Assistant (Reqursor Intelligence)

Reqursor Intelligence is the assistant built into the dashboard. Ask it a question in plain words, such as "Which tickets are stuck?" or "Why did this run fail?". It looks up the answer in your own tickets, runs and settings, or in the documentation that came with your release.

It can also draft a change for you. It never makes one on its own: nothing changes until you press Confirm.

It is available to every signed-in user and needs a valid license.

Open it

Select Ask Reqursor Intelligence at the bottom right of any page, or press Ctrl+J (Cmd+J on a keyboard with a Cmd key) anywhere in the dashboard. Press it again, or press Escape, to minimize it. Your conversation is kept while it is minimized.

• The chip at the top shows the page you are on, such as "Queue" or "Ticket RD-12". The assistant uses it to know what "this" means, and the suggestions change to fit. You can ask about anything else too.
• Maximize docks the panel to the side of the page. Restore floats it again.
• New Conversation starts a fresh one. Press it twice within a few seconds to confirm.
• History lists your recent conversations, up to 50.
• Think Harder answers the next question with a larger model. It is slower and uses more of your monthly assistant usage.
• The stop button ends an answer at any time. The part already written stays.

The first time an Owner or Admin opens it on a self-hosted install, they choose who pays for it. See Who pays.

What it can read

It reads what your own role may read, and nothing more. A Viewer gets Viewer answers. It can look at:

• projects, tickets, plans, runs, review verdicts, the queue and the run order;
• usage, budgets and the license summary, if your role may see them;
• Bug Hunt scans and findings, maintenance results and the audit log, if your role may see them;
• the settings pages, without any secret in them;
• the documentation that came with your release, and the public service status.

It never reads or shows passwords, tokens, keys, the license token, logs or diagnostics. It does not browse the web and it sends no mail.

What you can ask

Ask in your own words. A few examples of what it can answer from your own data:

• Where a ticket stands. Which stage it is in, how long each stage took, how many fix rounds it needed, what the reviewer decided, and which acceptance criteria have evidence.
• Why something failed or got stuck. Which stage and check went red, and what you can do next.
• What needs your attention across all your projects: tickets to review, changes waiting for approval, failed or paused tickets, failed Bug Hunt scans, findings waiting for you and license or connection notices.
• What a setting means. In plain words, what yours is now, what changing it would affect and which page changes it. It gives a safe suggestion and never changes the setting itself.
• What is connected. For each project, the ticket source, the repository host, whether pushing is on, which notification channels are set and whether Bug Hunt has a site and a saved sign in. It shows names and yes or no, never an address, a token or a password.
• Usage. What a project, model or stage uses most, and how this week or month compares with the one before. On your own key it shows counts such as tickets, runs and tokens, never a cost.
• Updates. Which version you run and whether a newer release is known. It reads the list your install already holds and fetches nothing.
• Whether your setup is complete, what to fix first, and a step by step walk through it.
• How to use the product, and whether Reqursor's own services are up.

When the answer is a set of figures, it shows a small card above its words, with a status next to each row written out as Good, Needs attention or Problem. The cards are built from the data it read, and they never show an amount of money.

Changes it drafts, and your Confirm

When you ask for a change, the assistant stages it as a card. The card shows exactly what will change, before and after, and who is affected. It expires after 10 minutes.

• Confirm makes the change as you, with your own sign in and permissions. It is recorded in the audit log as made by you via the assistant.
• Reject discards it.

If something changed since the card was staged, for example someone else edited the ticket, the confirmation is refused and you ask again. Text the assistant read from a ticket, such as an instruction hidden in its body, can never confirm anything. Only your click can.

It can stage, among other things:

• creating a draft ticket, and editing a ticket's title, text, acceptance criteria, labels, priority and dependencies;
• moving a ticket back to draft, reordering, archiving and restoring tickets;
• turning an error, meeting notes or a Bug Hunt finding into draft tickets that you confirm one by one, and splitting a ticket into smaller drafts;
• changing up to 20 tickets at once, by adding a label, setting the priority or moving them back to Draft, with every ticket shown before and after;
• recording time on a ticket, and dismissing a Bug Hunt finding;
• changing a project's Bug Hunt settings, its notification switch and levels, or its design system text.

Changes that start work

Marking a ticket ready, starting or retrying a run, switching Autopilot on and starting a Bug Hunt scan are separate. They are off until an Owner or Admin turns on Let the assistant start work in Settings → Intelligence. The page warns that starting work uses your model budget and can change code.

Even then the assistant only proposes them. Each one still needs your Confirm click, uses your model budget, and is refused when your license does not allow new work.

What it never does

• It never changes anything without your Confirm.
• It never manages users, roles, API tokens or any secret, the license or billing, updates or restores, the repository connection, the check commands, pushing, pull requests or the base branch. It tells you which page to use and links to it.
• It never deletes tickets in bulk, and it never changes your approval policy, mail server or budgets.
• It never sends mail or messages, opens a web page or runs code.

