GitHub Actions annotation and code frame caret: use the first frame that has a file, not a builtin frame on top

[robobun]

Problem

• With GITHUB_ACTIONS=true, an error thrown from inside a JS builtin is annotated with an empty file and the builtin's internal position. [].reduce((a, b) => a) in x.js prints ::error file=,line=1,col=11,title=TypeError: reduce of empty array with no initial value:: (release; line=12,col=9 on a debug build) instead of file=x.js,line=1,col=4.
• Same cause, other shapes of the top frame: file=native,line=1,col=11 once error.stack has been read (the frames are then parsed back out of the string, where the builtin is printed as native), file=unknown,line=1,col=1 for an error created below a native frame after error.stack was read, and file=node%3Aevents,line=59,col=31 for an error created inside one of bun's own modules (emitter.emit("error") with no listener). None of these name a file in the repository, so GitHub shows the annotation in the run summary only.
• The caret under the code frame has the same bug: for [].reduce((a, b) => a) the source line shown is the user's, but the ^ is placed at the builtin's column. Same for the native / unknown cases; node: frames were already skipped there.
• Cause: print_github_annotation (src/jsc/VirtualMachine.rs) takes frames[0] and only checks that it has a position, and the caret code in print_error_instance_body skips only bun: / node: frames. The code that picks which file to show in the code frame (remap_zig_exception) already skips empty, native, unknown, [unknown], [source:...], bun: and node: frames, so the three consumers disagreed about which frame the error is attributed to. Reproduces on 1.4.0 and on main.

Fix

• The skip list from remap_zig_exception becomes ZigStackFrame::has_user_source(); remap_zig_exception uses it unchanged, and the annotation and the caret now use it too, so all three point at the same frame: the first one whose source_url names a source the user can open. The annotation additionally requires a position, as before; if no frame qualifies it is emitted as ::error title=... without a location, which is what it already did for a positionless top frame.
• Why this is correct: the source lines shown in the code frame come from the frame remap_zig_exception picks, so the caret has to be placed by that frame's column or it points at nothing in particular; and file= is only useful to GitHub for a file that exists in the checkout, which none of the skipped source URLs are. The frames below the skipped ones are already source-mapped (the per-frame loop at the end of remap_zig_exception), which is why the TypeScript test files below report their original columns. The annotation is not gated on BUN_SHOW_BUN_STACKFRAMES like the code frame is: that flag is about listing bun's internal frames, and an empty file= is wrong with or without it.
• The annotation's frame list is unchanged; it still prints every frame, including the builtin one.
• Tests:
  • test/cli/test/bun-test.test.ts, "support for Github Actions": four rows for a builtin / node: module top frame, each fresh and after error.stack was read (covering the empty, native, node: and unknown source URLs), asserting the exact file=<test file>,line=,col= of the call in the test file; plus a stack with no file in any frame, asserting ::error title= and that the frame list is still printed.
  • test/js/bun/test/stack.test.ts: Bun.inspect of a reduce TypeError, fresh and after .stack was read, asserting the caret is under reduce on the user's line.
  • USE_SYSTEM_BUN=1 bun test on 1.4.0: all 7 new tests fail (file= / file=native / file=node%3Aevents / file=unknown, caret at the builtin's column). bun bd test with src/ stashed: the same failures, with the debug build's builtin position (line=12,col=9). With the change: bun-test.test.ts 86 pass / 6 todo, stack.test.ts 8 pass / 1 todo.
  • Also green on the debug build: test/js/bun/util/inspect.test.js, reportError.test.ts, test/js/node/v8/capture-stack-trace.test.js, test/regression/issue/23022-stack-trace-iterator.test.ts, test/js/bun/sourcemap/, test/js/bun/console/, test/js/node/vm/vm.test.ts, vm-sourceUrl.test.ts, test/js/bun/test/bun_test.test.ts. (inspect-error.test.js has two minified-file snapshots that already fail on main under a debug build because of the debug-only at require (51:24) frame; unchanged by this PR, error printer: keep parsing error.stack past frames without a function name #38308 and error printer: don't invent a 1:1 position for frames parsed out of error.stack without one #38328 fix the helper.)
• Related open PRs: error printer: don't invent a 1:1 position for frames parsed out of error.stack without one #38328 changes the same print_github_annotation block to key the location on the line being valid and print col= only when the column is; it fixes frames parsed without any position, while this PR is about frames that have a position but no file (reduce (native:1:11) still picks the builtin frame there). The two compose (the find here would use its line check); whichever lands second has a few-line rebase in that block. Bun.inspect: label builtin stack frames (native) instead of bare line:col #35177 only changes how the builtin frame is rendered in the list, not which frame the location comes from.

