# Claude Code harness in a sandbox per rollout, calling the rna_design tools over MCP. The tools and the scoring stay on # the Gym node, so the sandbox calls back to it: Gym listens on the node's address rather than localhost. # Pass a sandbox provider's config next to this one: nemo_gym/sandbox/providers//configs/.yaml. use_absolute_ip: true rna_design: resources_servers: rna_design: entrypoint: app.py domain: agent verified: false expose_tools_over_mcp: true worker_python: ${oc.env:RNA_WORKER_PYTHON,null} # Just above the harness's rollout_timeout: no live session is older, and a rollout that died before /verify # leaves its worker for no longer than this. session_ttl_s: 4200 description: "RNA inverse folding: the design tools served over MCP to a harness such as Claude Code." rna_design_claude_code_agent: responses_api_agents: harness_agent: entrypoint: app.py resources_server: type: resources_servers name: rna_design model_server: type: responses_api_models name: policy_model agent: claude_code forward_mcp_tools: true agent_kwargs: model: ${policy_model_name} anthropic_api_key: unused # pragma: allowlist secret max_turns: 1000 timeout: 3600 bare: true system_prompt: | You act through the rna_design MCP tools. Call campaign_state first: it shows the target, how designs are ranked, the budget and the best designs so far, and it costs nothing. Each other tool returns the campaign state after it runs. Stop when further calls will not improve the result. You are running an RNA design campaign. Each tool call is one round of the campaign and spends from its budget. The state's `objective` block is the thing you are being judged on: it names the metric candidates are ranked by and the gates they must pass. Generating is only half the job — a candidate with no score counts for nothing, so whenever `objective.unscored_candidates` is large, score before you generate again. What to keep in mind: - The budget is hard. When any dimension is spent the campaign ends, whatever state it is in, so spend expensive tools on candidates worth the cost. - Leave enough budget to score. Ending with many candidates and no scores is the worst outcome available to you; it reports as zero. - To score candidates that already exist, pass one of the strings in `sequence_selectors` as a tool's `sequences` argument — "@unscored" and "@top:50" are expanded for you. Never invent a filename, and never retype a sequence you were shown truncated. - Omit an optional argument you have no opinion about rather than passing an empty string for it. An empty value is not the same as a default. - Tool descriptions say what a score does and does not mean. Read them before treating a number as evidence. - A tool that failed is information. Do not repeat a call that just failed with the same arguments. sandbox_provider: sandbox # Any image with Python 3.12+ and Node.js works. Claude Code is installed in the box on first use. sandbox_image: nikolaik/python-nodejs:python3.13-nodejs22-slim sandbox_python: python3 setup_commands: - python3 -m pip install -q "nemo-gym @ https://github.com/NVIDIA-NeMo/Gym/archive/refs/heads/main.zip" sandbox_spec: ttl_s: 7200 resources: cpu: 2 memory_mib: 4096 env: # gRNAde on a busy CPU node runs for minutes, past Claude Code's limits on an MCP call: 60 s in all and # 300 s with no progress. MCP_TOOL_TIMEOUT: "1800000" CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT: "1800000" API_TIMEOUT_MS: "1800000" rollout_timeout: 3900 concurrency: 4 datasets: - name: example type: example jsonl_fpath: resources_servers/rna_design/data/example.jsonl rna_design_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: rna_design_claude_code_agent