On-prem 2026-11-10-003

Agents read Playwright traces through trace links and record their work as sessions, and projects open on the Overview, on-prem, with Helm chart 1.1.

In this release

NewSessions: record every agent task
Your agent saves its browser recording, or the files of any task, as a Currents session. You get one link: a page for people and a Markdown version for other agents. Post it in Slack, Linear or on the pull request.
NewProject Overview
Projects now open on the Overview, a page that shows whether the project is healthy, whether it is getting better or worse, and what to fix first.
NewTrace API for agents
Currents has a trace API made for agents. Instead of downloading a Playwright trace, an agent reads it one part at a time and picks each next call from what it just read.

The images are 2026-11-10-003, shipped as Docker Compose release on-prem/2026-11-10-003 and Helm chart 1.1.0.

With Docker Compose, pull the new images before you stop anything. This release changes the compose files. Pull while the current version keeps serving, then run docker compose up -d, without docker compose down. In testing, Currents was unreachable for 10–15 seconds, and CI runs in progress kept recording.
If you use AI explanations, turn on the AI worker in the same upgrade. AI tasks move to a new ai-queue-worker, which is off by default. Until it runs, AI requests wait and never finish.
The first start builds six small indexes and runs five migrations. They take seconds: on an install with 3.1 million spec file runs, startup finished in 18 seconds. With Helm, the upgrade first checks what the index builds would cost, and stops before any pod rolls if they would take long.

The steps for each install:

Trace links are served by Currents itself at <your Currents URL>/t, with nothing to set up. To serve them on a host of their own, set currents.domains.traceLinkHost with Helm, or TRACE_LINK_BASE_URL with Docker Compose.
The trace work runs in the API. In testing, an API with 2 CPUs and 2 GB answered about 2 agent investigations per second. See Trace Links for sizing.
With Helm, director and server pods keep serving for 10 seconds while a rollout replaces them, so upgrades after this one don't drop requests.