Skip to main content Module metrics Copy item path Source MetricsStatus Ownership of the process metrics recorder, not exporter connectivity. init Initialize once before recording any metrics. Never replaces a host recorder. metrics_handle Legacy handle accessor. When host-owned this is a detached compatibility
handle, NOT an activity/export signal. Use scrape_handle for HTTP scrapes.
It remains available for existing callers that require an infallible handle. record_cache_tokens Record prompt-cache token usage (write = cache miss tokens, read = cache hit tokens). record_eval_regression Increment the count of eval regressions detected against a baseline. record_eval_score Set the mean score for an eval suite + scorer (0.0–1.0). record_guardrail_flagged Increment the count of chat inputs flagged by an input guardrail, by category
(injection | pii). record_llm_call_latency Record the wall-clock duration of a single LLM driver call. record_llm_cost Record an estimated per-request LLM cost in USD. Recorded as a histogram so
the _sum series gives cumulative spend while preserving per-run distribution. record_llm_tokens Record LLM token usage. record_provider_health Record a provider’s current health (CH-03): 1.0 when available, 0.0
when in a failover cooldown window. record_request Record an HTTP request completion. record_sandbox_created Record sandbox creation. record_sandbox_error Record sandbox error. record_sandbox_execution Record sandbox execution duration. record_skill_activation Record a skill-activation decision: which skill was selected for an intent,
by which classifier backend, and whether it was accepted (vs. an override /
fallback). Enables per-skill / per-backend precision-recall accounting —
the prerequisite for measuring and improving activation accuracy
(fable §8, plan CH-08). record_skill_activation_outcome Record that an activated skill’s execution succeeded or failed — pairs with
record_skill_activation to distinguish “chosen” from “chosen and worked”. record_skill_invocation Exact host activation, distinct from matcher acceptance. record_skill_request_usage Attribution only. Never updates the ordinary token or cost totals. record_skill_shadow_recall One explicit activation is the ground-truth sample for Recall@10.
The histogram’s sum / count is recall; this does not filter any catalog. record_sycophancy_flagged Increment the count of responses flagged as sycophantic. record_sycophancy_score Record the sycophancy score (0.0 clean – 1.0 fully sycophantic) of a response. record_tool_call Record a tool call result. request_timer Create a request timer. Call .finish() on the returned value when done. sandbox_active_dec A sandbox execution finished — decrement the in-flight gauge. sandbox_active_inc A sandbox execution started — increment the in-flight gauge. scrape_handle Return a scrape handle only when UAR owns the recorder receiving writes.
HTTP exporters must use this API and report status when it returns None. set_a2ui_replay_runs Set the number of runs with retained A2UI replay state. set_active_sandboxes Set active sandbox count gauge. set_active_sessions Set the active session count gauge. set_mcp_server_status Record MCP server status. set_runs_retained Set the number of run records retained for replay and diagnostics. status Report ownership. A host-owned recorder receives UAR’s metric macros, but
UAR cannot infer whether or where that host exports them.