Agent health
Since v0.8.6, the desktop sidebar Agent health panel shows a read-only harness/tool telemetry report — the same aggregation as zagens doctor --tools.
Where to look
| Surface | Notes |
|---|---|
| Sidebar Agent health | Visual report (tool failure rates, gate/verify, hint coverage, …) |
| CLI | zagens doctor --tools (--json for scripts) |
Built from local kernel_events aggregates — nothing is uploaded.
What’s in the report
- Tool calls — volume, failure rate, top failing tools
- Loop guard — retry / loop-guard proxies
- Harness verify — pass / self-heal related counts
- Stage gate — e.g.
stage_gate_blockedcounts - Hint coverage — whether high-failure tools have failure hints
Use it after long engineering runs to sanity-check the tool surface — not as a substitute for unit tests or CI.
Tips
- Open health after a night-queue or large CRAFT run, then tighten gates if needed.
- In CI, snapshot with
zagens doctor --tools --jsonwhen local event logs exist. - For env/config issues, use plain
zagens doctor(without--tools).
Related: Gate-as-Code · Night queue · CLI · FAQ