# Chartography with tools: OpenCode answers each question in its own sandbox session, with a Python image toolkit # and OpenCode's own tools (bash, read, glob, grep, write, edit, todowrite; `read` returns images to the model). # The resources server lends the sandbox at each episode's seed and stops it at verify. The sandbox reaches only # the policy model server (model_only), so the model cannot look answers up. Grading is the same as in the # simple_agent setting (chartography.yaml): the benchmark's own judge over the final assistant message. # # Needs: CHARTOGRAPHY_OPENCODE_SANDBOX_IMAGE, an image built from resources_servers/visual_coding_opencode/sandbox_image # (OpenCode, its plugin dependencies and a Python image toolkit; the agent below pins the same OpenCode build), # OPENSANDBOX_DOMAIN and OPENSANDBOX_API_KEY, a host the sandbox can reach (use_absolute_ip), and a judge # (judge_model_openai.yaml, or your own judge_model with host: 127.0.0.1). config_paths: - nemo_gym/sandbox/providers/opensandbox/configs/opensandbox.yaml use_absolute_ip: true absolute_ip_server_types: [responses_api_models] chartography_opencode: # The simple_agent setting's server (chartography.yaml), plus the lent sandbox. It is not chained: the lent sandbox # may reach the policy model's host, so no no-tools server may listen there. The config test checks that the two # settings grade the same way. resources_servers: chartography: entrypoint: app.py host: 127.0.0.1 domain: other verified: false description: 'Professional chart reading with tools: OpenCode answers one chart question in a sandbox session with a Python image toolkit; graded against a golden answer by the Chartography benchmark''s own LLM judge' value: Train and measure tool-assisted, precise reading of specialized charts judge_model_server: type: responses_api_models name: judge_model judge_chat_create_params: messages: [] judge_prompt_template_fpath: prompt_templates/judge.txt judge_max_attempts: 5 judge_endpoint_max_concurrency: 64 # A run that does not complete (OpenCode's timeout, an output cap, a context limit or a failed run) scores 0, # as in visual_coding_opencode, instead of being graded on whatever it last wrote. score_unfinished_as_zero: true sandbox_provider: sandbox sandbox_config: image: ${oc.env:CHARTOGRAPHY_OPENCODE_SANDBOX_IMAGE,null} # Above the episode deadline (default_episode_timeout_seconds below) plus grading and cleanup. ttl_s: 7200 ready_timeout_s: 1200 resources: {cpu: 2, memory_mib: 8192} sandbox_workdir: /workspace sandbox_network_access: model_only sandbox_model_server: type: responses_api_models name: policy_model chartography_opencode_agent: responses_api_agents: opencode_agent: entrypoint: app.py host: 127.0.0.1 description: Chartography with OpenCode in a sandbox session that has a Python image toolkit resources_server: type: resources_servers name: chartography_opencode model_server: type: responses_api_models name: policy_model model: ${policy_model_name} opencode_version: 1.17.11 concurrency: 64 # The model reasons on every turn, and its earlier reasoning stays in the history OpenCode sends back. thinking: true # The chart is a prompt image; image_input also lets `read` return crops and plots the model makes. image_input: true # 1800 s per task and no turn cap. timeout: 1800 context_window: 131072 max_output_tokens: 32768 session_close_timeout_seconds: 90 sandbox_executable_path: /opt/opencode/bin/opencode sandbox_executable_sha256: 0254a429cd0e6cf0ba53fc01672cf98e4a8dc728f7fa94be88f1e4b3645e6ded # pragma: allowlist secret sandbox_plugin_dependencies_path: /opt/opencode/opencode-1.17.11-plugin-dependencies.tar.gz sandbox_plugin_dependencies_sha256: 77b03eba45c8d7fe3d0a58bfc215f6fb99a1db4dd18a7e3d91fd29981fddafb1 # pragma: allowlist secret opencode_config: tools: webfetch: false websearch: false task: false permission: bash: allow edit: allow external_directory: allow doom_loop: allow compaction: auto: false datasets: - name: example type: example jsonl_fpath: resources_servers/chartography/data/example.jsonl num_repeats: 1 license: Apache 2.0 chartography_opencode_environment_server: environment_servers: single_agent_turn_legacy: entrypoint: app.py host: 127.0.0.1 resources_server: type: resources_servers name: chartography_opencode agent_server: type: responses_api_agents name: chartography_opencode_agent # OpenCode works in the lent sandbox with its own tools; the server serves none. resources_tool_transports: [] # One sandbox per episode: admit as many episodes as the agent runs at once. max_concurrent_episodes: 64 queue_timeout_seconds: 86400 default_episode_timeout_seconds: 5400