Release Notes

Version 2.2.0

Current Version — Latest release

Added

elsai Agents SDK integration: Monitor elsai Agents applications via arms.elsai_agents_hook — a hook provider that bridges elsai-agents lifecycle events into ARMS trace storage.
Optional install extra: pip install 'elsai-arms[elsai-agents]' bundles elsai-agents>=0.2.1 for hook support (on-prem only).
Standalone agent tracing: Agent, generation, and tool spans with framework: "elsai_agents" and parent-child hierarchy.
Graph and Swarm orchestrators: Orchestrator and node chain spans, Swarm handoff instant spans, and nodes_executed rollup from metadata.elsai_node_id across ClickHouse, MongoDB, and DynamoDB.
Nested agent.as_tool() tracing: Full span trees for reject, queue, and spawn modes — manager → tool → child agent → generation.
Framework metadata on spans: Orchestrator root spans tagged elsai_graph or elsai_swarm; child spans carry elsai_agents with orchestrator context in metadata.
Hook-path run rollup: Token usage and cost aggregated from generation spans at end_run() via agent_metrics trace IDs (hook path keeps llm_calls empty by design).

Fixed

Parallel as_tool bridge resolution: Correct parent span assignment when multiple tools are registered before child agents forward (queue sequential and spawn parallel cases).
Graph async concurrency: Agents inside parallel Graph node batches parent under the correct node span instead of a stale active node.
Spawn worker tool spans: Inner tools on spawn child agents (for example skills) parent under the child agent span; orphaned as_tool session cleanup on outer tool failures only.

Changed

Default on-prem package version: Installation docs and wheels target elsai-arms==2.2.0.

See elsai Agents monitoring for setup and usage.

Version 2.1.0

Added

Trace completeness score: Run-level trace_completeness_score derived from span hierarchy integrity (orphan spans, broken chains).
Tool success rate: Run-level overall_tool_success_rate with per-tool breakdown from tool span status.
Per-agent-call metrics: per_agent_call_metrics map keyed by trace ID for drill-down into individual agent traces.
Thread-level analytics: Optional thread-level trace completeness and tool success aggregation (ARMS_THREAD_ANALYTICS_ENABLED).

Fixed

LangChain agent edge cases: Callback handler fixes for agent trace edge cases in LangGraph workflows.
Amazon Textract OCR: Resolved Textract extractor issues in OCR monitoring.

Version 2.0.4

Fixed

SDK Database Resolution: Hardened DB discovery to use the backend API as the sole source of truth; eliminated legacy env-var heuristics.
Configuration Validation: SDK now raises a clear error at startup for misconfigured DB settings instead of silently routing to the wrong database.

Version 2.0.3

Added

Support for Message-Based Parameters: Support for message-based parameters in LLM calls for compatibility with modern chat models.

Fixed

MongoDB Cost Calculation: Fixed MongoDB cost race condition issue.

Version 2.0.2

Added

Explicit Model Specification: Added the ability for developers to manually specify model names in decorators and agent metadata for highly accurate cost reporting.

Version 2.0.1

Fixed

Dynamic Model Configuration: Allowed model name to be passed from user input instead of hardcoding
Agent Output Serialization: Fixed AIMessage not serializable issue for agent observations

Version 2.0.0

Added

ClickHouse Integration: ClickHouse added as a database backend option alongside MongoDB and DynamoDB
Async Flushing: Asynchronous buffered flush for run data; configurable triggers (time, event count, buffer size) and optional manual flush for deterministic send

Version 1.3.3

Fixed

LLM Monitor in Agent Metrics: Fixes for LLM monitor behavior when used with agent metrics

Version 1.3.2

Added

OCR – Azure Computer Vision: OCR monitoring updated to support Azure Computer Vision service

Version 1.3.1

Added

Agent Cost Tracking: Extracts and tracks costs from LangGraph and agent framework traces

Version 1.3.0

Added

Web Search Monitoring: Comprehensive tracking for web search operations across supported providers

Version 1.2.2

Fixed

OCR Monitoring: Confidence score bug fixed in OCR monitoring

Version 1.2.1

Added

Agent Monitoring: Agent monitoring and distributed tracing capabilities

Version 1.2.0

Changed

Enhanced Security: Restructured backend to use API instead of direct save

Version 1.1.1

Added

OCR – AWS Textract: AWS Textract support added for OCR monitoring

Fixed

Bug fixes and stability improvements

Version 1.1.0

Changed

Database Structure: Restructured database to use a separate table for runs

Version 1.0.3

Changed

Logger: Improved logging behavior and output

Version 1.0.2

Added

LLM Streaming: Support for streaming LLM calls in monitoring

Version 1.0.1

Added

DynamoDB – Cognito: Cognito support for DynamoDB authentication
LLM: GPT-5 model support in LLM monitor

Version 1.0.0

Added

DynamoDB Support: DynamoDB added as a database backend option
OCR Monitoring: OCR operation monitoring and metrics
RAG Monitoring: RAG (retrieval-augmented generation) operation monitoring
Embedding Monitoring: Embedding operation monitoring
LLM Monitor: Improved LLM monitor with support for more models and basic content quality assessment

Version 0.1.1

Fixed

Bug fixes and stability improvements

Version 0.1.0

Internal Release

Added

LLM Monitoring: LLM call monitoring and metrics
Custom Metrics: Support for custom metrics logging