Studio & Datasets Makeover, Scoped Data Privacy, and TTFT

What You Can Do Now

Build in the studio

• Datasets got a makeover: a cleaner list, in-cell editing, and full editor parity.

• A new if/else node for branching workflows, a redesigned entry point, and an evaluate API.

• Rename code blocks, drag-sync code-node inputs, and run a workflow up to a chosen node with entered results.

• The prompt editor no longer render-loops and crashes the playground, and the studio stops falsely showing unsaved changes after a database load.

Protect sensitive data

• Unified, scoped data privacy: redact secrets and PII (API keys, database URLs, custom labels) across your traces, configurable at organization, team, or project scope.

• Privacy rules support an inherit option and a shareable drawer link.

Measure time to first token

• TTFT is now a first-class metric, with a p95 TTFT column, tooltips, and SDK timing canonicalization, captured from Vercel msToFirstChunk and OpenLLMetry chunk events.

Control trigger emails

• Trigger emails got spam prevention: a test-fire lockdown, an hourly cap, a per-project daily cap, and one-click unsubscribe.

• Thumbs-down alerts now require a real down-vote before they fire.

Sign in and manage access

• Credential sign-in now has a forgot and reset password flow.

• Fixed a sign-in loop on account-method collision, and restored project base and legacy API key rotation.

Feed richer inputs to models

• Workflows can fetch remote attachment URLs and deliver them to the model as content (text and images), with a clear error when a URL is bad.

• Experiments v3 handles multimodal image splitting and lets you abort interactive workbench runs.

Retention and onboarding

• Plan-based visibility windows: data older than your plan's window now shows as a redacted teaser instead of disappearing.

• Polished traces onboarding, and you can replicate online evaluators across projects.

Scenario

• A new realtime_langwatch_session context manager instruments live OpenAI Realtime apps.

• Per-role voice modality negotiation, a race-condition-free multi-tenant ScenarioConfig, and a context parameter on scenario.judge().

Screenshots

The datasets makeover: in-cell editing with editor parity.

Scoped data privacy: secrets and database URLs redacted automatically.

Time to first token as a first-class column.

Forgot password for credential sign-in.

Studio: run a workflow up to a chosen node.

Remote attachments delivered to the model as image content.

Notes

• Reliability: large offloaded trace IO over 64KB now resolves correctly on full detail reads, including past mixed-type OTLP attributes.

• Security: bumped Python infra deps (cryptography, tornado, python-multipart, aiohttp), pyjwt, the protobufjs floor, esbuild/vite/form-data floors, and hono/ws in Scenario.

• Performance: a large ClickHouse sweep this week, time predicates to stop cold scans on stored spans and logs, single-pass argMax folds for topic clustering and simulation dedup, and pruned event-log reads.