Human-Agent-Society/reef

Reef version Latest main. Commit f773bc07ab10eba98f718040c083bd73a0e0b719 Primary area Core and scenario lifecycle Environment Reproduced from a clean source checkout with CPython 3.14.7 on macOS. No accelerator or external provider is required. Minimal reproduction import math from reef.recipe.config_fields import _parse_float assert math.isnan(_parse_float("nan", "prm_temperature")) assert math.isinf(_parse_float(float("inf"), "session_ttl_s")) assert math.isinf(_parse_float("-inf", "prm_timeout_s")) The same values arrive through YAML (data.prm_temperature: .nan) and through the declared environment fallback, because both go through the field's parser. Expected behavior A declared float recipe setting resolves to a finite number. reef.service.wire.ReportPayload.from_dict already enforces this for the report score (#140); a recipe setting that reaches a sampler, a timeout, or a threshold deserves the same guarantee, and the parser is the single boundary every source passes through.

Primary area Training and runtime Problem Reef's production weight-training integration currently depends on Slime, Megatron, and CUDA-oriented runtime paths. A user with an Apple Silicon machine can run MLX-native inference, but cannot execute a supported Reef model-evolution loop locally without an NVIDIA or AMD deployment. This blocks Mac-only development, reward and recipe prototyping, small-model continual-learning experiments, and hardware-accessible integration testing. SGLang's MLX backend addresses inference, not training, while Slime does not provide an MLX training backend. Concrete use cases Run a small LoRA or QLoRA GRPO experiment from a clean checkout on an M-series Mac. Prototype Reef recipes and custom reward functions locally before moving the same workload to Slime on a GPU cluster. Consume a Reef TrainingBatch, produce a versioned adapter candidate, evaluate it, activate or reject it, and recover or roll it back without CUDA. Exercise the model-evolution lifecycle