Bring your own model (BYOM)

codingassist.bot is provider-agnostic by design. Every LLM-touching step calls a Provider interface; concrete adapters exist for OpenAI, Anthropic, Azure OpenAI, and Bedrock — and you can plug your own.

The contract

interface Provider {
  id: string;
  complete(prompt: Prompt, opts: { temperature: 0; seed: number }): Promise<Completion>;
  embed(text: string): Promise<Embedding>;
}

Three properties matter:

1. temperature: 0 is hard-coded — non-determinism breaks replay
2. A seed is supplied for any provider that supports it
3. The output is schema-validated before it leaves the plane

Why?

• Cost control — pin Stage 1 to a cheap model, Stage 2 to a strong one
• Compliance — keep regulated workloads on Azure OpenAI in your tenancy
• Resilience — survive a provider outage by routing to a fallback

Configuration

# codingassist.policy.yaml
providers:
  context: { provider: openai, model: gpt-4o-mini }
  intent:  { provider: anthropic, model: claude-sonnet-4-5 }
  security: { provider: azure-openai, deployment: gpt-4o, region: eu-west }
 
fallback:
  intent: { provider: openai, model: gpt-4o }

Related

• Use BYOM — OpenAI
• Multi-tenancy