Evidence
The maintainer’s own projects.
These figures are observed use and co-movement. They are not a controlled study, and nothing here claims fewer repeated failures.
Evidence
These figures are observed use and co-movement. They are not a controlled study, and nothing here claims fewer repeated failures.
Tool use
| OpenState tool uses in OpenCode | 35,579 |
| Structural tool uses | 30,782 |
| AST searches | 10,420 |
| Text edit previews | 8,102 |
| Text edit applies | 6,694 |
| Bounded text-region reads | 2,467 |
| Sessions that used it | 243 |
| Projects | 20 |
| Store records (10 October) | 82,586 |
| Resume capsule size | median about 500, at most about 1,600 tokens |
Direct calls count once; a Code Mode program using a tool counts once for that tool. The named structural categories are a subset of the structural total. Method and evidence on GitHub →
Reads
These are bytes returned per read in the maintainer’s sessions. How small a definition is next to its file depends on the code, and a careful agent can also read a range of lines.
Completion cost
In the maintainer’s OpenCode history, tokens per completed work item in the openstate project fell from about 16M to 10.6M (34%) and per commit from 16.0M to 11.05M (31%) between August and September 2026 while completions rose. A repo adopting the tools via Claude Code fell from 14.3M to 0.9M per commit. Non-adopting repos were flat or worse except one that improved in a lighter work phase.
No measurement yet shows fewer repeated failures. What OpenState guarantees is the record and its delivery to the agent.