Add Friendli as inference provider

[Lee-Si-Yoon]

Summary

FriendliAI provides an OpenAI-compatible Chat Completions API at https://api.friendli.ai/serverless/v1/chat/completions. The model list is available at https://api.friendli.ai/serverless/v1/models and includes models already in the Codebuff catalog (GLM 5.2, MiniMax M2.5) under org/model IDs (zai-org/GLM-5.2, MiniMaxAI/MiniMax-M2.5).

Why

Adding Friendli as a provider would let Codebuff route these models through Friendli's serverless endpoint directly, which may offer different pricing/latency characteristics compared to existing routes (OpenRouter, CrofAI, Fireworks).

Friendli-specific considerations

• Model IDs: org/model format (e.g. zai-org/GLM-5.2). Neither ALLOWED_MODEL_PREFIXES (common/src/constants/model-config.ts) nor providerModelNames currently covers a friendli prefix.
• Reasoning: Friendli handles reasoning differently from OpenRouter/DeepSeek. Instead of reasoning field or <think> tags, it uses:
  • parse_reasoning (boolean) — enables deterministic reasoning parsing
  • include_reasoning (boolean) — controls whether parsed reasoning appears in response
  • chat_template_kwargs.enable_thinking (boolean) — toggles reasoning for controllable models (e.g. GLM-5.2)
  • Response: message.reasoning_content (non-streaming) / delta.reasoning_content (streaming)
  • Docs: https://friendli.ai/docs/guides/reasoning
• API reference: https://friendli.ai/docs/openapi/model-apis/chat-completions

What's needed

Since provider routing lives in web/src/llm-api/ (private), I can contribute the public-side changes if a maintainer is open to adding the route:

• friendli to ALLOWED_MODEL_PREFIXES
• Friendli model IDs to model-config.ts (or use the existing zai-org/ prefix for GLM)

Is there interest in adding this route?