Harness² vs Junie CLI

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.

Junie CLI

Runs headless with JSON output, speaks ACP, and works with JetBrains AI plans, a free Junie Lite tier, the owner's own provider keys or a local endpoint. Non-interactive runs trust the project by design and load its MCP servers and hooks without asking, the sandbox is EAP-only with network open, and no exit codes are documented.

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

Junie CLI

1. Run headless only in repositories you trust. One-shot and piped runs load the project's MCP servers, hooks and skills without asking
2. Set JUNIE_API_KEY (from junie.jetbrains.com/cli) or a provider key such as JUNIE_ANTHROPIC_API_KEY in CI. The default sign-in opens a browser
3. Add --output-format json or --json-output-file to parse results. Check the output, since exit codes aren't documented
4. Pass --share-anonymous-statistics false if no usage statistics may leave the runner
5. Pin a build with --use-version and --skip-update-check. The CLI updates itself on start by default