Single non-retried EBADF socket error on registry.npmjs.org aborts the whole repository run

[eyalzek]

How are you running Renovate?

Self-hosted, official docker image renovate/renovate:44.49.0 (also observed similar socket-error aborts on 44.48.x), GitHub platform, large monorepo.

Describe the bug

A single write EBADF socket error on a request to registry.npmjs.org during dependency lookup aborts the whole repository run with Repository result: external-host-error. The process then exits 0, so from the outside the run looks successful.

Chain of events:

1. The npm datasource adds an implicit abortOnError: true host rule for https://registry.npmjs.org (lib/modules/datasource/npm/get.ts).
2. EBADF is not in got's default retryable errorCodes (unlike its transient-socket siblings ECONNRESET / EPIPE), so the request fails without any retry.
3. In Http.request the abortOnError check throws ExternalHostError before the stale-cache fallback (cacheProvider.bypassServer), so a cached packument that could have served the lookup is never used.
4. ExternalHostError aborts the repository.

The error consistently appears right after a long stretch of HTTP-cache hits, i.e. the first live request after keep-alive sockets sat idle — a stale/locally-closed socket being reused. It is transient: an immediate retry would almost certainly succeed.

Error: write EBADF
    at WriteWrap.onWriteComplete [as oncomplete] (node:internal/stream_base_commons:87:19)
  syscall: write, errno: -9, code: EBADF
wrapped as: Error: external-host-error
    at Http.request (lib/util/http/http.ts:255:15)
    at Http.requestJson (lib/util/http/http.ts:385:17)
    at getDependency (lib/modules/datasource/npm/get.ts:104:18)

For a large repo this means one bad socket write costs the entire run (which can be many hours), with exit code 0.

Expected behavior

One or both of:

• EBADF treated as retryable alongside ECONNRESET/EPIPE (add to got retry errorCodes in lib/util/http/got.ts).
• The stale-cache fallback consulted before abortOnError escalates a transient socket-level error to ExternalHostError (HTTP status errors still aborting as today).

Relevant debug logs

DEBUG: Host error (hostType=npm, packageName=@jest/globals) err: write EBADF
DEBUG: Aggregate error is thrown
INFO: External host error causing abort - skipping
INFO: Repository result: external-host-error, status: onboarded, enabled: true, onboarded: true

Observed on two consecutive runs (different packages each time: prettier-plugin-sh, @jest/globals). Node v24.20.0, got 14.6.6 (image defaults).