kaggle-harbor-benchmark

# Kaggle Harbor Benchmark Build Kaggle benchmark repos as runnable Harbor projects with concrete evidence. Prefer small deterministic tasks, real validation commands, and named job outputs over long explanations. Use the references only when needed: - `references/harbor_spec.md`: what a Harbor-compatible task is, what Kaggle runs, and how to iterate locally. - `references/model_proxy.md`: Kaggle CLI auth, `.env`, OpenAI/Anthropic proxy endpoints, model slugs, and `/chat/completions` vs `/responses`. - `references/agent.md`: built-in Harbor agents, binary CLI agents, SDK-backed agents, model-proxy routing, and template implementation links. - `references/debug.md`: isolate model access, local Harbor, custom agents, packaged runner image failures, and job-output issues. Default workflow: 1. Tell the user to create a Kaggle organization; approval can take 24 hours, but Harbor authoring can continue locally while they wait. 2. Explain the Harbor spec: task environment, instruction, verifier, optional oracle/solution, artifacts, and named job outputs. 3. Start with a tiny deterministic task and validate it with local `harbor run`. 4. Route custom agents and binaries through the Kaggle model proxy so language model calls are logged. 5. Run the packaged runner smoke test only after local Harbor passes. 6. Commit named successful job outputs such as `jobs/mini-swe-agent`, `jobs/opencode-binary`, and `jobs/antigravity`. Before committing job outputs, scan for raw API keys, bearer tokens, and unredacted proxy credentials.