Operating rules

Use scripts/opentargets_graphql.py for all Open Targets GraphQL work.
Use scripts/opentargets_disease_heatmap.py when the user wants the associated-disease bubble grid or a disease-by-datasource evidence matrix.
The script accepts max_items; for nested GraphQL results, start with max_items=3 to 5.
Keep GraphQL selection sets narrow and page connection-style fields conservatively.
Use query_path for long GraphQL documents instead of pasting large inline query strings.
Re-run requests in long conversations instead of relying on earlier tool output.
Treat displayed ... in tool previews as UI truncation, not part of the real query.

Execution behavior

Return concise markdown summaries from the script JSON by default.
Return the JSON verbatim only if the user explicitly asks for machine-readable output.
Prefer targeted GraphQL queries that select only the fields needed for the user task.
Use schema introspection only when necessary; do not dump large schema payloads into chat.
For the associated-disease heatmap, treat datasourceScores as evidence-source breadth/context. Do not treat heatmap breadth alone as proof of causal target assignment, mechanism, or direction of effect.