[diegosouzapw]

The release is red, and this is why

release/v3.8.50 has been failing 15 unit tests since #9275 merged on 2026-08-03. Every PR opened against it inherits those failures — #9425, which changes one YAML file and one markdown file, currently fails Fast Quality Gates and all four Unit Tests shards. In a queue sweep today, 67 open PRs were red on nothing but this shared base-red.

Root cause

#9275 widened CODEX_NATIVE_UNPREFIXED_MODELS from a single id to gpt-5.5 plus the gpt-5.6-sol/terra/luna tiers — a deliberate fix, so bare Codex CLI ids reach the ChatGPT subscription instead of fanning out. But the early return it added to resolveModelByProviderInference never consults the active-provider set:

if (CODEX_NATIVE_UNPREFIXED_MODELS.has(modelId)) {
  return { provider: "codex", model: modelId, extendedContext };   // ← before any provider lookup
}
const [activeProviders, ...] = await Promise.all([getActiveProviderSet(), ...]);

That made the guard 30 lines below unreachable for all 27 ids in the set — the one whose own comment reads "Route overlapping models through Codex only for Codex-only installations; when OpenAI is also active, preserve the historical OpenAI default below."

User-visible effect: an OpenAI-only install asking for bare gpt-5.5 is routed to codex — a provider it has no connection for — and fails with no active credentials for provider: codex on a model OpenAI serves fine. An install whose codex connection is merely inactive fails identically. It also silently reverted #5887's compatibility boundary.

The fix

The Codex preference may only preempt another provider when a codex connection is actually active:

• codex active → codex wins over OpenAI. This is exactly what fix(routing): bare model ids route to codex first; validate synced candidates #9275 wanted, and it is preserved.
• no codex active / codex inactive → falls through to normal inference, so OpenAI-only installs work again.
• ids only codex catalogs (codex-auto-review) → still resolve to codex with no connection at all. There is no alternative to preempt, and no codex credentials is the honest error.
• explicit openai/… prefix → still authoritative, in every case.

Tests

Of the 13 failing routing assertions, 10 were correct all along and pass again with no edit — they were reporting the regression, not lagging behind it. Only 3 encoded the intended #9275 change and were updated to expect codex, each with a comment explaining the moved boundary.

Added:

• #5887(b2) — an explicit openai/ prefix overrides the Codex-native set.
• a Codex-native bare id stays on OpenAI when no codex connection exists — the permanent guard for this regression.

Validation

codex-gpt55-routing-5887 + codex-synced-bare-model-routing + model-alias-provider-resolution
  + chat-helpers + plan3-p0 + lib/consoleInterceptor-writes   →  90 pass, 0 fail

Before the fix these same files were 13 fail / 77 pass. typecheck:core, scoped eslint, check-complexity, check-cognitive-complexity and check-changelog-integrity all pass.

Not covered here: the remaining 2 failures in vscode-token-routes.test.ts (a trailing period in the no-credentials message) belong to #9392 by @HouMinXi, which should land alongside this. Together they take the release back to green.

Also still open, and untouched by this PR: check-file-size reports src/sse/handlers/chat.ts 1846>1845 and open-sse/executors/base.ts 1623>1578 on the bare tip — baseline drift for the release captain.