Overview
Observe shows what agents load, what it costs, how much of your curation survives real permissions and real data, and what went wrong. Agents answer fluently even from degraded context — these pages are where degradation shows.
The seven pages
| Page | Answers |
|---|---|
| Trends | What changed since you last looked — the tab opens here |
| Usage | What’s being loaded, by whom, at what token cost |
| Quality | How much of what you designed actually arrives — and why not |
| Performance | How long assembly takes |
| Issues | What degraded, and where |
| Writes | Every write attempted through ContextWorks — the durable audit |
| Debug Logs | Exactly what one caller received, on demand |
The metrics pages share the same filters — Domain, Library, Artifact, Surface, Level, Format — and a timeframe selector. Library is the artifact’s library (Entities, Briefs, Skills, or Catalog), and the Artifact menu follows it: pick Skills and only skills are offered. Usage, Quality, and Performance export to CSV.
Two numbers on Quality do the heaviest lifting: FLS coverage % (a permissions question) and data completeness % (a data question). See Quality.
How it works
Every load is metered automatically — no configuration, no opt-in. The engine records the artifact and its library, surface, level, format, domain, token estimate, assembly time, and field-coverage counters, rolled up daily. All four libraries meter the same way: entity loads, brief retrievals, skill retrievals, and catalog reads each write the same daily rows.
The two logs are deliberately different animals. Debug Logs captures the exact text a caller received, time-boxed and self-expiring — a debugging tool, not an audit trail. Writes is the opposite: a durable record of every mutation attempted through write access — one row per field, grouped per call, covering commits, validation failures, and refusals alike. Nothing there expires.
When to check what
- First, always: Trends — week-over-week movement by domain, what’s new, and the freshest issues, each linking to the page that explains it.
- After any change: check Issues. A failed formula or enrichment is silent to the agent — this is the only place it shows.
- Weekly: check Quality for low-completeness fields and low FLS coverage.
- When cost matters: check Usage, sorted by tokens. The biggest line is usually an unfiltered related list.
- When something is wrong: enable debug logging for fifteen minutes, reproduce, read the payload, disable it.
- When write access is on: scan Writes for refusals and validation failures — a rejected write is an agent (or an instruction) that needs adjusting.