sjtu-sai-agents

MCTS-based autonomous physics problem solver with arXiv search, prior knowledge retrieval, and multi-agent reasoning. Use when you need to solve physics problems, search arXiv for relevant papers, or generate structured physics solutions with iterative refinement.

Use when solving problems involving Maxwell's equations, electrostatics, magnetostatics, electromagnetic waves, radiation, or relativistic electrodynamics.

Use when applying conservation of energy, momentum, angular momentum, charge, or other conserved quantities to constrain or solve a physical system.

Use when checking dimensional consistency, estimating physical scales, or deriving functional forms via the Buckingham Pi theorem.

Use when decomposing signals or fields into frequency/momentum components, applying Fourier transforms, or using spectral methods to solve differential equations.

Use when solving ordinary or partial differential equations numerically, including choosing integrators, discretization schemes, and stability analysis.

Use when solving a problem by expanding around a known solution in a small parameter, including regular and singular perturbation theory.

Use when solving quantum mechanical problems including the Schrodinger equation, angular momentum coupling, scattering theory, or many-body quantum systems.

Build, configure, and run EvoMaster autonomous AI agents. Use when you need to create single or multi-agent systems with tool calling, MCP integration, skill extensions, or self-evolving workflows for scientific discovery, Kaggle competitions, or general automation tasks.

Guide for uploading local files to Bohrium OSS when MCP tools require file URLs. Use this when you need to transmit local files through MCP tools that cannot accept local paths directly.

Feishu document read/write operations. Activate when user mentions Feishu docs, cloud docs, or docx links.

Feishu cloud storage file management. Activate when user mentions cloud space, folders, drive.

Feishu permission management for documents and files. Activate when user mentions sharing, permissions, collaborators.

Feishu knowledge base navigation. Activate when user mentions knowledge base, wiki, or wiki links.

Retrieval-Augmented Generation (RAG) skill for generic semantic search and knowledge retrieval. Use when you need to build or call vector-based search over your own documents using vector indexes (optional FAISS) and embedding models (local transformers or OpenAI embeddings), or when integrating LLMs with external knowledge bases in a project-agnostic way.

Dataset discovery and download toolkit covering HuggingFace, GitHub, Google, and academic paper search. Use when you need to find, inspect, or download ML datasets.

Tree-structured memory management for node-based exploration. Manages global memory, data link records, and per-node manifest (TL;DR + recordings). Use when agents need to persist, share, or retrieve exploration knowledge.

Code execution and submission tools for DataNode (Black/Red) agents. Manages DataLoader code, runs assembled scripts, validates and grades submissions.

Data loading, cleaning, external-data merge, and validation-diagnosis guidance for black-node style ML agents. Use when a coding agent must improve data quality or integrate external datasets without changing the core model logic, especially for image tasks but also for tabular and text tasks.

Retrieval-Augmented Generation (RAG) skill for generic semantic search and knowledge retrieval. Use when you need to build or call vector-based search over your own documents using vector indexes (optional FAISS) and embedding models (local transformers or OpenAI embeddings), or when integrating LLMs with external knowledge bases in a project-agnostic way.

When using Mat MCP tools (mat_sg_*, mat_dpa_*, mat_doc_*): (1) Use the URL returned by a previous tool (e.g. structure_paths) as input to the next tool—do not pass only a filename when the previous step returned an https URL. (2) For long-running DPA tasks use submit_* first, then query_job_status, then get_job_results when Done—do not use synchronous tools which may timeout. (3) When polling job status, wait 30–60 seconds between queries; call get_job_results once status is Done/Success.