Software
Software repositories are living systems. Commits, reviews, comments, releases and failed tests show how work moves, where it stalls and what is likely to happen next.
A world model can ask:
- Which pull requests will merge, and when?
- Whose turn is it?
- Where is work drifting from the way this repository normally operates?
- Which events or contributors look unusual?
- How does the forecast change if a pull request is split or reviewed now?
Project OSS
Project OSS learns from dated repository events and turns that history into practical tools: merge forecasts, resolution-time estimates, unusual-activity lists, next-actor predictions and proposed-split comparisons.
Research note: Project OSS ยท review access
Results: The July eight-repository model used 7.04 million events and transferred to OpenClaw without retraining: 0.859 AUC on 28-day merge and 0.711 time-to-merge rank correlation. The September scaling study covers 28.40 million events from 20 repositories, with OpenClaw held out of training; resolution-time rank correlation reached 0.679 on the August 10 OpenClaw capture, against 0.388 from current facts. The research note keeps these studies, their comparisons and their capture dates separate. Earlier scored results.
Watch it work: An agent triage comparison replays the same language model twice on the 794 open pull requests of nodejs/node, frozen on 17 August 2026, asking which ten a maintainer should deal with that day. One run starts from a world-model readout; the other has only the read tools. Their picks are scored against what maintainers later did.
In the Factory: GitHub event streams
Where this goes
The larger goal is not another repository dashboard. It is an agent that understands the software world it is responsible for: what matters, what is stuck, what changed, what should happen next, and whether its last action helped.
World Router, an earlier tool that routes AI coding work between models, is in Explorations. Overseeing the agents themselves is part of Oversight.