bun test: report directories the scanner cannot read, scan each directory once

[robobun]

Problem

• bun test skips a directory it cannot open, prints nothing, and exits 0. The cause is let Ok(child_fd) = opened else { continue } in src/runtime/cli/test/Scanner.rs:207.
• The scanner follows directory symlinks and keeps no record of them. One link back into the tree gives Ran 2 tests across 902 files. Two links to the root never finish.

Fix

• A directory that fails to open or read is reported: error: could not scan "<path>" for tests plus the errno line. The run exits 1.
• Every directory the scanner opens is recorded as (st_dev, st_ino). A directory seen before is skipped.
• Verified: test/cli/test/bun-test.test.ts (two new cases, stock 1.4.3 fails both). The two-link trees print Ran 1 test across 1 file.

Background

• The scanner walks breadth-first from the root and opens each child directory.
• A path filter (bun test packages/v2/) is a substring match on the path the scanner recorded.
• The unreadable-directory test runs only as root on Linux. It needs runuser.

Downsides

• Regression, the reason for the draft state: the visited set keeps the first path to a directory. With latest -> packages/v2, bun test packages/v2/ exits 1 with The following filters did not match any test files. Main runs the test.
• Behavior change: a tree with an unreadable directory exits 1. That exit code is a policy choice for a maintainer.
• Cost: +1 fstat per scanned directory, counted from the code. Not measured.

Notes

Measured on head 3ce7646 merged with main 95690fc (debug build, Linux x64), against stock 1.4.3-canary.1+367d939d9:

tree	stock 1.4.3	this head
examples/one/lib -> ../.. and examples/two/lib -> ../..	killed at the 5 s test limit	Ran 1 test across 1 file. (3 of 3 runs)
a/loop -> .. and alias -> a	killed at the 5 s test limit	Ran 1 test across 1 file. (3 of 3 runs)
packages/a/vendor/b <-> packages/b/vendor/a	Ran 2 tests across 82 files.	Ran 2 tests across 2 files. (3 of 3 runs)
packages/v2/ and latest -> packages/v2, bun test packages/v2/	runs 1 test, exit 0	no match, exit 1
v2/ and links/latest -> ../v2, bun test links/latest/	runs 1 test, exit 0	no match, exit 1
directory and link as siblings, filter on each name (8 cases)	0 of 8 fail	4 of 8 fail: the name that sorts first wins

The last three rows are the regression. The scanner is breadth-first, so a link nearer the root is opened before the real directory and the real directory is skipped. The suites below do not cover it: none uses a path filter on a tree with a directory symlink.

Other designs were not weighed before this was written, for example a cycle check against the ancestors of the current directory (src/glob/GlobWalker.rs keeps followed_links for that). The rework starts there.

Probe on bun 1.4.3, as a non-root user, with ok/a.test.ts readable and locked/b.test.ts under a chmod 000 directory:

$ bun test; echo rc=$?
ok/a.test.ts:
(pass) a
 1 pass
 0 fail
Ran 1 test across 1 file.
rc=0

Symlink loop probe on 1.4.3: sub/a.test.ts, sub/deeper/b.test.ts, sub/loop -> .., sub/deeper/self -> . gives Ran 2 tests across 902 files. The file count and the wall time grow with the path limit.

A directory that is gone when it is opened (ENOENT, a dangling link) is skipped without a report.

Root errors other than ENOENT and ENOTDIR were already logged under BUN_DEBUG_jest. They are now also reported to the user and counted.

On Windows fstat on a directory handle fills st_dev from the volume serial number and st_ino from the NTFS file index (read from src/sys/lib.rs, not run on Windows). A filesystem that reports st_ino == 0 gets no deduplication. If fstat fails the directory is scanned as before.

Self-review: 5 concerns raised, 4 addressed (Windows compile of the errno conversion, st_ino == 0, ENOENT on a dangling junction, non-errno resolver errors). The fifth is the exit-code policy. The self-review says node --test, Jest and vitest skip an unreadable directory and exit 0 (not rerun by me), and it recommends a split: the symlink handling and the error report in one change, the exit code in its own change. It did not find the filter regression above.

Suites run with the debug build on the head merged with main: test/cli/test/bun-test.test.ts, pass-with-no-tests.test.ts, test-shard.test.ts, path-ignore-patterns.test.ts, test-changed.test.ts (164 pass, 6 todo, 0 fail). Commit 3ce7646 was pushed before a local run. CI passed on it afterwards.

---

no test proof · iteration 1 · platform-specific test(s) that do not run on this machine, deferring to CI, which covers all platforms: test/cli/test/bun-test.test.ts

[coderabbitai]

Warning

Review limit reached

• Run on-demand review

_{On-demand reviews are free for the next 15 days. After that, they cost $0.25 per reviewed file.}

Or wait 6 minutes for your next included review.

Check out review usage here.

View limit details

Limit details: You’ve used all 10 included reviews currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Essentials

Run ID: 6ce2092f-ab8e-4854-941b-c4beb9d358c7

📥 Commits

Reviewing files that changed from the base of the PR and between f42e980 and 3ce7646.

