Results, files and Run Inspector.

Read the execution you actually started rather than assuming the current node preview still represents it. Retain its executionId, any relay actionId, result version, and the logical submission key.

#Read an omitted value without rerunning

The result tree can include omitted-value markers even when execution succeeded. Use the matching reader with expected_version, path and offset. Follow nextOffset and any nested omission paths. hasMore false ends that selected value, not necessarily the whole result. File retention and temporary download URL expiry differ. The illustration expands local example content only.

The retained image stays visible while the details branch is omitted. Expanding that branch reveals illustrative width, height and caption values from the same result version; it does not generate a new image or execute a run.

#Follow durable execution state

Use get_graph_run_status(execution_id), get_graph_run_result(execution_id), and get_graph_run_events(execution_id,limit) for that owned execution. Events are a bounded newest-event window, not a promise of complete history. A result can still be pending.

Terminal results can include version, valueTool, resultComplete, and omission information. Retrieve omitted retained values through get_execution_value with the exact expected_version, path and offsets. Follow nested omission paths and continuation offsets; omitted or redacted does not mean an output is empty.

#Open displayable outputs

get_displayable_outputs accepts the supported execution or node selector and returns bounded displayable media, text or JSON with output identities and warnings. Prefer the execution selector when recovering a particular run; the node's current output may have changed.

Run Inspector can open inline when the host supports MCP Apps and the deployment has its ready UI bundle. On unsupported or degraded hosts, use structured outputs. The viewer is an additional presentation layer, not proof of rerun, permanent storage or a universal client UI feature.

#Distinguish display, retention, and cancellation

Read: use executionId and the exact result version, with value readers for omitted data. Display: Run Inspector is conditional; structured outputs remain the recovery path. Retain: copy or download useful files through the supported surface; signed URL expiry differs from retained-result loss. Cancel execution and cancel relay action use different IDs and do not undo completed work. No result, file, or cancellation is requested here.

#Keep files through the right surface

Result retention and signed-media URL expiry are different. Request current access or use retained value readers while the record exists; an expired URL does not necessarily mean the retained file is gone. A missing retained result cannot be repaired by pretending its omitted value was empty.

Attached file retention uses studio_file_assets_action operations such as copy_to_library, add_library_file_to_graph, or import_from_url. Use the current attachment and expected revision, and inspect separate copy, Library, graph insertion and adoption receipts. Library IDs, storage filenames and REST fileIds are not interchangeable. Generic MCP upload_file(url,extension) imports a public URL; it is not the REST upload-session API.

#Stop the intended thing

cancel_graph_run(execution_id) can report canceled, requested, already canceled or already requested. Running work may only accept a cancellation request; this is not an undo or refund. cancel_studio_action(action_id) targets a relay action instead. Read the final execution state and preserve usable results before deciding to start again.

For large Studio captures, finish all needed paths/pages through get_studio_capture_value before release_studio_capture. A completed page is not necessarily the complete capture.