---

name: medrax description: "Use MedRAX for chest-X-ray reasoning workflows, selective model-tool orchestration, DICOM/image preparation, Gradio interaction, and bounded ChestAgentBench evaluation with explicit resource and safety checks." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MedRAX

Use this repo skill when a task names MedRAX, ChestAgentBench, chest-X-ray agent reasoning, or asks to combine chest-radiograph tools such as classification, segmentation, report generation, visual QA, phrase grounding, DICOM conversion, or a Gradio interface. MedRAX is a LangGraph/LangChain agent that delegates to independently loaded imaging tools; it is not a clinical diagnostic system.

Route the request

• Agent setup, tool selection, prompts, model endpoint, graph state, or logs: read agent-orchestration.
• CXR classification, anatomy masks, reports, CheXagent VQA, MAIRA-2 grounding, LLaVA-Med, or RoentGen: read chest-xray-analysis.
• DICOM conversion, window/level, image validation, visualization, or original/display path handling: read image-data-utilities.
• Gradio uploads, chat threads, streamed tool results, or server configuration: read web-interface.
• ChestAgentBench/Eurorad data, bounded multimodal evaluation, or JSONL result logs: read benchmark-evaluation.

For a multi-stage task, start with image-data-utilities for DICOM/image preparation, then chest-xray-analysis for the selected model tool, and use agent-orchestration or web-interface only for the surrounding control plane. Do not initialize every tool by default.

Installation and first checks

Install the package from the target MedRAX checkout with its published metadata, preferably in an isolated Python 3.10/3.11 environment. A CUDA PyTorch build is the practical path for model-backed tools; CPU is sufficient for import, schema, LangGraph-with-fake-model, DICOM, and visualization checks. The repository declares a pinned Transformers Git dependency and broad runtime dependencies, so use the package's resolved compatibility set rather than blindly upgrading individual ML libraries.

python -m pip install -e .
python -c "import medrax, medrax.agent; print('MedRAX import OK')"
python sub-skills/agent-orchestration/scripts/check_medrax_import.py

The bundled checker is read-only: it reports import/signature facts without loading model weights or contacting a provider. Read references/runtime-compatibility.md before changing Transformers, Diffusers, Torch, Gradio, or LangGraph versions.

Non-negotiable preflight

1. Identify whether the input is JPG/PNG or DICOM. Convert DICOM for display or model input with the utility route, but preserve the original path and its metadata when it matters.
2. Choose the smallest tool set that answers the question. Utility tools do not require model weights; model tools may download large Hugging Face assets or require CUDA, bfloat16, bitsandbytes, and substantial VRAM.
3. Declare device, model/cache location, temporary-output location, model availability, and whether network or credentials are authorized before constructing a weight-backed tool.
4. Treat outputs as assistive observations. Preserve analysis_status, errors, image paths, coordinate spaces, and uncertainty; do not present scores, generated reports, masks, or synthetic images as a diagnosis or calibrated clinical measurement.
5. Keep API keys, DICOM/PHI, model caches, logs, temporary outputs, and uploaded files outside the generated skill tree.

Shared troubleshooting

Read references/troubleshooting.md for cross-cutting install/import compatibility, missing optional dependencies, model-cache and CUDA failures, path/config validation, provider setup, and output/privacy boundaries. Each sub-skill has more specific recovery guidance.

Provenance and scope

Read references/repo-provenance.md before using this skill against a different checkout or deciding whether a refresh is needed. The generated graph deliberately excludes large demo media and medical-data binaries, exploratory experiments, benchmark-generation scripts, and vendored LLaVA serving internals; their reusable behavior is distilled into the linked references and safe helpers.