Background

• GitHub Actions workflow command: a line ::error file=F,line=L,col=C,title=T::body on a job's output. The runner attaches it to file F in the checkout when it exists; otherwise it is listed without a file. Bun prints one per uncaught error / failed test when GITHUB_ACTIONS is set, from print_github_annotation, which runs at the end of the error printer (print_error_instance_body) on the same remapped ZigException the terminal output was printed from.
• ZigStackFrame.source_url for the frames of an error: an absolute path or URL for user code; empty for JSC's JS-implemented builtins (Array.prototype.reduce and friends are JS inside JSC, so unlike C++ natives they have a line:column, into the builtin's own source, but no URL); node:<name> / bun:<name> for bun's built-in modules; native / unknown when the frames were rebuilt by parsing error.stack (that string prints builtins as native and positionless native frames as unknown, and the parser gives the latter position 1:1); [unknown] / [source:N] are placeholders from the prepareStackTrace call-site path.
• Code frame: the numbered source lines, caret and name: message printed above the stack. remap_zig_exception picks the frame whose file to read (skipping the URLs above), source-maps it and reads the lines; print_error_instance_body later prints them and places the caret from a frame it picks itself, which is the second of the two loops unified here.
• Structured frames vs. parsed frames: JSC keeps an error's captured frames until error.stack is first read, then formats the string and drops them. After that, bun's printer rebuilds frames by parsing the string (fromErrorInstance in ZigException.cpp), which is why "after error.stack was read" is a separate shape in the tests.

Before / after on the probes

$ printf '[].reduce((a, b) => a);\n' > reduce.js && GITHUB_ACTIONS=true bun reduce.js

# before (1.4.0)
1 | [].reduce((a, b) => a);
              ^
TypeError: reduce of empty array with no initial value
      at reduce (1:11)
      at /tmp/reduce.js:1:4
::error file=,line=1,col=11,title=TypeError: reduce of empty array with no initial value::%0A      at reduce (1:11)%0A      at /tmp/reduce.js:1:4

# after
1 | [].reduce((a, b) => a);
       ^
TypeError: reduce of empty array with no initial value
      at reduce (1:11)
      at /tmp/reduce.js:1:4
::error file=reduce.js,line=1,col=4,title=TypeError: reduce of empty array with no initial value::%0A      at reduce (1:11)%0A      at /tmp/reduce.js:1:4

Other top frames, annotation location only, before (1.4.0) -> after; the call is on line 2 of each probe file:

[].reduce(...) after e.stack was read:          file=native,line=1,col=11         -> file=<probe>.js,line=2,col=10
new EventEmitter().emit("error"):               file=node%3Aevents,line=59,col=31 -> file=<probe>.js,line=2,col=20
same, after e.stack was read:                   file=unknown,line=1,col=1         -> file=<probe>.js,line=2,col=26
e.stack assigned with only a native frame:      file=native,line=1,col=11         -> (no file/line/col)

The first row is a CommonJS probe; in an ES module the top-level frame is printed without parentheses and the stack-string parser currently stops there (#38308), so that variant has no frame with a file and gets no location.

[robobun]

Reproduced on bun 1.4.0 and on main (e697804): printf '[].reduce((a, b) => a);\n' > reduce.js && GITHUB_ACTIONS=true bun reduce.js prints ::error file=,line=1,col=11,...; the same frame choice also produces file=native, file=unknown and file=node%3Aevents headers and puts the code frame caret at the builtin's column. Fixed in this PR by picking the first frame with a file, via the skip list the code frame already used; tests in test/cli/test/bun-test.test.ts and test/js/bun/test/stack.test.ts fail on 1.4.0 and pass with the change.

