Changelog
What's New in CognitiveStackTech.
Every release, every improvement, every fix. We ship every week. Here's the log.
v1.4.2November 6, 2026
Faster replay + tool versioning
- NEWTool versioning — pin a tool schema to a specific deploy so old traces still replay correctly.
- IMPROVEDReplay trace load time is 4× faster for traces over 20 steps.
- FIXFixed a race condition that could cause concurrent eval runs to cross-contaminate.
v1.4.1October 28, 2026
SDK v2.3 + Anthropic Sonnet 4.5
- NEWNative support for Claude Sonnet 4.5 in the model router.
- IMPROVEDFull type hints across all tool schemas.
- FIXFixed a bug where retries on tool calls could double-charge. (Invoiced credits were applied automatically.)
v1.4.0October 14, 2026
OTel-native observability
- NEWOpenTelemetry tracing for every request. Native Datadog and Honeycomb integrations.
- NEWFull trace export to any OTLP-compatible backend.
- IMPROVEDReduced default trace payload size by 40%.
v1.3.8September 30, 2026
Eval suite improvements
- NEWSynthetic data generator for rare edge cases.
- IMPROVEDLLM-as-judge now supports custom rubrics.
- FIXFixed UI bug where large eval results could hang the dashboard.
v1.3.0September 12, 2026
Deterministic replay
- NEWEvery request is now replayable. Debug any failure with the exact original inputs and state.
- NEWLocal replay CLI for offline debugging.
v1.2.0August 15, 2026
Retrieval v2 + hybrid search
- NEWHybrid BM25 + vector search as the default retriever.
- NEWAuto-chunking with semantic boundaries.
- IMPROVEDEmbedding refresh latency reduced by 3×.
v1.0.0June 1, 2026
General availability
- NEWCognitiveStackTech leaves beta. Production-grade from day one.
- NEWSOC 2 Type II certification achieved.
- NEWEnterprise tier with on-prem deployment option.