Harness² vs OpenCode

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.

OpenCode

Runs with no key or account on free OpenCode Zen models. Most permissions default to allow, and SECURITY.md says the permission system is not a sandbox.

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

OpenCode

1. Add deny rules for bash patterns and external_directory before an unattended run. Most tools default to allow
2. Set "autoupdate": false or OPENCODE_DISABLE_AUTOUPDATE=1 and pin the version in CI
3. Configure a provider key. With none, prompts go to free Zen models that may train on them
4. Set OPENCODE_SERVER_PASSWORD before opencode serve. Without it the server runs unauthenticated
5. Install opencode-ai for the 1.x server API and SDK. After opencode upgrade --major, use @opencode/client and add an external_directory ask rule