xAI's Grok uses an OpenAI-compatible endpoint, making it a one-row addition to the provider registry introduced in #413.

Blocked by #413 — the _PROVIDERS registry must land first.

Background

xAI's Grok API is OpenAI-compatible:

• Endpoint: https://api.x.ai/v1/chat/completions
• Auth: Bearer token (identical to OpenAI/Moonshot/Mistral pattern)
• Response: choices[0].message.content (identical)
• Env var: VERA_XAI_API_KEY
• Default model: grok-3-mini-fast-beta (cheap/fast tier — analogous to Haiku, GPT-4o-mini, mistral-small-latest; their flagship is grok-4-0709 but that's the expensive tier)

Once the _PROVIDERS registry exists, this is literally one dict entry:

"xai": _ProviderConfig(
    env_key="VERA_XAI_API_KEY",
    url="https://api.x.ai/v1/chat/completions",
    default_model="grok-3-mini-fast-beta",
    auth_style="bearer",
    response_style="openai",
),

Auto-detection, error messages, and the dispatch logic all inherit from the registry with no further changes.

Change list

vera/codegen/api.py

• Add "xai" entry to _PROVIDERS (after Add Mistral AI provider to the Inference effect #413 lands)

Tests — tests/test_codegen.py

• test_inference_xai_auto_detect — copy pattern from other provider auto-detect tests
• test_xai_provider — unit test for the xAI config entry

Documentation sweep

• spec/09-standard-library.md — one row in the provider table
• SKILL.md — one row in the provider table
• README.md — one row in the env var table
• CHANGELOG.md — entry for xAI support

Out of scope

• Grok-specific features (real-time web search, etc.) — Inference.complete is plain String → Result<String, String>
• Browser runtime — separate PR, same pattern as other providers