Nightly tests are failing

[github-actions]

Tests against nightly failed, see: https://github.com/nodejs/undici/actions/runs/31252119887

[RaphaelFakhri]

The nightly jobs are not failing, they are wedging, which is also why this issue only gets filed some days.

Every nodejs-nightly run since 28 June 2026 has ended cancelled rather than failure: 40 cancelled and 21 failure in the last 70 runs, with the last success on 2026-06-28. report-failure is gated on needs.test.result == 'failure', so a cancelled run never reports, and the issue only appears on days when an unrelated job also fails outright.

The wedge is visible in the logs. On today's ubuntu job (97763666164) the last test output is 10:08:51 ✔ connect through proxy (with pool), then 19 minutes of silence, then The job has exceeded the maximum execution time of 20m0s and Terminate orphan process: pid (2566) (npm run test:unit).

I ran the suite locally against the same v27.0.0-nightly202608254b5e86c4e2 build and found two causes.

1. req.headers now has a null prototype, from nodejs/node#62900. Every deepStrictEqual(req.headers, { ... }) fails on the prototype:

+ [Object: null prototype] {
- {

In test/proxy-agent.js that assertion throws inside the server's request handler, so the response is never written, the in-flight fetch never settles, and the file hangs: # Interrupted while running: test/proxy-agent.js. That is exactly where the ubuntu run goes quiet, since proxy.js is the file before it in the runner's order.

Spreading the headers fixes it: deepStrictEqual({ ...req.headers }, { ... }). With that change test/proxy-agent.js is 38/38 on both v24.16.0 and the nightly, and stops hanging. The same pattern appears in at least test/client.js, client-keep-alive.js, client-post.js, content-length.js, headers-as-array.js, header-function-crlf.js, pipeline-pipelining.js, test/fetch/request.js and test/fetch/issue-rsshub-15532.js, most of which CI never reaches because it wedges first.

2. h2 streams reset before the readable side is consumed now throw ERR_HTTP2_STREAM_ABORTED ("The stream was reset before the readable side was fully consumed", from Http2Stream.onStreamClose). test/http2-stream.js expects undici's own UND_ERR_INFO with "HTTP/2: stream half-closed (remote)"; on the nightly the raw Node error surfaces instead, the assertion fails and the file wedges in teardown. That is the stall point on both macOS and Windows. This one is not just a test edit: undici's public error for a half-closed h2 stream changes under Node 27 unless client-h2.js maps the new code.

With several workers each parked on a wedged file the pool drains and the whole run stalls, which matches the per-OS stall points.

Happy to send the headers fix as a PR if that is useful. The h2 error mapping looks like a call for you rather than something to guess at, and separately, report-failure not firing on cancelled is why this has been silent since June.