Filed an Agent Canvas issue for an ACP/Laminar observability mismatch: github.com/OpenHands/agent-canvas/issues/1663 ACP conversations can complete successfully, but Laminar marks the trace as failed because a child

litellm.completion

span errors while trying to resolve ACP/subscription model IDs like

opus[1m]

or

gpt-5.5

. The ACP-specific spans succeed (

acp_agent.astep

,

conversation.arun

,

conversation.send_message

,

acp.notification

), and the final response is correct. The confusing part is that the overall trace still turns red. Question: should ACP runs avoid/suppress LiteLLM completion instrumentation for ACP model IDs, or should those handled model-resolution errors be treated as non-fatal?