Feat/add litellm provider

[RheagalFire]

Summary

Adds LiteLLM as a first-class provider, enabling access to 100+ LLM providers (OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Mistral, Cohere, and more) through Mercury's existing provider system via a LiteLLM proxy.

Changes

• src/utils/config.ts - Added litellm to ProviderName union, config defaults (LITELLM_API_KEY, LITELLM_BASE_URL, LITELLM_MODEL, LITELLM_ENABLED), keyless proxy support in isProviderConfigured()
• src/providers/registry.ts - Registered litellm as OpenAICompatProvider
• src/providers/openai-compat.ts - Fixed isAvailable() for keyless proxies (checks baseUrl instead of apiKey for litellm/ollama), added formatApiError() with specific messages for 401/404/429
• src/utils/provider-models.ts - LiteLLM-specific error message for model fetch failures, added litellm to model ID filter
• src/web/api/providers.ts - Fixed /api/providers/:name/test endpoint for keyless proxy testing

Tests

Tests (66/67 passing):
66 pass. 1 pre-existing failure in tokens.test.ts (locale-dependent number formatting, unrelated to this change).

Error handling:

• 401: [litellm] Authentication failed - check your API key or proxy credentials.
• 404: [litellm] Model not found - verify the model name and base URL.
• 429: [litellm] Rate limited - wait a moment before retrying.
• Keyless proxy: isAvailable() now checks baseUrl instead of apiKey
• Model fetch failure: Mercury could not fetch models from the LiteLLM proxy. Please check the base URL and ensure the proxy is running.

Risk / Compatibility

• Additive only. Reuses existing OpenAICompatProvider class.
• Opt-in via LITELLM_ENABLED=true env var.
• Keyless proxy fix also benefits Ollama local provider.

Example usage

1. Start a LiteLLM proxy (manages models + API keys server-side):

pip install litellm
litellm --model openai/gpt-4o --model anthropic/claude-sonnet-4-20250514
# Proxy running at http://localhost:4000

2. Configure Mercury via env vars:

export LITELLM_ENABLED=true
export LITELLM_BASE_URL=http://localhost:4000/v1
export LITELLM_API_KEY=sk-your-litellm-master-key  # optional for keyless proxies
export LITELLM_MODEL=openai/gpt-4o

3. Use in code - Mercury routes through the LiteLLM provider:

import { createProvider } from './providers/registry.js';
import { getConfig } from './utils/config.js';

const config = getConfig();
const provider = createProvider('litellm', config.providers.litellm);

// Non-streaming
const response = await provider.generateText(
  "What is the capital of France?",
  "You are a helpful assistant."
);
console.log(response.text);     // "The capital of France is Paris."
console.log(response.provider); // "litellm"
console.log(response.model);    // "openai/gpt-4o"
console.log(response.totalTokens); // 42

// Streaming
for await (const chunk of provider.streamText(
  "Tell me a joke",
  "You are a funny assistant."
)) {
  if (!chunk.done) {
    process.stdout.write(chunk.text);
  }
}

4. Under the hood - Mercury calls the LiteLLM proxy via Vercel AI SDK:

// OpenAICompatProvider creates an @ai-sdk/openai client pointing at the proxy
const client = createOpenAI({
  apiKey: 'sk-your-litellm-key',
  baseURL: 'http://localhost:4000/v1',
});

// generateText/streamText call the proxy's /v1/chat/completions endpoint
// The proxy handles routing to the right provider based on the model name
const result = await generateText({
  model: client('openai/gpt-4o'),
  system: "You are a helpful assistant.",
  prompt: "What is 2+2?",
});