Obtain an inspectable execution

Read the operating agreement and capture recipes.

For bundled local tools and exact commands, read toolkit setup. Run the workspace installer for this skill; it preserves existing customized files. For the managed route, use concrete service operations after confirming the connected schema.

Inspect before instrumenting

Find the actual app entrypoint, framework, provider clients, streaming lifecycle, tools, retrieval, and current logging/OTel setup. Inspect one existing trace if possible. Do not replace a working exporter or add duplicate provider wrappers.

If there is no usable trace, instrument the smallest representative execution locally. Reuse app-owned logging hooks or an injected local exporter. Do not connect to a new hosted service yet. Exercise the real app with a safe input; a fake SDK response may smoke-test capture but is not the promised real trace.

Verify the evidence

Inspect input, final output, model calls, tool arguments/results, relevant context, parent relationships, errors, and available usage. Check that all steps belong to the same request. For conversations preserve turn order and thread identity. For streaming finish or explicitly close the stream before flushing.

If a field is absent, identify whether the app, provider, or capture policy omitted it. Do not reconstruct precise usage from text or invent private reasoning. Save available public reasoning only when actually returned and permitted. Aggregate leaf model-call usage; do not add already-aggregated parent totals.