Feature Engineering

A fixed Harbor feature-engineering task on public cryptocurrency minute data, with hidden scoring after submission.

The agent phase has two containers:

• main: the public Parquet, three context Markdown files, and Python libraries
• mcp-server: asynchronous train, backtest, and submission tools

Both images derive their public data from one canonical Parquet snapshot. The main image materializes only the agent-visible columns as compressed Parquet; the two scoring columns remain available only to the MCP runtime.

Harbor collects /app/submission directly from the MCP sidecar, tears the agent environment down, and starts a separate verifier. The verifier owns hidden data and sends only feature matrices and the submitted bundle to a networkless, read-only model-worker container. The verifier imports the submitted source, refits it once on all public 2022–2023 rows, serializes and reloads that new estimator, and calls .predict() on hidden 2024 features. Hidden labels never enter the worker; the estimator artifact fitted during public research is not used for the hidden score.

Run locally

Prerequisites

Use Harbor 0.20.0 and Git LFS. The codex and claude-code adapters are bundled with that Harbor release; they are not separate Python packages.

git lfs install
git lfs pull
git lfs ls-files
uvx harbor==0.20.0 --version

git lfs ls-files must include environment/mcp-server/runtime_public.parquet and tests/hidden_data/hidden.parquet. Docker must have at least the task's declared 4 CPUs, 8 GiB memory, and 32 GiB storage available.

Data provenance and license terms are documented in DATA_PROVENANCE.md.

Tests

From the repository root, run the complete Python test suite with:

uv run --project environment/mcp-server test

The failure-path tests intentionally print exception traces. A successful run ends with OK.

This command covers the local Python tests. Harbor's Docker-based task and verifier runs remain separate and use the commands below.

Run the configured lint checks from the repository root with:

uvx ruff check environment/mcp-server/feature_engineering \
  environment/mcp-server/evalenv_shared environment/mcp-server/tests solution

Local oracle

uvx harbor==0.20.0 run -p . -a oracle -e docker -n 1

Run with Harbor's Kimi adapter by selecting kimi-cli and a provider/model:

ALLOWED_TOOLS='["kimi_cli.tools.agent:Agent","kimi_cli.tools.todo:SetTodoList","kimi_cli.tools.shell:Shell","kimi_cli.tools.background:TaskList","kimi_cli.tools.background:TaskOutput","kimi_cli.tools.background:TaskStop","kimi_cli.tools.file:ReadFile","kimi_cli.tools.file:Glob","kimi_cli.tools.file:Grep","kimi_cli.tools.file:WriteFile","kimi_cli.tools.file:StrReplaceFile"]'
uvx harbor==0.20.0 run -p . -a kimi-cli -m <provider/model> \
  --agent-kwarg version=1.49.0 \
  --agent-kwarg "allowed_tools=$ALLOWED_TOOLS" \
  --agent-kwarg max_context_size=<tokens> \
  -e docker -n 1

Harbor automatically registers the task's streamable-HTTP MCP server. Kimi's context window can be set with the adapter's max_context_size argument shown above; otherwise Harbor derives it from model metadata.

The allowed_tools argument requires a Harbor build whose kimi-cli adapter supports that structured option. Harbor converts it to an ephemeral Kimi agent file inside the running agent container; neither task image contains Kimi configuration. Kimi's default subagent definitions remain inherited. The root agent receives the listed built-in tools plus the task's five MCP tools; subagents retain their native tool policies and do not receive the task MCP configuration.

Runtime and reference result

The configured limits are 30 minutes for environment build, four hours for the agent, and one hour for verification. Cached local builds are much faster. A representative Harbor run on 2026-07-29 took 20 minutes 53 seconds end to end: 14 seconds for environment setup, 28 seconds for agent setup, 16 minutes 40 seconds for agent execution, and 3 minutes 11 seconds for verification.

Troubleshooting

• Parquet files are tiny text pointers or missing: run git lfs pull, then confirm both hashes with git lfs ls-files. Do not build until both objects are present.
• Authentication fails: export the provider key in the shell that launches Harbor. Do not add keys to task files, Dockerfiles, or committed .env files.
• The agent setup changes between runs: retain the Harbor and agent CLI version pins shown above.
• Docker reports insufficient space: inspect usage with docker system df and remove only disposable, unrelated build artifacts before retrying. The task requests 32 GiB of environment storage.
• The MCP server never becomes healthy: inspect the mcp-server container logs and verify port 8000 is reachable inside the Compose network.
• Verification fails after the agent exits: inspect the trial's verifier/test-stdout.txt, verifier/metrics.json, and verifier/reward.json. Submitted artifacts are collected from the mcp-server service, not the agent container.

The verifier writes Harbor-compatible numeric reward fields to reward.json: primary_score, reward, sharpe, cagr, max_drawdown, and pearson_ic. primary_score, reward, and sharpe are the same hidden after-cost Sharpe value. metrics.json is a separate diagnostic artifact containing the full official scoring payload, including the complete backtest metrics, causal-audit result, and refit diagnostics.