Permissions, approvals and your rules
Choose how much a chat may do, see what each level still asks, answer approvals, know that deleting a database's data asks twice and write rules for what matters.
You decide how much each chat may do. When an agent needs something its permission does not cover, it asks you. With «Your rules» you tell it ahead of time what to do about buying, sending, deleting and more.
The chat's permission
In the composer, the «+» button lists the three levels (and the permission button next to it, which shows only when the chat is not on the usual level, opens the same menu; the usual level is chosen in «Settings» › «Agents & helpers»). Each one says what it does and what it asks:
| Level | What it does |
|---|---|
| «Read only» | «Reads and searches. Changes nothing, so it asks nothing.» |
| «Project only» | «Edits and runs commands in the project without asking. Asks before acting outside it: browser, your Mac, sending, installing, merging.» |
| «Full access» | «Does what you order in your business's folders without asking. Asks only before paying, deleting data from a database, and deleting files outside your projects (they go to «To delete»).» |
- Press the «+» in the composer (or the permission button, when it shows).
- Choose the level. It holds for the whole chat, its subagents too, and you can change it while it works (see Change the permission while it works).
With «Project only», the menu has a fourth choice: «Let your rules decide». Turned on, the level's own questions (the browser, your Mac, sending, installing, merging) stop waiting for you: they run at once and the agents never wait. Only what your rules keep for you asks you: paying, deleting a database's data, your own rules, deleting or destroying files outside the app, a Nobreinier's command leaving its confined terminal, and a team member's line. What leaves is still checked by its content (see What leaves is checked, not asked). Turned on it says «Your rules decide»; press it again to decide every step yourself.
- No model reviews your actions before they run: your rules decide, and they are clear and cost neither time nor money. Before 2.0.141 an automatic reviewer, the Guardian, sent each action to a model here.
- In «Steps», an action your rules let through is marked «Your rules».
- When one of your rules or a hook declines an action, only that action is declined: the agent reads the reason and goes on with the rest of the task. No turn is stopped after declined actions in a row.
Full access
With «Full access», your order is enough and you see no permission cards. Only three things still ask you, always:
- Paying or placing an order.
- Deleting a database's data: two cards, your yes on each.
- Destroying files outside your projects: deleting or overwriting outside your working folder, a forced push, deleting a remote branch,
git reset --hardoutside it.
Everything else goes without asking, and the chat having read a web page or other outside content first does not change that.
The folders do not change with the level: with «Full access» too, the agents open only their business's folders and what you attach to the chat. See Several businesses.
- Choose «Full access» in the permission menu.
- The app asks «Turn on Full access?» and explains what it means.
- Leave «New chats too» on so that every new chat starts with Full access and you are not asked again. Turn it off to keep it to this chat.
- Press «Turn on».
What each level asks
| Level | What still asks |
|---|---|
| «Read only» | Nothing. It only reads and searches, so there is nothing to approve. |
| «Project only» | The level's own questions when an agent acts outside the project (the browser, your Mac, connections, installing, merging), plus everything in the table below. With «Let your rules decide» on, only what is in the table below, and a Nobreinier leaving its confined terminal. What you ask for in your message counts as your permission: see What you asked for does not ask again below. |
| «Full access» | Only paying, deleting a database's data, destroying files outside your projects, administrator rights, making a picture or a video with one of your paid models and Breini entering «Settings». |
What asks with «Project only»
With «Project only», this asks even when your message asks for it. With «Full access» only the first three rows remain, plus the ones about administrator rights and your paid models, and the one about someone else in the chat. Each card says its reason:
| It asks about | The card's line of why |
|---|---|
| Paying on a page | «Paying always asks you first.» It is its own card, which only you answer. |
| Deleting a database's data | «Deleting data from a database always asks you twice.» Two cards, your yes on each. See below. |
| Using administrator rights (sudo) | «Using administrator rights asks you in every mode, even with full access, until you allow that exact command always: only you allow it.» Your yes runs that one command once, inside your business's folders, and your Mac asks for your password in its own window; the agent never sees it. «Always allow this command in «…»» keeps that exact command for that business only. |
| Making a picture or a video with one of your paid models | «Using one of your paid models for a picture or a video asks you in every mode, even with full access, until you allow that model always for that work.» The card names the model and, when the service says it, what it will cost. «Always allow … for … in «…»» keeps that model for that kind of work in that business (Organization › What they may use takes it back); «No» uses nothing, and the agent is told what else you have. |
| Changing or removing things outside your project | «It changes or removes things outside your projects, so it asks you first.» |
| Someone else in the chat (a network's channel or another business) | «Someone else writes in this chat: this decision is yours.» |
| Your rules | «Your rule «…» asks you first.» |
| One of your own automatic steps (hooks) | «One of your automatic steps asks you first.» |
| Automatic steps a project brings, or settings that let its commands run and its sites open | «They came with the project's files: they run on your Mac only after your yes.» |
| Your Mac's secrets | «It reaches private files of your Mac, such as keys or passwords.» |
| Deleting things inside the app (a project, a chat, a card…) | The «Confirm deletion» card. See below. |
| A program you brought from another tool, the first time it runs | «It asks only this once.» |
| The line of a member of your team | «It comes from a member's line: this decision is yours.» A member's line never has Full access. |
- What goes without asking in Full access: everything that is not one of the three things. For example creating an account, merging, installing, a Nobreinier's command leaving its confined shell, commands that are hard to undo but destroy no file (restarting the Mac, changing its settings, uninstalling, running what was just downloaded), reading your Mac's secrets, deleting things inside the app, a project's automatic steps, a program brought from another tool and what your rules or hooks mark to ask. A
sudoor killing a process adds no question of its own. - A program your agent wrote is checked like a command: if your agent writes a file with a program and then runs it, the app reads that file and asks the same questions it asks of a command written in one line. Whatever a command asks, the saved program asks too.
