Title: Issue referencing LLM profiles directly in file-based agents (API key/decryption error?) Hi everyone, I'm encountering an issue with my agent setup in Agent Canvas and could use some help or best-practice advice. Setup & Steps: 1. Configured an LLM profile in Agent Canvas. 2. In my file-based agents, instead of using inheritance (

inherit

), I directly referenced the profile name configured in Step 1. Issue: An error occurs during the chat session. Based on initial observations, it seems Agent Canvas might be encrypting the API key, but the Agent Server isn't decrypting it before making the model call (though this is just a guess). Questions: 1. What is the root cause of this issue, and how can it be resolved? 2. Alternatively, is there a better design or architecture pattern to replace my 5 file-based agents (each currently has its own dedicated set of skills)?

Hey @bright-car-87888 Which version of agent-canvas and agent-server are you using? Could you share the trajectory of the conversation if it is not private? Thanks

Environment Details: • `agent-canvas`: v1.20.0 (built from source) Agent Canvas Analysis & Root Cause: After diving into the behavior, here is why

inherit

works while explicitly referencing the profile name fails: • Inherit mode: Reuses the parent (main conversation) LLM object directly and completely bypasses the profile store, which is why it stays functional. • Referencing by name: Triggers

store.load

without decryption (

cipher

). As a result, execution strictly depends on whether the API key stored in the profile file is plain text or encrypted. Since the key is encrypted in the profile store,

agent-server

receives the raw encrypted string instead of the decrypted API key during model invocation.

Thanks I opened an issue: github.com/OpenHands/software-agent-sdk/issues/4558