LangWatch
Sign Up
← Changelog

Traces v2 Leveled Up, and Model-Cost Preview

Model-cost rules preview the spans they match before you save them. Traces v2 adds an AI prompt view, a flame graph and an evaluator drilldown.

LangWatch Team · June 14, 2026

What You Can Do Now

Trace your agents with more fidelity

  • Codex and opencode traces now come through clean: reasoning effort surfaced as a pill, infra-span noise filtered out, and a tidy input/output tree.

  • Traces v2 rounds 3 and 4: an AI prompt view, flame graph, chip labels, evaluator drilldown, span-level costs, a fixed cost filter, and calmer live updates.

  • Update trace metadata after the fact via synthetic span injection.

  • Large-trace IO now reads from the event log as the source of truth, with leaner projections.

See what a model-cost rule will hit before you save it

  • Cost rules now show a live preview of the spans they match, suggest costs for unmapped spans, and let you click a model chip to fill the regex.

  • Org and team-scoped custom model costs are applied at ingestion, and bedrock and prefixed model ids get priced correctly.

Parameterize your prompts

  • Prompt versions now support versioned runtime parameters, exposed in the Python SDK facade and documented.

Govern your AI estate

  • Governance adds departments and a tool catalog, plus CLI path selection.

  • RBAC fixes: team membership and MCP authorization resolve via RoleBindings, and the EXTERNAL role no longer caps org permission resolution.

Build on the API

  • New REST endpoints to edit and move projects and to manage groups.

  • A new events module with POST /api/events/track (alias for /api/track_event).

  • The traces search API gained a from/select projection DSL and an updated-axis.

Langy

  • Langy is now a real, user-managed VirtualKey, so it shows up and is managed like any other key.

  • Multi-turn chat works properly now (the conversation id threads through every send), and the panel resets cleanly when you delete the active chat.

Scenario: voice testing

  • Voice scenarios support agent-initiated turns and echo-safe transcript surfacing, plus a voice-mode user simulator that speaks in real sentences instead of telegraphic fragments.

  • A long run of voice hardening: tool-only realtime turns, keepalive-aware audio receive, and attributable first-chunk timeouts.

  • ScenarioResult.messages now returns the actual conversation instead of the judge context.

And more

  • The usage page shows the real monthly event count for metered seat plans.

  • Route chunk-load failures after a deploy are now recoverable instead of a permanent white screen.

  • The LangWatch AI Gateway got a 14-minute provider timeout and fault-attributed error logging.

  • Simulations play sequential audio automatically, one clip at a time.

  • The CLI defaults login to a project (no more personal-project confusion), and a stale ingestion-key cache is invalidated after a platform revoke.

Screenshots

Codex and opencode traces, cleaned up: a tidy agent IO tree with infra-span noise filtered out.

screenshot

Model-cost rules now preview the exact spans they match before you save.

screenshot

Prompt versions with runtime parameters.

screenshot

Governance: departments.

screenshot

Governance: the tool catalog.

screenshot

Notes

  • Security: closed CRITICAL shell-quote and HIGH grpc-js alerts across many manifests this week, plus uv and npm sweeps across the SDKs and examples.

  • Performance: more ClickHouse cold-scan and out-of-memory fixes (storage metering, the span-tree drawer, analytics facets, and the model-cost preview).