test: keep --watch alive when --bail stops a run

[robobun]

What

bun test --watch --bail exits the whole watch process on the first bail-out. The edit that would fix the failing test never re-runs. Without --bail the same watcher survives and reruns on the next edit.

Repro

mkdir d && cd d
cat > b.test.ts <<'EOF'
import { test, expect } from "bun:test";
import { v } from "./v";
test("t1", () => { expect(v).toBe(1); });
test("t2", () => { expect(2).toBe(2); });
EOF
echo 'export const v = 0;' > v.ts
echo '{}' > package.json
bun test --watch --bail
# => prints 'Bailed out after 1 failure' and exits 1 instead of watching

Cause

CommandLineReporter::handle_test_completed calls Global::exit(1) directly when the fail count reaches the bail threshold, inside the per-test completion callback. Under --watch that kills the process before control reaches run_event_loop_for_watch. The module-load-rejection bail path in TestCommand::run has the same hard exit.

The parallel coordinator already handles this correctly: it sets a bailed flag and drains, returning normally.

Fix

Give the serial runner the same flag-based bail:

• Execution.aborted stops the step loops (step_group, step_group_one, step's per-group loop) so no further tests in the current file run.
• CommandLineReporter.bailed stops the per-file loop in run_all_tests and the --rerun-each repeat loop in TestCommand::run.

Under --watch the bail path now prints the banner, sets the flags, and lets the run unwind back to exec, which prints the full summary and enters run_event_loop_for_watch. When the watched file changes, reload_process re-execs as before. When not watching, the existing Global::exit(1) is unchanged.

Verification

# fail-before (released bun): process exits, never sees the fixing edit
USE_SYSTEM_BUN=1 bun test test/cli/test/bun-test.test.ts -t 'keeps watching after a bail'
# pass-after
bun bd test test/cli/test/bun-test.test.ts -t bail

The test spawns bun test --watch --bail with a failing test, waits for the bail banner and the end-of-run summary on stderr, then fixes the dep file and waits for 2 pass from the re-run. Skipped on Windows where --watch uses a respawning parent instead of exec-in-place (same as the existing --changed --watch test).

---

no test proof · iteration 2 · 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

Supersedes #19918 (same approach against the pre-Rust Zig sources, which no longer exist).

Fixes #20318
Fixes #6453

[coderabbitai]

Walkthrough

Changes

Watch-mode bail handling

Layer / File(s)	Summary
Execution abort propagation src/runtime/test_runner/Execution.rs	Adds an aborted state and stops top-level, group, and concurrent test sequencing when set.
Watch-aware reporter bail handling src/runtime/cli/test_command.rs	Tracks bail state, aborts the current execution, exits in non-watch mode, and unwinds for watch-mode reruns.
Watch and bail integration test test/cli/test/bun-test.test.ts	Validates that --watch --bail stops the failed run and reruns successfully after a file fix.

Possibly related PRs

• oven-sh/bun#34443: Both changes modify test-runner early-exit and cleanup behavior.

Suggested reviewers: jarred-sumner

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The changes implement watch-aware bail handling and a regression test, matching the linked issues' requirements.
Out of Scope Changes check	✅ Passed	The patch stays focused on watch-and-bail behavior and the related regression test, with no obvious unrelated changes.
Title check	✅ Passed	The title clearly summarizes the main change: keeping watch mode alive when bail stops a test run.
Description check	✅ Passed	The description covers what changed and how it was verified, though it uses custom headings instead of the template's exact section names.

---

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

[robobun]

^{Updated 5:18 AM PT - Jul 22nd, 2026}

❌ @robobun, your commit e16c003 has 2 failures in Build #77648 (All Failures):

• test/js/node/test/parallel/test-gc-http-client-connaborted.js - timeout on 🐧 13 x64
• test/js/node/test/parallel/test-net-connect-memleak.js - code 1 on 🐧 25.04 x64
• test/js/node/test/parallel/test-net-connect-memleak.js - code 1 on 🐧 13 x64

---

🧪   To try this PR locally:

bunx bun-pr 35094

That installs a local version of the PR into your bun-35094 executable, so you can run:

bun-35094 --bun

[github-actions]

Found 2 issues this PR may fix:

1. bun test: --watch and --bail should work together #20318 - Requests --watch and --bail to work together; this PR prevents bail from killing the watch process
2. Allow bun test to support both the --watch and --bail flags at the same time. #6453 - Duplicate of bun test: --watch and --bail should work together #20318, same request for --watch and --bail compatibility

If this is helpful, copy the block below into the PR description to auto-close these issues on merge.

Fixes #20318
Fixes #6453

🤖 Generated with Claude Code

[github-actions]

This PR may be a duplicate of:

1. fix(test): bail correctly halts current run in watch mode #19918 - Same fix for --bail exiting the process instead of keeping the watcher alive in bun test --watch --bail, but targets the old Zig sources (pre-Rust rewrite)

🤖 Generated with Claude Code

[claude]

All three prior nits are now addressed (the fixture tests were renamed to "test #1"/"test #2", and the two bail sites now share the hoisted let watching = ... shape with summary → banner ordering preserved for non-watch). No new issues found. Deferring to a human for the control-flow change itself — flag-based unwind through Execution.aborted and CommandLineReporter.bailed across three step loops, the per-file loop, and the --rerun-each loop.

