Refactor TranslationConfig to embed ModelConfig per chatbot role

[MaleicAcid]

Refactor TranslationConfig to embed ModelConfig per chatbot role

Motivation

LeanTranslator (PRs #132–#139) introduced three independent chatbot roles: primary translator, retry fallback, and Context Review. Each may need a different endpoint, proxy, and extra_body (e.g. disabling Qwen3.6 thinking mode). The current flat design shares base_url_config, proxy across all roles and has no way to pass extra_body from config files at all.

Proposal

Make ModelConfig serialization-friendly, then embed it directly in TranslationConfig.

Convert ModelProvider from Enum to StrEnum so ModelConfig becomes JSON/YAML-native. For Python 3.10 compatibility, use str, Enum multiple inheritance with a __str__ override (verified on 3.10–3.13):

import sys
if sys.version_info >= (3, 11):
    from enum import StrEnum as _StrEnum
else:
    from enum import Enum
    class _StrEnum(str, Enum):
        def __str__(self) -> str:
            return self.value

class ModelProvider(_StrEnum):
    OPENAI = "openai"
    ANTHROPIC = "anthropic"
    GEMINI = "gemini"

ModelConfig.provider typed as ModelProvider | str to allow custom providers (e.g. "litellm") without modifying the enum.

Refactored TranslationConfig:

@dataclass
class TranslationConfig:
    chatbot: ModelConfig | None = None
    retry_chatbot: ModelConfig | None = None
    cr_chatbot: ModelConfig | None = None
    translate_mode: str = "standard"
    enable_cr: bool = True
    chunked_guideline: bool = False
    fee_limit: float = 0.8
    consumer_thread: int = 4
    glossary: str | None = None
    is_force_glossary_used: bool = False

Removed fields: chatbot_model, base_url_config, proxy, retry_model, cr_model — all replaced by per-role ModelConfig with base_url, proxy, extra_body.

Side benefit: provider2chatbot dict becomes dict[str, type[ChatBot]], enabling custom provider registration (provider2chatbot["litellm"] = LiteLLMBot).

Backward compatibility

This is a breaking change. The LeanTranslator fields (translate_mode, cr_model, enable_cr, chunked_guideline) have not been in any PyPI release. The older fields (chatbot_model, base_url_config, retry_model) require mechanical migration. Carrying deprecated fields alongside new ones adds significant __post_init__ complexity for marginal benefit.

Propose releasing as 1.7.0 with a migration guide in the changelog.

[MaleicAcid]

@zh-plus Since this issue was brought in by the LeanTranslator, I'm thinking of just dropping backward compatibility and rolling straight into version 1.7.0. What do you think?
Also, I'd suggest dropping retry_chatbot in this refactor. The retry logic that actually matters already lives inside the chatbot layer — GPTBot._create_chat retries on rate limits, timeouts, and connection errors automatically. What retry_chatbot does is something different: it swaps to a completely different model when an entire chunk fails. But in practice, if a chunk fails, it's usually a prompt issue or the model just can't handle it, and throwing a different model at the same prompt doesn't reliably fix that. It might even make things worse since different models format their output differently.

Looking at the codebase, I don't think anyone is actually configuring retry_model in production. And it adds real complexity — every place that creates a chatbot has to handle the retry variant too, plus the fee tracking with retry_fee_start indexing is the most convoluted part of the whole billing system.

On the LeanTranslator side, the anchor alignment mechanism already handles failures through atomic fill and full retry, so a backup model adds nothing there.
If we're going to make a clean break for 1.7.0, this is the right time to simplify.

[zh-plus]

Thanks for the detailed proposal. I agree with the main direction here.

Refactoring TranslationConfig from globally shared model parameters to per-role embedded ModelConfig makes sense, especially now that LeanTranslator has separate chatbot roles that may need different base_url, proxy, and extra_body settings. I’m also fine with treating this as a breaking change and targeting it for 1.7.0, with the old flat fields removed and documented in a migration guide.

However, I’d prefer to keep the retry fallback model mechanism for now. There are real-world cases where the primary model may refuse to translate certain subtitle content, such as political or violent material, while another fallback model can still handle it. In that scenario, retrying with a different model is not just generic error recovery; it is a useful escape hatch for model-specific refusal behavior.

So my preference is:

• Proceed with the TranslationConfig -> per-role ModelConfig refactor.
• Make it a clean breaking change for 1.7.0.
• Keep retry_chatbot / fallback model support, but adapt it to the new ModelConfig structure.

[MaleicAcid]

@zh-plus All done. Could you release 1.7.0a1 at soon? Having this alpha version would really help me run my tests better.

[zh-plus]

@MaleicAcid I will release 1.7.0a1 soon today.

[MaleicAcid]

@MaleicAcid I will release 1.7.0a1 soon today.

@zh-plus Come on!

[zh-plus]

@MaleicAcid Just released at https://pypi.org/project/openlrc/1.7.0a1.