Python agent release notes-Application Real-Time Monitoring Service(ARMS)-阿里云帮助中心

This topic describes the release notes for the Python agent of Application Real-Time Monitoring Service (ARMS) Application Monitoring.

View the agent version

Log on to the ARMS console.
In the navigation pane, choose Application Monitoring > > Agent Management, and then click the Agent Release Notes tab.

Agent versions

Version	Release date	Release notes
2.10.3	June 22, 2026	New features Added support for anthropic-sdk-python. For more information, see Supported Python components and frameworks. Added support for collecting the model reasoning process for large language model (LLM) calls. For more information, see LLM trace field definitions. Added support for collecting request IDs for LLM calls. For more information, see LLM trace field definitions. Fixed issues Fixed an issue where the AgentScope framework swallowed exceptions in certain scenarios. Fixed an issue where the installation and reporting of continuous profiling failed in scenarios where pip was not installed.
2.9.0	June 12, 2026	New features Added support for the Qwen-Agent, QwenPaw, and Hermes frameworks. For more information, see Supported Python components and frameworks. The LangChain plugin supports observability for Rerank calls. Enhanced the observability of the SGLang inference engine. For SGLang engine versions 0.5.9 and later, the agent supports fine-grained performance tracing, including the latency of components and stages such as the tokenizer, scheduler, Prefill, and Decode. For more information, see Enable observability for vLLM and SGLang inference engines. Fixed issues Fixed an issue where span states could interfere with each other in AgentScope ReAct scenarios involving multi-turn calls, nested agents, and concurrency. Fixed an issue where span end times were inaccurate for Google ADK streaming and concurrent calls within the same session. Fixed the GenAI semantic collection for AI agent frameworks such as CrewAI, Google ADK, and LiteLLM. Fixed an issue where ARMS health check requests were captured in application traces.
2.8.3	June 4, 2026	New features Added support for custom metrics using the OpenTelemetry Python SDK. Fixed issues Fixed an issue where system metric collection printed error stack traces to the user's standard output in certain scenarios. Fixed an issue with metric dimension divergence in multi-process scenarios. Fixed an issue where the DashScope plugin failed to collect output content in certain scenarios.
2.7.0	May 18, 2026	New features Added support for viewing observability data grouped by AI agent in the AI Agent Observability feature of Cloud Monitor 2.0. Added support for integrating AI Profiling in ECS scenarios. Added support for Prefill/Decode (PD) disaggregation scenarios for observability for vLLM and SGLang inference engines. Fixed issues Fixed an issue where error metrics were not collected for failed LLM calls in the OpenAI plugin. Fixed slow agent startup caused by network issues in certain scenarios.
2.6.4	April 21, 2026	New features The vLLM plugin supports Token-level analysis. For more information, see Enable observability for vLLM and SGLang inference engines. Added support for Python 3.14 and discontinued support for Python 3.8. For more information, see Supported Python components and frameworks. Fixed issues Fixed an issue where the time to first byte (TTFB) metric was not reported for the AgentScope framework. Added compatibility for associating `session_id` and `user_id` from LangGraph. Fixed an issue where the `gen_ai.provider.name` attribute was not recorded for OpenAI plugin Embedding spans. Fixed an issue where the server crashed on requests in gevent scenarios.
2.5.4	April 2, 2026	New features Added support for the Dify ReACT plugin. For more information, see Supported Python components and frameworks. Added support for observing each iteration of ReACT for LangChain, AgentScope, and Dify ReACT. For more information, see LLM trace field definitions. The vLLM plugin supports engine concurrency analysis and fine-grained, Token-level engine observability. For more information, see Enable observability for vLLM and SGLang inference engines. Added observability for Embedding and Rerank operations in the vLLM engine. For more information, see Enable observability for vLLM and SGLang inference engines. Fixed issues Fixed an issue where continuous profiling could not be enabled over the internet. Fixed an issue where Dify plugin enhancement failed. Fixed incorrect upstream and downstream associations in metrics. Fixed an issue where console redirection failed for some database plugins. Fixed an issue where fetching console configurations failed in certain scenarios. Fixed an issue where requests were repeatedly counted in Dify in certain scenarios. Fixed an agent registration failure for Python 3.9 applications in certain scenarios. Fixed a request replay issue in urllib3 scenarios. Fixed incorrect status code statistics in FastAPI. Fixed an issue where `user_id` and `session_id` were not collected for the LangChain framework.
2.4.3	February 9, 2026	New features The OpenAI SDK and Dashscope SDK support multimodal data collection. For more information, see the example outputs for fields such as `gen_ai.input.messages` and `gen_ai.output.messages` in the LLM section of the Attributes document. Added support for the Claude Agent SDK framework. For more information, see Supported Python components and frameworks. Fixed issues Fixed a null pointer issue when using the OpenTelemetry API in certain scenarios. Fixed a package path conflict issue when used with other Alibaba Cloud SDKs. Fixed an issue where configuration updates did not take effect when using older versions of Protobuf. Fixed a duplicate Token recording issue between the OpenAI and LangChain plugins in asynchronous scenarios. Fixed a failure in linking traces between LangChain framework Chain nodes and other microservice call nodes. Fixed an issue where health check failures caused reporting to return a 405 status code in certain scenarios.
