feat: OpenAI-compatible API endpoint — serve gateway as an API for external tools

[rexlunae]

Inspiration: Unsloth

Unsloth exposes loaded models via both Anthropic-compatible (/v1/messages) and OpenAI-compatible (/v1/chat/completions, /v1/responses) API endpoints. This lets Cursor, Continue, Cline, Open WebUI, and any OpenAI-compatible tool use your local models. It also supports API key management with sk-unsloth-... keys.

What RustyClaw could do

Expose the RustyClaw gateway as an OpenAI-compatible API endpoint:

1. POST /v1/chat/completions — OpenAI chat completions endpoint
2. GET /v1/models — List available models/providers
3. API key management — Generate/revoke sk-rustyclaw-... keys with scoped access (which models, tools allowed, rate limits)
4. Streaming support — SSE streaming responses
5. Tool calling passthrough — Map OpenAI tool call format to/from RustyClaw's internal representation

Why this matters

RustyClaw already connects to providers. This makes RustyClaw be a provider — so any OpenAI-compatible tool (VS Code extensions, chat UIs, other agents) can use RustyClaw as their LLM backend, with all of RustyClaw's security, credential management, and tool ecosystem available.

Scope

• rustyclaw-core: /v1/chat/completions + /v1/models HTTP handlers in the gateway
• rustyclaw-core: API key management (CRUD, scoping, revocation)
• Feature flag: openai-api (off by default, since it exposes the gateway to external requests)
• Security: keys are scoped, the endpoint is opt-in, default bound to localhost

Prior art

• Unsloth API: https://unsloth.ai/docs/basics/api.md
• Ollama's OpenAI-compatible endpoint
• vLLM's OpenAI-compatible server
• LiteLLM proxy