COFR helps AI coding agents maintain the live state of a long-running research project: what the project believes, what evidence supports or opposes it, what has been decided, what remains open, and what is at risk.
It is a concrete implementation of the LLM Research Manager pattern. You direct the research, your agent interprets the source material, and COFR preserves the resulting structured state across files and sessions.
COFR requires Python 3.12 or newer. Install the current version directly from GitHub:
pip install git+https://github.com/GPTomics/cofr.gitThen open your preferred terminal-capable coding agent in the root of the research project. You describe the research task; the agent operates COFR.
Paste this prompt into the agent:
Run `cofr --help` to learn how it works and remember it. Then inspect this research project, propose its initial COFR state, and ask for my approval before saving anything.
Review or correct the proposal, then paste:
Save the approved initial project state using COFR.
Paste this prompt at the beginning of a later session:
Use COFR to catch me up on this project and tell me what needs attention next.
Replace the bracketed text and paste:
Ingest [the new paper, result, note, or changed files] using COFR and tell me what it changes.
When your interpretation changes, paste:
Use COFR to reflect this change in the project's belief: [the new conclusion and why].
Paste this prompt:
Based on COFR's outline of the project, create an HTML overview of the current project state.
If you repeat the same prompts often, you can create—or work with your agent to create—COFR-based agent skills that capture your preferred shortcuts and project conventions. COFR itself does not ship or require agent skills; they are optional user-created helpers, and COFR's CLI help remains the source of truth.
Most users should not need to operate COFR directly. The complete CLI, schemas, file layout, authoring format, and development instructions are in the Technical reference.
MIT