test: pin route-GOODBYE reconnect classification before the SDK flip
One test. It pins behaviour that is currently correct, unpinned, and about to change through a dependency rather than through this repo.
What's unpinned
A module restart sends GOODBYE to every route on the connection, so a mid-request call rejects with a bare SubcError — the raw request() path never produces a managed SubcCallError.
isConsumerReconnectTransient reaches its err instanceof SubcCallError branch first, so a bare SubcError falls through to non-transient. Today that means: route dropped, client kept, next call re-opens on the same connection. That's the behaviour every seat here relies on across ck module restart aft.
Nothing asserts it. The nearest test — a plain timeout (non-transient SubcError) KEEPS the client — covers the same class, but a timeout and a GOODBYE can be classified differently, so it would stay green through the change below.
Why now
cortexkit/subconscious#6 proposes putting kind on SubcError itself. If the classifier then reads it, GOODBYE flips false→true and the dropClient decisions consuming this classifier change behaviour — on an SDK bump alone, with no diff in this repo.
I think that flip is probably correct: a route GOODBYE genuinely is the transient class. This isn't an argument against it. It's that a reconnect-policy change arriving through a dependency should land as a red test you adopt deliberately, not as a reconnect storm someone diagnoses later. The comment says so, and says to flip the assertions rather than delete the test when it happens.
Verification
Red control, since a characterization test that can't fail is worse than none:
green 55 pass / 0 fail
mutate SubcError → SocketClosedError (verified applied via cmp, not assumed)
↳ simulates exactly the flip: GOODBYE becomes transient
red 54 pass / 1 fail — only the new test
restore (verified via cmp)
green 55 pass / 0 fail
Full packages/aft-bridge suite: 474 pass / 0 fail. Lint and format clean.
The mutation targets the classification, not the assertion, so it fails for the reason the test exists — expect(client.closed).toBe(0) and expect(madeClients).toBe(1) are what break.
Test-only; no production code touched.
Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.
Summary by cubic
Pins the current reconnect classification for route GOODBYE to prevent an SDK-only change from silently changing reconnect policy. Today a mid-request GOODBYE yields a bare SubcError that isConsumerReconnectTransient treats as non-transient: the route drops, the client is kept, the next call reopens on the same connection, and there is no in-place retry.
- Adds a characterization test in
packages/aft-bridge/src/__tests__/subc-transport.test.tsasserting: first call rejects withSubcError(notSubcCallError), client stays open, second call succeeds on the same client, and there is no auto-retry. - The fixture carries
kind: "outcome_unknown"even though today’sSubcErrorhas onlycode, so the test will turn red when the classifier starts readingkind. - Test-only; no production code.
When the SDK flips
- When
cortexkit/subconscious#6addskindtoSubcErrorand the classifier consumes it, GOODBYE will classify transient and this test should fail. - Action: flip the assertions to expect client drop and reconnect; do not delete the test.
Greptile Summary
This test-only PR pins the current reconnect policy for route GOODBYE failures before a planned SDK classifier change.
- Models a route GOODBYE as a bare
SubcErrorcarryingkind: "outcome_unknown". - Verifies the client remains open, the route reopens on the same connection, and the failed call is not retried in place.
Confidence Score: 5/5
The PR appears safe to merge.
No blocking failure remains.
Important Files Changed
| Filename | Overview |
|---|---|
| packages/aft-bridge/src/tests/subc-transport.test.ts | Adds a focused reconnect-policy characterization test and now explicitly models the structured GOODBYE discriminator requested in the previous review. |
Reviews (2): Last reviewed commit: "test: pin route-GOODBYE reconnect classi..." | Re-trigger Greptile
Context used:
- Knowledge Base — AFT SubC tool-surface bridge
- Knowledge Base — TypeScript AFT bridge: native and daemon-backed transport