Summary

The providers config section was previously limited to OpenAI-compatible endpoints. This evolves it into a universal provider grouping mechanism that works with any provider type and supports shared model-level defaults.

Example

providers:
  my_anthropic:
    provider: anthropic
    token_key: MY_ANTHROPIC_KEY
    max_tokens: 16384
    thinking_budget: high

agents:
  root:
    model: my_anthropic/claude-sonnet-4-5

Code changes

• Add provider field to ProviderConfig (defaults to openai for backward compatibility)
• Add model-level defaults: temperature, max_tokens, thinking_budget, top_p, frequency/presence_penalty, parallel_tool_calls, track_usage, provider_opts
• Update applyProviderDefaults to propagate all new fields (model-level overrides take precedence)
• Make base_url required only for OpenAI-compatible providers
• Update validation, env var gathering, JSON schema, and examples

Documentation

• Rewrite providers/custom doc as "Provider Definitions" covering all provider types with inheritance examples
• Add "Inheriting from Provider Definitions" section to model config docs
• Update configuration overview, nav, and introduction

Backward compatibility

Fully backward-compatible: when provider is not set, it defaults to openai, and base_url remains required for that case.