feat: multi-model support — multiple endpoints from one server

[yossiovadia]

Summary

Support running multiple model identities from a single Katan server, so gateways can test routing between different "models" without spawning separate processes.

Proposed Interface

# Multiple models with different behaviors
llm-katan --models gpt-4o,claude-sonnet,llama-3-70b --backend echo

# Different response characteristics per model
llm-katan --models "gpt-4o:fast,claude-sonnet:verbose,llama-3-70b:slow"

Behavior

• /v1/models returns all configured model names
• Requests with model: "gpt-4o" get responses with model: "gpt-4o"
• Each model could have different simulated characteristics (response length, latency, token usage)
• Unknown model names return 404 (tests error handling)

Use Cases

• Test model routing in AI gateways (VSR, LiteLLM, etc.)
• Test fallback chains (primary → secondary → tertiary)
• Test load balancing across models
• Test model-specific error handling

[yossiovadia]

🎳 OpinAI — Bug Reproduction Report

Verdict: ✅ NOT A BUG — Feature Request

Analysis: This requests multi-model support from a single server instance (--models gpt-4o,claude-sonnet,llama-3-70b). Currently llm-katan runs with a single --model identity. This is a new feature to support multiple model names with different behavior profiles. No existing behavior is broken.

Testing: N/A — requests new multi-model capability that does not exist yet.

---

"That's just, like, your opinion, man." — OpinAI