Download Latest Version v1.3.4 - Tool call cost, improved detector judge source code.zip (16.1 MB) Google Add to Preferred Sources
Home / v1.3.1
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-10-16 6.3 kB
0
v1.3.1 - Widget cards in chat, sandbox reclaim source code.tar.gz 2026-10-16 14.0 MB
0
v1.3.1 - Widget cards in chat, sandbox reclaim source code.zip 2026-10-16 15.3 MB
0
Totals: 3 Items   29.4 MB 0

This release draws widget and dashboard reads as live cards in the agent chat panel, and stops the agent leaking sandbox containers on the Docker host. Detector scan quota now counts only runs that reached the model, legacy embedding calls record their token usage, and the trace list pages stably when traces share a start time. The docs gain a generated REST API reference, an Authentication page, and a consistent template across every integration page.

New Features

  • Widget and dashboard cards in the chat panel. When the agent reads a saved widget or a dashboard, the panel renders the widget's own chart, or the dashboard on its stored grid scaled to the card, over the window the read answered for, instead of only a tool line and prose (#2606).

Bug Fixes

  • Sandbox containers leaked on the Docker host. Sandbox ownership lived only in a map inside the agent process, so traceroot-sandbox-* containers accumulated with nothing able to reach them. Containers now carry session and owner labels, the agent reclaims orphans at startup, and idle sandboxes are torn down after SANDBOX_IDLE_TTL_MINUTES (#2124).
  • Failed detector runs consumed scan quota. The usage meter counted every detector_runs row, including failures that never reached a model, so Free workspaces hit the scan cap early and the paid overage denominator was diluted. Each run now collapses to its latest version and counts only when it completed or used inference (#2607).
  • Legacy embedding usage was dropped. Legacy embed() / embedMany() provider calls lost their reported token usage during ingestion. Their model and usage fields are now recognized, so input tokens record once on the provider call; spans without reported usage stay uncounted rather than estimated (#2608).
  • Trace list pages skipped or repeated tied traces. The list ordered by trace_start_time only, so traces sharing a millisecond had no fixed order across OFFSET pages. trace_id now breaks the tie (#2570).
  • Self-traces picked up SDK default instrumentation. The agent service and detector worker initialized self-tracing without instrumentModules, so the SDK installed provider hooks whose spans bypassed the services' redaction and budget callbacks. Both now pass an empty module map and keep only their explicit captures (#2576).
  • Save buttons wrote no-op versions. Dataset and evaluation edit surfaces used mixed labels and let you save with nothing changed, publishing redundant versions. Every edit-commit button says "Save" and stays disabled until a field changes (#1955).

Improvements

  • Docs. A REST API reference generated from the public OpenAPI spec plus an Authentication page covering which credential each route accepts (#2569, [#2612]); every integration page follows one Install, Set up, Usage, Verify, Limitations, Next steps template (#2450); the CLI page documents browser login, project scoping and --fields projections (#2478); onboarding covers evaluations, dashboards, threshold alerts and workspace roles (#2381); the agent self-tracing runbook states the deploy order (#2572); the Mistral example uses the current SDK import path (#2379).

Upgrade notes

  • One migration runs on upgrade. ClickHouse 016 adds used_inference Bool DEFAULT false to detector_runs; it is outside the sort key, so the ALTER is metadata-only. Legacy failed runs keep the old behaviour and are not counted (#2607).
  • Agent sandbox sweep. On start the agent removes every docker sandbox it did not create, assuming one agent per Docker host. Set SANDBOX_RECLAIM_ON_START=false when several agents share a host. SANDBOX_IDLE_TTL_MINUTES defaults to 30; 0 disables the sweep (#2124).
  • Deploy order for self-tracing. Deploy the agent service before the RCA worker. A newer worker reading an older agent's stream records executions as failed even when the RCA completes, and finishing the rollout does not repair those records (#2572).

What's Changed

New Contributors

Full Changelog: https://github.com/traceroot-ai/traceroot/compare/v1.3.0...v1.3.1

Source: README.md, updated 2026-10-16