Observe

Follow ../../shared/policy.md. Run dev-skills:project-profile if not loaded this session. Reads only. Tool notes: ../../shared/references/observability.md.

1. Frame

• Target: turn the feature/ticket into concrete signals: routes, controllers, job classes, event names, flag names, log messages (read the code to find them).
• Environment: default prod; match the profile's per-environment tools.
• Window: given, or since the feature's deploy (find it: merge time, deploy annotation, release list), or the last 24h. Baseline: the same window a week earlier, or the window before the deploy.

2. Collect

For target and baseline:

• Traffic, error rate, p95/p99 latency for the routes/jobs.
• Error tracker: new issues and their counts, first-seen vs deploy time.
• Logs: error/warning patterns for the target, volume change.
• Traces: slowest and failing spans for the target.
• Jobs/queues: failures, retries, queue depth.
• Analytics/experiments (if connected): event counts, funnel steps, variant metrics with sample sizes. Run the collection inside one Agent call, model: haiku (policy "Delegating big reads"), returning per signal: numbers for target and baseline, verbatim decisive lines, and the queries and labels used. A source not connected is listed under Gaps with the exact query the user could run.

3. Judge

Separate real change from noise: compare with baseline, check whether a change is only this target or system-wide, and line up with deploys and incidents in the window.

Output

<target> — <env> — <window> vs <baseline>
Verdict: healthy | degraded | broken | not enough data
- Traffic: <n> (<±%>)
- Errors: <rate> (<±%>); new issues: <title — count — first seen>
- Latency p95: <ms> (<±%>)
- Analytics: <metric — change>
Findings: <one line each with numbers and a link>
Next: <nothing | run dev-skills:investigate on <finding> | rollback candidate>
Gaps: <missing source — query to close it>

When called by deploy, return the verdict and findings only.