Documentation Index
Fetch the complete documentation index at: https://docs.selftune.dev/llms.txt
Use this file to discover all available pages before exploring further.
Usage
selftune watch [skill]
selftune watch [skill] --ignore-watch-alerts
Flags
| Flag | Type | Default | Description |
|---|
--skill | string | — | Required. Skill name to monitor |
--skill-path | string | — | Required. Path to the skill’s SKILL.md |
--window | number | 20 | Number of recent sessions to evaluate |
--threshold | number | 0.1 | Trigger-rate regression threshold below baseline |
--grade-threshold | number | 0.15 | Grade score regression threshold below baseline |
--no-grade-watch | boolean | false | Disable grade-based regression detection |
--auto-rollback | boolean | false | Automatically roll back when a regression is detected |
--sync-first | boolean | false | Refresh telemetry before reading watch inputs |
--sync-force | boolean | false | Force a full rescan during --sync-first |
--help | boolean | false | Show command help |
Trust score
Every watch cycle produces a trust score between 0 and 1 that summarizes skill health:
| Score | Meaning |
|---|
1.0 | No regressions, sufficient check data |
0.5–0.99 | Minor issues or limited data |
< 0.5 | Active regressions or rollbacks detected |
How the score is calculated
| Signal | Effect on score |
|---|
| Trigger pass rate regression | −0.5 |
| Grade regression (scaled by delta, max) | −0.3 |
| Active alert without specific regression | −0.2 |
| Recent rollback | −0.2 |
| Insufficient check data | Capped at 0.5 |
[0, 1]. A skill with no regressions and enough data scores 1.0.
The current trust score is visible in the skill report on the dashboard.
Output
watch emits structured JSON by default. The key fields are:
snapshot.pass_rate and snapshot.baseline_pass_rate for measured trigger
deltaalert for any trigger or grade regression messagerecommended_command for a machine-readable follow-up, usually rollback when
watch detects a regressiongradeAlert and gradeRegression for grade-specific evidence
create publish --watch now returns the nested same-shape watch_result
payload too, so agents and the local dashboard can inspect measured post-deploy
watch evidence without reparsing raw terminal output.
Grade watch
Publish gate
Before a skill is published to the registry, SelfTune evaluates the most recent watch results. This gate is advisory — it produces warnings, not hard blocks — but publishing with active warnings is not recommended.
What triggers a warning
- Low trust score — score below
0.70
- Active alerts — any unresolved alert from recent watch cycles
- Recent rollback — the skill was rolled back and the issue may not be resolved
- No watch data — skill has never been watched; consider running
selftune watch first
Bypassing warnings
If you are confident the alerts do not apply, pass --ignore-watch-alerts:
selftune watch my-skill --ignore-watch-alerts
Alerts
When watch detects a problem, it sets an alert on the result. Alerts flow through to the publish gate and are visible in the dashboard. Resolve the underlying regression before publishing to avoid distributing a degraded skill.
selftune watch --skill my-skill --skill-path path/to/SKILL.md --auto-rollback
Auto-rollback is irreversible without re-running selftune evolve. Use it
only in automated pipelines where you have a clear re-evolution path.
Examples
Run watch on a specific skill:
Publish despite active watch warnings (expert use):
selftune watch summarize --ignore-watch-alerts
