Model identity / cutoff inconsistency: requested gpt-6-astra replies with June 2024 knowledge cutoff through a custom Responses gateway

Reported problem

The user selected Astra and is concerned that a much older model (GPT-4o) answered. The concrete record we recovered is a reply claiming a June 2024 knowledge cutoff. We have not established that GPT-4o executed, and do not claim silent model substitution as a proven fact. Please help identify the upstream execution model or the source of the unsupported cutoff claim.

Environment

• Windows 11 x64, build 26200.
• Codex Desktop MSIX 26.901.6511.0; bundled app 26.901.51231; session CLI 0.153.4.
• The session metadata identifies provider codex_local_access.
• The configured provider is a local Cockpit Responses gateway, wire_api=responses, requires_openai_auth=false. It is not a direct connection from Codex to the official OpenAI API. The local listener is owned by cockpit-cliproxy.exe.

Matched evidence

At 2026-09-07T11:17:41.611Z, the user asked in Chinese for the model's knowledge cutoff. At 2026-09-07T11:17:44.669Z, the saved assistant response said: “我的知识截止日期是 2024 年 6 月。” (My knowledge cutoff is June 2024.)

• Corresponding saved turn context: model=gpt-6-astra.
• Codex request trace: model_client.stream_responses_api{model=gpt-6-astra wire_api=responses transport="responses_http" http.method="POST" api.path="/responses"}.
• The loopback /v1/responses request returned HTTP 200 with SSE content type.
• That response's x-request-id was 8df73d29-6b22-499b-91cb-dbabd474e3b5. This is an observed gateway response header, not a verified OpenAI-owned trace ID.
• The gateway's local request ledger contains a same-time entry at 11:17:44Z: model_id=gpt-6-astra, gateway mode sidecar, success 1, HTTP 200, latency 2849 ms. Time correlation is not a proven one-to-one request-ID join.
• An earlier reply in the same session described itself as Codex based on GPT-6, while saying it could not independently verify the precise variant. We found no assistant claim of “GPT-4o” in that recovered exchange.

Evidence boundary

The selected label, request model and local ledger agree on gpt-6-astra. None is cryptographic attestation of the model that executed upstream. The inspected local evidence does not expose a verified upstream response model field for this request. The local sidecar request-log/file-log settings are disabled; no settings were changed to capture new traffic. This investigation does not establish the true knowledge cutoff for Astra either.

Possible explanations remain an incorrect self-description, provider-side alias/routing behavior, or another upstream issue. No one of these is confirmed. This has not been reproduced using a direct official provider, and the gateway is a material confounder.

Requested behavior

Please investigate the unsupported cutoff response where the trace can be resolved, and provide a supported way to distinguish selected/requested model from provider-reported execution model. If a provider reports a different model, show a clear discrepancy rather than leaving the user to infer identity from self-description. If execution identity cannot be verified, expose that limitation accurately.

The user-facing failure is loss of trust in model selection. This report deliberately preserves the uncertainty rather than alleging that a 2024 model was proven to replace Astra. No credentials, full transcript, account identity or user filesystem paths are included.

Start at model_client.stream_responses_api with model=gpt-6-astra, wire_api=responses, and transport=responses_http; inspect how the /responses SSE response and its metadata are handled. Use the supplied gateway trace and ledger context, while noting that the upstream model is not verified. Done means exposing a provider-reported execution model when available, or clearly reporting that execution identity cannot be verified.