📒 Files selected for processing (3)

• src/runtime/cli/test/Scanner.rs
• src/runtime/cli/test_command.rs
• test/cli/test/bun-test.test.ts

---

_{Comment @coderabbitai help to get the list of available commands.}

[robobun]

^{Updated 12:13 AM PT - Sep 6th, 2026}

✅ @robobun, your commit 3ce764607a1987f2be8779d20835f970d9b39e78 passed in Build #110739! 🎉

---

🧪   To try this PR locally:

bunx bun-pr 41465

That installs a local version of the PR into your bun-41465 executable, so you can run:

bun-41465 --bun

[claude]

Findings marked 🟡 are optional suggestions and need no follow-up push.

[claude]

I re-reviewed after the latest push — all five of my earlier inline findings are addressed (from_code_int for the Windows errno type, st_ino == 0 falls through without dedup, ENOENT is skipped in report_unreadable_dir, non-Sys resolver errors are now reported, and the /etc/passwd read is guarded). No new issues found. A maintainer should still weigh in on the exit-code policy the PR flags (fail vs. warn on unreadable dirs), and note that both new tests skip on Windows so the (st_dev, st_ino) dedup and error-reporting paths there rely on rust:check-all only.

What was reviewed

• report_unreadable_dir / report_dir_read_error: ENOENT early-return, non-Sys fallthrough, from_code_int cross-platform (Windows SystemErrno is #[repr(u16)], E::ENOENT alias exists at windows_errno.rs:258).
• mark_visited: st_ino != 0 guard; root seed via fstatat inserting a single (dev, 0) on FAT is harmless since children go through the guarded path.
• test_command.rs: unreadable_dirs is captured before drop(scanner) and gates exit code alongside existing failure conditions.
• Tests: hasNobodyUser() try/catch, finally restores chmod, output asserted before exit code.

Extended reasoning...

Overview

The PR adds two behaviors to the bun test file-discovery scanner: (1) (st_dev, st_ino) tracking to avoid re-scanning directories reached via symlink loops, and (2) reporting + nonzero exit when a directory cannot be opened or read. Changes are in src/runtime/cli/test/Scanner.rs (~60 lines added), a small hook in src/runtime/cli/test_command.rs to surface the count and set exit code 1, and two new tests appended to the existing scanner block in test/cli/test/bun-test.test.ts.

Security risks

None identified. This is local filesystem traversal for test discovery; no network, auth, or untrusted-input parsing is involved. The added fstat/fstatat calls use existing bun_sys wrappers and their errors degrade to "scan anyway".

Level of scrutiny

Moderate. My prior review raised five findings (three blocking, two optional); commits 3fc9c3b and 59cea25 address all of them, verified against the current diff and the underlying types (bun_resolver::Error is Copy, Windows SystemErrno is #[repr(u16)] so the as c_int cast is valid, and E::ENOENT is an associated const on Windows so get_errno() == E::ENOENT holds cross-platform). I did not find new issues on re-read.

Other factors

Two things keep this from an approve: the PR author explicitly leaves the exit-code policy (fail vs. warn on unreadable directories, where Node/Jest/vitest exit 0) for a maintainer to decide, and both new tests are gated off Windows — so the Windows-specific concerns I raised earlier (dangling junctions, FAT/exFAT st_ino == 0) are addressed in code but have no test coverage on that platform. No CODEOWNERS entry matches these paths.

[claude]

Code review found no issues

No high-confidence issues detected in this change.

[robobun]

Status: draft. The symlink half of this PR has a regression. Do not merge head 3ce7646.

Measured on this head merged with main 95690fc (debug build, Linux x64), against stock 1.4.3-canary.1+367d939d9.

The head does end the walk that never finishes:

tree	stock 1.4.3	this head
examples/one/lib -> ../.. and examples/two/lib -> ../..	killed at the 5 s test limit	Ran 1 test across 1 file. (3 of 3 runs)
a/loop -> .. and alias -> a	killed at the 5 s test limit	Ran 1 test across 1 file. (3 of 3 runs)
packages/a/vendor/b <-> packages/b/vendor/a	Ran 2 tests across 82 files.	Ran 2 tests across 2 files. (3 of 3 runs)

The regression: the visited set keeps only the first path that reaches a directory. A path filter that names another path to the same directory matches nothing.

mkdir -p r/packages/v2 && cd r
echo 'import {test,expect} from "bun:test"; test("s",()=>expect(1).toBe(1));' > packages/v2/t.test.js
ln -s packages/v2 latest
bun test packages/v2/

• Stock 1.4.3: Ran 1 test across 1 file., exit 0.
• This head: The following filters did not match any test files: packages/v2/, exit 1.

The scanner is breadth-first. It opens latest (depth 1) before packages/v2 (depth 2), so it records the file as latest/t.test.js and skips the real directory. With the link deeper than its target (v2/ and links/latest -> ../v2), the filter links/latest/ fails and v2/ works. For a directory and a link that are siblings, the name that sorts first wins (4 of 8 probe cases fail on this head, 0 of 8 on stock).

The suites I ran do not cover this. bun-test.test.ts, pass-with-no-tests, test-shard, path-ignore-patterns and test-changed pass on this head (164 pass, 6 todo, 0 fail). None of them uses a path filter on a tree with a directory symlink.

The unreadable-directory half is not affected.

Next: rework the symlink handling so that a filter matches every path it matches on main, with these trees as tests.