What shipped, and when.

Business plans can now record the thirty seconds of the session before a throw and play them back on the issue itself — the clicks, the field the customer was in, the request that did not come back.

• Recording starts on the error, not on the page load, so a quiet session costs nothing.
• Inputs are masked by default; you opt a field in, never out.
• The replay is scrubbed by the same rules as your events, in the SDK, before anything leaves your servers.

It is off until you turn it on, per project, and there is a hard cap so an incident cannot run up a bill.

A rule used to be able to count events. It can now watch a route's p95 against a number you set, which is how most teams find out about the slow failures nobody files a ticket for.

Existing rules are untouched. New condition, same escalation: page after fifteen minutes over budget, acknowledge in the channel, and the rule quiets itself when the route comes back under.

Written after watching a checkout route sit at 2.9 seconds for a week with a perfectly green error rate.

When a new issue's stack points at a file, Artisan already named the last few commits that touched it. It now also reads CODEOWNERS and assigns the issue to the team that owns the path.

• Assignment happens once, at first sight, and a human can always take it.
• Unowned paths route to the project's default, as before.
• The rule that made the assignment is written on the issue, so nobody has to guess why it landed on them.

Filters are now typed rather than clicked: release:4.15.* route:/checkout is:unresolved does what it reads like, and every tag your code attaches is searchable the moment it arrives.

The old filter chips still work and now write the query as you click, which turned out to be the fastest way to learn the syntax.

Saved searches live on the project, not the person, so "the ones that wake us up" means the same thing to everyone on the rotation.

Trace sampling decided at the start of a request, which meant the slowest request of the day was exactly as likely to be thrown away as the fastest. It now decides at the end.

• A request that threw is always kept.
• A request slower than its route's p95 is always kept.
• Everything else is sampled at your rate, default 20%.

The SDK holds the spans for the length of the request to do this, which costs about 38 ms at p95. Charts drawn from sampled data are marked with their rate, and a count is never drawn from a sample.