Release notes · Observability

Observability Release Notes

Changelog for @cognipeer/observability. For usage, see the Observability documentation.

All notable changes to this project are documented here. The two packages — @cognipeer/observability (npm) and cognipeer-observability (PyPI) — share one version number and one changelog.

The format follows Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

[0.2.0] - 2026-08-17

Added

• Session-level metadata (init({ metadata }) / client default, and a per-startSession/trace override merged on top) — free-form attribution tags reported alongside agent on every session payload (batch, stream /start, and the streaming reopen path). The Console now supports grouping and reporting on these as a dynamic group_by/group_by_entity=metadata.<key> dimension, so callers can slice spend by anything they tag a session with (e.g. { complexity: "complex" }) without a schema change on either side. Distinct from the existing per-event metadata — that one is free-text content (redacted, size-capped); this one is short, structured attribution tags and is never redacted or capped client-side.

0.1.0

First release.

Added

• Core, in Python and TypeScript, with no required dependencies in either: session/event/section model, background delivery with retry, secret redaction, base64 stripping, content capping, and three delivery modes (auto, stream, batch).
• LangChain callback handler, working from langchain-core 0.1 through 1.x and @langchain/core 0.1 through 1.x. Captures prompts, completions, per-call token usage with the cache-read breakdown, tool calls with arguments and results, and the tool definitions bound to each model call.
• LangGraph support on the same handler, plus graph_config / langgraphConfig and trace_graph / withCognipeerTracing. Interrupts are recorded as control flow rather than failures, and a conversation's runs group by thread id.
• OpenAI Agents SDK tracing processor for both SDKs, mapping every span kind and counting tokens only from leaf model spans.
• Claude Agent SDK message-stream tracer and a trace_query / traceQuery drop-in for query(), pairing each tool_use with its tool_result and translating Anthropic's cache-exclusive token accounting.
• Vercel AI SDK integration with three routes — the native telemetry integration (ai 6+), a language-model middleware (every version), and experimental_telemetry with a bundled tracer (ai 3–6).
• n8n integration: a polling bridge with a cognipeer-n8n CLI that works on n8n Cloud and Community, and a workflow.postExecute external hook for self-hosted installs.
• OpenTelemetry span exporter in both languages, normalising OpenInference, current OTel GenAI and legacy OpenLLMetry attributes — which is what makes CrewAI, LlamaIndex, Pydantic AI, Google ADK, AWS Strands, Semantic Kernel, smolagents, Haystack and DSPy work without a bespoke integration.
• Manual instrumentation: @observe / observe() and trace() for agents no framework covers, with sync, async, generator and async-generator support in Python.
• Runnable examples per framework and a full guide per integration in the Console documentation.