Tracing

--trace instruments proxied requests with padwan-ai's OpenTelemetry GenAI telemetry. It needs the trace extra:

uvx 'padwan-proxy[trace]' --backend-url https://api.example.com/v1/ -m my-model --trace

Environment	Export
`LANGFUSE_PUBLIC_KEY` set	Langfuse
Langfuse + `OTEL_EXPORTER_OTLP_ENDPOINT` or `OTEL_EXPORTER_OTLP_TRACES_ENDPOINT`	Langfuse and OTLP
No Langfuse credentials	OTLP traces and metrics

With Langfuse, requests carrying a Claude Code session id in metadata.user_id are grouped into one Langfuse session.

--trace-content also records prompts and completions on the spans. A Claude Code turn ships its whole context (system prompt, tool schemas, history), so spans get large.

Request log

Flag	Logs
`-v/--verbose`	One line per request: models, tokens, duration, output tok/s, tools a `tool_use` turn asked for
`-vv/--timings`	Plus the timing split: backend wait, request-translation time, proxy overhead
`--breakdown`	Plus a prompt split: system, tool schemas (by MCP server, heaviest first), message history
`--rich`	Colours and aligned columns; with `--breakdown`, a tree of every tool source

Lines are prefixed with the Claude Code session id when the request carries one.