Harness² CLI Reference
Last updated: October 19, 2026
Reference for all six Harness² commands — harness2, harness2-setup, harness2-check, harness2-judge, harness2-report and harness2-demo — with usage and flags.
Six commands ship with the package. All of them are documented in the repository README; this page collects them in one place.
harness2-setup — prepare a benchmark.
harness2-setup labTargets: jb (JobBench), lab (LAB), wb (WorkBuddy-Bench, also accepts --subset office|web|code).
harness2-check — verify the benchmark is ready before spending a run.
harness2-check --bench labharness2 — run the Propose–Probe–Compose step.
harness2 --bench lab --k 1 --mode parallel \
--only-domains antitrust-competition --max-tasks 1| Flag | Meaning |
|---|---|
--k 1 | One Propose–Probe–Compose step |
--k K --mode parallel | Probe K general and K domain-specific candidates, then produce K compositions |
--k K --mode sequential | Refine the harness through K successive steps |
--substrate cx | Use Codex instead of the default OpenCode executor |
--dry-run | Preview task selection without calling a model |
harness2-judge — score the run.
harness2-judge --bench lab --k 1 --mode parallelharness2-report — turn judged results into a report.
harness2-report --bench lab --k 1 --mode parallelharness2-judge and harness2-report must use the same benchmark, executor, subset, --k, --mode and seed as the refine run, or they have nothing to read.
harness2-demo — see the loop with no model credentials.
harness2-demo --output runs/demo --k 2Runs both recursion modes with a deterministic editor and real Python execution. This is the fastest way to confirm an install is sound.
Runs and resumption. Output goes under runs/, containing the task summary, deliverables, the adopted harness, reflection notes and evaluation results. Repeating a command resumes the same run; set HARNESS2_RUNS_DIR to a new directory when you change models or task selection.
Flags not listed here are not documented in the README and are omitted rather than guessed at.