Error monitoring

Errors you don't have to read yourself

Production errors pile up. Sentry sends you fifty emails and you stop reading them after the third one because most are the same null reference in different contexts. The important one, the auth token crash affecting 800 users, gets buried.

Inside Claude Code, type /sentry. An analyst agent connects to your Sentry project, pulls recent issues, and does the triage you don't want to do. It categorizes by severity, maps errors to the features that caused them, identifies root causes, and writes fix recommendations with specific file and line references.

How it works

The agent reads your docs/INDEX.md to understand what features exist, then queries Sentry for unresolved issues. If a previous report exists, it compares: what's new, what's recurring, what got resolved since last time.

Each issue gets a severity rating, a root cause analysis, and a fix recommendation. The output isn't "something went wrong in auth." It's "the JWT refresh check in lib/auth/utils.ts:47 doesn't handle expired tokens from sessions created before the March migration."

What you get

A report in dev/reports/sentry/ with a summary table and detailed findings. If you have email configured, you get it in your inbox too.

The report feeds directly into /fix-issue for interactive fixing, or into /heal for automatic fixing on a branch. You can also just read it and fix things yourself.

Running it

/sentry                # analyze since last report (or last 24h)
/sentry --days 7       # look back a week
/sentry --critical-only

Set it up on a daily schedule with /monitor and it runs without you thinking about it. When something critical shows up, you get an email.