Agent Skills: Sentry Incident Runbook

Sentry Incident Runbook

Overview

Structured incident response framework built on Sentry's error monitoring platform. Covers the full lifecycle from alert detection through severity classification, root cause investigation using Sentry's breadcrumbs and stack traces, Discover queries for impact analysis, stakeholder communication, resolution via the Sentry API, and postmortem documentation with Sentry data exports.

Prerequisites

• Sentry account with project-level access and auth token (SENTRY_AUTH_TOKEN)
• Organization slug (SENTRY_ORG) and project slug (SENTRY_PROJECT) configured
• @sentry/node (v8+) or equivalent SDK installed in the application
• Alert rules configured for critical error thresholds
• Notification channels connected (Slack integration or PagerDuty)

Instructions

Step 1 — Classify Severity

Assign a severity level based on error frequency and user impact. This determines response time and escalation path.

| Severity | Error Criteria | User Impact | Response Time | Escalation | |----------|---------------|-------------|---------------|------------| | P0 — Critical | Crash-free rate below 95% or unhandled exception spike >500/min | Core flow blocked for all users, data loss risk | 15 minutes | PagerDuty page to on-call engineer | | P1 — Major | New issue affecting >100 unique users per hour | Key feature degraded, no workaround | 1 hour | Slack #incidents channel, tag team lead | | P2 — Minor | New issue affecting <100 unique users per hour | Feature degraded but workaround exists | Same business day | Slack #alerts-production | | P3 — Low | Edge case, cosmetic error, staging-only issue | Minimal or no user-facing impact | Next sprint | Add to backlog, assign owner |

Decision logic for classification:

Alert fires →
├── Check crash-free rate (Project Settings → Crash Free Sessions)
│   └── Below 95%? → P0
├── Check unique users affected (Issue Details → Users tab)
│   ├── >100/hr on core flow? → P1
│   └── <100/hr or workaround exists? → P2
└── Staging-only or edge case? → P3

Step 2 — Triage and Investigate

Execute this checklist within the first 15 minutes of a P0/P1 alert.

Initial triage (Sentry UI):

1. Open the Sentry issue link from the alert notification
2. Check the error frequency graph — determine if the rate is spiking, steady, or declining
3. Read the "First Seen" and "Last Seen" timestamps to determine if this is new or a regression
4. Check the "Users" count on the issue to quantify impact
5. Verify the environment filter — confirm this is production, not staging
6. Check the "Release" tag — identify which deployment introduced the error
7. Open "Suspect Commits" to find the likely-causal changeset

Deep investigation (stack trace and breadcrumbs):

8. Read the full stack trace — identify the failing function and line number
9. Expand the breadcrumbs panel — trace the sequence of events leading to the error (HTTP requests, console logs, navigation, UI clicks)
10. Check the user context panel for device, browser, OS, and custom user tags
11. Review the "Tags" panel for patterns (specific release, region, browser)

API-based investigation:

# Fetch issue details programmatically
curl -s -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/issues/$ISSUE_ID/" \
  | python3 -c "
import json, sys
issue = json.load(sys.stdin)
print(f'Title:      {issue[\"title\"]}')
print(f'First Seen: {issue[\"firstSeen\"]}')
print(f'Last Seen:  {issue[\"lastSeen\"]}')
print(f'Events:     {issue[\"count\"]}')
print(f'Users:      {issue[\"userCount\"]}')
print(f'Level:      {issue[\"level\"]}')
print(f'Status:     {issue[\"status\"]}')
print(f'Platform:   {issue.get(\"platform\", \"unknown\")}')
" || echo "ERROR: Failed to fetch issue — check SENTRY_AUTH_TOKEN and ISSUE_ID"

# Fetch latest events for the issue (most recent 5)
curl -s -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/issues/$ISSUE_ID/events/?per_page=5" \
  | python3 -c "
import json, sys
events = json.load(sys.stdin)
for e in events:
    release = e.get('release', {})
    ver = release.get('version', 'N/A') if isinstance(release, dict) else 'N/A'
    print(f'Event {e[\"eventID\"][:12]} | {e.get(\"dateCreated\", \"N/A\")} | Release: {ver}')
" || echo "ERROR: Failed to fetch events"

Sentry Discover queries for impact analysis:

