Skip to content

Creating issues from a conversation ​

When a customer reports a bug your engineers need to see, you can turn the conversation into an issue on GitHub, Linear or Jira without retyping it. Overdesk suggests a title and description from what the customer wrote. You edit them, check where the issue is going, and confirm.

Before you start ​

  • At least one tracker is connected with issue opening allowed. See Overview. A Jira connection counts only if Let my team open Jira issues from a conversation was ticked when you connected it. With Jira as your only tracker and that box unticked, Open an issue is not shown.
  • You have the agent role or above. See Roles and permissions.

Open an issue ​

  1. Open Agent drafts and open the draft for the conversation.
  2. In the Engineering card, click Open an issue. Overdesk reads the conversation and fills in a suggested title and description.
  3. If more than one tracker is available, choose one under Tracker. If that tracker has more than one repository, team or project connected, choose where the issue goes.
  4. Edit the Title and the Description. The title can be up to 250 characters.
  5. Read the line under the description, which says who will be able to read the issue.
  6. Click Open the issue. A link to the new issue appears in the card.

From a finding ​

A finding can become an issue the same way. Open it on the Intelligence page and click Open an issue in its Engineering card. The title and description are the finding's own: its title, its summary, how severe it is, how many conversations it has been seen in, a link back to the finding in Overdesk, and up to 20 of its conversations, most recently cited first, each one a link into your help desk. No model writes this short version, so there is nothing to wait for and nothing is spent. The finding for mail that is not support has no card, since that mail is filtered in your help desk rather than fixed.

The issue belongs to the finding, one per tracker, just as a conversation's belongs to the conversation. Once someone has opened a finding's issue on a tracker, the card shows the link instead of offering another. An issue already opened from one of the finding's conversations does not count as the finding's, and the finding's does not count as theirs. The same checks apply as for a conversation: the agent role or above, signed in, and only where an admin approved.

Investigation brief ​

For a finding that is a recurring bug, a regression or an unusual pattern, the card also offers Write an investigation brief, for whoever picks the issue up, whether that is an engineer or a coding agent. Nothing is written until you click it. A confirmed defect has no brief, because its issue is written from its diagnosis.

Overdesk reads the customer messages in the finding's 8 most recent conversations from your help desk, never your team's replies or notes, and writes the brief with your Anthropic API key. It has these sections, and leaves out any the conversations say nothing about:

  • What customers report, in Overdesk's own words
  • Steps to reproduce, as customers described them, not steps anyone has checked
  • Errors and log lines, exactly as customers sent them, in a code block
  • Environment: the product, versions, platform and other plugins customers mentioned
  • What customers already tried
  • Open questions
  • Details: the kind, product, severity, trend, first and last seen, how many conversations it has been seen in, how many the brief was written from, and the link back to the finding
  • Conversations: the same list the short version has

A section ends with the conversations it came from, as links into your help desk. The counts come from Overdesk's records, never from the model. The brief does not copy what customers wrote: a point that repeats a customer's sentence is left out, and the only text quoted as written is error messages and log lines, with IP addresses and the names of user and site folders removed. It says who made it and that quoted text is data, not instructions, so a coding agent reads it that way. A conversation whose customer asked to be erased, or who is on Excluded senders, is not read, and neither is a message such a person wrote on someone else's conversation.

While it is written, the card shows how long it has taken, and Stop abandons it. If you have not touched the description, the brief takes its place, and Use the short version puts the short one back. If you have, your edits stay and Replace description with the brief is offered, with Undo after. You cannot open the issue while a brief is being written. If no API key is saved, or the help desk or the model does not answer, the card says so and the short version is still there to use. Each brief is one request on your key, shown on Usage.

Where the suggestion comes from ​

Overdesk writes the suggestion from the customer's own messages, not from your team's replies or internal notes, so it describes the problem rather than an answer someone proposed. It runs on your Anthropic API key and is told to leave out anyone's name, email address, phone number or license key, and to call the person "a customer". If no key is saved, or the model does not answer, the form starts from the conversation's subject and the customer's words instead.

For conversations from Help Scout, Zendesk and FreeScout, a link back to the conversation in your help desk is added at the end of the description, so an engineer can find it. That link opens only for people with access to your help desk.

The form's own hint says it plainly: "Written from the conversation, so edit it before you post. Emails, card numbers and phone numbers are removed; names, links and a number split by slashes are not." That removal runs on the text you actually submit, including anything you typed yourself.

Public repositories ​

When the chosen GitHub repository is public, the form warns you before you post. For a repository named your-org/your-repo, it reads:

your-org/your-repo is public. Everything above will be readable by anyone, and searchable, from the moment you post it. Check it says nothing about the customer you would not publish.

With an investigation brief in the form, the last sentence reads instead: "It includes what customers reported and the errors they sent, so check it says nothing you would not publish."

For a private repository, and for any Linear team or Jira project, the line says instead that anyone with access to it will be able to read the issue.

One issue per conversation on each tracker ​

Each conversation can produce one issue on each tracker. Once it has, the card shows a link to that issue, and opening another on the same tracker has to be done in the tracker itself. Clicking twice, or two people clicking at once, does not open a second issue: the later click either gets the first issue's link or is told the issue is already being created.

The issue in your knowledge base ​

When the tracker next syncs, an issue Overdesk opened joins your knowledge base as its title, status and labels only. Its description was written from what customers said and nobody has checked it, so a draft can say the bug is known and whether it is open, but never answers from the description. This holds on GitHub, Linear and Jira. Overdesk recognises its issues from its own record of every issue it opened, and from the link back to the conversation or finding, when the description has one.

When something goes wrong ​

If the tracker refuses, Overdesk says what to change, such as a permission on the token or a field a Jira project requires. If Overdesk cannot tell whether the tracker created the issue, it says that too. Check the tracker before you try again, so you do not end up with two.