Skip to content

fribnd/ensdf-agent

v0.3.9MIT

ENSDF-Agent for nuclear data evaluation

GitHubOpen Website

ENSDF-Agent

Part of the AI/ML Technical Innovation at the Nuclear Data Group at the Facility for Rare Isotope Beams (nucleardata@frib.msu.edu).

Supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under Award Number DE-SC0016948.

Open-source release: github.com/FRIBND/ENSDF-Agent

Overview

The first AI Agent designed for Evaluated Nuclear Structure Data File (ENSDF) workflows. Developed and refined through daily evaluation tasks at FRIB Nuclear Data Group.

Built on the open-source platforms Microsoft Visual Studio Code and GitHub Copilot, ENSDF-Agent employs Harness Engineering and Semantic Structuring and integrates the power of rapidly advancing Large Language Models (LLMs) into the routine workflows of nuclear data evaluators.

Development Timeline

  • 2026-08-13: ENSDF-Agent Harness was introduced at the 2026 Low Energy Community Meeting at ANL.
  • 2026-08-11: Migrated ENSDF-Agent Plugin to Agent Plugins 1.0 open standard format.
  • 2026-05-07: ENSDF-Agent Harness was refined for improved token efficiency, consuming <30K input tokens in one round.
  • 2026-03-25: ENSDF-Agent Version 0.0.1, with 3 Agent Hooks and 24 Agent Skills, was released as an Agent Plugin via the Microsoft VS Code Plugin Marketplace.
  • 2026-03-04: ENSDF-Agent became available as an open-source repository at github.com/FRIBND/ENSDF-Agent.
  • 2026-02-23: Agent Skills were introduced as modular, portable capabilities that can be dynamically loaded into ENSDF-Agent to perform specific tasks within ENSDF workflows.
  • 2025-11-14: The FRIBND Custom Agent Chat Mode was upgraded to ENSDF-Agent.
  • 2025-10-30: The FRIBND AI Agent was introduced at the 2025 U.S. Nuclear Data Program Meeting.
  • 2025-08-14: The FRIBND AI Agent was first introduced at the 2025 Low Energy Community Meeting at TAMU.
  • 2025-08-06: The initial version of the FRIBND Custom Agent Chat Mode was posted within github.com/sunlijie-msu/ENSDF.

Prerequisites

  • Python 3 on PATH — hooks and ENSDF validation scripts require it.
  • Windows users: Python must be invokable as python from terminal (the hooks also ship .ps1 fallbacks for block-root-file-creation and block-git-revert; validate_ens requires Python on all platforms).

Installation

  1. Configure plugin marketplaces by clicking "Add Item" and entering FRIBND/ENSDF-Agent in the setting(chat.plugins.marketplaces) setting.
  2. Open the Extensions view (kb(workbench.view.extensions)) and enter @agentPlugins ENSDF-Agent in the search field.
    • Alternatively, select the More Actions (three dots) icon in the Extensions sidebar and choose Views > Agent Plugins.
  3. Click Install to install the ENSDF-Agent Plugin to your user profile.
  4. Troubleshooting: https://code.visualstudio.com/docs/agent-customization/agent-plugins#_troubleshooting

Hooks

Three hooks in hooks.json use the same base format as workspace hooks. All fire automatically for any active agent while the plugin is enabled.

HookEventBlocksAllows
block-root-file-creationPreToolUsecreate_file at workspace root, A<N>/, or XUNDL/.github/temp/; non-create_file tools
block-git-revertPreToolUsegit restore/git checkout on .ens or non-temp paths; bare checkouttemp-scoped restore/checkout; git switch; unrelated commands
validate_ensPostToolUseNon-ASCII .ens content; data-record ruler failuresComment-only edits; non-.ens files

Caveats

  • VS Code may not allow you to enable ENSDF-Agent from the Agents menu. If you encounter this issue, restart VS Code and select ENSDF-Agent again.

  • VS Code does not support installing Agent Plugins in a specific workspace.

  • VS Code Agent Plugins do not support shipping a standalone copilot-instructions.md alongside an agent; the plugin's sync tooling merges its content into agents/ENSDF-Agent.agent.md instead.

  • Agent Skills performance and reliability vary based on the underlying LLM capabilities and the complexity of the task.

  • VS Code extension-contributed agents and skills will negatively impact the user experience of the ENSDF-Agent. Extensions register these agents via their package.json contribution points. When an extension is enabled, its agents and skills are automatically discovered and added to the chat menu. Currently, there is no setting in VS Code to disable or hide these extension-contributed agents while still keeping the main extension enabled. One workaround is to execute Workspace-Level Disable: Click the Gear icon (Manage) directly on the extension list item. Select Disable (Workspace) from the dropdown menu.

Disclaimer and Usage Notice

ENSDF-Agent is designed to assist in nuclear data evaluation, not a substitute for human evaluators. It does not possess true understanding or expertise in nuclear physics.

AI-generated content in data sets require review and judgment by human evaluators for scientific and technical accuracy.

The human evaluator remains the sole authority and is responsible for the validity of any datasets submitted to NNDC, IAEA, or other organizations.

ENSDF-Agent developers and affiliated institutions (FRIB/MSU) disclaim all responsibility for errors, data loss, or scientific inaccuracies.

Use of this software constitutes acceptance of these terms.

Contact: nucleardata@frib.msu.edu

License

MIT

ENSDF-Agent Architecture