[laplaque]

Nice — Prometheus/Grafana is the right call for token spend visibility. The harder part I've found is seeing what's actually in the traffic, not just how much. I ended up building an open-source MITM proxy for that — Prometheus metrics endpoint is on the roadmap but not there yet.

[kisamoto]

You can get user prompts and tool calls in the logs (but not system prompts) although you do have to set a couple of extra environment variables.

It's under the `prompt` and `tool_parameter` fields in the logs.

    {service_name=~"claude-code"} | prompt != ""  | line_format `prompt={{.prompt}}`