observability-context

name	observability-context
description	Fetch production observability signals from Sentry — error, exception, stack trace, regression, and alert data. Auto-loads on Sentry references. Returns message/culprit, first-seen, frequency, and a condensed stack trace, always labeled as Phase-1 framing, never a profile.

When to invoke

Auto-loads when the caller references:

• Sentry issue URL: sentry.io/organizations/<org>/issues/<id>/ or <org>.sentry.io/issues/<id>/.
• Sentry event ID: a raw event ID mentioned alongside a Sentry context clue (e.g. "Sentry event abc123").
• Free-text production signal: text describing a production error, exception, stack trace, regression, or alert — WHEN a Sentry MCP tool (mcp__sentry__*) is reachable in-session.

If no reference is present, exit cleanly.

Adapters

Sentry

• Trigger: Sentry issue URL (sentry.io/organizations/<org>/issues/<id>/ or <org>.sentry.io/issues/<id>/), a raw Sentry event ID paired with a Sentry context clue, or free-text mention of a production error/exception/stack trace/regression/alert, gated on a reachable mcp__sentry__* tool.
• Fetch: mcp__sentry__get_issue with the issue ID or URL (aspirational — this plugin ships no Sentry MCP integration; this recipe activates only if the user has one connected). For a bare event ID, first resolve it via mcp__sentry__get_event.
• Extract → block fields: title/message and culprit → Message / culprit; firstSeen → First seen; event count and rate over the query window → Frequency; top stack frames or breadcrumb trail (condensed) → Stack / breadcrumb summary; issue permalink → Source.
• Degrade: If no mcp__sentry__* tool is reachable, emit observability-context: Sentry MCP unavailable; proceeding without context. Never fabricate or guess at error rates or stack content. On auth error, surface the raw error. On empty/404, say so; do not guess.

Output format

One block per reference, prepended to caller context:

## Observability context: <issue ref or URL>

**Message / culprit:** ...
**First seen:** ...
**Frequency:** ... (event count / rate over window)
**Stack / breadcrumb summary:** ... (condensed, not a full raw dump)
**Evidence class:** production signal (Phase-1 framing) — NOT a profile; does not satisfy the profile-evidence gate.
**Source:** <URL>

Absolute rules

• Strip secrets and PII (API tokens, user emails/IPs in stack traces or breadcrumbs) before returning.
• Cap at ~400 words per reference; summarize long stack traces without truncating mid-frame in a misleading way.
• Do not editorialize or infer root cause. Output the signal; downstream agents (e.g. debugger, workflow-performance-investigation) interpret it.