node: Node-style uncaught-error output — Error: prefix, quoted props, ResolveMessage surface (+4 tests)

[cirospaciari]

Important

Deliberate user-visible output change, shipped for explicit review/veto. Bun's uncaught-exception printer historically wrote a lowercase error: prefix. This PR makes the uncaught/unhandled path print the error's own name the way Node does (Error: boom, TypeError: x, Error: ENOENT: no such file…) and single-quotes string properties in the property dump (code: 'ENOENT'). bun test failure rendering and console.log(err) / Bun.inspect(err) are unchanged. This one prefix structurally blocked ~17+ reachable vendored Node tests that spawn a child and regex /Error: .../ on its stderr — no amount of feature work reaches them.

Stacked on claude/node-v26-fix-tls. +5 upstream tests (4 newly vendored, 1 pre-existing fixture family restored to verbatim). test-internal-modules.js was vendored at first and dropped again: the test harness sets BUN_FEATURE_FLAG_INTERNAL_FOR_TESTING, which since the quic work exposes internals, so require('internal/freelist') resolves under the harness and the test cannot pass in CI.

What this does

1. Node-style uncaught-error rendering (src/jsc/VirtualMachine.rs, src/jsc/ConsoleObject.rs, src/runtime/jsc_hooks.rs)

A new Formatter.node_uncaught_style flag is set only on the uncaught/unhandled print funnel (print_exception + the jsc_hooks funnel), and explicitly not under bun test, so the runner's classic failure output is untouched. In that mode:

• the error's name prints verbatim — no lowercasing of Error, no promoting a CODE: -prefixed message's code into the name slot (Error: ENOENT: no such file or directory, open 'x', matching Node);
• string properties print single-quoted for simple ASCII strings (code: 'ERR_LOAD_SQLITE_EXTENSION', permission: 'FileSystemWrite'), falling back to the JSON writer for anything needing escapes;
• a rethrow from inside an uncaughtException handler now unwraps the JSC::Exception wrapper before printing, so the error's own creation stack (at throwException (...)) and property dump render — previously only the rethrow-site frame printed, with no properties (exit code 7 was already correct).

2. ResolveMessage gets Node's name/toString()/.stack surface (src/jsc/ResolveMessage.rs)

String(err) on a module-not-found was ResolveMessage: ...; Node needs Error: Cannot find module 'x' (CJS), Error [ERR_MODULE_NOT_FOUND]: ... (ESM), Error [ERR_UNKNOWN_BUILTIN_MODULE]: ... (missing node: builtin) — all with err.name === 'Error' (.code was already right). name becomes a getter so Bun.build logs and non-module-resolution uses keep reporting ResolveMessage. The uncaught printer renders these Node-style too (Error: Cannot find module 'x' + Require stack:). is_bare_esm_specifier no longer classifies URL-scheme specifiers (file://…, pkg:bar) as bare packages — Node never says Cannot find package 'file:', and the module-form message keeps the full specifier (all message shapes oracled against node v26.3.0).

3. Node's inspector permission gate (src/runtime/jsc_hooks.rs::ensure_debugger)

Under --permission without --allow-inspector, --inspect/--inspect-wait are silently skipped and --inspect-brk raises uncaught ERR_ACCESS_DENIED (resource PauseOnNextJavascriptStatement), exit 1 — verified against node v26.3.0's observed behavior (plain --inspect runs fine there; only the pause request throws).

4. Node's ERR_MISSING_OPTION guard: any --allow-* flag without --permission exits 1 with TypeError [ERR_MISSING_OPTION]: --permission is required (initializePermission, pre_execution.js).

5. util.callbackify falsy-rejection message drops the stray "a": Promise was rejected with falsy value, Node's exact wording.

