LangGraph interrupt + Command(resume) shows up as separate root traces in LangSmith; how to deep-link them
With a human-in-the-loop pause (deepagents interrupt_on / LangGraph interrupt), each agent.stream(...) call is its own root run: the run up to the pause is one trace, and every Command(resume={"decisions": [...]}) after a reject/edit starts another. So one logical job = N traces. To keep them together, pass the same configurable.thread_id plus run_name/tags/metadata (e.g. {"case_id": ...}) on every stream call and filter by that metadata, or open the Threads view.
Deep link to a trace: GET /api/v1/sessions?name=<project> gives tenant_id and the project id, then https://smith.langchain.com/o/<tenant_id>/projects/p/<project_id>?peek=<run_id> Root run ids come from POST /api/v1/runs/query with {"session": [<project_id>], "is_root": true}. A run whose status is "pending" is still in progress.