Research Harness
A file-first framework that turns a research goal into a recoverable, inspectable run, organizing focused skills into explicit workflows and preserving intermediate artifacts, evidence and decisions at every step.
Research Harness turns a research goal into a Run that produces a readable deliverable together with its evidence trail, rather than only a polished document with no record of which sources support which claims. It organizes focused Skills into explicit Workflows, checks observable contracts at checkpoints, and points a failed contract back to the smallest surface that needs repair, so work can resume the next day without reconstructing a chat history.
Its command-line tool creates a goal against a chosen workflow and workspace, starts a run that advances automatically until it finishes or reaches an unmet prerequisite, and lets the user inspect status, approve a checkpoint, resume the run, and inspect the underlying evidence; a diagnose command says where to make a repair when a contract fails. Workflows are chosen by the deliverable a user wants, such as a topic snapshot, a single-paper review, or a synthesis of studies under an approved protocol, with the Skills and Units behind each workflow treated as implementation detail until something needs inspecting or fixing.
It runs from a source checkout with Python 3.10 or later and uv, and the project describes itself as infrastructure for making agent-assisted research inspectable and resumable, not a claim of autonomous scientific discovery.