[etang4]

A Codex session configured for Ultrafast or Flex can start with the correct tier, but codex-acp reduces that tier to a Fast on/off boolean. Its first prompt then sends serviceTier: null, clearing the selection.

Retain the service-tier string returned by Codex through new, resume, load, and fork operations and subsequent prompts. Explicit ACP Fast On/Off changes still override the selection. Both fast and Codex's canonical priority name are recognized as Fast, with the existing model-support check preserved. Codex remains responsible for account and model eligibility; the adapter does not force a tier that Codex rejected.

Reopening a live session before its first prompt also preserves its tier: the no-rollout fallback leaves the tier unknown instead of reporting a clear. A tier explicitly returned by Codex, including null, still takes precedence over the existing session value.

The change uses the existing app-server string field and removes casts to the older generated fast | flex type. It does not change ACP or generated protocol schemas.

Related work

• Fast mode is disabled when Codex reports the canonical priority service tier #285 reports the canonical Fast/priority initialization bug; this change also covers that case.
• fix: Map priority mode to fast #327 handles that Fast alias but still stores a boolean and emits only fast or null, so it does not preserve Ultrafast/Flex. There is overlap in Fast alias recognition.
• fix: recognize priority service tier as Fast mode #648 independently improves Fast capability detection for gateway models advertising serviceTiers. This PR leaves that model-discovery change to fix: recognize priority service tier as Fast mode #648.

Validation

At 52d47d7a022525ed19f3757191254e77d8a44310:

• npm run typecheck
• Full npm test with Node 24: 1,911 passed, 36 skipped (the existing suite's normal skips)
• Full npm run test:integration: 10 passed using the real Codex app-server and a local mock Responses API
• npm run build
• npm exec --yes --package=bun@1.3.11 -- npm run bundle:all: Linux, macOS, and Windows, each x64 and arm64
• Repository run-codex smoke helper against real Codex and a local mock model: Flex configured in config.toml reached the model as service_tier: "flex"

Regression coverage exercises the actual new, resume, load, and fork paths, repeated prompts, explicit Fast overrides, both Fast aliases, and empty-session resume/load. It distinguishes an unavailable tier from an explicit null or replacement tier. The new empty-session tests reproduced four unit-test failures before the fix. The real local-gateway tests use Flex because the anonymous model catalog does not advertise Ultrafast; Ultrafast is covered at the app-server request boundary.

Earlier validation at 0a37c04 included a built-adapter protocol probe preserving Ultrafast with Fast disabled. An authenticated Codex 0.164.0 smoke test at 8ccbf8a accepted Ultrafast for GPT-6 Astra and completed one short prompt. These earlier probes were not rerun against the final commit. Cross-platform bundles were built locally; they were not executed on every target OS.