Add Weights & Biases as a provider

[shariqriazz]

• Add wandb provider implementation with OpenAI-compatible API
• Add wandb models with full context window support (maxTokens = contextWindow) - EDIT: As requested this has now been changed to the documented max tokens from providers ( although the wandb provider offers context window as max context - input )
• Add comprehensive error handling with localized messages
• Add thinking token support for reasoning models
• Add wandb API key settings to all language locales
• Add wandb error translations to all supported languages
• Support for multimodal models and reasoning effort
• Proper token usage tracking and cost calculation

---

Important

Add Weights & Biases as a provider with full API support, error handling, and validation updates.

• Provider Addition:
  • Add wandb provider with OpenAI-compatible API in wandb.ts.
  • Support for multimodal models and reasoning effort.
  • Implement token usage tracking and cost calculation.
• API Key Management:
  • Add wandbApiKey to global-settings.ts and provider-settings.ts.
  • Update ApiOptions.tsx and Wandb.tsx for API key input and management.
• Error Handling:
  • Add error translations for wandb in all supported languages in locales.
  • Implement comprehensive error handling in wandb.ts.
• Validation:
  • Update validate.ts to include wandb API key validation.
• Testing:
  • Add tests for WandbHandler in wandb.spec.ts.

^{This description was created by for 862f0390b070a926fc5899ba7847a729af6db303. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Thank you for your contribution! I've reviewed the changes and have some suggestions for improvement. Overall, the implementation follows existing patterns well and includes comprehensive i18n support. The code is clean and well-structured.

[roomote]

Is the 20% maxTokens calculation intentional for all W&B models? I noticed this differs from other providers like OpenRouter which use varying percentages. Could we document why this specific ratio was chosen, or should it vary per model?

[roomote]

Consider at least logging these parsing errors for debugging purposes instead of silently ignoring them:

[roomote]

These token estimates seem quite rough (4 chars per token). Could we use a more accurate tokenizer, or at least document why these specific ratios were chosen? This could impact cost calculations significantly.

[roomote]

The test coverage could be improved. Consider adding tests for:

• Thinking token stripping functionality ()
• XmlMatcher integration for reasoning models
• Image content handling in messages
• Actual streaming response parsing with realistic data

These are critical features that should have test coverage.

[roomote]

The temperature is clamped to 0-2 range with a comment "W&B typically supports 0-2 range". Should this be configurable per model or documented in the model definitions? Different models might have different optimal ranges.

[roomote]

For consistency, should this also use the localized error format like in ?

[roomote]

Consider adding JSDoc comments to the main class and its public methods. This would improve maintainability and help other developers understand the API better.

[daniel-lxs]

Is this intentional? I think these models have documented maxTokens.
Can you change these so that it matches the actual value of the models?

[daniel-lxs]

Is the failing unit test legit?

[shariqriazz]

Is the failing unit test legit?

Unrelated to my changes i don't know why this keeps happening

[hannesrudolph]

We are not adding this provider at this time.