[feature] OrcaRouter provider support for Molt #104

User problem

Molt is an agentic-first RL framework for research — a ~9.2K-line PyTorch-native stack over Ray, vLLM, and NVIDIA AutoModel that takes a model from SFT to fully-async, multi-turn agentic RL at frontier MoE scale. Its design center is that the agent is the program: reward is any Python inside an Env or ChatAgent, with LLM-as-judge named in the README as a valid reward source.

On Molt's ChatAgent path, any OpenAI-compatible endpoint already works as a target — a stock SDK client pointed at a session-scoped base_url with a standard API key. For users who grade rollouts with hosted frontier judges, OrcaRouter could add one useful, strictly optional choice: a single endpoint in front of many chat/reasoning models, with routing, failover, caching, and usage budgets.

This issue proposes adding OrcaRouter as an optional provider for that hosted judge/eval slot. It would not replace or change Molt's local vLLM rollout, NVIDIA AutoModel training path, or any existing setup.

Desired outcome

Evaluate adding OrcaRouter support as an optional OpenAI-compatible provider for hosted judge/reward/eval calls. OrcaRouter exposes an OpenAI-compatible API and uses standard API-key authentication, so it should fit Molt's existing client wiring without a new protocol. Concretely, this could look like a documented ChatAgent/Env reward example that points a judge client at an OrcaRouter base_url — reusing the same AsyncOpenAI(base_url=..., api_key=...) pattern Molt already documents — plus, if maintainers want it, a small helper so one run can reach several chat/reasoning models through a single endpoint. This is a proposal only; no code exists yet.

The most relevant capabilities for a research lab running many judge-heavy experiments are automatic model routing with provider failover, prompt caching for repeated grading calls, and usage tracking/budgets shared across the team.

Affected area

area:llm

Extra context

OrcaRouter is already integrated with open-source projects such as opencode (models.dev), CAMEL, and promptfoo — external agent/eval tooling similar in spirit to the OpenAI-wire callers Molt's README names (opencode, Claude Code, AgentScope).

For transparency: OrcaRouter runs an optional open-source partner program in which approved OSS projects can receive a 5% revenue share from OrcaRouter usage attributed to their integration. Participation is not a prerequisite for integration, and I'd be glad to follow whatever disclosure or governance rules Molt prefers. See https://www.orcarouter.ai/built-with for how other projects present it.

I'm an engineer on the OrcaRouter team. If this direction seems useful, I'd appreciate the maintainers' view on where it fits best, and I'd be happy to contribute the example and docs (or a small provider helper) as a PR once approved.