flux-pr-1076
graphql-go-tools (Go) · W2 · GPT-5.3 Codex
Tests passed. 1/1 commands passed. Strength: strong.
go test -C v2 ./... -count=1 -timeout=300sPartial score: 1/1
Trajectory
codex · partial order onlyprovider-native trajectory captured; validation and decision steps are appended with coarse ordering only
Quality
Equivalence Reasoning
behavioral
The patch covers major resolver/concurrency intent (per-subscription worker serialization, per-sub heartbeat, and client-close updater fix), but it does not implement the CI race-detector change as intended. It leaves `make test` non-race (no Makefile switch to `go test -race`) and keeps CI using `make test-race` instead of making race the default test path; it also changes the `CI` step to Windows-only, altering workflow behavior beyond the stated intent.
Code Review
The patch appears to implement much of the subscription concurrency redesign, but it likely does not fully satisfy the intended PR because CI/race-detector integration diverges materially and test coverage for trigger cleanup was weakened.
Both workflow files gate the `CI` step with `if: runner.os == 'Windows'`, so Linux/macOS jobs no longer run `make ci`; they only run the race step. This behavior diverges from the intended change and can miss checks included in `ci` but not in `test-race`.
The workflows continue to call `make test-race` rather than switching to `make test` with race enabled by default in Makefiles, leaving behavior inconsistent with the stated goal of making race detection the default CI/test path.
The test no longer asserts trigger map size before/after unsubscribe, reducing detection of cleanup regressions (e.g., leaked triggers/subscriptions) in a refactor that heavily changes shutdown and worker lifecycle.