feat: add LiteLLM as AI gateway provider

[RheagalFire]

Summary

• Adds LiteLLM as a first-class AI gateway provider, giving T3MP3ST access to 100+ LLM providers (AWS Bedrock, Azure, Vertex AI, Groq, Mistral, Cohere, etc.) through a single OpenAI-compatible proxy.
• Additive only - zero changes to existing providers.

Changes

• src/types/index.ts - Added 'litellm' to the LLMProvider union type
• src/llm/index.ts - Added LiteLLMAdapter (extends OpenRouterAdapter since LiteLLM proxy speaks the same OpenAI-compatible wire protocol), factory function createLiteLLMBackbone(), registered in createAdapter() switch and createBestAvailableBackbone() priority chain
• src/config/index.ts - Added LiteLLM config block (baseUrl, defaultModel), LITELLM_API_KEY/LITELLM_BASE_URL/LITELLM_MODEL env vars, AVAILABLE_MODELS entry, hasLiteLLMProxy() helper, and wired through all provider helper methods (getLLMConfig, getConfiguredProviders, buildFallbackChain, setDefaultProvider, setDefaultModel, exportConfig, createEnvTemplate)
• src/setup.ts - Added interactive setupLiteLLMProxy() wizard (prompts for proxy URL, optional API key, default model; tests connection with save-anyway fallback)
• src/index.ts - Exported createLiteLLMBackbone

Tests

1. TypeScript type check: npx tsc --noEmit passes clean (zero errors).

2. Full test suite: All 328 tests pass across 28 test files.

 Test Files  28 passed (28)
      Tests  328 passed (328)
   Duration  3.53s

3. Live E2E against LiteLLM proxy (Azure Foundry backend):

Proxy config: litellm --config config.yaml --port 4111 routing anthropic/claude-sonnet-4-6 through Azure Foundry.

=== T3MP3ST LiteLLM E2E Test ===
Provider: litellm
Model: anthropic/claude-sonnet-4-6
Validation: {"valid":true}

--- Test 1: Simple chat ---
Content: "4"
Model: anthropic/claude-sonnet-4-6
Usage: {"promptTokens":20,"completionTokens":5,"totalTokens":25}
Finish reason: stop

--- Test 2: Streaming ---
hello world
Stream content: "hello world"

--- Test 3: Tool calling ---
Content: "I'll look up all the DNS records for example.com..."
Tool calls: [
  { "id": "toolu_01C4B2vpP8AXiYinmZUDwqra", "name": "dns_lookup", "arguments": { "domain": "example.com", "type": "A" } },
  { "id": "toolu_014UHdpekiLj1NbiordiW8Bn", "name": "dns_lookup", "arguments": { "domain": "example.com", "type": "AAAA" } },
  { "id": "toolu_01HnyTDQHgsFjckXqyi2TT6w", "name": "dns_lookup", "arguments": { "domain": "example.com" } }
]
Finish reason: length

=== All tests passed ===

Full chain verified: LLMBackbone('litellm') -> LiteLLMAdapter -> fetch(proxy/v1/chat/completions) -> LiteLLM proxy -> Azure Foundry -> Claude Sonnet 4.6 -> parsed LLMResponse. Chat, streaming, and tool calling all work.

Risk / Compatibility

• Additive only. Existing providers untouched.
• LiteLLM is detected via LITELLM_BASE_URL env var or explicit config - no auto-detection from a default localhost URL (avoids false positives).
• Sits in the createBestAvailableBackbone() priority chain between Venice and Anthropic.
• Participates in the model fallback ladder when TEMPEST_MODEL_FALLBACK is enabled.

Example usage

Environment variables:

export LITELLM_BASE_URL=http://localhost:4000/v1
export LITELLM_API_KEY=sk-...        # optional, depends on proxy config
export LITELLM_MODEL=anthropic/claude-sonnet-4-6  # optional, defaults to gpt-4o

CLI:

npx t3mp3st --provider litellm --model anthropic/claude-sonnet-4-6

Programmatic:

import { createLiteLLMBackbone } from 't3mp3st';

const llm = createLiteLLMBackbone(
  'sk-my-proxy-key',              // optional API key
  'anthropic/claude-sonnet-4-6',  // model (any litellm model string)
  'http://localhost:4000/v1',     // proxy base URL
);

const response = await llm.prompt('Plan a recon sweep of example.com');

Setup wizard:

npx t3mp3st setup
# Select "LiteLLM Proxy" from the provider list
# Enter proxy URL, optional API key, and default model

[RheagalFire]

cc @elder-plinius - would love your review when you get a chance!

[jmagly]

Reviewed the wiring — it's logically sound (LiteLLM via the OpenRouter-compatible adapter, base URL http://localhost:4000/v1 → ${baseUrl}/chat/completions, and treating the provider as configured by base-URL presence rather than a key is a nice touch). Main thing is it's currently conflicting — needs a rebase onto main. It also collides with the other provider PRs on llm/config/types (#55, and #57), so sequence it with whichever provider PR lands first and rebase so the union / Record<LLMProvider> picks up both keys.