Adding a tool a ticket needs

Sometimes a ticket stops because the sandbox lacks a program. Ask the assistant to add it. It first checks which tools are built in. When it is sure of the vendor's official download address, it stages one card to add the tool. Nothing is installed until you press Confirm.

For this you need the right to add toolchains (Owner, Admin or Developer), the Let the assistant start work setting must be on, and the automatic toolchain install must be on. When the address is not one it knows, it tells you to add the tool yourself in Settings → System, so you read the checksum yourself. See also Missing toolchain vs. a real failure.

Activity and undo

Show Activity in the panel menu lists every change the assistant made for you, with the time and a link to the audit entry. An Owner or Admin who may read the audit log can also see everyone's activity.

An Undo button appears for changes that have an exact way back, for 7 days: new drafts, edits to drafts, labels, order, status moves and project settings. Undo does not change anything by itself. It prepares one more card for you to confirm. If the item changed again since, Undo is refused so nothing is overwritten. Changes that started work, such as a run or a scan, cannot be undone from here. Use the audit entry.

Digest and heads-up cards

• Digest. A short summary of what shipped, what failed and what needs you, for the last day or the last week. Ask for it in the panel, or open Show Digest in the panel menu. To get it by email, turn on Email me a digest in Settings → Intelligence. An Owner or Admin sets how often it goes out and at what hour (UTC) once the mail server is set up. Each person turns their own email on or off. Nothing is sent when nothing happened.
• Heads-up cards. When you open the panel, it can show up to three short cards about things that need you, such as new high severity Bug Hunt findings, tickets that failed or paused, a license that ends within 14 days or a change about to expire. A dot on the assistant button tells you there is a card. Dismiss hides a card until the thing it is about changes.

Both are built from your own data and your own role, and they use no model work.

Who may use it for what

An Owner sets a level for Admins, Developers and Viewers in Settings → Users & roles, in the table of who can do what.

The defaults are Read Only for Viewers, Drafts for Developers and Start Work for Admins. The Owner always has Start Work. A level can only narrow what a role may do. It never gives anyone a right their role does not already have, because every change runs as the person who confirms it.

Who pays

The assistant never shows an amount of money. It shows how much of the month's assistant usage is left as a band, and the date it resets.

On a self-hosted install, an Owner or Admin chooses one of two ways to pay, the first time they use it:

• Reqursor pays, up to a monthly allowance that depends on your plan. Nothing is charged to you, and the unused part does not carry over. Because this sends your questions, with the ticket details they need, to Reqursor's AI service, it needs an explicit Confirm Choice. It is never a silent default.
• Your own key. The assistant uses your own key and your provider's bill. Reqursor does not see the questions and sets no limit. Your own budgets still apply.

Other users see "Ask an Owner or Admin to choose" until the choice is made. An Owner or Admin can change it later, and can turn on Continue on my own key when the allowance is used up, which is off by default.

On an install that Reqursor hosts, there is no choice to make. The assistant is paid with your Reqursor Credits. Each month it also has a limit that depends on your plan, so it cannot use up the credits your tickets need.

What is sent where

Each question is a model request, so ticket text leaves your install. The panel says where, in a line under the box. The question, the page you are on, the last turns of the conversation and the ticket, run and project details it needs are sent, after stored secrets are masked.

• With your own key, they go to your own model provider, like a run.
• When Reqursor pays, they go to Reqursor's AI service, which passes them to the model provider and keeps neither question nor answer.

Answers from the documentation use an index that ships with your release and send nothing. Masking cannot know a password someone typed into a ticket, so keep passwords out of tickets. For the full picture see Security and data flows.

Your conversations

• Conversations are stored as text in your install, for you. Other users, an Owner included, cannot read them in the dashboard. A backup of the database contains them.
• They are kept for 30 days by default. An Owner or Admin can change the period.
• You can delete one conversation, or all of yours, from History.
• The audit log records which changes were confirmed, never what was asked or answered.
• Answers can be wrong. Check anything important. Helpful and Not Helpful under an answer send an anonymous rating: only up or down, your role and the product version, never your name, your question or the answer. An Owner can turn this off.

Settings and turning it off

An Owner or Admin opens Settings → Intelligence to turn the assistant on or off for the whole install, choose who pays, allow or stop it from drafting changes, set how long conversations are kept, set the digest email and see how much of the month's usage is used. Everyone who can read settings sees the page.

• To switch it off completely, turn off Turn on Reqursor Intelligence. The button stays in place and says the assistant is switched off, and no question is sent anywhere.
• To keep it for reading but stop it from proposing changes, turn off Let the assistant draft changes instead.

If it does not answer

• "The assistant needs a license." Activate a license in Settings → License & billing.
• "The assistant cannot answer right now." The background worker is not running. Start it and try again. See Troubleshooting.
• "Assistant usage is used up." It resets on the date shown.
• Anything else: nothing was changed. Try again, and see How to get help if it keeps happening.