CI: claude-sdk-oauth binding-lifecycle security test fails only inside the coding-agent 1/3 shard (reproduces on main)

[code-yeongyu]

Summary

test/claude-sdk-oauth-binding-lifecycle-security.test.ts > forgets process binding and invalidates/deletes durable state when selecting a non-Claude provider fails inside the Test (coding-agent 1/3) shard while passing in isolation. It reproduces on main itself, so it blocks unrelated PRs whose only red check is that shard.

Evidence

Failure on a main run (no PR involved) — run 35055501202, head fef17495, job Test (coding-agent 1/3):

 FAIL  test/claude-sdk-oauth-binding-lifecycle-security.test.ts > Claude SDK OAuth binding lifecycle security > forgets process binding and invalidates/deletes durable state when selecting a non-Claude provider
AssertionError: expected { …(11) } to be undefined
 ❯ test/claude-sdk-oauth-binding-lifecycle-security.test.ts:200:33
    200|   expect(getBinding(sessionId)).toBeUndefined();

 Test Files  1 failed | 468 passed | 2 skipped (471)
      Tests  1 failed | 3934 passed | 6 skipped (3941)

The received value is a live binding record (sdkSessionId, systemPromptHash, toolsetHash, …) for session id model-select-non-claude, i.e. the model_select handler did not forget the process binding.

Contrast runs I made locally on the same file:

Scope	Result
the file alone, at main-derived base 30b5ea8024	Test Files 1 passed, Tests 2 passed
the file alone, at a feature head	Test Files 1 passed, Tests 2 passed
the whole test/claude-sdk-oauth-*.test.ts cluster (61 files)	Test Files 61 passed, Tests 493 passed | 3 skipped
shard 1/3 in CI	fails as above

So the interference comes from a file outside the claude-sdk-oauth cluster that shares this module's process-global state. The suite keeps the binding registry in module scope: the test registers one with registerSessionRegistry(extension.api) and clears it in afterEach via resetSessionRegistryBoundary() — a leftover registration from an earlier file in the shard would leave getBinding and the handler reading different boundaries.

Expected

The test passes in every shard composition, because the binding registry boundary cannot leak across test files. The fix is isolation at the seam (establish the boundary per test rather than trusting the previous file to have cleared it), not a retry, a skip, or a loosened assertion.

Acceptance criteria

• The exact shard composition that fails today (vitest run --shard=1/3) passes.
• The cause is named in the PR: which state leaks and which file leaves it.
• test/claude-sdk-oauth-binding-lifecycle-security.test.ts still fails if the model_select handler stops forgetting the binding (the assertion keeps its teeth).

Related

Found while verifying #1726/#1729 (#1746), whose only red check was this shard.

[code-yeongyu]

Additional evidence and a lead.

Three-way local contrast on the same machine, same command, for the two files that fail there (agent-session-compaction, interactive-mode-status) shows those are a separate, environment-dependent story — pure main a7dee82bb1, a feature base, and a feature head all produce an identical 2 failed | 91 passed (93). Not related to this issue, recorded so nobody chases it twice.

For this issue specifically: a local vitest run --shard=1/3 did not reproduce the claude-sdk-oauth-binding-lifecycle-security failure — it failed on the two files above instead. Local collected 469 files against CI's 471, so the shard composition differs and the interfering neighbour is not in the local slice. Reproduction likely needs the CI file set.

Lead on the mechanism: resetSessionRegistryBoundary() (session-registry.ts:74) restores only the injected boundary — now, queryFactory, scheduleReap. It does not clear resident registry entries or bindings. The failing test's afterEach explicitly closes and forgets the ids it created, so its own cleanup is sound; what is unguarded is state a neighbouring file leaves behind under a different id, plus registerSessionRegistry(...) being re-registered per file against module-scope state. A per-test boundary (fresh registry instance, or an explicit registry reset alongside the boundary reset) would make shard composition irrelevant.

Acceptance criteria are unchanged: the fix must keep the assertion able to fail if model_select stops forgetting the binding.