Flaky api and ocpp tests fail CI on unrelated PRs #1712

On 17 October 2026, flaky tests on master failed CI on four PRs whose changes did not touch the failing packages (#1701, #1703, #1706 and the #1674 follow-ups). Master requires an up-to-date branch, so each flaky run costs a full CI cycle before a merge. Re-running the failed jobs passed every time.

Seen today

go/internal/api, in the race detector job (and so in the required go test + vet):

• TestAViewerMayStillRead, api_viewer_writes_test.go:212: "a viewer's read of /api/system/info answered error, want a head". One run gave the same for /api/mode.
• TestReadingTheConfigIsAllowedForAnyRole/owner, api_passthrough_test.go:588: "the answer was not the handler's JSON: unexpected end of JSON input".
• TestAViewerIsRefusedAtEveryWrite, api_viewer_writes_test.go:143: "POST /api/ev/chargers was refused with "E_UNAVAILABLE", which is not a refusal to write".

Runs: 39045010631 (#1703), 39045384675 attempt 1 (#1701), 39047533237 attempt 1 (#1706).

go/internal/ocpp, in core (Go):

• panic: send on closed channel in github.com/lorenzodonini/ocpp-go@v0.19.0/ws.(*Client).Write, reached from ocppj.(*Client).SendError ← ocpp1.6.(*chargePoint).notSupportedError ← readPump. The client's read pump answers a request after the client has closed. Run 39045946853 (fix(web): switching the import ceiling on applies the box's last value (#1674) #1703).

Proposed fix in flight

Draft #1709 (fix-api-flaky-race-tests) addresses the api failures: the box frees its passthrough-queue slot before it sends an answer's last frame. That changes go/internal/appproto/passthrough.go, so it needs review as a product change, not a test fix. Its session reports 5 failures in 4,000 runs on master and 0 on the branch. It does not cover the ocpp panic.

Done when

• The api tests above pass reliably under -race (for example 4,000 runs with no failure), with fix(appproto): free the passthrough slot before an answer's last frame #1709 or another fix reviewed and merged.
• The ocpp test no longer panics: stop the test's charge point before its client closes, or avoid writing on a closed client, with a repeat count that shows it.
• Until then, re-running failed jobs is the workaround. Do not add skips or retries to hide the failures.

🤖 Generated with Claude Code