Agent evals
Success rate and latency per agent.
Not instrumented
No eval set is versioned for the agents yet, so there is nothing to chart. This page stays empty rather than showing invented numbers.
Review chainobservational
Generated Oct 22, 2026/pr-review → /pr-fixer, extracted from PR history (koydas/ai-dev-tools, 30 latest closed PRs): 1 of 30 PRs carry a report, 1 report in all.
Only 1 reviewed PR: too few to read a rate.
| Figure | Value |
|---|---|
Blocking findings touched Share of blocking findings whose cited file a later commit changed. A proxy for fixed, not proof. | 100% |
Non-blocking findings touched Same measure for non-blocking findings and nits. | 100% |
Converged Share of reviewed PRs whose last report is DONE. | 0% |
Reports per reviewed PR Review rounds before the PR ended. | 1 |
Human threads beyond the review Threads from trusted humans, after the first report, on files no report cited. A proxy for what the review missed. | 0 |
These are proxies from PR history, not eval scores: there is no ground truth, and "touched" does not mean "fixed".