feat: 'unsloth start'-style bridge — one-command connect local model to RustyClaw

[rexlunae]

Inspiration: Unsloth

unsloth start claude is Unsloth's killer onboarding flow:

1. Start Unsloth, load a model
2. Open your project folder
3. Run unsloth start claude
4. Claude Code is now using your local model — endpoint, API key, model name, context length all auto-configured

It works with Claude Code, Codex, Hermes, OpenCode, OpenClaw, and Pi. The server auto-stops when the agent exits.

What RustyClaw could do

The inverse bridge — make it dead simple to connect a local model running in Unsloth/Ollama to RustyClaw:

# Auto-detect a running Unsloth instance and wire it up
rustyclaw connect unsloth

# Auto-detect a running Ollama and add its models
rustyclaw connect ollama

# Manual: point at any OpenAI-compatible endpoint
rustyclaw connect --url http://localhost:8888/v1 --name "my-local-model"

The command should:

1. Auto-detect: Find running Unsloth, Ollama, llama.cpp, vLLM instances on the machine
2. Auto-configure: Add the provider to RustyClaw's config with correct base URL, models list, capabilities
3. Test: Send a quick ping to verify it works
4. Ephemeral mode: --temporary flag that doesn't persist to config (one-shot use)

Why this matters

The current flow for using local models with RustyClaw is manual config editing. A one-command bridge dramatically lowers the barrier. It also creates a symbiotic relationship — Unsloth users can use RustyClaw as their agent frontend.

Scope

• rustyclaw-cli: rustyclaw connect subcommand
• rustyclaw-core: Auto-detection of local model servers (port scanning, known defaults)
• Integration with ollama_manage tool for Ollama detection

Prior art

• Unsloth Start: https://unsloth.ai/docs/integrations/unsloth-start.md
• claude mcp add for MCP server discovery
• gh cs cp for Codespace connection