v0.1.0-rc2

The pre-1.0 breaking wave: rc.2 freezes the wire contract and finalizes the authoring API. It does not interoperate with rc.1, so upgrade every environment, model, and runtime together.

Breaking changes

• Observation and action values now travel as spec-directed bytes, and protocol generation (rlmesh-wire-v1) is matched by exact equality. An rc.1 peer cannot connect to an rc.2 peer.
• Roles are required on ImageTag, StateTag, TextTag, Image, and Text, passed as the first argument: ImageTag(IMAGE_PRIMARY), Text(INSTRUCTION).
• Adapter spec classes use their final names. Rename the old input and action classes in your imports, for example ImageInput to Image, StateComponent to Concat, ActionComponent to Actuator, ActionLayout to Action.
• action_horizon= on run() and session() is now execution_horizon=, and execute_horizon is gone from Action. The horizon is a runtime setting, not part of the spec.
• predict_chunk(self, observation) no longer takes a horizon; return your native chunk and the runtime replays it one action per step. An autoregressive head may add execution_horizon: int = 1 to decode exactly that many.
• on_reset is removed from Model and Session. Use on_episode_end, which fires at every episode boundary on both the local and served paths; a subclass reset() is wired to it.
• Sandbox config is grouped. Pass build=SandboxBuild(...) and runtime=SandboxRuntime(...) instead of flat kwargs: SandboxEnv("CartPole-v1", build=SandboxBuild(packages=["gymnasium==1.3.0"])).

Added

• Action chunking and batched prediction in the runtime. Implement the most general of four corners and the runtime derives the rest: predict, predict_chunk, predict_batch, predict_chunk_batch. The runtime owns chunk replay and the execution horizon, so one action still reaches the environment per step.
• session() and Session for manual reset/predict/step control alongside run(), plus Session.reader() and Session.read() for read-only, role-addressed views of an observation through the model's adapter.
• PyTorch and JAX environments served with framework tensors: EnvServer(env, framework="torch", device="cuda:0"), with rlmesh.torch and rlmesh.jax factory, model, and sandbox classes. A model's device moves observations before predict.
• Declared construction parameters: EnvFactory.params = ParamSpec(Param(...)) validates make() arguments and adds a Vector type. enumerate_variants() lists a factory's concrete sub-environments.
• describe(), describe_json(), and python -m rlmesh.describe emit a JSON metadata envelope for a factory or model without constructing it.
• More adapter conversions: rotation accept-sets, image fit, normalize ranges, channel validation, optional cameras, per-actuator clip, role-less actuators, and adapter.advisories() for data-loss notes. A newer environment contract also parses against an older core and fails only when a model input needs the missing piece.
• A live debug viewer: pass view="terminal", view="http:9000", or view="both" to run()/session(), or configure rlmesh.View(...). It is best-effort and never breaks an eval.
• Sampling and ergonomic helpers on the space wrappers.

Changed

• The adapter engine and chunk replay moved into the Rust core for vectorized, stateful serving. The Python authoring path is unchanged.
• Actuator scale, invert, and threshold apply on both sides and compose in order, so a model can bridge its output convention at the boundary.

Fixed

• Sandbox startup waits longer before timing out, so slow environments such as LIBERO task suites start reliably.
• Debug viewer fixes for SSH terminals and observation-fed frames.