Proof.
The incident list is what agents have already done in public. This page is the measured record of Agent-Causality on live OpenCode runs.
The runs
On 11 Oct 2026, five tasks ran once each on OpenCode 1.19.0. Every run completed. Capture coverage was 1.0, both as a micro-average and as a macro-average.
The scores
Cause identification averaged 0.3 across the five runs. Provenance resource recall was 1.0. Joint-branch ancestry was 0.0 on the two tasks that check it.
The limit
A capture score of 1.0 means the expected event classes were present. It does not mean every cause was named. Several runs skipped the injected actions, so the scorer had no ground-truth events to find.
Live traces also cannot show causal minimality or a counterfactual interaction. The writeup marks those as not measurable and unsupported.
The writeup
- OpenCode results The five runs, the scores, and the limits above.
- Benchmarks How the offline suite and the OpenCode runs are scored.
- Earlier snapshot A historical note. The numbers on this page are the current record.