Add Crusoe provider

[acheamponge]

Adds Crusoe Managed Inference as a provider.

Provider

• OpenAI-compatible API at https://api.inference.crusoecloud.com/v1 (@ai-sdk/openai-compatible, CRUSOE_API_KEY)
• Docs: https://docs.crusoecloud.com/managed-inference/overview
• Official logo SVG from Crusoe's docs site

Models (14)

All chat models from the current catalog. Each entry references an existing canonical models/ definition via base_model (zhipuai GLM-5.1/5.2, the nvidia Nemotron 3 family, google/gemma-4-31b-it, deepseek V4 Flash/Pro, meta/llama-3.3-70b-instruct, openai/gpt-oss-120b, alibaba/qwen3-235b-a22b, moonshotai/kimi-k2.6).

Also adds one new canonical entry: models/deepseek/deepseek-v3-0324.toml (no existing provider defined it).

Pricing intentionally omitted. Excluded: nvidia/Nemotron-3-VoiceChat (speech-to-speech).

Context

I maintain Crusoe's ecosystem integrations (LiteLLM provider: BerriAI/litellm#23195, SGLang: sgl-project/sglang#20475). Happy to adjust anything to match registry conventions.

[github-actions]

Action items

• [high] [violation] providers/crusoe/models/{deepseek-ai/DeepSeek-V4-Flash,deepseek-ai/DeepSeek-V4-Pro,google/gemma-4-31b-it,moonshotai/Kimi-K2.6,nvidia/Nemotron-3-Nano-30B-A3B,nvidia/Nemotron-3-Nano-Omni-Reasoning-30B-A3B,nvidia/Nemotron-3-Super-120B-A12B,nvidia/Nemotron-3-Ultra-550B,openai/gpt-oss-120b,qwen/Qwen3-235B-A22B,zai/GLM-5.1,zai/GLM-5.2}.toml - Check: Reasoning models must declare reasoning_options. Why: Each of these 12 files inherits reasoning = true from its base_model metadata (verified in the corresponding models/... entries), but reasoning_options is provider-specific and is never inherited — model-metadata files do not even carry the field, and generate() only deep-merges inheritable metadata. The merged provider model therefore has reasoning = true with reasoning_options undefined, which the schema refine rejects ("Must set reasoning_options when reasoning is true"), so bun validate fails for all 12. Action: Add a reasoning_options array to each file. Crusoe is an OpenAI-compatible endpoint, so audit whether it forwards reasoning controls (e.g. reasoning_effort, chat_template_kwargs.enable_thinking, thinking.budget_tokens); if no caller control is verified for that model, reasoning_options = [] is the correct declaration.
• [high] [violation] providers/crusoe/logo.svg:1 - Check: New provider logos must use currentColor, have no fixed size, and preferably a square viewBox. Why: The SVG declares width="121" height="30" (fixed size) and every path uses fill="black" (hardcoded color), so it will not adapt to light/dark themes and will not scale; the viewBox="0 0 121 30" is also non-square. A compliant logo is a hard blocker for new providers. Action: Remove the width/height attributes, replace each fill="black" with fill="currentColor", and prefer a square viewBox.
• [low] [possible mistake] .pr-review/pull-request.json:4 (PR body) vs providers/crusoe/models/**/*.toml - Check: PR body claims must match the final diff. Why: The body advertises "pay-as-you-go pricing from https://www.crusoe.ai/cloud/pricing#Managed-Inference-pay-as-you-go (input/output/cache_read per 1M tokens)" with specific per-model prices and states Nemotron-3-Ultra-550B was excluded, but the second commit ("Remove pricing; add Nemotron-3-Ultra-550B") deleted every [cost] block and added that model. The final diff ships 14 Crusoe models with no pricing and includes the model the body said was excluded. Action: Reconcile the body with the diff — either restore the cited per-1M-token pricing on each model (mapping the pricing-page citation to each value) or update the body to explain the pricing removal and the Ultra-550B inclusion.

[github-actions]

Action items

• [medium] [violation] providers/crusoe/models/deepseek-ai/DeepSeek-V4-Flash.toml:2 (and the 11 other reasoning-model TOMLs) - Check: Rationale/citation comments must live in a leading comment block above the first key. Why: The three-line comment explaining reasoning_options = [] is placed after base_model (the first key) and above reasoning_options. AGENTS.md requires source citations/rationale to be in a leading header block above the first key because sync serialization silently deletes comments placed between keys. This affects all 12 reasoning-model files added in patch 3. Action: Move the comment block above base_model so it becomes a leading header, or move the rationale/citation into the PR body (per the audit-reasoning-options skill, citations belong in the PR body unless kept as a leading TOML header).
• [low] [possible mistake] providers/crusoe/provider.toml / PR body - Check: Data-changing PRs should cite direct provider sources for material factual claims that cannot otherwise be reviewed. Why: The 14-model catalog and exact model IDs/namespacing (e.g. zai/GLM-5.1, nvidia/Nemotron-3-Ultra-550B) are material claims not verifiable from the general docs-overview link alone, and the reasoning_options = [] claim needs a provider reasoning-API citation per the audit skill's evidence standard. Action: Add a direct citation to the Crusoe catalog/models listing that enumerates these models and confirms their IDs, and cite the specific reasoning API documentation (or confirm its absence) supporting the empty reasoning_options arrays.