xai provider advertises xhigh variant its own SDK schema rejects

[guitaripod]

Version

1.18.18 (Linux, x86_64)

What happens

Picking the xhigh effort variant for xai/grok-4.6 (offered in the picker, present in GET /config/providers) makes every prompt in that session fail instantly with:

AI_InvalidArgumentError: invalid xai provider options
AI_TypeValidationError: Invalid option: expected one of "none"|"low"|"medium"|"high" at path reasoningEffort

Payload opencode built: {"store":false,"promptCacheKey":"...","reasoningEffort":"xhigh"}

Why

Two of opencode's own halves disagree:

1. The catalog advertises variants: { "xhigh": { "reasoningEffort": "xhigh" } } for grok-4.6 (and other post-2025-12 xAI models) — correctly, since xAI's API does support reasoning_effort: "xhigh" on grok-4.6 (docs.x.ai/docs/guides/reasoning).
2. The bundled @ai-sdk/xai provider's zod schema accepts only none|low|medium|high for reasoningEffort, so the request dies at local type validation before any network call.

Expected

Either the schema accepts xhigh (xAI supports it on grok-4.6+) or the catalog stops advertising it. As shipped, the picker offers an option that can never run.

Repro

1. opencode 1.18.18, xAI connected.
2. /models → xai/grok-4.6, effort xhigh.
3. Send any prompt → "invalid xai provider options".

[opencode-agent]

Thanks for the report. Support for passing through xhigh reasoning effort was fixed in #42160 and is included in v1.18.18. That release patches both the xAI Chat and Responses schemas and includes explicit coverage for xhigh, so we can’t reproduce this with the official v1.18.18 build.

This likely came from a stale or different binary, or an unpatched source installation. If it still reproduces, please share the installation method, output of which -a opencode, and the full stack trace so we can identify which build is running. Closing as not reproducible.