Триаж инцидентов
Трейс, логи, подозрительный деплой, severity, решение пейджить или нет.
Обзор
Do the mechanical first minutes of incident response before a human is woken up. A sweep spawns a fresh, read-only session; this skill turns a raw alert into a trace, a suspect deploy, a correlated log window, and a severity call — then decides whether that diagnosis is enough to close the loop or whether it has to page.
Reactive and read-only; covers every new or still-unresolved alert in {{sentry_project}} found in a sweep, nothing carried over between sweeps. Handle each alert as an independent unit: a failure or inconclusive diagnosis on one alert never blocks triage of the others found in the same sweep.
Когда подключать
- The alert-triage sweep fires on its cadence.
- A human asks the agent to re-triage a specific alert or explain why one was paged (or wasn't).
Workflow
Step 1 — Pull the alert from Sentry (read-only)
Via the Sentry connector, for every alert in {{sentry_project}} that is new or still unresolved since the last sweep:
- The full stack trace and exception type.
- Event frequency (spike vs. steady trickle) and the first-seen timestamp.
- Whether Sentry has already grouped it under an existing, previously-seen issue fingerprint.
Step 2 — Correlate the logs (read-only)
Via the Datadog connector, pull the log lines in a window around the first-seen timestamp (default: 10 minutes before to 5 minutes after), scoped to the service(s) named in the stack trace. Line the log entries up against the trace to confirm the failure mode — a matching upstream timeout, a bad payload, a dependency erroring — rather than trusting the exception message alone.
Step 3 — Check recent deploys on GitHub (read-only)
gh pr list --repo {{target_repo}} --state merged --limit 20 \
--json number,title,mergedAt,url,author \
--search "merged:>=<first-seen-minus-2h>"
git log --since="<first-seen-minus-2h>" --until="<first-seen>" --onelineRank candidate deploys by how closely they precede the first-seen timestamp and whether the changed files touch the failing service/module. Treat the closest matching deploy as the suspect, not a certainty — say so in the diagnosis.
Step 4 — Classify severity and benign vs. real
| Signal | Classification |
|---|---|
| Matches a known flaky/noisy pattern (repeated transient timeout, third-party rate limit, a fingerprint seen before with no user impact) | Benign — close with reasoning, no page |
| New error type, affects one user or a narrow code path, no matching deploy | Real, low severity — post diagnosis, no page unless it recurs |
| Spike in frequency, matches a recent deploy, or affects a shared/critical path | Real, elevated severity |
Marked fatal/critical in Sentry, or severity at or above {{page_severity_threshold}} | Page immediately regardless of diagnosis confidence |
| Trace, logs, or deploy history are inconclusive after Steps 1–3 | Unresolved — page, diagnosis attached as partial context |
Step 5 — Post the first-pass diagnosis
Post one message per alert to {{incident_channel}}: the error and where it's from, the first-seen time and frequency, the correlated log lines, the suspect deploy (PR/commit + author, or "no matching deploy found"), and the severity classification from Step 4.
Step 6 — Page or close
- Page — at or above {{page_severity_threshold}}, or unresolved per Step 4: @-mention the on-call rotation in {{incident_channel}} on the same message, with the diagnosis attached as context, not a substitute for looking at it.
- Close — benign or low severity with a confident diagnosis: mark it resolved in the post with the reasoning, no page.
Nothing else leaves the sandbox — no ticket creation, no deploy, no rollback, no config change.
Ограничения
- Read-only across every connector. Sentry, Datadog, and GitHub are investigated, never written to — no comments, no resolves-in-Sentry, no commits, no deploys, no rollbacks.
- One diagnosis per alert. Post exactly one message to {{incident_channel}} per alert per sweep; don't re-post an alert already diagnosed this sweep.
- Independent per alert. Treat every alert found in a sweep as its own case; a failure or inconclusive result investigating one never stops triage of the others found in the same sweep.
- Page on doubt. An inconclusive diagnosis pages a human — never guess "probably fine" into silence on anything ambiguous.
- Severity floor is non-negotiable. Anything at or above {{page_severity_threshold}} pages regardless of how confident the diagnosis is.
- Scoped secrets. Sentry, Datadog, and GitHub credentials are brokered server-side through connectors and the GitHub token is injected at runtime; scoped to this agent's grant.
- No memory required between sweeps. Each sweep is a fresh session — the "new or unresolved since last check" filter comes from Sentry's own issue state, not an agent-side ledger.
