## Problem
The `danger-js` job in **CI Utils** has been failing across most PRs.
The failure is not a real Danger violation — it's a transient network
error fetching the PR diff/commits from the GitHub API:
```
Failed to fetch GitHub pull request files: FetchError: Invalid response body while trying to fetch
https://api.github.com/repos/twentyhq/twenty/pulls/XXXXX/files?page=1&per_page=100: Premature close
at Gunzip.<anonymous> (.../node_modules/node-fetch/lib/index.js:400:12)
errno: 'ERR_STREAM_PREMATURE_CLOSE',
code: 'ERR_STREAM_PREMATURE_CLOSE'
```
GitHub closes the gzipped HTTP response mid-stream, and Danger's bundled
`node-fetch` has **no retry** on a dropped connection — so any single
blip fails the whole check.
## Why is this happening now?
Nothing on our side changed at the boundary where failures started. The
Node 24.16 bump landed Jun 8 (the job stayed green for 2+ weeks after),
Danger has been pinned at `13.0.4` for months, and there are **zero
commits** to `.nvmrc`, `twenty-utils/package.json`, or the
`yarn-install` action since Jun 22.
What changed is GitHub's API reset rate, and it changed abruptly:
| Day | Failures | Successes | Failure rate |
|-----|----------|-----------|--------------|
| Jun 23 | 1 | 48 | ~2% (green) |
| Jun 24 | 38 | 204 | ~16% |
| Jun 25 | 23 | 25 | **~48%** |
The same PR passes on one run and fails on the next (e.g. one PR shows
up as both pass and fail; another failed 3 runs in a row) — a code bug
can't flip outcomes on identical input, only an infrastructure flake
can. When GitHub's connection-reset rate was ~0% we never noticed; now
that it's in the tens of percent, roughly half of all PRs trip it.
## Fix
Wrap the Danger invocation in a small retry loop (3 attempts, 5s
backoff) so the check absorbs these transient fetch errors instead of
red-flagging the PR. Applied to both the `danger-js` and `congratulate`
jobs since they share the same failure mode.
This is the correct mitigation rather than a code revert — there's no
change on our side to revert. 3 attempts drop a ~48% single-shot failure
rate to ~11%, and a less-degraded ~16% rate to well under 1%. If
GitHub's reliability recovers, the retries simply stop firing and cost
nothing.
This is a CI-only change — no application code is touched.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/22151?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-light.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
Charles Bochet
f416f81548
