Expose per-request latency breakdown (upload / server queue / generation) for diagnosing slow turns

Problem: When a turn is slow, there's no way to tell whether the bottleneck is local (upload bandwidth — the Messages API is stateless, so the full growing conversation, including any images, is re-sent on every request) or remote (server queue time, prefill, or output generation). This matters especially for users on asymmetric connections, where a fast download can mask a slow upload as the real cause.

Request: A --verbose/--debug mode (or a /latency command) that reports, per turn:

• request payload size
• upload duration / time-to-first-byte
• server-side processing time (and cache hit/miss if available)
• output generation time
• total wall-clock time

Why it matters: without this, a user has no way to self-diagnose "is it my connection or the model" — I checked and there's currently no skill, plugin, or local telemetry that exposes it either.

🤖 Generated with Claude Code