Vendored Node.js compatibility suite #

Status: active Last reviewed: 2026-08-29 Owner: theMackabu

Goal #

Run a pinned external Node.js compatibility corpus directly under Ant, with reproducible provenance and explicit expectations instead of copying failures into Ant-owned tests.

Scope #

• Bun's portability-adjusted copy of Node's parallel tests.
• The matching common helpers and fixtures required by those tests.
• A Node-hosted process runner with selection, concurrency, timeouts, output capture, expectations, and JSON reports.

Node's sequential, native-addon, V8-internal, pseudo-TTY, pummel, and other specialized groups remain out of scope until the parallel corpus has a useful baseline and their scheduling or build contracts can be represented honestly.

Decisions #

• Pin Bun commit ed950b88ab2ec6b58bccdfe7d310731b8ca13c4d.
• Preserve the imported files rather than patching them for Ant.
• Treat success as process exit 0, matching the upstream direct-execution contract.
• Pass upstream // Flags: directives through to Ant and record unsupported CLI behavior as a result rather than silently discarding it.
• Enter through an Ant-owned bootstrap that supplies Node harness metadata such as process.config; do not patch the imported common implementation.
• Keep generated reports outside version control.

Tasks #

1. Inspect and pin the portable upstream subtree.
2. Vendor parallel, common, fixtures, licenses, and provenance.
3. Add the Ant runner and expectations format.
4. Run a focused initial corpus and classify harness blockers.
5. Establish a reviewed baseline before enabling broad CI coverage.

Initial results #

• The first direct run exposed two harness prerequisites: the imported .js files need CommonJS package scope, and Node's shared harness expects the build-artifact property process.config. Both are now supplied outside the imported files.
• A 12-test node:path selection reached real assertions: 2 passed and 10 failed on behavior including Win32 paths, extname('..'), module identity, and missing matchesGlob() / toNamespacedPath().
• A sorted 100-test smoke selection produced 19 passes, 79 ordinary failures, one crash, and two timeouts. Results are intentionally not copied into the expectations file until each family is classified; the current empty baseline keeps every failure visible.
• The complete 3,236-test parallel corpus produced 1,410 passes (43.57%), 1,439 ordinary failures, 383 timeouts, and 4 crashes. All 330 tests carrying upstream Node CLI flags failed before execution because Ant does not accept those flags; excluding that harness/CLI bucket, the observed pass rate is 1,410 / 2,906 (48.52%). Strong large families include HTTPS (56/59), Web Crypto (36/38), crypto (112/120), TLS (173/186), HTTP/2 (235/256), and diagnostics_channel (32/35). Weak large families include dgram (0/75), VM (3/95), cluster (3/80), DNS (1/24), Buffer (7/63), HTTP (49/386, including 244 timeouts), and path (2/15).

Validation status #

• node tests/node-compat/test.mjs: passed.
• JSON report generation, concurrency, upstream flags, timeout handling, and crash classification were exercised by the initial samples.
• A full-run attempt exposed orphaned long-lived fixture subprocesses after timeout. The runner now gives each test a Unix process group and terminates that group on timeout or completion; the invalid partial run was discarded.