What problem does it solve? Slow time-to-first-token in an LLM agent harness is usually caused by a broken prompt prefix cache or oversized requests, and this Skill walks the exact code paths that decide both so you can find and fix the bottleneck. ## Core Features & Use Cases - Prefix-cache stability audit: Verifies the system prompt is built from stable sources and that transient work-state messages stay at the tail so they never bust the cached prefix. - Per-provider caching check: Confirms implicit caching on OpenAI-compatible paths and detects missing explicit cache_control breakpoints on the Anthropic path, the single biggest TTFT lever. - Cache-hit measurement: Uses already-captured metrics like cached_tokens, cache_read_input_tokens, and ttft_ms to diagnose cache health without writing new code. - Use Case: When chat turns feel slow to start responding, run this audit to discover that the Anthropic request builder never sets cache_control, then add explicit breakpoints on the stable system block and last persisted message. ## Quick Start Ask the agent to audit the TTFT and prompt-caching health of the provider request path and report the one concrete fix that would improve first-token latency.
