Framework examples (adk/langgraph/openai) crash with SpawnSafetyError — tool functions defined inside main() instead of module level

[deeptireddy-lab]

Summary

Several framework example files fail immediately with SpawnSafetyError because they define their tool callables inside main() (local/nested functions). The SDK's spawn-safety probe requires tools to be importable by qualified name, and its own error message states the fix. This is an examples defect — the tools should be defined at module level.

Environment

• conductor-oss/python-sdk @ main (also repros on the released conductor-agent-sdk)
• conductor-oss server 3.32.0-rc.10, model anthropic/claude-sonnet-4-6, macOS

Repro

cd examples/agents
export CONDUCTOR_SERVER_URL=http://localhost:8080/api \
       AGENTSPAN_SERVER_URL=http://localhost:8080/api \
       AGENTSPAN_LLM_MODEL=anthropic/claude-sonnet-4-6 \
       PYTHONPATH=$PWD
python3 adk/14_callbacks.py

Actual

AttributeError: Can't get local object 'main.<locals>.lookup_customer'
...
conductor.ai.agents.runtime._worker_entries.SpawnSafetyError: worker 'lookup_customer'
is not spawn-safe (AttributeError("Can't get local object 'main.<locals>.lookup_customer'")).
Define the callable at module level (importable by qualified name).

Cause

The tool functions are declared inside main(), so they can't be pickled/imported by worker processes → the SDK's probe_spawn_safety (intentionally) rejects them. The SDK behavior is correct and even prints the remedy: "Define the callable at module level." Passing examples already define tools at module scope.

Affected files (17)

adk/14_callbacks.py
adk/15_global_instruction.py
adk/16_customer_service.py
adk/17_financial_advisor.py
adk/18_order_processing.py
adk/19_supply_chain.py
adk/20_blog_writer.py
langgraph/03_memory.py
langgraph/13_multi_turn.py
langgraph/27_persistent_memory.py
openai/02_function_tools.py
openai/04_handoffs.py
openai/05_guardrails.py
openai/07_streaming.py
openai/08_agent_as_tool.py
openai/09_dynamic_instructions.py
openai/10_multi_model.py

Suggested fix

Move each example's @tool/tool functions from inside main() to module level so they're importable by qualified name (matching the pattern used by the examples that run successfully).

[nickorkes]

Worth reading before fixing this one: #448 hits the same transport code through a different trigger.

There the tool is already defined at module level and it still fails, with PicklingError: ... it's not the same object as __main__.get_inventory rather than AttributeError: Can't get local object.

That matters for the fix here. Moving the example tools out of main() will make these examples pass, but any OpenAI @function_tool at module level still fails. @function_tool rebinds the module global to a FunctionTool, and _CONTAINER_ATTRS only unwraps .func and .coroutine, so the container hop that covers LangChain does not cover OpenAI.

#448 also has a Linux reproduction, which shows this is not platform-specific, and a suggested by-value transport with cloudpickle that would cover both cases.

[deeptireddy-lab]

correct for openai it has nothing to so with the module level. for langgraph and adk, it does.