2.3.5	January 16, 2026	New features Introduced continuous profiling to support continuous diagnosis of CPU and memory for Python applications. For more information, see Continuous profiling for Python applications. Added support for database call metric statistics for the PyMySQL and psycopg2 frameworks. For more information, see Supported Python components and frameworks. Added support for the mem0 framework. For more information, see Supported Python components and frameworks. Added support for instrumentation of newer MCP framework versions (up to and including 1.25.0). For more information, see Supported Python components and frameworks. Added support for the openai-agents framework. For more information, see Supported Python components and frameworks. Fixed issues Fixed an issue where streaming call metrics were empty for the gRPC framework. Fixed an issue with incorrect error count statistics on the application overview page in certain scenarios. Fixed an issue where agent registration caused Dify workflow trace data reporting to fail. Fixed performance degradation caused by potential metric cardinality explosion in HTTP frameworks. Fixed performance degradation caused by potential metric cardinality explosion in Redis frameworks. Fixed an issue where setting baggage using the `opentelemetry-api` failed. Fixed an issue where no data was reported for OpenAI framework calls. Fixed an issue where HTTP 400 status codes were incorrectly counted as 5xx errors in certain scenarios.
2.2.2	December 22, 2025	New features Added support for the CrewAI framework. For more information, see Supported Python components and frameworks. Added support for the LiteLLM framework. For more information, see Supported Python components and frameworks. Added support for the AgentScope framework. For more information, see Supported Python components and frameworks. Added support for compressing duplicate spans. For more information, see Connect LLM applications and services to ARMS. Added support for OpenTelemetry API versions later than 1.35.0. For more information, see Supported Python components and frameworks. Synced with the latest open-source code from LoongSuite Python Agent. See the change details. Fixed issues Fixed an issue where the time to first byte (TTFB) metric was not collected for the LangChain plugin. Fixed an issue where the tool name was not collected for Tool operations in the LangChain plugin. Fixed an issue where the time to first byte (TTFB) metric was not collected for the OpenAI plugin. Fixed incorrect input/output and Token statistics for the Dashscope plugin in streaming scenarios. Fixed an issue where Rerank operations were not collected for the Dashscope plugin.
2.1.3	December 9, 2025	Fixed issues Fixed an issue where plugins such as LangChain did not work with Python 3.9. Fixed an incorrect Token count issue in the LangChain plugin for streaming calls. Fixed a memory leak issue in the LangChain plugin in certain scenarios. Improved agent registration in multi-Python interpreter scenarios. Optimized the detection logic for LLM inference service frameworks. Fixed an issue with inaccurate agent instance counts.
2.1.2	November 24, 2025	New features Supports the Milvus vector database. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports observability data collection for Dify Plugin processes. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports the BlackSheep framework. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports metric aggregation for the Falcon, Starlette, gRPC, and MySQL frameworks. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports the RagFlow framework. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports associating logs with TraceIDs when you use the Loguru framework. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Fixed issues Fixed an issue of repeated token counts when using the LangChain framework with the OpenAI or DashScope frameworks. Fixed an issue where the plugin switch in the console did not work. Fixed an issue where some plugin enhancements failed in Python 3.8 environments. Fixed an issue where the garbage collection (GC) count for instance monitoring was a negative value.
2.0.0	October 13, 2025	New features Added support for multiple plugins, such as Google ADK, Tornado, and vLLM v1. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports controlling the collection of model conversation history in logs. For more information, see Control the collection of conversation history for LLM applications. Aligned the input/output data collection of the OpenAI plugin with the latest semantic specifications and supports configuring the maximum length for large model content collection. For more information, see Semantic specification change history for LLM applications. Differentiates Dify applications by workspace when splitting them. For more information, see Register LLM applications or inference services with ARMS Fixed issues Optimized agent log printing to avoid conflicts with user loggers. Fixed an issue with agent workspace calculation.
1.8.3	September 19, 2025	New features Supports limiting the length of message content fields using environment variables. For more information, see Limit the length of message content fields. Fixed issues Unified OpenAI-related loggers to improve the consistency of Log Management. Fixed the abnormal behavior that occurred when ARMS_WORKSPACE was empty.
1.8.2	September 12, 2025	New features Optimized the input/output display format of the Langchain plugin for LLM spans. For more information, see LLM trace field definitions. The OpenAI plugin supports collecting metrics such as the number of LLM requests, number of errors, latency, and token consumption. For more information, see LLM metrics. The OpenAI plugin supports non-UTF-8 character input. Added a switch to the Langchain plugin to enable or disable input/output content collection. For more information, see Register LLM applications or inference services with ARMS. Fixed issues Filtered out observability data generated by requests related to ECS metadata in external calls to prevent this data from interfering with user queries. Fixed an issue where the Python agent failed to redirect to instance monitoring in Cloud Monitor 2.0. Fixed an issue where spans could not be reported when the Python agent was used with a later version of the Langfuse SDK. Fixed an issue where error logs were printed to standard output in non-cloud environments. Fixed package conflicts related to aliyun in non-intrusive registration scenarios. Fixed a potential resource leak issue when the agent runs for a long time. Fixed a slow startup issue in some scenarios with network connectivity problems.
