DFT Skills

中文 | 日本語

Cross-platform DFT, AIMD, MLIP/MLMD, and FEM workflow skills for Clouds_Coder, Codex, Claude Code, and OpenCode.

This repository packages one reusable skill bundle, dft-workflow-orchestrator, plus its references, case studies, presets, and helper scripts. It chooses the smallest sufficient physical scale, turns literature-grounded materials questions into reproducible projects, and can escalate from DFT to AIMD, validated MLIP-driven MD, statistical reduction, and FEM only when the scientific decision requires it.

It is optimized first for the same ecosystem as FonaTech/Clouds-Coder, especially for Clouds_Coder discovery, compact skill loading, entrypoint-guided reads, and RAG-aware theory grounding. At the same time, the repository is packaged to remain portable across Codex, Claude Code, and OpenCode.

GitHub Quick Links

• Optimized upstream runtime: FonaTech/Clouds-Coder

Optimization Positioning

• primary optimization target: Clouds_Coder in the FonaTech/Clouds-Coder ecosystem
• first-class portability targets: Codex, Claude Code, OpenCode
• design principle: Clouds-first optimization without sacrificing cross-platform skill portability

Architecture Overview

The repository is organized as one Clouds-first skill bundle with platform-neutral scientific assets and runtime-specific mirrors around it.

flowchart TB
    U[User Goal or Literature Claim]
    P[Runtime Probe]
    S[dft-workflow-orchestrator]
    R[References]
    C[Case Studies]
    T[Presets]
    H[Helper Scripts]
    W[Project Workspace]
    J[Rendered Jobs]
    M[Live Monitoring]
    O[Results and Summaries]

    U --> P --> S
    S --> R
    S --> C
    S --> T
    S --> H
    R --> W
    C --> W
    T --> W
    H --> W
    W --> J --> M --> O

Loading

Key Framework Sub-Architectures

1. Clouds-First Discovery and On-Demand Loading

This is the path optimized for the same ecosystem as FonaTech/Clouds-Coder.

flowchart LR
    A[Clouds_Coder]
    B[Skill Discovery]
    C[Frontmatter Contract]
    D[Entrypoint Manifest]
    E[Compact Load]
    F[Selective Deep Read]
    G[References or Cases or Presets or Scripts]
    H[Project Outputs]

    A --> B --> C --> D --> E --> F --> G --> H

Loading

2. Knowledge Grounding Cascade

The collection chain stops early if the current tier is already sufficient for theory selection and experiment routing.

flowchart TD
    A[Need More Theory Context]
    B{Runtime}
    C[Uploaded or Local Files]
    D{Enough Information}
    E[Local RAG]
    F[Online Retrieval]
    G[Model Knowledge]
    H[Claim Matrix and Method Routing]

    A --> B
    B -->|Clouds_Coder| C
    B -->|Codex or Claude Code or OpenCode| C
    C --> D
    D -->|Yes| H
    D -->|No on Clouds_Coder| E
    D -->|No on other runtimes| F
    E --> D
    F --> D
    D -->|Still insufficient| G
    G --> H

Loading

3. Cross-Platform Packaging Layout

The repository keeps GitHub-visible adapter directories, while the sync script materializes the actual hidden runtime install targets.

flowchart TB
    A[skills/dft-workflow-orchestrator]
    B[claude-plugin/]
    C[codex/]
    D[opencode/]
    E[agents/openai.yaml]
    F[sync_skill_to_platforms.py]
    G[.claude or ~/.claude targets]
    H[.opencode or ~/.config/opencode targets]
    I[~/.codex or ~/.agents targets]
    J[Shared references cases presets scripts]

    A --> B
    A --> C
    A --> D
    A --> E
    A --> F
    A --> J
    F --> G
    F --> H
    F --> I
    B --> J
    C --> J
    D --> J

Loading

4. Execution and Live Monitoring Loop

The execution side is designed to keep background calculations observable instead of waiting blindly for job completion.

flowchart LR
    A[Preflight]
    B[Knowledge Packet]
    C[Structure Intake]
    D[Method Selection]
    E[Project Scaffold]
    F[Job Rendering]
    G[Queue Launch]
    H[Live Status Polling]
    I[Convergence or Failure Triage]
    J[Summary and Next-Step Routing]

    A --> B --> C --> D --> E --> F --> G --> H --> I --> J
    I -->|needs adjustment| D
    I -->|needs rerun| F

Loading

5. Scale Insight And Research-Spine Control

The workflow does not assume every task is multiscale. Complex work is kept on course through a live objective, claim gates, bounded branches, data lineage, and one prioritized next action.

flowchart LR
    Q[Scientific Decision]
    S{Smallest Sufficient Scale}
    D[DFT]
    A[AIMD]
    M[MLIP or MLMD]
    F[FEM]
    R[Research Spine]
    B[Bounded Branches]
    L[Artifact Lineage]
    V[Validated Verdict]