# Count total events and unique affected users in last 24 hours
curl -s -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/events/" \
  --data-urlencode "field=count()" \
  --data-urlencode "field=count_unique(user)" \
  --data-urlencode "query=issue.id:$ISSUE_ID" \
  --data-urlencode "statsPeriod=24h" \
  -G | python3 -c "
import json, sys
data = json.load(sys.stdin)
if 'data' in data and data['data']:
    row = data['data'][0]
    print(f'Events (24h):       {row.get(\"count()\", \"N/A\")}')
    print(f'Unique users (24h): {row.get(\"count_unique(user)\", \"N/A\")}')
" || echo "ERROR: Discover query failed"

# Check p95 transaction duration for affected endpoint
curl -s -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/events/" \
  --data-urlencode "field=transaction" \
  --data-urlencode "field=p95(transaction.duration)" \
  --data-urlencode "field=count()" \
  --data-urlencode "query=has:transaction event.type:transaction" \
  --data-urlencode "statsPeriod=1h" \
  --data-urlencode "sort=-count()" \
  --data-urlencode "per_page=5" \
  -G | python3 -c "
import json, sys
data = json.load(sys.stdin)
if 'data' in data:
    for row in data['data']:
        txn = row.get('transaction', 'unknown')
        p95 = row.get('p95(transaction.duration)', 0)
        cnt = row.get('count()', 0)
        print(f'{txn}: p95={p95:.0f}ms, count={cnt}')
" || echo "ERROR: Transaction query failed"

Step 3 — Resolve, Communicate, and Document

Identify the root cause pattern:

| Pattern | Diagnostic Signal | Immediate Action | |---------|------------------|-----------------| | Deployment regression | "First Seen" aligns with latest deploy timestamp | Rollback via sentry-cli releases deploys $PREV_VERSION new --env production | | Third-party failure | Breadcrumbs show failed HTTP calls to external hosts | Enable circuit breaker, add retry logic, monitor dependency status | | Data corruption | Event context contains malformed input samples | Add input validation, fix data pipeline upstream | | Resource exhaustion | Error rate correlates with traffic spikes (OOM, pool exhaustion) | Scale horizontally, add connection pooling, implement rate limiting | | SDK misconfiguration | Events missing context, breadcrumbs, or release info | Review Sentry.init() options, verify source maps uploaded |

Resolve the issue via Sentry API:

# Mark issue as resolved (closes the issue)
curl -s -X PUT \
  -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"status": "resolved"}' \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/issues/$ISSUE_ID/" \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(f'Status: {d.get(\"status\",\"unknown\")}')" \
  || echo "ERROR: Failed to resolve issue"

# Resolve in next release (auto-reopens on regression)
curl -s -X PUT \
  -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"status": "resolved", "statusDetails": {"inNextRelease": true}}' \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/issues/$ISSUE_ID/" \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(f'Status: {d.get(\"status\",\"unknown\")} (regression detection enabled)')" \
  || echo "ERROR: Failed to resolve issue"

# Ignore with threshold (snooze until count exceeds limit)
# Use 100 as the re-alert threshold for noisy low-severity issues
curl -s -X PUT \
  -H "Authorization: Bearer $SENTRY_AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"status": "ignored", "statusDetails": {"ignoreCount": 100}}' \
  "https://sentry.io/api/0/organizations/$SENTRY_ORG/issues/$ISSUE_ID/" \
  | python3 -c "import json,sys; d=json.load(sys.stdin); print(f'Status: {d.get(\"status\",\"unknown\")} (snoozed until 100 events)')" \
  || echo "ERROR: Failed to ignore issue"

Stakeholder communication templates:

Initial alert (send within 15 minutes of P0):

INCIDENT — [Service Name]
Status: Investigating
Impact: [Description of user-facing symptoms]
Started: [Timestamp from Sentry "First Seen"]
Sentry Issue: [Link to issue]
Incident Lead: @[on-call engineer]
Next update: 30 minutes

Resolution notice:

RESOLVED — [Service Name]
Duration: [Total incident time from first alert to resolution]
Root Cause: [One-line description from investigation]
Fix Applied: [What changed — rollback, hotfix, config change]
Postmortem: [Link — due within 48 hours]

Postmortem template with Sentry data:

