菜单
Read, edit, commit, and diff experiment code; coordinate shared worktrees and preserve frozen branch history. Use before branch work, when starting work alongside other sessions, comparing nodes, preparing a run, repairing a provisional node, or diagnosing stale code.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/alphaXiv/OpenResearch --skill orx-git命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use the `orx` CLI to run local OpenResearch projects from a terminal — create experiment branches, launch and supervise compute, inspect logs and evidence, and manage the experiment tree. Read this before driving `orx` programmatically.
Publication-quality figures in matplotlib or TikZ: learning curves, scaling laws, benchmark and ablation comparisons, Pareto trade-offs, heatmaps and confusion matrices, method diagrams. Covers the shared style module, sizing, uncertainty, and vector export. Use whenever you plot, chart, or visualize results, add a figure to a paper or report, or one looks unpolished; then read one reference.
Draft an academic paper or preprint as LaTeX. Create a .tex file in the project working tree, where it renders for the user and compiles to PDF. Use for a paper, preprint, manuscript, arXiv or submission draft, or a section of one; generic reports and result summaries belong to `orx-reports`.
正在显示 SKILL.md
| name | orx-git |
| description | Read, edit, commit, and diff experiment code; coordinate shared worktrees and preserve frozen branch history. Use before branch work, when starting work alongside other sessions, comparing nodes, preparing a run, repairing a provisional node, or diagnosing stale code. |
Git records every experiment. GitHub publication may be enabled for collaborator visibility, but it is never part of compute transport. Follow the project playbook's publication status; do not push merely to launch compute and do not fetch from or publish back to the paper's upstream repository.
Other chat sessions may work in sibling worktrees of the same clone. Before
starting, inspect git branch -a, orx runs <projectId>, and relevant orx exp desc <expId> notes so you do not duplicate their work. Branches and remotes are
shared even though worktrees are separate. Keep experiment notes current as you
learn so sibling sessions can orient from them.
One branch has one worktree owner. If checkout says a branch is already checked out, keep working only when the named path is your worktree; otherwise leave that branch to its owning session. Session worktrees start detached on the baseline, so check out the experiment branch before editing.
Each experiment node has a local orx/<slug> branch. orx create-experiment creates it from its parent. Work in the session worktree,
check out the printed branch, make only that experiment's change, and commit it:
git checkout orx/<slug>
git status --short
git add <changed files>
git commit -m "describe the experiment change"
The runner builds an immutable source archive from the recorded commit, so
committed work is sufficient on every backend. Uncommitted files are never
included in a run. Before launching, confirm git status --short is empty and
inspect the recorded commit with git show --stat --oneline HEAD.
To compare a child with its parent, use local refs only:
git diff <parent-branch>...orx/<child-slug>
git log --oneline <parent-branch>..orx/<child-slug>
Once a run answers an experiment, its branch and history are immutable. Never merge or rebase it. To incorporate other work, create a child and put the merge commit on the child's branch. Never rebase experiment history; it records the exact code that ran.