Support to fix: from a bug report to a code fix

Support to fix turns a customer's bug report, which arrived by email, into a coding session that fixes it, and then into a reply to the customer. The AI screens the email first, because a customer's text is untrusted. Two things stay yours: the click that opens the session, and the send of each reply. No session starts without that click, and nothing reaches a customer unless you send it.

Partly built. It is built and was tried with the real AI models on sample emails. It has not been tried on a real mailbox with a real customer, and a trial with hostile emails is still owed. Only email is covered; Telegram and a chat on your site come later.

From the email to the screening

The AI that reads your mail says whether an email reports a fault. You can say so yourself with This is a bug report, or take it back with Not a bug report. The mailbox needs Read by AI on; without it the email is not screened. An email that holds a bug report opens with a panel headed Bug report above it.

The screening comes first. Code looks for text no reader sees, hidden HTML, encoded runs and risky attachments. Then a strong model reads the email alone and gives a verdict: Clean, Suspicious or Hostile, with each finding quoted under What the screening found. It looks for instructions to an AI, requests for secrets or data, actions beyond a fix, hidden text and social engineering. The fault is then restated as plain facts under The fault, in the AI's words: steps, expected, what happened, version, device and error text.

What the session is given

The customer's own words reach the session only when the report is clean, or suspicious with every flagged span taken out. A hostile report never reaches it: the session gets only the AI's restatement, and the studio asks for a second yes, I read the findings and still want a session, before it will open one.

Reports of the same fault are grouped, and the group shows 3 reports of this fault. Move a report with Move this report if the AI grouped it wrongly.

Fix it

1. Open the email. The panel shows the verdict and the AI's proposal: the Program, the Model, the effort, a line Why and a difficulty from 1 to 5. Change any of them.
2. Write a note to the session if you want to. It is marked as yours.
3. Leave Plan first on, or turn it off. On, the session writes its plan and waits for your yes before it touches a file.
4. Set the Next update date if you know it. The reply will tell the customer the fix arrives around that day, and without a date, with the next update.
5. Click Fix it.

A session named Fix: ... opens in that project, on the computer where the project lives. It reproduces the fault, fixes it and runs the project's tests. It is told never to push, publish, deploy or send, and to treat the customer's text as data. When more reports of the same fault arrive while it works, Give 2 new reports to the session sends them on.

The reply

When the session ends it reports fixed, no fault or not fixed. For each customer a reply is drafted in that customer's own language and waits as the suggested answer in their email, labeled as from the session's report. You read it, change it and send it. If the session found no fault, the reply explains how to do it right. If it could not fix the fault, no reply is written.

An alert tells you when a report arrives, in red if it is hostile, and when the work ends. The iPhone app shows a ladybug in the color of the verdict in its email list, and the same panel inside an email.

Setting up the mailbox is in Mailboxes for each project. For the sales and support tools around it, see Sales and support.

Small tips

Questions