Base repairs (the branch did not compile)

claude/node-v26-fix-tls's merges dropped whole generations of already-merged work (kept one side of each conflict):

• reject_bad_negations initializer missing from Arguments.rs → the base does not build;
• all Arguments.rs consumers of the Node-style CLI work were dropped while the bun_clap half survived: -e/-p/--eval/--print were completely dead (every invocation printed help), -c/--check panicked (--check is not a parameter), --inspect-port/--debug-port unregistered, NODE_OPTIONS validation and node-style missing-value errors gone;
• node_fs.rs call sites of from_js_required/BUFFER_EXPECTED_TYPES dropped while types.rs kept the helpers (dead-code hard errors).

These are restored as the first two commits (re-applied from the original commits, conflicts resolved by hand; test/cli/run/run-eval.test.ts 88→91 passing again, the five test-fs-* files from the fs-validation commit re-verified green). Every other agent branching from this base is currently broken the same way.

Tests

Newly vendored (byte-verbatim, preflight-verified; each fails on the unfixed build, passes with this PR, and fails again with a throw appended):

• parallel/test-internal-modules.js — anchors /^Error: Cannot find module 'internal\/freelist'/ against String(err) (no longer included, see above)
• parallel/test-permission-inspector.js — in-process Session.connect() denial + child /Error: Access to this API has been restricted/
• parallel/test-permission-inspector-brk.js — --inspect-brk under --permission exits 1 with the same line
• parallel/test-permission-sqlite-load-extension.js — /Error: Cannot load SQLite extensions…/ + /code: 'ERR_LOAD_SQLITE_EXTENSION'/
• parallel/test-unhandled-exception-rethrow-error.js — exit 7, /Error: boom/, /at throwException/, /rethrow: true/

Restored to byte-verbatim upstream: fixtures/require-resolve.js (its /^Error: Cannot find module/ anchors now hold; test-require-resolve.js re-verified green, and fails with the old binary).

test-util-callbackify.js (pre-existing adapted vendored file on the base): updated its error:-prefix and falsy-message assertions to the upstream spellings; still not fully verbatim — upstream additionally asserts errLines.length === 7 (Bun's multi-line source excerpt makes this 10) and at process.processTicksAndRejections frame naming. This is the single remaining preflight drift, inherited from the base and reduced by this PR.

Not vendored, with reasons:

• test-node-output-{errors,console,vm,sourcemaps}.mjs — byte-exact snapshot comparisons requiring V8 frame naming (at Object.<anonymous> (...)), node:internal frames in stacks, V8 column numbers, and a Node.js <version> footer; all 30 error fixtures still fail. Unreachable without stack-shimming JSC to V8's frame model.
• test-dotenv-node-options.js — needs NODE_OPTIONS from --env-file files applied at startup (permission grants included); separate feature.

Bun-side assertions on the old rendering updated in the same PR (each verified to fail on the old build): hot.test.ts, process.test.js, sourcemap.test.js, reportError.test.ts, stack.test.ts, serve.test.ts, serve-direct-readable-stream.test.ts, async-iterator-stream.test.ts, run-autoinstall.test.ts, tsconfig-override.test.ts, run-eval.test.ts, resolve.test.ts, resolve-error.test.ts, resolve-test.js, import-meta.test.js. Suites re-run green: resolve/ (37 files), http server suites, worker_threads, web/workers, hot, process, run-eval, regression set.

Known pre-existing issue surfaced while sweeping (not touched here): worker terminate() leaks the per-worker node_fs_binding::Binding under LeakSanitizer (worker-terminate-lifetime.test.ts dns case) — unrelated to this diff.

[robobun]

^{Updated 11:05 PM PT - Aug 21st, 2026}

@robobun, your commit c51429f is building: #103290

[robobun]

Heads-up for a rebase: #38278 makes print_exception unwrap the JSC::Exception cell to the thrown Error before printing. Once it lands, the err.to_error().unwrap_or(err) at the rethrow-inside-handler call of run_error_handler in this PR is redundant for an Error, and for a non-Error value it would drop the throw-site frames that the cell path keeps.