| name | sentinel-stats |
| description | Show Sentinel telemetry stats — rule performance, latency, violations, and hottest targets |
| user-invocable | true |
Sentinel Stats
Show a performance dashboard from Sentinel's JSONL telemetry log.
Process
Run the stats script from the repo root (it auto-discovers .claude/sentinel/config.yaml and reads log_file from it):
python3 ${CLAUDE_PLUGIN_ROOT}/skills/sentinel-stats/sentinel-stats.py
If the script prints an error about no log file, tell the user:
Telemetry logging is not enabled. Add log_file: "telemetry.jsonl" to your .claude/sentinel/config.yaml and re-run your workflow to collect data.
A specific log file path can also be passed directly: python3 ... /path/to/log.jsonl
Present the output to the user. Then provide a brief interpretation highlighting:
- Noisy rules: High eval count with 0% violation rate — candidates for scope narrowing or
fast_pattern pre-filters.
- Flaky rules: Violation rate between 20-80% on the same target — the model is inconsistent, the prompt needs work.
- Slow rules: P95 or max latency near the timeout — risk of silent fail_open bypass.
- Skipped/timeouts: Any skipped entries indicate rules that silently passed without evaluation.
- Hot targets: Files or commands that trigger many evaluations — may need scope exclusions or rule consolidation.
If the user wants machine-readable output, run with --json.
Next steps
Based on the stats interpretation, suggest the most relevant action:
/sentinel-rule — edit or create rules to fix noisy, flaky, or missing coverage/sentinel-config — tune confidence_threshold, timeout_ms, or switch models if latency is high/sentinel-learn — discover new conventions from documentation if rule coverage is thin
