plat 416

← Platform issue index

PLAT-416 — Codex gpt-5.3-codex-spark refused for ChatGPT accounts; allowed-models 422 on a stale saved model

Coordination	Value
State	fixed on main; deploy pending (Excellence and the other servers)
Severity	P2 (a chat failed with a 422 although the agent had answered)
Date	2026-10-04
Owner	coding-agent-bridge
Related	PLAT-407 (allowed models per account)

Problems

1. Codex signed in with a ChatGPT account answers The 'gpt-5.3-codex-spark' model is not supported when using Codex with a ChatGPT account. The platform still offered Spark: the provider's visible Codex catalog, the auto-published fast models, the web-search provider list, and the image analysis and generation lists.
2. On Excellence the Admin-managed Codex account was restricted to gpt-6-luna. A Code chat whose project still had Spark saved answered correctly (the terminal showed GPT-6 Luna) but the chat showed Request failed with status code 422: gpt-5.3-codex-spark is not allowed on Admin-managed account; allowed: gpt-6-luna.

Cause (2)

constrainProductChatModel treated a model that differed from the conversation's bound model as a new pick and refused it. After the first turn fell back to Luna the conversation was bound to Luna, while the browser kept re-sending the saved Spark, so every later message counted as a new pick of a disallowed model.

Done

• Spark is gone from the visible Codex catalog (multi-llm-provider-go daffefd; its pricing metadata stays so a saved selection still resolves), from the auto-published models (now gpt-5.4-mini only), and from the web-search, image analysis and image generation lists. Contract tests and docs use gpt-5.4-mini. agent_go/go.mod and go.sum pin that provider commit.
• constrainProductChatModel never refuses: a model the account does not allow runs on the first allowed one (logged). TestAllowedModelsProductChatTurn covers a picked and a re-sent model, no model, and an account without a list.

Left

• Deploy. The Code project that had Spark saved keeps it until its next turn; it then runs on the allowed model.
• The browser still shows the saved model in its picker; syncing the shown selection to the allowed one when the account's list loads is not done.
• The Codex tier shortcuts High, Medium and Low (high is the Codex default model) appear next to the concrete models they resolve to (High = GPT-6.1 Sol, Medium and Low = GPT-6 Luna), so a model shows twice in the catalog. They are the tier system the defaults and a test rely on, so they were not hidden; an allowed list of a concrete model does not cover its shortcut (a pick of low is not matched against gpt-6-luna).
• A failing pkg/codingagentmodels test in the provider repo (a retired Cursor model still listed) fails on a clean main too.