Repository navigation
Fix span/event normalization and multi-hop agent ownership - #33
Conversation
- Backfill agent ownership regardless of record arrival order - Merge lifecycle events with authoritative completed spans - Preserve sessionless MCP/A2A ancestry in private trace context - Correct MAS identity and reconcile graph nodes with observed agents - Fix same-named Agent/Tool edge typing and session boundary timing - Share reconciliation between batch normalization and the realtime worker Signed-off-by: Pavan Sudheendra <pavan0591@gmail.com>
Signed-off-by: Pavan Sudheendra <pavan0591@gmail.com>
Signed-off-by: Pavan Sudheendra <pavan0591@gmail.com>
Neo4j views must exclude private normalization metadataThe realtime normalizer at the moment persists These are internal bookkeeping records—not ontology entities or execution-graph nodes. Their presence does not change the KG ontology. Unfiltered queries can expose them as disconnected nodes and inflate node counts. Update graph visualizations, entity counts, label selectors, and ontology exports to exclude them: For relationship-based views, exclude both endpoints: Since, this is a correlation state, maybe we can also persist them in worker memory. The trade-off being that it loses context on restart and is unsafe when run in a distributed setting. |
Signed-off-by: Pavan Sudheendra <pavan0591@gmail.com>
|
Switched trace context to in-memory storage; ontology nodes and relationships still use Neo4j.
|
Description
Make span and event normalization converge on the same execution graph,
independent of arrival order.
Previously, child calls arriving before their agent could remain disconnected.
MCP server-side tool spans also pointed to intermediate transport spans rather
than directly to agents. Participating services could become extra MAS nodes,
and arbitrary execution-graph nodes could be misclassified as Agents.
This change:
_NormalizationTracerecords.Type of Change
Checklist