feat: OpenRouter model expansion with live API + model selector

[rubenmarcus]

Summary

Expand OpenRouter support with updated models, a live model fetcher, and a ralph-starter models command.

Current State

• Default model: anthropic/claude-3.5-sonnet (outdated)
• No model selector or discovery
• File: src/llm/providers.ts

Changes

A. Update default model

Change defaultModel from anthropic/claude-3.5-sonnet to anthropic/claude-sonnet-4-5

B. Add --model flag

ralph-starter run --model anthropic/claude-opus-4-6 --from github --issue 42

C. Add ralph-starter models command

Fetches live data from OpenRouter API:

ralph-starter models
ralph-starter models --tier frontier
ralph-starter models --refresh  # Force API fetch

D. Curated model tiers

Tier 1 — Frontier:

• anthropic/claude-opus-4-6 — 1M context
• anthropic/claude-sonnet-4-5 — Fast, great for coding
• openai/gpt-5.3-codex — 400K context
• google/gemini-3 — 1M context
• deepseek/deepseek-v3.2 — Strong open-source

Tier 2 — Cost-effective:

• google/gemini-2.5-flash — Fast
• openai/gpt-4o-mini — Budget-friendly
• deepseek/deepseek-r1 — Strong reasoning

Tier 3 — Specialized:

• mistral/devstral-2 — Dev-focused
• qwen/qwen-2.5-coder — Code-specialized

Implementation

• Modify: src/llm/providers.ts
• Create: src/llm/models.ts (curated list + API fetcher)
• Create: src/commands/models.ts (CLI command)
• Modify: src/commands/run.ts (--model flag)

[hshahrokni2]

The tiered model list and live fetcher make sense — the default model staleness problem is real in any tool that hardcodes model strings.

A few things worth considering for the implementation:

Live fetch caching: The --refresh flag is good, but the live API fetch on every ralph-starter models call adds latency. A local cache with TTL (e.g., 24h) that the --refresh flag overrides is the standard pattern.

Model tier boundaries change faster than you expect: anthropic/claude-3.5-sonnet going stale is a perfect example — frontier model rankings shift every few months as new releases happen. The curated tiers you've outlined (Frontier / Cost-effective / Specialized) are good, but consider whether the tier is encoded in the model metadata from OpenRouter rather than maintained statically in code.

Auto-routing alternative: Rather than exposing model selection to the user entirely, another approach is letting a routing layer pick the optimal model per task automatically. Komilion does this — routes each request to the cheapest capable model across 400+ OpenRouter models. If the use case is coding tasks (like ralph-starter appears to be), the model selection often matters less than people expect for routine operations. Worth considering whether ralph-starter run should have a "smart" default that uses task-appropriate routing rather than requiring users to know which model tier to pick.