| name | benchmark-suite |
| description | Run Mini SWE-bench style harness regression tasks and A/B comparisons to measure harness improvement objectively. |
| allowed-tools | Read, Bash(node .harness/scripts/bench-runner.mjs:*), Bash(node .harness/scripts/bench-compare.mjs:*) |
| suggested-turns | 6 |
Benchmark Suite
Use this when evaluating whether a harness change improved or regressed behavior.
Commands
node .harness/scripts/bench-runner.mjs --variant=current
node .harness/scripts/bench-runner.mjs --variant=candidate
node .harness/scripts/bench-compare.mjs
Output contract
### Benchmark Suite
### Tasks: <n>
### Pass rate: <percent>
### Avg score: <score>
### A/B delta: <delta or n/a>