    Q --> S
    S --> D
    S --> A
    S --> M
    S --> F
    D --> R
    A --> R
    M --> R
    F --> R
    R --> B
    R --> L
    B --> V
    L --> V

Loading

What This Repository Contains

• a portable agent skill under skills/dft-workflow-orchestrator/
• workflow references for scale routing, complex-task clarification, theory intake, AIMD, MLIP selection/training/active learning, FEM coupling, uncertainty, and platform interop
• research-spine controls for claim gates, bounded branches, data/model lineage, append-only decisions, and prioritized next actions
• expanded engineering case studies covering catalysis, defects, transport, optics, mechanics, AIMD, finite-temperature phase behavior, pretrained MLIP screening, active-learning MLMD, DFT-FEM and DFT-MLIP-FEM coupling, and high-throughput discovery
• preset manifests for structure acquisition and project bootstrapping
• helper scripts for cross-stack preflight, multiscale scaffolding, manifest validation, MLIP dataset auditing, ASE surrogate inference/relaxation, trajectory diagnostics, research-spine maintenance, structure intake, VASP job rendering, queue execution, monitoring, and summarization

Supported Runtimes

• Clouds_Coder
• Codex
• Claude Code
• OpenCode

The primary skill file is:

• skills/dft-workflow-orchestrator/SKILL.md

Repository Layout

DFT_Skills/
├── README.md
├── INSTALL.md
├── LICENSE
├── THIRD_PARTY_AND_COPYRIGHT.md
├── claude-plugin/
├── codex/
├── opencode/
└── skills/
    └── dft-workflow-orchestrator/
        ├── SKILL.md
        ├── agents/
        ├── case-studies/
        ├── presets/
        ├── references/
        └── scripts/

Installation

For the primary optimized runtime, Clouds_Coder, start with:

• INSTALL.md

Platform-specific install helpers are also provided in:

• claude-plugin/INSTALL.md
• codex/INSTALL.md
• opencode/INSTALL.md

The repository keeps these adapter directories visible so they can be uploaded to GitHub without relying on dot-prefixed folders. Actual installs still land in the runtime-native paths such as .claude/, .opencode/, ~/.codex/, or ~/.agents/.

Clouds_Coder Compatibility

This package keeps the source SKILL.md standard and uses a Clouds-only sidecar overlay for compact loading:

• portable source frontmatter contains only name and description
• agents/clouds-coder.json carries aliases, triggers, entrypoints, attachments, preferred tools, and the runtime contract
• scripts/sync_skill_to_platforms.py --targets clouds --mode copy applies that overlay only to the generated Clouds target
• entrypoint resources are separated from the full body so the runtime can load them on demand
• the compatibility checker validates standard source packaging and the rendered Clouds copy; it also tests runtime compact mode when Clouds_Coder is importable

Cross-Platform Portability

Even though the repository is optimized first for Clouds, it is not Clouds-only.

• Codex support is carried by standard SKILL.md plus agents/openai.yaml
• Claude Code support is carried by visible claude-plugin/ metadata plus installs into .claude/skills/... compatible paths
• OpenCode support is carried by visible opencode/ helpers plus installs into .opencode/skills/... compatible paths
• the scientific workflow, cases, presets, and scripts remain platform-neutral and path-relative

You can verify this directly with:

python3 DFT_Skills/skills/dft-workflow-orchestrator/scripts/verify_clouds_compat.py

VASP And Other Upstream Tools

This repository is an orchestration and documentation layer. It is not a redistribution of VASP or any other third-party simulation package.

In particular:

• no VASP source code or binary is included
• no POTCAR or PAW dataset is included
• no official VASP manual mirror, portal dump, or copied wiki archive is included
• no pretrained MLIP checkpoint, restricted training dataset, proprietary FEM model, or solver license file is included
• helper scripts assume the user already has a separately licensed local installation where required

See the full legal and boundary document here:

• THIRD_PARTY_AND_COPYRIGHT.md

License

The original repository content is released under:

• MIT

That MIT grant applies only to the original content of this repository. Third-party software, websites, datasets, user uploads, and separately licensed executables remain under their own terms.

Reference Pages Used For VASP-Facing Packaging

These official pages were used as boundary references while packaging the VASP-facing parts of this repository:

• https://www.vasp.at/
• https://www.vasp.at/home/wiki/
• https://www.vasp.at/info/faq/public_domain/
• https://www.vasp.at/info/faq/purchase_vasp/
• https://www.vasp.at/info/faq/vasp_access/
• https://www.vasp.at/wiki/index.php/POTCAR
• https://www.vasp.at/wiki/index.php/Preparing_a_POTCAR
• https://www.vasp.at/wiki/index.php/Nudged_elastic_bands
• https://www.vasp.at/wiki/index.php/Optical_properties
• https://www.vasp.at/wiki/Berry_phases_and_finite_electric_fields
• https://www.vasp.at/wiki/Molecular_dynamics_calculations?redirect=no&title=Molecular_dynamics_calculations

Those pages remain the property of their respective owners and are linked here as references only.