xAI policy 403 aborts Codex turns and retry loops the same user message #5277

Summary

xAI sometimes refuses a turn with HTTP 403 and the body I can't help with that request. instead of HTTP 200 + finish_reason: content_filter.

Codex treats that 403 as a transport failure. The user message is never recorded as a completed assistant turn. Retrying the same thread resends the same pending user text and the 403 repeats. A new Codex thread on the same account and xai/grok-4.6 succeeds.

Grok Build (native Grok CLI) does not get stuck: the refusal lands as a normal assistant message and the next user turn proceeds.

This is not a plan/subscription gate. Subscription 403s use different wording (need a Grok subscription, credits, not allowed to use this model). Those should stay errors.

Reproduction

1. Codex → OpenCodex → xai/grok-4.6 (OAuth or key).
2. In a long thread, send a request xAI's policy classifier refuses.
3. Observe unexpected status 403 Forbidden: Provider error 403: I can't help with that request., url: http://127.0.0.1:10100/v1/responses.
4. Retry / “continue” in the same thread → same 403 (~1s, no tokens).
5. New thread, trivial prompt → 200.

OpenCodex usage log for the failing request: status: 403, errorCode: permission_denied, upstreamError: Provider error 403: I can't help with that request., closeReason: non_stream.

Dashboard copy currently labels this as a plan/subscription gate. That label is wrong for this body.

Expected

The refusal is recorded as a completed (incomplete/content_filter) assistant turn. The next, different user message proceeds.

Actual

Codex surfaces a 403 transport error. The user message stays pending. Retries loop.

Version

OpenCodex 2.31.0 (reproduced). Latest 2.60.0 still maps this 403 through the generic error path (#5200 makes refusals terminal errors, which is correct for retry, but does not rewrite them into an in-band turn).

Proposed fix

Allowlist the refusal sentence. On HTTP 403 with that body, rewrite to a Chat Completions 200 with finish_reason: content_filter so the existing openai-chat adapter emits incomplete_details.reason = content_filter. Leave subscription/entitlement/model-access 403s unchanged.