[coderabbitai]

Warning

Review limit reached

@robobun, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 14 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: ec1988eb-678f-4a52-a72e-df1c61b1f887

📥 Commits

Reviewing files that changed from the base of the PR and between e697804 and 2ae77aa.

📒 Files selected for processing (4)

• src/jsc/VirtualMachine.rs
• src/jsc/ZigStackFrame.rs
• test/cli/test/bun-test.test.ts
• test/js/bun/test/stack.test.ts

---

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

[claude]

LGTM — extracts the existing frame-skip list into ZigStackFrame::has_user_source() and applies it to the two other consumers so all three agree on which frame the error is attributed to.

What was reviewed:

• has_user_source() is an exact extraction of the 7-case skip list from remap_zig_exception; that call site is behavior-identical.
• Caret loop now skips empty/native/unknown/[source: in addition to bun:/node:, aligning it with the frame whose source lines are printed; annotation now .find()s the first user frame instead of unconditionally taking frames[0].
• top_frame.is_some() → !frames.is_empty() is equivalent (old top_frame was frames.first()), so the annotation body still lists every frame.
• Tests cover all four top-frame shapes (empty / native / node: / unknown) with exact computed line/col, the no-location fallback, and the caret column against the source line.

Extended reasoning...

Overview

Touches error-printing diagnostics only: src/jsc/VirtualMachine.rs (three sites in remap_zig_exception, print_error_instance_body's caret loop, and print_github_annotation) plus a new has_user_source() predicate on ZigStackFrame. Tests added in test/cli/test/bun-test.test.ts (GitHub Actions annotation) and test/js/bun/test/stack.test.ts (caret placement).

Security risks

None. This is formatting of diagnostic output — which frame's file/line/col is printed in the code-frame caret and the ::error file=... GHA workflow command. No new parsing of untrusted input, no allocation, no FFI; has_user_source() is a pure predicate over an already-held &BunString.

Level of scrutiny

Low-to-medium. The change is small and mechanical: it lifts an existing 7-case skip list verbatim into a helper and reuses it at two sites that were using strict subsets of that list, so the three consumers (source-line lookup, caret placement, annotation location) now agree. The only observable effect is on diagnostic text, and the worst-case regression would be a mis-placed caret or annotation — not a runtime behavior change. The remap_zig_exception site is behavior-identical after the extraction; the caret site widens its skip set to match; the annotation site switches from "first frame with a position" to "first frame with a user file and a position", falling back to the same location-less ::error title= it already emitted for a positionless top frame. The annotation deliberately isn't gated on hide_bun_stackframes (an empty file= is wrong regardless of that opt-in), which the PR description calls out.

Other factors

Test coverage is thorough: a test.each matrix over the four source_url shapes (empty builtin, native after .stack read, node:events, unknown after .stack read) each asserting the exact file=,line=,col= computed from the fixture's own source; a no-user-frame case asserting the location-less form still prints the frame list; and two caret-column tests asserting ^ lands on reduce in the printed source line. The PR reports all new tests fail on 1.4.0 and pass with the change, and lists the adjacent suites re-run green. The top_frame.is_some() → !frames.is_empty() rewrite is a no-op given the old top_frame was frames.first(). No prior review comments to address; the PR notes the small overlap with #38328 in the same block and how they compose.

[robobun]

Heads-up on the caret half of this: #38349 makes the code frame's caret column a property of the excerpt (ZigStackTrace.source_lines_caret_column, written by whoever fills source_lines[0], so on the source-map path it is the column of the frame remap_zig_exception picked), which was needed for node:vm's columnOffset and as a side effect also puts the [].reduce caret under reduce. The annotation fix here is independent of that, and the two diffs do not touch the same lines of VirtualMachine.rs, so they should land in either order; after both, the printer's own frame loop only decides whether a caret is drawn at all.

[robobun]

^{Updated 1:05 AM PT - Aug 14th, 2026}

❌ @robobun, your commit 2ae77aa has some failures in Build #95344 (All Failures)

---

🧪   To try this PR locally:

bunx bun-pr 38335

That installs a local version of the PR into your bun-38335 executable, so you can run:

bun-38335 --bun