Trace every model and tool call with the same request id

"The bot gave a weird answer yesterday" cannot be investigated unless you logged the call. Trace every model and tool call under one request or conversation id.

For each call, record:

• model id and prompt version
• input and output token counts
• latency and cost
• stop reason
• tool calls and their results

For agents, keep the whole chain of steps as one trace, so you can see where a task went wrong.

With this in place you can replay the exact call, add it to your eval set and confirm the fix.