[kosarko]

Problem description

The e2e step fails before any Cypress spec runs:

Waiting on "http://127.0.0.1:8080/server/api/core/sites, http://127.0.0.1:4000" with timeout of 120 seconds
...
##[error]Socket closed before the connection was established

This happens on runners with Actions Runner 2.338.0 (Node 24.21), which GitHub is rolling out gradually. With cypress-io/github-action@v6, the wait-on check crashes on the first refused connection instead of retrying until the timeout. Upstream issue: cypress-io/github-action#1873, fixed in v7.4.4.

Example: #188, run 39619750707 (runner 2.338.0). The last green runs, the clarin-v7 nightly on 2026-10-21 and #185 on 2026-10-20, were on runner 2.337.0. Once the rollout is complete, every PR will fail this way.

Analysis

• build.yml: cypress-io/github-action@v6 → @v7.
• v7's only breaking change is the action runtime moving from Node 20 to Node 24 (v6 on Node 20 is deprecated). The inputs used here (browser, start, wait-on, wait-on-timeout, component, record, publish-summary) are unchanged.

Problems

• The rerun-on-failure job doesn't work. Every failing run ends with cannot be rerun; Resource not accessible by integration, because the workflow token only has contents: read and packages: read. Not fixed here: granting actions: write may not be enough, since the job reruns the run it is part of. That needs its own change.

Copilot review

• Requested review from Copilot