Mistral's API is OpenAI-compatible, making it a straightforward fourth provider for Inference.complete. With four providers (three structurally identical OpenAI-compatible endpoints), this is also the right moment to refactor _call_inference_provider() to a table-driven registry rather than bolting on another elif branch and deferring the cleanup to a fifth.

Background

The _call_inference_provider() helper in vera/codegen/api.py currently has three elif branches. The Moonshot branch is the OpenAI branch with a different URL, key, and default model. Mistral is the same shape again:

• Endpoint: https://api.mistral.ai/v1/chat/completions
• Auth: Bearer token (identical pattern)
• Response: choices[0].message.content (identical to OpenAI/Moonshot)
• Env var: VERA_MISTRAL_API_KEY
• Default model: mistral-small-latest (consistent with the cheap/fast defaults: Haiku, GPT-4o-mini, kimi-k2-0905-preview)

At four providers — three of which are structurally identical — the elif chain wants to become a table. At five or six it becomes actively painful to extend.

Proposed design

Replace the elif chain and the per-key function parameters with a _ProviderConfig dataclass and a _PROVIDERS registry dict:

@dataclass(frozen=True)
class _ProviderConfig:
    env_key: str        # e.g. "VERA_MISTRAL_API_KEY"
    url: str            # chat completions endpoint
    default_model: str  # cheap/fast default
    auth_style: str     # "anthropic" | "bearer"
    response_style: str # "anthropic" | "openai"

_PROVIDERS: dict[str, _ProviderConfig] = {
    "anthropic": _ProviderConfig(
        env_key="VERA_ANTHROPIC_API_KEY",
        url="https://api.anthropic.com/v1/messages",
        default_model="claude-haiku-4-5-20251001",
        auth_style="anthropic",
        response_style="anthropic",
    ),
    "openai": _ProviderConfig(
        env_key="VERA_OPENAI_API_KEY",
        url="https://api.openai.com/v1/chat/completions",
        default_model="gpt-4o-mini",
        auth_style="bearer",
        response_style="openai",
    ),
    "moonshot": _ProviderConfig(
        env_key="VERA_MOONSHOT_API_KEY",
        url="https://api.moonshot.ai/v1/chat/completions",
        default_model="kimi-k2-0905-preview",
        auth_style="bearer",
        response_style="openai",
    ),
    "mistral": _ProviderConfig(
        env_key="VERA_MISTRAL_API_KEY",
        url="https://api.mistral.ai/v1/chat/completions",
        default_model="mistral-small-latest",
        auth_style="bearer",
        response_style="openai",
    ),
}

_call_inference_provider() becomes a generic dispatcher (~20 lines) that looks up the config and delegates to one of two request builders: _call_anthropic_style() or _call_openai_compat_style().

Auto-detection in execute() simplifies to iterating _PROVIDERS in priority order and checking which env_key is set — no more per-key parameters on _call_inference_provider().

Adding provider five (xAI Grok, tracked in #425) becomes a one-row addition to _PROVIDERS.

Change list

vera/codegen/api.py

• Add _ProviderConfig dataclass and _PROVIDERS registry dict
• Replace _call_inference_provider(provider, prompt, model, anthropic_key, openai_key, moonshot_key) with _call_inference_provider(provider, prompt, model, env) that looks up the provider's key from env via _PROVIDERS
• Extract _call_anthropic_style() and _call_openai_compat_style() helpers
• Simplify auto-detection loop: iterate _PROVIDERS, pick first with a set key
• Update "no provider configured" error to enumerate all _PROVIDERS keys

Tests — tests/test_codegen.py

• test_inference_mistral_auto_detect — copy the Moonshot auto-detect test, change key name and assertion
• test_mistral_provider — unit test for the Mistral config entry in _PROVIDERS

Documentation sweep

• spec/09-standard-library.md — one row in the provider table
• SKILL.md — update the Inference effect provider table
• CHANGELOG.md — entry noting refactor + Mistral addition
• README.md — update the env var table / provider list

Out of scope

• Mistral-specific features (function calling, vision, etc.) — Inference.complete is a plain String → Result<String, String> operation
• The embed operation — already deferred (Host-import float array return: alloc_result_ok_float_array infrastructure #373)
• Browser runtime — already has its own provider dispatch; keep in sync but separate PR