If you run phoenix and export our OpenInference otel in orchestration mode you will see some very hard to read initial worker prompts. This makes it pretty hard to debug the coordinator to worker delegation.
{"role":"user","content":[{"type":"text","text":"BACKGROUND (read-only, do not act on this): Create /app/hello.txt with the content \"Hello, world!\"\n\nYOUR TASK: Run the following command to create the file /app/hello.txt with the content \"Hello, world!\":\n\n```\necho 'Hello, world!' > /app/hello.txt\n```\n\nAfter running the command, capture the terminal output to confirm it completed without errors.\n\n\n\nUse your tools when the task requires fetching, generating, or computing new data. If your task is to analyze, summarize, or transform results already provided above, reason directly. When a tool exists for the operation, always prefer the tool. Large tool outputs may be saved to artifact files — if you see a `[Tool output saved to artifact: ...]` reference, use `read_artifact` to retrieve the full content. When you have your answer, call `submit_result` with a concise summary, your complete findings, and your confidence level.\n"}]}
If you run phoenix and export our OpenInference otel in orchestration mode you will see some very hard to read initial worker prompts. This makes it pretty hard to debug the coordinator to worker delegation.
Example: