feat: add Novita AI as a native LLM provider

[Alex-wuhu]

Summary

• Adds Novita AI as a native LLM provider via its OpenAI-compatible endpoint (https://api.novita.ai/openai)
• New NovitaCompletion class that extends OpenAICompletion — minimal code, inherits all chat completions, streaming, tool-calling, and structured output support
• Registers novita in the provider routing (SUPPORTED_NATIVE_PROVIDERS, provider_mapping, _get_native_provider, etc.)
• Supported models: moonshotai/kimi-k2.5 (default), zai-org/glm-5, minimax/minimax-m2.5

Usage

from crewai import LLM

# With provider prefix
llm = LLM(model="novita/moonshotai/kimi-k2.5")

# Direct vendor/model ID (auto-detected)
llm = LLM(model="moonshotai/kimi-k2.5")

# Explicit provider kwarg
llm = LLM(model="zai-org/glm-5", provider="novita")

API key is read from NOVITA_API_KEY environment variable or passed as api_key constructor parameter.

Design decisions

• Extends OpenAICompletion instead of duplicating code — Novita's API is fully OpenAI-compatible
• Forces api="completions" — the Responses API is OpenAI-specific and not supported by Novita
• Vendor-prefixed model IDs (e.g. moonshotai/kimi-k2.5) are auto-detected via NOVITA_MODELS lookup before the generic /-split logic kicks in, so users don't need the novita/ prefix

Test plan

• Verify LLM(model="novita/moonshotai/kimi-k2.5") routes to NovitaCompletion
• Verify LLM(model="moonshotai/kimi-k2.5") auto-detects Novita provider
• Verify NOVITA_API_KEY env var is picked up correctly
• Verify chat completion and tool calling work end-to-end with a Novita API key
• Verify existing providers are unaffected (no changes to existing provider files)

---

Note

Medium Risk
Introduces a new native provider and modifies LLM provider inference/routing, which could change how some model strings are resolved and instantiated at runtime. Low algorithmic complexity, but affects core LLM initialization paths and requires correct API key/base URL configuration.

Overview
Adds Novita AI as a native LLM provider backed by Novita’s OpenAI-compatible endpoint.

Updates LLM provider routing to recognize novita (including vendor-prefixed model IDs like moonshotai/kimi-k2.5), adds NOVITA_MODELS constants, and registers context window sizes for the supported Novita models.

Introduces NovitaCompletion, a thin wrapper over OpenAICompletion that defaults the base URL to https://api.novita.ai/openai, requires NOVITA_API_KEY, and forces use of the OpenAI Completions API.

^{Written by Cursor Bugbot for commit 2bef65c. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Novita models get wrong context window size

Medium Severity

NovitaCompletion inherits get_context_window_size() from OpenAICompletion, which only checks OpenAI-specific model prefixes (e.g. gpt-4, o1). Since Novita models like moonshotai/kimi-k2.5 don't match any OpenAI prefix, the method returns the default ~6,963 tokens instead of the intended ~111,411 tokens (131072 × 0.85). The entries added to LLM_CONTEXT_WINDOW_SIZES are dead code here — that dict is only consumed by the LLM (litellm) class's override, not by native provider classes. Every other native provider (Anthropic, Gemini, Bedrock) overrides get_context_window_size() with its own model-specific windows; NovitaCompletion needs the same.

Additional Locations (1)

• lib/crewai/src/crewai/llm.py#L314-L318

 

[github-actions]

This PR is stale because it has been open for 45 days with no activity.