Open WebUI 0.12.5 Adds Conversation Execution Tracing and Fixes Duplicate Counting of Streamed Token Usage

Open WebUI released version 0.12.5 on October 29, adding a Sessions view to Analytics in the admin panel and bringing monitoring down to the individual conversation level. Administrators can filter by model, token range, and sub-agent, then inspect saved replies and the duration and results of tool calls. For teams self-hosting multiple models and agent workflows, this provides a way to investigate failed steps from usage statistics. Release announcement

The feature has a clear access boundary: ENABLE_ADMIN_CHAT_ACCESS must be enabled for Sessions to appear. The official documentation also states that the relevant API returns 403 when access is disabled. Because this interface can expose other users’ conversation titles, tool calls, and error information, deployers should treat it as part of administrators’ ability to read conversation data. Analytics documentation

The release also fixes a streaming usage aggregation issue. When a provider resends usage information in each stream chunk, earlier versions could add the repeated reports together, inflating the token count for an individual reply and in the admin panel. The fix instead uses the value from the provider’s final report. This affects internal usage interpretation, but does not support an inference that providers had overcharged on bills. Release announcement

Operationally, several metrics still need to be distinguished. A Sessions Span includes idle time between messages, so it should not be treated as inference latency. Token fields indicate when usage is unrecorded or partially recorded, and detailed timing is saved only for new replies. Upgrading therefore does not automatically fill in execution timelines for older conversations. Analytics documentation

For operations teams, useful follow-up checks include whether the model endpoints in use consistently return usage data, and whether tool errors retain enough detail in traces to aid investigation. Based on the publicly described features, the new interface can help narrow troubleshooting; whether it is sufficient for cost allocation or service quality monitoring still needs to be confirmed against data from each deployment.