loongsuite-instrumentation-openhands 0.9.0

What is covered

This package wraps the V0 python -m openhands.core.main execution path:

• openhands.core.main.run_controller for the ENTRY span.

• openhands.core.loop.run_agent_until_done for the AGENT span fallback.

• AgentController.__init__ / AgentController.close for lifecycle-bound ENTRY and AGENT spans that survive python -m from-import binding.

• AgentController._step for ReAct STEP spans.

• Runtime.run_action for TOOL spans.

• LLM.__init__ to bridge the current OpenHands context into LiteLLM calls.

Span tree

ENTRY  enter openhands
`-- AGENT invoke_agent codeact
    |-- STEP  react step [xN]
    |   |-- LLM   chat {model}
    |   `-- TOOL  execute_tool {tool_name}
    `-- STEP  react step [...]

python -m and from-import binding

When OpenHands V0 is launched via python -m openhands.core.main, Python executes main.py as __main__. Symbols imported with from ... import can be bound before module-level wrappers are installed, so patching openhands.core.main.run_controller is not enough by itself.

To keep ENTRY and AGENT spans reliable, this instrumentation primarily opens them from AgentController.__init__ and closes them from AgentController.close. The module-level wrappers remain as a fallback for programmatic invocations.

Cross-thread context bridge

OpenHands V0 may execute controller steps and runtime tool calls in worker threads with fresh asyncio loops. The instrumentation stores the active OTel context by session id and re-attaches it in STEP, TOOL, and LLM bridge wrappers so the trace remains:

ENTRY -> AGENT -> STEP -> (LLM / TOOL).

Semantic-convention I/O capture

STEP spans emit input.value / output.value alongside GenAI semantic attributes. ENTRY and AGENT spans use only GenAI-native attributes — they do not mirror OpenInference input.value / output.value. TOOL spans never set input.value / output.value; they always set gen_ai.tool.call.arguments (JSON object string, "{}" when empty) and gen_ai.tool.call.result.

• ENTRY emits gen_ai.input.messages and gen_ai.output.messages using the ARMS parts-based message schema.

• AGENT emits gen_ai.input.messages, gen_ai.output.messages, gen_ai.system_instructions, and gen_ai.tool.definitions.

• STEP emits recent input history and the pending assistant/tool-call output for the ReAct round.

• TOOL emits gen_ai.tool.name, gen_ai.tool.type, gen_ai.tool.call.id, gen_ai.tool.description, gen_ai.tool.call.arguments, and gen_ai.tool.call.result.

Usage

from opentelemetry.instrumentation.openhands import OpenHandsInstrumentor

OpenHandsInstrumentor().instrument()

Configuration

Environment variables:

• OTEL_INSTRUMENTATION_OPENHANDS_ENABLED (default true)

• OTEL_INSTRUMENTATION_OPENHANDS_OUTER_SPANS (default true)

• OTEL_INSTRUMENTATION_OPENHANDS_AUTO_INSTRUMENT_LITELLM (default true)

I/O capture is always on and content is emitted in full.