1.8.0	August 15, 2025	New features Compatible with later versions of opentelemetry-api (<=1.35.0). For more information, see Supported OpenTelemetry API versions. Supports associating K8s/ECS metadata to enhance monitoring capabilities in cloud environments. Supports distributed tracing (propagate feature) for the MCP protocol. Supports MCP version 1.12.4. For more information about supported versions, see Supported plugin versions. Supports SGLang version 0.4.9.post6. For more information about supported versions, see Supported plugin versions. Supports using LangChain to obtain user information such as sessionid and userid. The OpenAI/LangChain plugins now support filtering Base64-encoded image data to avoid reporting large amounts of image data. Fixed issues Fixed a memory leak issue in the Flask plugin in multi-process scenarios. Fixed errors and repeated injection issues when using gevent with Dify. Fixed an issue where SGLang could not be used with version 0.4.6. Fixed an error caused by a None value during registration under specific conditions. Fixed a compatibility issue with the OpenAI plugin for Python versions earlier than 3.11. Pinned the cramjam version to 2.10.0 to fix a dependency library conflict in Python 3.8. Fixed a potential trace breaking issue in LangChain single-node mode. Unified the log level for agent loading exceptions to improve the troubleshooting experience.
1.6.1	July 28, 2025	Fixed issues Fixed an issue where the agent unexpectedly reported self-monitoring logs over the Internet when health checks failed.
1.6.0	July 22, 2025	New features Supports the SGLang large language model inference framework. For more information, see Enable observability for vLLM/SGLang inference engines. Supports the MCP framework. For more information, see Supported Python components and frameworks for ARMS Application Monitoring. Supports Dify version 1.4.0. For registration instructions, see Register Dify applications with ARMS Application Monitoring. Supports specifying a log path for the agent. For more information, see FAQ about the Python agent. Supports specifying the Python application type. For more information, see Register LLM applications or inference services with ARMS. Fixed issues Fixed an issue where the latency statistics of the Flask plugin were inaccurate. Fixed an issue where the exception count of the Flask plugin was inaccurate. Fixed a failure in detecting the LLM application type. Fixed an issue where incorrect exception information was displayed when the OpenAI plugin call was abnormal. Fixed an occasional token calculation error in the LangChain plugin.
1.5.1	July 9, 2025	New features The Dify plugin supports the latest version of the OTel specification. For more information, see the PR for the specification to be merged. Supports reporting data to Cloud Monitor 2.0. Fixed issues Fixed a trace breaking issue for Chat/Completion applications in the Dify plugin. Fixed an incompatibility error in the Dify plugin when starting a service using gevent. Fixed a failure in connecting upstream and downstream traces in some scenarios. Fixed package conflicts with other Alibaba Cloud third-party dependencies in some scenarios.
1.4.0	June 23, 2025	New features The Dify plugin supports version 1.3.0 and includes a switch to enable or disable input/output content collection. Fixed issues Removed grpc-exporter-related dependencies to reduce the possibility of dependency conflicts. Fixed an issue where user calls failed in DashScope streaming scenarios. Fixed an attributes warning caused by None and prevented Nacos error logs from being printed to standard output by default. Fixed an issue where the span.kind task was lost in Dify workflows. Fixed a duplicate input/output issue in the Dify plugin. Fixed an issue where the response was unexpected during OpenAI asynchronous invocations.
1.3.1	May 23, 2025	Fixed issues Fixed an issue where metadata, such as the service name, was occasionally not obtained when the agent reported data. Fixed an issue where the Python agent version number was displayed incorrectly.
1.3.0	April 23, 2025	New features Supports instance-level and application-level tags. Supports distinguishing between LLM applications and LLM services. Fixed issues Limited the supported scope of the Dify framework to prevent call errors when users use a Dify framework version that is not supported by the agent.
1.2.0	March 13, 2025	New features Added the vLLM plugin. Added LLM metrics to the Langchain plugin. Supports Python 3.12. Fixed issues Resolved a compatibility issue with the OpenTelemetry API. Fixed an issue where the Langchain plugin changed the log level. Fixed an issue where prompt information was not collected by the Langchain plugin.
1.1.0	January 06, 2025	Supports the Dify plugin. Supports the HTTPX plugin. Supports associating application logs with trace IDs. For more information, see Associate trace IDs with Python application logs. The OpenAI plugin supports large language model (LLM) application metrics. The LlamaIndex plugin supports large language model (LLM) application metrics. The LlamaIndex plugin supports the time-to-first-packet metric. Fixed incorrect labels for Flask metrics. Fixed an issue where TraceId and SpanId values were duplicated when using the Random operator. Fixed compatibility issues between the agent and the OpenTelemetry package.
1.0.0	September 24, 2024	Supports non-intrusive registration of Python applications with ARMS Application Monitoring to provide features such as Tracing Analysis, performance monitoring, and intelligent Alerting.