Tip: Use the telemetry CLI to surface cost and quality signals early
Use the telemetry subcommands to spot problems before they compound — a spike in Opus fallbacks or a drop in prompt cache hits is much cheaper to fix when you catch it in the same session that caused it.
Why: Telemetry data ages quickly. CoordinationSignal TTLs default to 60 seconds, and HeartbeatCoordinator marks agents stale after 60 seconds by default, so signals you don't inspect soon are gone.
Recommended commands
| Command | What it shows |
|---|---|
attune telemetry show |
Recent telemetry entries |
attune telemetry savings |
Accumulated cost savings |
attune telemetry cache-stats |
Prompt caching hit rate |
attune telemetry sonnet-opus-analysis |
Sonnet → Opus fallback frequency and cost impact |
attune telemetry agent-performance |
Per-agent performance metrics |
Run attune telemetry --help to see the full list of subcommands exposed by main().
Tradeoff
Polling these commands frequently gives you faster feedback, but each call reads from the underlying store. In high-throughput runs with many active agents, prefer get_recent_events() on EventStreamer with a bounded count rather than open-ended queries, to avoid pulling large Redis streams into memory.
Related topics
- Concept: Template feedback loop — explains how
record_template_feedback()andget_template_confidence()feed quality signals back into template ranking.
Tags: telemetry, metrics
Unresolved references
Auto-generated by attune-author fact-check. Review and either fix the source code, fix this doc, or add an override.
| Location | Severity | Issue |
|---|---|---|
| Line 35 | error | [Concept: Template feedback loop](concepts/feedback-loop.md) — target does not exist |