Repository navigation
Description
Summary
quality / Security (composer) fails the whole job when packagist.org's advisory endpoint is briefly unavailable, because the outage-tolerance branch only recognises npm's wording. A transient 502 is reported identically to a real vulnerability.
Evidence
scholiq development, push run 31337136985, job quality / Security (composer):
composer audit --format=plain --abandoned=report
##[error]The "https://packagist.org/api/security-advisories/" file could not be downloaded (HTTP/2 502 )
##[error]Process completed with exit code 100.
No advisory was reported. The job is red purely because packagist returned a 502. The previous push run on the same branch had this job green, and nothing in composer.lock changed between them.
Root cause
The workflow already anticipates this failure mode — for npm only. The tolerance test is:
if [ "$rc" -ne 0 ] && printf '%s' "$out" | grep -qiE \
"audit endpoint returned an error|Invalid package tree|ECONNRESET|ETIMEDOUT|ENOTFOUND|being retired|Service Unavailable"; thenComposer's outage does not use any of those phrasings. It says:
The "https://packagist.org/api/security-advisories/" file could not be downloaded (HTTP/2 502 )
so the guard does not match, and exit 100 propagates.
Why it matters beyond one red job
Security (composer) failing is exactly the signal that should stop a merge. Once it can go red for reasons unrelated to the dependency tree, the honest reading of a red security job becomes "probably packagist again" — and the next real advisory lands in a check people have learned to skip. The cost is not the false red; it is the true red that gets waved through after it.
It also breaks the merge rule the fleet uses: comparing a PR's failing set against development's is only meaningful if a job's colour is a property of the code. A job that flips on upstream availability makes the baseline non-deterministic — a PR opened during an outage inherits a failure its diff did not cause.
Proposed fix
Extend the existing tolerance to composer's phrasing, keeping it narrow enough that a genuine advisory still fails:
if [ "$rc" -ne 0 ] && printf '%s' "$out" | grep -qiE \
"audit endpoint returned an error|Invalid package tree|ECONNRESET|ETIMEDOUT|ENOTFOUND|being retired|Service Unavailable|could not be downloaded \(HTTP/[0-9.]+ (429|5[0-9][0-9])|packagist\.org.*could not be downloaded"; thenRestricting the HTTP branch to 429 and 5xx matters: a 404 on that endpoint would mean the advisory API moved, which should stay red rather than be tolerated.
Two things worth doing alongside:
- Emit a
::warning::when the tolerance fires, naming the endpoint and status, so a tolerated outage is visible in the run rather than silently green. A tolerance nobody can see is indistinguishable from a check that stopped running. - Retry once before tolerating. A 502 from packagist is usually seconds long; one retry converts most of these into a real verdict instead of a skipped one. Tolerating on the first failure gives up the measurement more often than necessary.
Activity