fix(service-tier): make /fast work on ChatGPT-subscription Codex models

[eddieparc]

Fixes #545. Reverses the conclusion of my own #503: the ChatGPT Codex backend does offer a priority ("Fast") tier to subscription accounts, so /fast should work there.

Why the previous notice was wrong

/fast is registered for openai-codex but switches through an openai-codex/<id>-fast catalog sibling that generate-models.ts only emits for the direct openai provider. #503 turned that dead end into a notice claiming priority tier is unavailable on a ChatGPT subscription. Its evidence was the SSE service_tier echo — and that echo carries no information:

sent	response.created echo	response.completed echo	status
priority	auto	default	200
(absent)	auto	default	200
default	auto	default	200
auto / flex / scale	—	—	400 Unsupported service_tier

Identical echoes for priority and for sending nothing, so the echo cannot distinguish the tier that served the request.

What the backend and the first-party client actually do

GET https://chatgpt.com/backend-api/codex/models?client_version=0.145.0 with originator: codex_cli_rs and a live ChatGPT Pro token (chatgpt_plan_type: pro) returns, per model:

model	service_tiers	additional_speed_tiers
gpt-5.6-sol / -terra / -luna, gpt-5.5, gpt-5.4	[{"id":"priority","name":"Fast","description":"1.5x speed, increased usage"}]	["fast"]
gpt-5.4-mini, gpt-5.3-codex-spark, codex-auto-review	[]	[]

"1.5x speed, increased usage" is the backend's own description of the tier, returned to a subscription account.

Official @openai/codex 0.145.0 on ChatGPT OAuth (requires_openai_auth = true), routed through a local logging proxy to chatgpt.com:

• service_tier = "priority" in config.toml → POST /backend-api/codex/responses body contains "service_tier": "priority", turn completes (200)
• line removed → body has no service_tier

So this is not an API-key-only field.

The change

When an openai-codex model has no compatible -fast catalog sibling, /fast now flips a session-scoped priority tier instead of reporting the feature unavailable. The existing before_provider_request handler injects it — addServiceTierToPayload already whitelists openai-codex-responses and api/openai-codex-responses.ts already forwards options.serviceTier — so there is no core or provider change. The catalog-variant path for the direct openai provider is untouched.

Semantics:

• session-only, never persisted, reset on session_start
• an explicit model/scoped serviceTier (models.json, provider/id:tier) still wins
• non-Codex providers keep the existing "only available for OpenAI Codex models" warning, and a mid-session hop to a non-OpenAI model stops sending the field

Tests

test/suite/service-tier-extension.test.ts (11 cases, all green) now covers the toggle payload, a mid-session switch to another Codex model keeping the tier, a hop to an Anthropic model dropping it, explicit-tier precedence, and the session_start reset. Each new assertion was mutation-proved — removing the injection fails 3 cases, leaking the tier to every api fails 2, inverting precedence fails 1, dropping the reset fails 1.

Manual QA

Real TUI, subscription OAuth, provider baseUrl pointed at a logging proxy, request bodies decoded:

turn 1 "reply with exactly: one"    -> one    | gpt-5.6-sol  service_tier=<absent>
/fast                               -> Fast mode enabled: gpt-5.6-sol
turn 2 "reply with exactly: two"    -> two    | gpt-5.6-sol  service_tier=priority
/model openai-codex/gpt-5.5         -> Model: gpt-5.5
turn 3 "reply with exactly: three"  -> three  | gpt-5.5      service_tier=priority
/fast                               -> Fast mode disabled: gpt-5.5
turn 4 "reply with exactly: four"   -> four   | gpt-5.5      service_tier=<absent>

npm run check clean; packages/coding-agent test/suite green apart from a pre-existing app-server-thread-handlers-archive timestamp-monotonicity flake that also fails with this change stashed.

Open question

getServiceTierCostMultiplier() in api/openai-codex-responses.ts multiplies displayed cost by 2 (2.5 for gpt-5.5) for priority. On a subscription the dollar figure is notional ((sub) in the footer) but the backend does say "increased usage", so I left the multiplier alone rather than guessing a subscription-specific number. Happy to change it if you have a preference.

---

Summary by cubic

Enables Fast mode on ChatGPT-subscription Codex models by toggling a session-scoped priority tier when no -fast catalog variant exists. This makes /fast work on openai-codex models and replaces the incorrect “unavailable on subscription” notice. Fixes #545.

• Bug Fixes
  • /fast now toggles a session priority tier for openai-codex models without a -fast sibling, with notify messages for enabled/disabled.
  • Injects the tier in before_provider_request for openai-codex-responses; other providers unchanged.
  • Semantics: session-only and reset on session_start; explicit model serviceTier still wins; switching between Codex models keeps the tier; moving to non-OpenAI drops it.
  • Tests cover payload injection, mid-session model switches, dropping on non-OpenAI, explicit-tier precedence, and session reset.

^{Written for commit 144048a. Summary will update on new commits.}