[bug] LangChain JS instrumentation exceeds OpenTelemetry’s 128 span-attribute limit for long conversations

Describe the bug The JavaScript LangChain instrumentation creates an unbounded number of span attributes as conversation history grows. formatInputMessages() includes every input message, and flattenAttributes() converts every message field into indexed attributes: - llm.input_messages.0.message.role - llm.input_messages.0.message.content - llm.input_messages.1.message.role - llm.input_messages.1.message.content The resulting object is passed to span.setAttributes(). The OpenTelemetry Node SDK defaults to 128 attributes per span. After that limit is reached, subsequent attributes are dropped. To Reproduce Use LangChain with a conversation containing ordinary role-and-content messages and enable: @arizeai/openinference-instrumentation-langchain Observed results: - 60 input messages → 127 total span attributes → 0 dropped - 61 input messages → 129 total span attributes → 1 dropped - 62 input messages → 131 total span attributes → 3 dropped - 64 input messages → 135 total span attributes → 7 dropped - 70 input messages → 147 total span attributes → 19 dropped This was reproduced with versions 4.0.6 and 4.0.17. Tool calls and tool schemas generate additional attributes, so agent conversations can reach the limit earlier. Relevant implementation: https://github.com/Arize-ai/openinference/blob/main/js/packages/openinference-instrumentation-langchain/src/utils.ts Expected behavior The instrumentation should produce bounded attribute cardinality while preserving important attributes. Possible solutions include: 1. Add a configurable maximum input-history length. 2. Retain only the current conversation turn instead of the complete history. 3. Apply an attribute budget before flattening messages. 4. Record session, model, token-count, and metadata attributes before verbose message attributes. Additional context The entire span is not dropped. Attributes beyond the OpenTelemetry limit are dropped. Because message attributes are added before token counts, metadata, and session.id, important later attributes can be lost. Deleting old attributes in an OpenTelemetry Collector does not solve the problem because the Node SDK applies the 128-attribute limit before exporting the span. Environment: - OS: macOS - @langchain/core: 1.x - @arizeai/openinference-instrumentation-langchain: 4.0.6 and 4.0.17 - OpenTelemetry Node SDK using default span limits