Observability: Dashboards

Document ID: OBS-004
File Path: docs/14-observability/dashboards.md
Version: 1.1.0
Status: Draft (aspirational multi-service catalog) — a real, working starter dashboard now exists at deployment/observability/dashboard.json (RM-GA-P4 OBS-803): RED-per-route + LLM cost/token panels over the actual metrics the single-node server emits, not the full 7-dashboard catalog §2 describes. Import it into a Grafana instance directly; see that directory’s README.md for scope and caveats (never rendered against a live Grafana in this dev environment).
Owner: Platform Operations Team
Last Updated: 2026-07-09

1. Purpose

This document defines the standard dashboards for operating the Wovyr AI Platform — the curated views over metrics, traces, and logs that operators rely on.

These power both Grafana and the native dashboard monitoring surfaces.

2. Dashboard Catalog

Dashboard	Audience	Shows
Platform Health	On-call	Service up/down, golden signals, SLOs
Service Detail	Owners	Per-service RED/USE deep dive
Workflow & Agent	Operators	Executions/runs, durations, failures
LLM Cost & Usage	FinOps/admins	Tokens, spend, cache savings, by model/tenant
Tool Runtime	Operators	Queue, start latency, sandbox kills, pools
Memory Engine	Operators	Retrieval latency, index size, tiers
Capacity	Operators	Saturation, autoscaling, headroom

3. Golden-Signal Layout

Each service dashboard follows a consistent layout:

┌── Latency (p50/p95/p99) ──┬── Traffic (req/s) ──┐
├── Errors (rate, by code) ─┴── Saturation (CPU/mem/queue) ─┤
└── Exemplars → traces · recent error logs ────────────────┘

Panels link to traces via exemplars and to filtered logs, so an operator drills from a spike to root cause in a click.

4. Cost Dashboard

Built on cost metrics/events:

Spend over time by tenant / project / agent / model
Token mix (prompt/completion/cached)
Cache savings (LLM + Memory)
Budget/quota utilization vs. project quotas

This is the operator-facing twin of the in-product cost explorer.

5. SLO Dashboard

Shows SLO attainment and error-budget burn for each service (e.g. Tool Runtime SLOs), feeding the alerting strategy.

6. Provisioning as Code

Dashboards are version-controlled (JSON/Grafana provisioning) and deployed with the platform (deployment), so they are reproducible and reviewed like any other artifact.

7. Native vs. Grafana

Surface	Use
Native dashboard	In-product, RBAC-scoped, tenant views
Grafana	Infra-level, cross-cutting operator views

The native dashboard can embed Grafana panels for infrastructure metrics.

8. Access & Scoping

Native dashboards are tenant/project scoped by RBAC (monitoring §9); operator Grafana is access-controlled separately for platform staff.

9. Dependencies

11. Revision History

Version	Date	Description
1.1.0	2026-07-09	Added a status note pointing to the real starter Grafana dashboard at `deployment/observability/dashboard.json` (RM-GA-P4 OBS-803) — one dashboard covering RED + LLM cost/tokens, not this doc’s full 7-dashboard catalog
1.0.0	2026-06-27	Initial Dashboards specification