Provider-side prompt-cache visibility: cached tokens are the subset of input tokens served from the provider’s cache (Gemini explicit content cache / OpenAI automatic prompt cache); cache_status reports the Gemini content-cache lifecycle ("created", "skipped:<reason>"), "automatic" for OpenAI, or "" when no cache is in play.

total_cached_tokens

uint64

cache_status

string

frames_dropped_pending_request

uint64

Backpressure visibility: frames_dropped_pending_request counts sampled frames discarded because an LLM request was still in flight (a slow provider silently lowering the effective sample rate); calls_skipped_by_gate counts interval ticks where a call would have been made but the evaluation gate was closed; average_call_interval_ms is the observed interval between LLM calls (compare with the configured analysis interval).

calls_skipped_by_gate

uint64

average_call_interval_ms

double

consecutive_state_failures

uint32

State-loop health: consecutive resets on every successful _update_state application.

total_state_failures

uint64