Crush vs Harness²

Crush

Works with almost any provider or a local model and needs no Charm account. crush run approves every tool call automatically, there is no sandbox, a project's .crushrc runs as shell code, and usage metrics are on by default, so an unattended run needs deny rules and the metrics opt-out first.

Harness²

Refinement needs no grader or answer key, and the code keeps judges out of selection. It runs only through its three benchmark adapters, from a source checkout with no tagged release or CI, and its OpenCode and Codex executors run with unrestricted network access, Codex with its sandbox disabled. The benchmark gains are the authors' claims.

Crush

1. Add permissions deny rules or a PreToolUse hook before crush run. It approves every permission request in its session
2. Read .crushrc, crushrc and crush.json in a cloned repository before running Crush there. They execute as shell code at startup
3. Set CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1 to stop usage metrics to data.charm.land
4. Wrap crush run in your own timeout. It has no turn, time or cost limit flag
5. Configure a provider by environment variable or crushrc first. With none, crush run exits with an error asking for interactive setup

Harness²

1. Don't plan to call Harness² as a service. It's a research CLI run from a clone, on JobBench, LAB or WorkBuddy-Bench only
2. Ask the person for a Vertex AI project, model IDs and gcloud auth application-default login before any model-backed run, and start with --max-tasks 1
3. Run executors behind an egress firewall and block huggingface.co and github.com after setup, as the docs advise, so rubrics can't be fetched
4. Write .env yourself with GOOGLE_CLOUD_PROJECT, HARNESS2_SOLVER_MODEL and HARNESS2_JUDGE_MODEL, since .env.example isn't shipped, and never commit it
5. Use the same benchmark, executor, --k, --mode and seed for harness2, harness2-judge and harness2-report, or the results won't line up