feat(xai): wire reasoning effort through the Grok adapter - #6627

Summary

• Declare capabilities.reasoningEffort on the four Grok models that actually accept it, with per-model values
• Forward reasoning_effort from providers/xai/index.ts, so the flag stops being dead for xAI
• Fix grok-4.5's missing cachedInput ($0.30/M) and add a provider TSDoc header

Before this, xAI declared no reasoningEffort capability and the adapter only ever sent temperature and max_completion_tokens. The Agent block's effort dropdown never appeared for Grok, and any effort value was silently dropped.

Values are live-verified, not doc-derived

I probed every live Grok model × every effort value against POST /v1/chat/completions with the staging hosted key. xAI's docs are wrong in three places:

none is a real vendor value on the two models that list it (reasoning_tokens: 0), and is rejected by 4.5/4.6 — hence the per-model lists rather than one shared array. auto is rejected by every model (Invalid reasoning effort), which is why Sim's auto sentinel must never be forwarded.

Also confirmed live that reasoning_effort co-exists with everything the adapter sends: tools + tool_choice auto/forced/none, response_format: json_schema, temperature, and streaming. And that xAI streams raw reasoning_content deltas, which is what the generated docs table now claims.

Placement

The guard sets reasoning_effort on basePayload, which every derived payload spreads — so it survives the tool loop, the response-format pass, and the streaming request. The new test asserts exactly that on all four paths; I verified all four fail when the guard is removed.

Forwarding an undeclared value is intentional

sanitizeRequest drops reasoningEffort when a catalogued model lacks the capability, but does not check the value against that model's values list. That is a deliberate, documented policy, not a gap — providers/index.test.ts pins it:

an unrecognized level is forwarded rather than dropped: the provider answers with an error naming the values it accepts, instead of Sim silently substituting the model default and quietly corrupting a sweep.

This PR is itself evidence for that policy: xAI's docs omit xhigh for grok-4.5 and don't list grok-4.3 as supporting the parameter at all, yet the API accepts both. A stricter sanitizer would have silently discarded working values.

The practical consequence is that switching a block from grok-4.3 (which takes none) to grok-4.6 (which rejects it) leaves the stored none in place — dependsOn refetches the dropdown options but does not clear the value — and xAI answers with a clear 400 naming the accepted values. That is the intended behavior.

Type of Change

• New feature

Testing

New providers/xai/index.test.ts (5 tests) covering forward-when-set, omit-on-auto, omit-when-unset, and persistence across the tool loop / response-format / streaming payloads. Full provider suite green (1483 tests), plus lint, type-check, and agent-stream-docs:check.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)