Observability and Metrics¶
Target
This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.
Filled 2026-10-09
This page was empty. It maps onto the final-state Runtime Center and the prototype's Execution Trace, Observability and Health Ops Center destinations. Requirements: FUX-04 (observable state and freshness), FUX-08 (usage and uncertainty), FUX-07 (recovery visibility).
Views¶
| View | Shows | Authority |
|---|---|---|
| Run trace | The AI-native trace: resolve → memory → knowledge → context → model call (model, provider, tokens, latency, cost) → tool call → approval → verification → result, correlated by run/step/attempt | Trust & Operations BC-21 over OpenTelemetry backends |
| Metrics | Task and agent success, cost and latency per task, tokens, tool failure rate, fallback rate, intervention rate, sandbox failures, retrieval quality, evaluation score, backlog age, stuck approvals and leases | BC-21 |
| Dependency health | Identity, stores, messaging profile, model providers, tool servers, sandboxes; degraded modes | BC-21 |
| Incidents and SLOs | Incidents, error budgets, SLO attainment (targets only after owner-approved workload evidence, Q-08) | BC-21 |
| Generated-product runtime | Health and incidents of generated products, linked, not owned | Generated SaaS pack operators |
Rules¶
- Telemetry is a projection; run state comes from the Control plane. If they disagree, the journal wins and the discrepancy is shown.
- Raw prompts, retrieval queries and tool arguments are not displayed unless the scope's retention policy explicitly captures them, redacted.
- Every chart states its freshness and the window it covers.
Current implementation¶
OpenTelemetry, Serilog and a Factory meter exist as libraries; no run attribution, dashboards, alerts or SLOs exist; the prototype's observability views are fixtures. Label: Simulated.