OpenAI subscription runtime

AI OS includes a local, script-only provider named codex-account. It sends text-analysis requests through a locally authenticated OpenAI account and the ChatGPT-backed Codex Responses transport.

When the account can use the requested model, this path uses the account’s ChatGPT/Codex subscription allowance. It is separate from API-key billing. OpenAI still controls model entitlement, quotas, and reset windows.

Provider status

Provider	Current behavior
`codex-account`	Implemented live text runtime through stored account authentication
`mock`	Deterministic credential-free runtime for local checks and tests
`disabled`	Explicit no-analysis state with fallback guidance
`openai-api`	Readiness-only placeholder; it does not send analysis requests

Quick start

Run from the repository root:

bun run auth:openai -- login
bun run auth:openai -- status
bun run codex:smoke -- --model gpt-5.6-sol --json

Login starts a browser-account flow with a loopback callback and stores the result in ignored local state. A successful login proves authentication was stored; it does not prove that every model is included in the account plan.

The standalone smoke command does not read a Dream or Trend Finder model setting. Pass the intended model explicitly when you need the smoke to match a consumer.

Authentication lifecycle

Command	Purpose
`bun run auth:openai -- status`	Report a redacted missing, expired, invalid, or ready state
`bun run auth:openai -- refresh`	Refresh stored credentials for diagnostics
`bun run auth:openai -- reauth`	Replace a rejected grant with a fresh login
`bun run auth:openai -- logout`	Remove the configured local authentication state
`bun run runtime:check-private`	Confirm local runtime artifacts remain ignored and untracked

Runtime consumers try to refresh expired credentials automatically. A request that receives an authentication rejection can refresh and retry once. If the refresh grant is rejected, use reauth.

Consumers and non-consumers

The shipped runtime can be used by Dream and Trend Finder. It is not the runtime for Hermes chat, Floating Oracle, Voice, or the Intelligence portal.

Configured reasoning effort is currently validated runtime metadata. The Codex transport does not yet forward a reasoning field, so the setting must not be treated as proof of a backend control.

Recovery

State	Recovery
Authentication required	Run `bun run auth:openai -- login`
Expired authentication	Let automatic refresh run, then use `refresh` or `reauth`
Unsupported model	Select a model available to the signed-in account and rerun the smoke
Usage limit	Wait for the provider-reported reset or use another implemented runtime
Timeout or server error	Check connectivity and retry; repeated failure remains a provider issue
`openai-api` selected	Choose `codex-account`, `mock`, or `disabled`

Privacy boundary

Credentials, account identifiers, request material, and provider bodies stay outside browser state, demo fixtures, scheduler summaries, and normal command output. Requests use store: false, but provider processing still leaves the machine. Submit only material appropriate for the consumer’s intended privacy boundary.