---

name: setup-and-ops description: "Install, inspect, configure, and safely operate the RD-Agent package and its CLI, UI, logs, and environment checks." metadata: disco-role: operating parent-skill: rd-agent license: MIT disable-model-invocation: true

RD-Agent setup and operations

Use this sub-skill for installation, import/CLI failures, health checks, configuration discovery, log/UI operation, and safe maintenance. For an actual research run, route onward to the domain sub-skill after the runtime is proven.

Install and identify the active package

Use a supported Python 3.10 or 3.11 environment. From a checkout, prefer:

python -m pip install -e .
python -m pip check
python -c "import rdagent; print(rdagent.__file__)"

The last line is intentionally not an RD-Agent command; use the bundled scripts/verify_install.py instead. For a published installation, python -m pip show rdagent and the import path must agree. Do not mix a globally installed rdagent with a different checkout through PYTHONPATH.

Probe the CLI in increasing cost order

rdagent --help
rdagent health_check --no-check-env --no-check-docker
rdagent data_science --help
rdagent fin_quant --help
rdagent fin_factor_report --help
rdagent llm_finetune --help
rdagent general_model --help
rdagent ui --help
rdagent server_ui --help
rdagent ds_user_interact --help
python -m rdagent.scenarios.rl.autorl_bench.run --help

Use scripts/verify_install.py to collect these results as JSON. A passing --help check proves registration only; it does not prove model access, Docker, Qlib data, a GPU, or an external benchmark.

Configuration and processes

• Keep secrets in an uncommitted .env or environment manager. Never paste API keys into prompts, generated code, or logs.
• Choose a run-specific output/log directory and record the resolved configuration before starting an agent loop.
• For the generic UI, use rdagent ui --log-dir <log-dir> --port <local-port> after checking its help. Bind locally and stop it after inspection.
• server_ui and the fine-tune UI may need extra assets or a long-running process. Start them only when the user explicitly wants a UI smoke test.
• Health checks that skip environment/Docker checks are lightweight package checks, not a deployment readiness proof.

Failure handling

• Import/command not found: inspect pip show, import rdagent, and the active interpreter first.
• Health check failure: separate missing optional tools from hard package/import errors; preserve the exact check output.
• UI startup failure: check port ownership, log directory readability, and optional frontend dependencies.
• Long or silent process: inspect the run log and child processes before retrying; do not launch duplicate listeners.

Read operational-checklist.md for a compact preflight and recovery checklist. Use the parent troubleshooting guide for cross-scenario failures.