What was reviewed:

• Non-watch --bail path is byte-for-byte behavior-preserving at both sites (handle_test_completed and the module-load-rejection arm in TestCommand::run).
• aborted guards in step_group, step_group_one, and step's per-group loop return Complete before starting the next sequence — same effective skip set as the old Global::exit(1).
• New test polls stderr for observable conditions (no sleeps), asserts test #2 never ran, then edits the dep and awaits 2 pass from the re-run; await using proc covers cleanup.
• The two unrelated CI failures (test-gc-http-client-connaborted.js, test-net-connect-memleak.js) are pre-existing flakes on x64 Linux, not touched by this diff.

Extended reasoning...

Overview

The PR changes bun test --watch --bail so that hitting the bail threshold no longer calls Global::exit(1) and kills the watcher. Instead it sets two flags — CommandLineReporter.bailed (stops the per-file loop in run_all_tests and the --rerun-each repeat loop) and Execution.aborted (stops step_group, step_group_one, and the per-group loop in Execution::step) — so the run unwinds back to exec, which prints the summary and enters run_event_loop_for_watch. The non-watch path is unchanged: it still prints summary → banner → junit → Global::exit(1). The module-load-rejection bail arm in TestCommand::run gets the same treatment. A new test in test/cli/test/bun-test.test.ts spawns bun test --watch --bail, waits for the bail banner and the end-of-run summary, asserts the second test never ran, edits the failing dep, and waits for 2 pass from the re-run.

Security risks

None. This is test-runner control flow; no untrusted input, no auth/crypto/permissions surface.

Level of scrutiny

Moderate. The change is small (~60 lines of Rust across two files, plus a test) and follows the same flag-and-drain pattern the parallel coordinator already uses, but it threads a new early-exit through five loop sites in the serial runner. Getting the unwind order wrong could leave the watcher in an odd state or change what runs after bail. I traced each guard: aborted is set inside handle_test_completed (called from on_sequence_completed inside advance_sequence), and every loop that could start a next sequence/group checks it at the top before doing so, so no additional test bodies run — matching the old hard-exit's effective behavior. Under --watch the next file-change triggers reload_process (exec-in-place), so any skipped afterAll groups on the bailed run don't leak into the next run. That all looks correct, but it's the kind of multi-site control-flow change a maintainer should eyeball.

Other factors

I left three nits on earlier revisions (non-watch output-order regression at the module-load bail site, duplicated banner block, and a ~0.13% flake from not.toContain("t2") colliding with the mkdtemp suffix). All three are fixed in the current diff — the two bail sites now bind watching once and print summary before the banner when not watching, and the fixture tests are named "test #1"/"test #2" with not.toContain("test #2"), matching the sibling --bail tests. The 30s per-test timeout was flagged and ruled out (it's a spawned --watch subprocess awaiting a filesystem-triggered rerun; a deadline is appropriate and await using proc guarantees kill on failure). The two CI failures are unrelated Node-parallel GC/memleak flakes on Linux x64. Given this is a real control-flow change to the test runner rather than a mechanical edit, I'm deferring rather than approving.

[robobun]

Diff is green. The only red CI lanes across the last three builds (77612, 77626, 77648) are two unrelated Node.js GC-timing tests on Linux x64 that don't go through the test runner at all:

• test/js/node/test/parallel/test-net-connect-memleak.js
• test/js/node/test/parallel/test-gc-http-client-connaborted.js

Both are bun <file> style (not bun test), so test_command.rs / Execution.rs aren't in their path. They fail the same way on all three builds regardless of the commit here.

The new --bail with --watch test passes on every completed lane. All review threads addressed. Ready for a maintainer.

[claude]

Beyond the inline nit, I traced a few things and they check out: the non-watch --bail path is behavior-identical (same summary→banner→Global::exit(1) order after f2d9e07); Execution.aborted's StepResult::Complete reaches Phase::Done via the same _advance path as natural completion; and neither bailed nor aborted leak across watch reruns (watch reloads via re-exec, and each file's Execution::init() starts aborted: false regardless).

Extended reasoning...

Verified the three earlier review comments were addressed in f2d9e07 and e16c003. Traced the new aborted checks in step_group/step_group_one/step — they're read-only early-returns at loop heads that funnel into the existing StepResult::Complete → _advance completion path, and the fresh .get() reborrow in step_group_one matches the file's existing stacked-borrows pattern. The run_all_tests per-file check and the --rerun-each break correctly unwind to exec, which prints the summary at line 2878 and enters run_event_loop_for_watch. The two CI failures (test-gc-http-client-connaborted.js, test-net-connect-memleak.js) are unrelated to this change.

[robobun]

Closing as part of a cleanup of stale pull requests. This PR has had no new commits since 2026-07-22, it conflicts with main, and its last CI run failed. This is not a judgment on the fix itself. The linked issue (#6453) stays open. If the problem still reproduces on a current build, reopen this PR after a rebase or open a new one against main.