Analyzes and refactors TensorCircuit-NG code to achieve peak time and memory performance. It enforces advanced JAX vectorization, intelligent JIT staging, optimal tensor network contraction, and memory-efficient autodiff strategies.
Tune tensor-network contraction path search and slicing for TensorCircuit-NG workloads, especially OMECO and cotengra hyperparameters, memory targets, total FLOPs/write, slice counts, and large-circuit amplitude or expectation contractions.
Review TensorCircuit-NG code diffs for correctness, performance, backend agnosticity, and minimal high-quality fixes.
Performs a comprehensive sanity check on the codebase to reduce entropy, increase readability, and improve maintainability. Identifies issues with imports, comments, dead code, DRY violations, exception handling, magic numbers/secrets, docstrings, duplicated implementations, and test sufficiency.
Maintain TensorCircuit-NG development memory in `.agents/memory/`. Use `update` mode to save durable lessons from the current session into the right memory file, and use `dream` mode to scan and refactor the whole memory set for lower redundancy, clearer taxonomy, and better long-term usefulness.
Autonomously reproduces quantum computing arXiv papers using TensorCircuit-NG. It creates standardized repository structures, generates meta.yaml, writes and runs JAX-accelerated code, and strictly enforces code quality (black/pylint) before saving final figures.
Autonomously translates quantum scripts from other frameworks (Qiskit, PennyLane) into TensorCircuit-NG. It performs end-to-end intent understanding, applies JAX vectorization/JIT, and outputs a strict before-and-after execution time benchmarking report.
Autonomously explores the space of quantum circuits and optimization strategies to solve specific physical problems (VQE, QAOA, QML, etc.). It uses a budget-constrained, multi-frontier search to discover high-performance solutions and records all experiments in a reproducible registry.