[rudycelekli]

Bug

Guesser.checkNPMPackageWithClient (internal/experiments/guesser.go) writes its result to the cache manager on every exit path, including the failure paths: request errors, timeouts, non-200/404 statuses (5xx, 429), and an unparseable body. Those entries use the default 2h TTL, and the cache is checked before the network on the next call.

A failed lookup tells us nothing about whether the package exists, but it gets cached as exists: false. For two hours, every registry search that involves that repository loses its npm install_cmd / repository_info.

Trigger

The runtime's registry search (Runtime.SearchRegistryServers) builds a guesser with the shared cache manager. Batch lookups use a 3s client timeout. One slow npm response, a brief network drop or an npm 5xx during a search caches the failure. Refreshing the registry (RefreshRegistryCache) does not help, because it only drops the registry-servers: keys, while the guess entries live under npm:<package>.

Fix

Cache only the definitive answers: 200 (package found) and 404 (package does not exist). Failed lookups are returned without being cached, so the next search tries again.

Test

guesser_transient_cache_test.go drives checkNPMPackageWithClient with a scripted http.RoundTripper, so it makes no real network calls:

• TestCheckNPMPackage_TransientFailureNotCached: the first response is a network error, a 503 or a 429, and the second is a 200. The second call must reach the transport and report the package as found.
• TestCheckNPMPackage_DefinitiveAnswersCached: a 200 and a 404 are each still served from the cache on the second call.

Commands run

• CGO_ENABLED=0 go test -count=1 -run 'TestCheckNPMPackage_(TransientFailureNotCached|DefinitiveAnswersCached)|TestGuesser_CacheEntriesAreStampedInternal|TestCaching|TestNilCacheManager' ./internal/experiments/: ok
• CGO_ENABLED=0 go vet ./internal/experiments/: ok
• On current main (and again with guesser.go reverted), all three TransientFailureNotCached subtests fail: expected: 2, actual: 1 transport calls, because the second call returned the cached failure.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[Dumbris]

Verified: the PR test fails on main for the reason described and passes on this branch merged with main; race tests for the touched packages and CI lint (bare and server tags) are clean. Also built and tested together with the rest of this batch, independently reviewed, and cross-reviewed with a second model.

[Dumbris]

Merged. Thank you, @rudycelekli! A transient npm lookup failure is no longer cached as a permanent miss.