Triage and fix Issues

Work the Issues inbox, hand an Issue to your coding agent, then monitor it with a judge or dismiss it.

New Issues arrive in Needs review. Triage means deciding, for each one, whether it’s real, who fixes it, and whether a judge should keep watching for it.

Work the inbox

The inbox sorts by Severity by default, highest first. You can also sort by Latest or Evidence.

Filter by Last seen, Trend (First seen, Ongoing, Regressed), Severity (Critical, High, Medium, Low) or Status (Needs review, Monitored, Inactive, Dismissed). The inbox shows only Needs review until you add a Status filter.
Group by Trend, Status or Severity.
Search matches the title and description. Use j and k to move between Issues.
Select several Issues to Ask Judgment about them together, or to Dismiss them in one step (only when every selected Issue is in Needs review).

Start with Regressed Issues. Those are problems you thought were gone.

Read an Issue

Open an Issue and read it top to bottom:

Description, Impact and Consider: the failure, who it affects, and directions for a fix.
Evidence: one row per trace, with what happened, where in the trace, and when. View all N traces expands the list. Open a trace to see the exact span.
Incidents: occurrences over time.
Activity: status changes and your team’s comments.

If the Severity is wrong for your product, change it at the top of the Issue. Before you accept a diagnosis, open at least two evidence traces and check that they fail for the same reason.

Fix with sends the Issue to a coding agent. The button remembers the last agent you used; the arrow next to it lists Open in Codex, Open in Claude Code, Open in Cursor and Open in Devin, plus Copy as prompt.

The handoff prompt includes the Issue’s IDs, severity, trend, description, impact, suggestions and the cited traces. It tells the agent to read the Issue and its traces, make the smallest fix, and recommend a review action. If the app doesn’t open, the prompt is already on your clipboard. Judgment records each handoff on the Issue.

For the best handoffs, set up Connect your coding agent.

Judgment doesn’t apply the fix or mark the Issue resolved. You ship the change; if the same failure keeps happening, detection keeps adding evidence to the Issue.

Review: monitor it or dismiss it

On a Needs review Issue, select Review. It offers two options:

Monitor with Judge (“Reliably monitor for this in production.”) opens a Judgment Agent chat that creates a judge for this failure, or picks an existing one, and the label it should apply.
Dismiss issue (“Future occurrences of this issue will be ignored.”) takes an Optional comment, then Submit review.

Opening the Monitor with Judge chat doesn’t change the Issue’s status. The Issue becomes Monitored only when the judge and label are linked to it. The link requires that judge’s production version to run on incoming traces with sampling above 0.

Check the Issue afterward for Monitored and the linked judge and label. A Monitored Issue can’t be reopened or linked to a different judge. From then on, watch the judge’s label rate, and alert on it with a judge action.

Dismissed and Inactive Issues have a Reopen button that puts them back in Needs review.

Turn an Issue into a test

Add evidence traces to a dataset, or have the dataset agent build test cases from the Issue. See Build a dataset.

From your editor or the Judgment Agent

With the Judgment MCP server connected, your coding agent can call list_issues, get_issue, and update_issue. update_issue can:

Dismiss an Issue (user_status: "ignored").
Mark an Issue Monitored with the judge and label IDs.
Set user_status: "merged". A merged Issue no longer appears in the inbox or in list_issues, detection stops updating it, and its status can’t change again.

Only Judgment sets Inactive. create_issue exists for detection runs, so let detection create Issues. Full arguments are in MCP tools.

In the app, ask the Judgment Agent to root-cause an Issue or rank open Issues by impact. The cookbook walks through a full example.