respan-instrumentation-agentscope

Respan instrumentation plugin for AgentScope.

This package patches AgentScope 2.x agent, chat-model, and toolkit execution surfaces and emits Respan-compatible OpenTelemetry spans directly.

Configuration

1. Install

pip install respan-instrumentation-agentscope

2. Set Environment Variables

Quickstart

import os

from dotenv import load_dotenv
from respan import Respan
from respan_instrumentation_agentscope import AgentScopeInstrumentor

load_dotenv()

respan = Respan(
    api_key=os.environ["RESPAN_API_KEY"],
    base_url=os.getenv("RESPAN_BASE_URL", "https://api.respan.ai/api"),
    instrumentations=[AgentScopeInstrumentor()],
)

# Build and run AgentScope agents as usual.

respan.flush()

For custom model classes defined outside AgentScope's public model modules, configure all instances on one instrumentor so Respan activates one lifecycle identity while patching every distinct custom model class:

AgentScopeInstrumentor(models=[planner_model, reviewer_model, fallback_model])

Use model=... for a single custom model. Do not pass model and models together.

What Is Captured

• Agent reply() and reply_stream() calls as agent spans.
• Chat model calls as chat spans with canonical gen_ai.*, llm.*, and traceloop.* fields.
• Toolkit call_tool() execution as tool spans.

The instrumentor does not emit off-contract shortcut attributes such as tools, tool_calls, model, prompt_tokens, or respan.span.tool_calls.

Further Reading

• AgentScope examples in respan-example-projects
• AgentScope documentation