---
title: "Set up issue detection"
description: "Turn on issue detection, tell it what to look for, and set its schedule and Slack reports."
sidebar:
  label: "Set up detection"
seo:
  title: "Set up issue detection | Judgment Docs"
  description: "Turn on issue detection, tell it what to look for, and set its schedule and Slack reports."
---

Detection needs traces in the project and nothing else. If you went through project setup, a first run already happened and detection is on; skip to [Change detection settings](#change-detection-settings). You need a developer role or higher to change detection. Viewers can read Issues.

## Turn on detection

1. **Open Issues and turn it on**

    If detection is off, Issues shows **Discover issues autonomously**. Choose **Turn on**.

2. **Tell it what to look for**

    Under **What issues have you seen?**, write your **Custom instructions**. They're required. The **Prompt suggestions** chips (such as **Tool call failures**, **User friction** or **Cost and tokens**) add starter text you can edit.

3. **Start the first run**

    Choose **Start Detection**.

4. **Choose a schedule and where reports go**

    While **Detection is running**, set a **Schedule**, turn on **Send reports to Slack**, and decide whether to **Mark stale issues inactive**. Choose **Save**, or **Skip** to set these later.

## Write instructions that find real problems

Your instructions add priorities on top of the detector's own checklist. They don't replace it, so detection still looks at errors, tool failures and outliers you didn't mention. Describe what your agent is for and which failures cost you the most. For the [lunch-order agent](/documentation/start/example-agent):

```text
This agent orders catered lunch for office teams. A good run ends with a
restaurant that can handle the headcount, an order total under the user's
budget, and a delivery time before the requested time.

Prioritize:
- Tool calls (get_menu, check_catering_capacity, check_delivery_eta) that
  fail or time out, and what the agent does next.
- Recommendations that ignore dietary notes from get_user_context.
- Users who have to repeat or correct their request.

Ignore wording and formatting preferences that don't change the order.
```

If you name a time window in your instructions ("look at the last 3 days"), detection uses it instead of the schedule.

## Change detection settings

Open Issues and choose **Configure**. The **Detection settings** dialog has four parts:

| Section | Setting | What it does |
|---|---|---|
| **What should Judgment look for?** | Your instructions (required) | **Show suggestions** brings back the chips. |
| **Automation** | **Run on a schedule** | Runs detection on a recurring cadence. Choose **Daily/weekly**, **Interval**, or **Advanced** for a cron expression. Times are Pacific. |
| **Issue lifecycle** | **Mark stale issues inactive** | Off until you turn it on. Marks an Issue **Inactive** after 1 to 365 days with no new occurrence (the field starts at 10). Inactive Issues return to **Needs review** if they happen again. |
| **Notifications** | **Send reports to Slack** | Posts an issue report after each detection run to the channels you pick. Select at least one channel before saving. |

Choose **Save**. To run detection right away, choose **Run now**; it stays disabled until instructions are set.

{/* SCREENSHOT: Detection settings dialog with all four sections expanded, dark mode */}

Slack channels come from your connected workspace. If Slack isn't connected, see [Connections](/documentation/administration/connections).

## Watch a run, or stop it

The status line at the top of Issues shows where detection is: **Queued for investigation…**, **Analyzing traces…**, or when it last checked and how many Issues it created or updated. It also shows the schedule (**Manual** when none is set). To stop a run in progress, choose **Stop**. To stop future runs, turn off **Run on a schedule** and save.

## Environments

Detection has no environment setting today. It reviews the project's traces. If you only care about production, say so in your instructions.

## Cost

Each detection run is an agent run and uses Judgment Units (JU), so a frequent schedule on a busy project costs more than a daily one. See [Billing and plans](/documentation/administration/billing-and-plans).

## If something looks wrong

- If a run finds no Issues, that can be a real result: detection only writes Issues it's confident about. Send more varied traffic, or make your instructions more specific.
- If the inbox looks empty but runs found Issues, remember that it shows only **Needs review** by default. Add a **Status** filter to see Monitored, Inactive or Dismissed Issues.
- If Slack reports don't arrive, check that **Send reports to Slack** is on, at least one channel is selected, and the Slack connection is active.