and auditing its coverage.

1. Regression coverage added

The original PR changed 8 production files but only added URI unit tests — none of the Dir.glob hardening was covered. Added 10 targeted tests, each verified to fail when the fix is reverted (not just to pass with it):

Test	Covers
addon_test — project add-ons under [id]/{slug}/	addon.rb discovery (URI encode/decode and glob, in one path)
references_test — bracketed workspace	references.rb
rename_test — bracketed workspace	rename.rb
resolve_test_commands_test — real dir, unstubbed	test_style.rb test_dir expansion
go_to_relevant_file_test ×3 — impl→test, test→impl, bracketed parent dir	go_to_relevant_file.rb
completion_test — require_relative "[id]/"	glob escaping in completion.rb
store_test — client-encoded URI vs server-built URI	the actual #4023 / #3503 failure
test/fixtures/[id].rb	expectations_test_runner.rb

The existing test_dir tests stub Dir.glob, so they bypass the base: semantics being fixed; the new one hits the real filesystem.

Without the fix, the fixture makes every expectation suite fail to parse:

> 73 | def test_hover__[id]__does_not_raise
 | ^ unexpected '[', ignoring it

2. Pre-existing bug found in the expectation harness

Adding that fixture surfaced something unrelated to brackets. Seven expectation test helpers built their document URI as:

uri = URI("file://#{@_path}") # @_path is RELATIVE, e.g. "test/fixtures/foo.rb"

file://test/fixtures/foo.rb parses test as the URI authority, so path came out as /fixtures/foo.rb — the test/ segment was silently swallowed. Consequence: code_lens.rb:213 adds -Itest when the path matches **/test/**/*, and because the segment was missing, 5 committed code_lens expectations asserted commands without -Itest — the opposite of what a real user gets for a file under test/.

Switching to URI::Generic.from_path(path: File.join("/", @_path.to_s)) (machine-independent, correctly escaped) fixes it. The regenerated expectations differ only by the restored /test/ segment and the -Itest flag it enables — verified line-by-line across all 324 changed lines.

3. Still outstanding: URI parity is only partial

@vinistock's criterion in #3503 was that the URI must be escaped "the same exact way that the editor would escape it, otherwise we would end up with duplicate entries."

Checked against the real vscode-uri package rather than by reading its source — 5/17 paths match, 12 mismatch:

/some/path/[id].rb %5Bid%5D ok
/some/path/{slug}.rb %7Bslug%7D ok
/some/path/(id).rb %28id%29 vs (id) MISMATCH
/some/path/file (copy).rb %28copy%29 vs (copy) MISMATCH
a&b a,b a+b a=b a@b a;b it's a!b a*b a$b MISMATCH (10 more)

vscode-uri percent-encodes everything outside the unreserved set (A-Za-z0-9-._~ plus /). Narrowing the safe set to %r{[^\-_.~a-zA-Z\d/]} yields 17/17, keeps every round-trip green, preserves Windows drive handling, and additionally fixes URI::Generic.from_path(path: "/a/what?.rb") raising URI::InvalidComponentError.

Two caveats on "exact" parity: vscode-uri lowercases Windows drive letters (c%3A vs our C%3A), and RubyLsp.escape_glob_metacharacters doesn't escape ,, which breaks the one call site that interpolates into a brace group (require_relative "a,b" → glob returns []). Both are pre-existing and arguably separate.

4. Related issues

• Linux file system with square brackets in the directory/filename are failing to be indexed #3503 (open, bug/help-wanted/pinned) — the live issue; acceptance criterion above lives here
• Dir.glob in configuration.rb vulnerable to bracket/brace characters in workspace path #4024 (closed as stale) — deliberately deferred configuration.rb globs. The justification holds: Dir.glob("{}/**/*.rb", base: dir) escapes the base and walks the filesystem from / (reproduced on Ruby 4.0.1: Errno::EPERM on /Library/Application Support/com.apple.TCC). A bracketed workspace root still indexes 0 files, so that gap is real and worth reopening.
• Duplicate entries when receiving a textDocument/didOpen before indexing is complete #3639 (open, bug/pinned) — "Duplicate entries when receiving a textDocument/didOpen before indexing is complete". This is the same failure mode @vinistock described: the indexer writes under its own URI, the editor writes under the editor's, and mismatches produce two entries for one file. The remaining parity gap in §3 keeps a permanent instance of this alive for any path containing ( ) & , + = @ ; ' ! * $.
• textDocument/definition returns duplicate results with inconsistent file paths #3842 (closed as duplicate of Duplicate entries when receiving a textDocument/didOpen before indexing is complete #3639) — concrete user report of the symptom under Neovim: two identical definition results differing only by URI (// vs /). Different proximate cause, same category — the index keyed on non-canonical URIs.
• Improve GoToRelevantFile performance #3327 (closed) — GoToRelevantFile performance; Escape URIs like the editor does and harden Dir.glob against metacharacters in paths #4022 touches the same find_relevant_paths glob.

Fixes #4023 in #4022 points at an already-closed issue; #3503 is the one that matters for merge.