Research Harness CLI
A research loop that keeps evidence: a literature review, paper review or survey runs in steps with checks and resume.
- 517 ★ on the repository
- Last commit
Extended access
Why this level
- The `rh` commands run scripts, write workspaces and fetch papers over the network.
- LaTeX workflows call system PDF build tools.
Install
- Works in
Manual install
git clone https://github.com/WILLOSCAR/research-units-pipeline-skills.git
cd research-units-pipeline-skills
uv sync --lockedThe start of the first README block: clone and dependency sync. Then `uv run rh goal create` and `uv run rh run start`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The `rh` commands run scripts, write workspaces and fetch papers over the network.
- LaTeX workflows call system PDF build tools.
This is third-party code. Review the repository files before installing.
What it does
Research Harness turns a research goal into a Run: a workspace with files, steps and checks. The `rh` command creates the goal and drives one workflow: `research-brief`, `paper-review`, `evidence-review`, `arxiv-survey`, `arxiv-survey-latex`, `idea-brainstorm` or `source-tutorial`. Every step leaves artifacts, hashes and decisions, so work can resume the next day and a failed check can be traced. In Codex and Claude Code a workflow starts with one sentence such as "Use research-brief to map ...". The repository holds more than a hundred skills from which the workflows are assembled.
Who it is for. Researchers and students who hand an agent a literature review or paper review and want sources and checks behind every paragraph.
Good fit when
- You need to understand a topic and decide which papers to read first.
- A manuscript review is written with every concern tied to a place in the text.
- A course literature survey is delivered as LaTeX and PDF.
- A long run was interrupted and work resumes without retelling the earlier chat.
Not a fit when
- A finished scientific paper is expected without review: in the author's own words this is not an autonomous scientist.
- `paper-review` has no manuscript and `source-tutorial` has no source pack: the workflow will not invent them.
- Python 3.10 and `uv` are unavailable: it runs only from a repository clone.
Example request
Use research-brief to map test-time adaptation for robotics and tell me what to read first.Limitations
For now it runs only from a repository clone on Python 3.10 or newer with `uv`. PDF delivery needs `latexmk`, XeLaTeX, BibTeX and `pdfinfo`, and PDF ingestion also needs `pdftotext`. A PASS confirms run integrity and the workflow contract, not scientific correctness. Published examples cover one topic each and stability across topics is not measured. Each run accumulates artifacts and events in `.harness/`. No licence is stated in the repository.
How to disable. Remove the skills from the agent directory and the repository clone. Delete result workspaces and the `.harness/` service folder separately.
FAQ
Which workflow should I pick?
Pick by outcome: `research-brief` for a topic map, `paper-review` for a review, `evidence-review` for protocol-based synthesis, `arxiv-survey` for a literature survey, `idea-brainstorm` for directions, `source-tutorial` for teaching material.
What does PASS mean?
Run integrity and agreement with the observable workflow checks. It does not confirm the quality of the scientific conclusion or exhaustive retrieval.
When does a human step in?
At checkpoints: in `research-brief`, after reviewing the paper set, taxonomy and outline, run `rh run approve --checkpoint C2` and `rh run resume`. Approval is bound to artifact hashes, so editing the outline revokes it.
What if a check fails?
`uv run rh improve diagnose --workspace <folder>` shows where to repair: diagnosis separates the observable defect from the layer that owns it.
Which utilities does PDF work need?
LaTeX and PDF delivery need `latexmk`, XeLaTeX, BibTeX and `pdfinfo`, and reading PDF sources also needs `pdftotext`.
Discussion
No questions or reviews yet.
Related
CLI toolsA marketplace of five Claude Code plugins for researchers: literature search, research design, project memory and manuscript work
Seven Codex skills carry research from topic choice and paper reading to reviewer replies and an editable presentation.
Claude Code skills for the academic pipeline: research, paper drafting, peer review and citation checks
Takes an agent from a topic to an arXiv review paper in LaTeX IEEEtran: approved plan, every citation checked against a live source, PDF build
Research environment for Claude Code and Codex: 87 skills, 15 reviewer agents and shared context for LaTeX papers.
A Claude Code skill for scientific literature reviews: an interview, Scopus and OpenAlex search, DOI checks and a final docx file
README in short
The README sets the principle of a trail, not just an answer: goal, run, evidence, artifact. A table maps the desired outcome to a workflow and main file. It then explains pinned contracts, what PASS means at three levels, and published examples with their limits.