npx @buildinternet/releases get langfuseEvaluations now support multimodal inputs, and the in-app agent adds prompt cache breakpoints for Claude behind an OpenAI Responses gateway. Also fixed truncated output tokens showing as blank agent turns and cached input token/cost tracking in the trace table.
Span processing now skips the span formatter when debug logging is off, and log calls use lazy %-style formatting, reducing overhead in hot paths. API spec updated from langfuse/langfuse.
In-app agent features now support OpenAI-compatible APIs and default their worker surfaces to LANGFUSE_IN_APP_AGENT_ENABLED. Also tracks evaluator creation configuration, prevents prototype-chain clobbering in OTel attribute mapping, and treats elkjs stack overflow as a too-large graph.
New evaluation UX with evaluator model filtering is available. Annotation queues can be created directly from the trace and events tables, and Anthropic Messages is now a supported model provider for Langfuse Assistant (OSS).
Fixed handling of NaN/Inf floats nested in pydantic models during serialization. Updated API spec from langfuse/langfuse and fixed a PyPI hero image aspect ratio issue.
Ask AI is now supported on self-hosted deployments. Also added LinkedIn, Reddit, and X click-id attribution for signups, more span-list filters, and in-app agent run outcome metrics; fixed step-limit truncation surfacing and PII redaction in Sentry session replays.
Fixed two automation issues: repository-dispatch tokens are now stripped from the read path, and request headers are preserved when regenerating a webhook secret.