## Incident Postmortem: [Title from Sentry Issue]

### Timeline
- [HH:MM] Alert fired — Sentry issue [ISSUE_ID] created
- [HH:MM] On-call engineer acknowledged
- [HH:MM] Root cause identified via [breadcrumbs / stack trace / suspect commits]
- [HH:MM] Fix deployed — [rollback / hotfix description]
- [HH:MM] Error rate returned to baseline, issue resolved in Sentry

### Impact (from Sentry Discover)
- **Duration:** [X hours Y minutes]
- **Total events:** [count() from Discover query]
- **Unique users affected:** [count_unique(user) from Discover query]
- **p95 latency during incident:** [p95(transaction.duration) from Discover]

### Root Cause (5 Whys)
1. Why did the error occur? [Direct cause from stack trace]
2. Why was that code path triggered? [From breadcrumbs]
3. Why was it not caught in testing? [Gap analysis]
4. Why did the alert take [X] minutes? [Alert rule review]
5. Why is this class of error possible? [Systemic cause]

### Action Items
- [ ] [Preventive measure] — Owner: @[name] — Due: [date]
- [ ] Update Sentry alert rules to catch [pattern] earlier
- [ ] Add regression test covering [scenario from breadcrumbs]
- [ ] Review and tighten ownership rules for [component]

Output

• Severity classification (P0-P3) based on error frequency and user impact
• Completed triage checklist with root cause identification
• Sentry Discover query results quantifying incident impact
• API-driven issue resolution with regression detection enabled
• Stakeholder communication messages (initial alert + resolution)
• Postmortem document populated with Sentry data exports

Error Handling

| Error | Cause | Solution | |-------|-------|----------| | 401 Unauthorized from Sentry API | Auth token expired or lacks org-level scope | Regenerate token at Settings > Developer Settings > Internal Integrations with event:read, issue:write scopes | | Alert fatigue — too many P2/P3 alerts | Alert rules trigger on every event instead of thresholds | Change alert condition to "New issue" or "Event frequency > N in M minutes" | | Suspect Commits shows wrong commit | Release association not configured | Run sentry-cli releases set-commits --auto in CI pipeline | | Missing breadcrumbs in events | SDK not capturing HTTP/console/navigation breadcrumbs | Verify Sentry.init({ integrations: [breadcrumbsIntegration()] }) and check maxBreadcrumbs setting | | Issue keeps regressing after resolve | Root cause not fully addressed, only symptom fixed | Use "Resolve in next release" for auto-reopen, add regression test | | Discover query returns empty data | Wrong time range or missing event.type filter | Expand statsPeriod to 7d, verify query syntax in Sentry Discover UI first |

Examples

Example 1 — P0 payment failure spike:

An alert fires: PaymentProcessingError with 200 events in 5 minutes. Triage reveals crash-free rate dropped to 91%. Breadcrumbs show the Stripe webhook handler receiving malformed payloads after a Stripe API version change. Suspect Commits points to a dependency update merged 20 minutes ago. Resolution: rollback the deployment, resolve the Sentry issue with inNextRelease, file a postmortem with the 5 Whys showing the missing Stripe API version pin.

Example 2 — P2 intermittent 503 from upstream API:

Sentry shows ServiceUnavailableError affecting 40 users/hour. Discover query reveals count() = 180, count_unique(user) = 40, p95(transaction.duration) = 8200ms. Breadcrumbs show the third-party geocoding API returning 503. Resolution: enable the circuit breaker fallback to cached results, ignore the Sentry issue with ignoreCount: 100, create a backlog item to add a secondary geocoding provider.

Example 3 — P1 new unhandled exception after deploy:

A TypeError: Cannot read properties of undefined appears immediately after a release tagged v2.4.1. First Seen matches the deploy timestamp. Stack trace points to a renamed API response field. Suspect Commits identifies the exact PR. Resolution: deploy hotfix renaming the field access, resolve the issue tied to v2.4.2, update the postmortem with Discover data showing 320 affected users over 45 minutes.

Resources

• Sentry Issue Details — anatomy of an issue page
• Sentry Alerts — configuring alert rules and thresholds
• Sentry Discover Queries — building impact analysis queries
• Issues API — programmatic issue management
• Ownership Rules — auto-assigning issues to teams
• @sentry/node SDK — Node.js SDK configuration

Next Steps

For configuring Sentry alerts and error capture, see sentry-error-capture. For CI/CD integration with Sentry releases, see sentry-ci-integration. For performance monitoring and tracing, see sentry-performance-tracing.