- What you refuse stays refused: a rule or hook of yours that says no, a connection set to «Never» and the commands the app never runs are refused in Full access too, and the agent reads why.
- A private file asks, however its path is written. The question the app asks before a private file of your Mac is used (a key, for example) no longer skips an unusual way of writing the path to that file: one place has one answer.
- The app protects itself. No agent touches the app's install, its state, its keys, its licence, its keychain items or its settings. It is not a question: the step is not done and the agent knows why. Reading or using what the app ships stays free.
- Hidden instructions. If a page, a mail or a file the chat read in this turn carries instructions aimed at the agent, whatever is not one of the three things is not done, with no card: the agent tells you what stopped it. Your next message starts a new turn and lifts it. Reading a page, by itself, changes nothing.
- Destroying files outside your project asks in every level: deleting or overwriting a file or folder outside your working folder, a forced push, deleting a remote branch,
git reset --hardin a repository outside it. Inside your project it asks nothing: the chat's own undo keeps it. - What an agent deletes goes into «To delete». Its rm, rmdir, unlink, trash and mv commands, and its file tools, do not delete: they set the file aside in the «To delete» folder at the top of its working folder, and only you empty it. Deleting for good another way (rm by its path or under sudo, find -delete, git clean, git rm, rsync's deletions, shred) and emptying the Trash are refused at every level, with no card, and the agent reads the way that keeps the files. See Terminal, Files and Changes.
- Links are only links. Opening a link is not read as paying or sending, whatever words it holds; a button that only opens a new message is not a send.
What leaves is checked, not asked
Since 2.0.140 the app does not ask before you upload pictures or post: it looks at what leaves. Uploading a file to a page never asks, in any level, and a post or a send is no longer held by a question about what the chat read. Instead, before an agent posts, sends, uploads a file, calls a connection or sends a mail, the app checks what it carries: a key, a password, private data going to a strange site, a secret you typed in the chat, or instructions hidden in a page the chat read.
- Nothing found: it goes, with no card and nobody reviewing it first, in every level, also after the chat read a web page.
- Something found: it stops with a plain line to the agent, who tells you in the chat. You do not get a card; your next message starts a new turn. When it is a file, only that file stays out and the others go.
The permission of new chats
- Open «Settings» (⌘,) › «Agents & helpers».
- In «Permission of new chats», choose «Read only», «Only this project’s files» or «Full access».
It is what every new chat starts with; each chat can change its own in its message box. If «Full access» is already the permission of new chats, choosing it in a chat does not ask again.
Closed folders
In «Settings» (⌘,) › «Agents & helpers» › «Advanced permissions», the «Folders with their own access» list lets you mark a folder as «Closed»: it is neither read nor changed. This holds in «Read only» and in «Only this project’s files», for an agent's commands and for its file tools alike.
With Full access, closed folders do not apply, to commands or files; they count in Read only and Only this project’s files.
The same page has the switch «With a closed folder, the agent does not read that project's git history either», which starts on. On, git does not work for the agent there. Turn it off only if you need it: it could then read what was already saved of the closed folder.
What is worth knowing about that switch:
- It holds for any git project with a closed folder inside, even when git never saved anything of that folder: the app does not ask git, it looks at whether the folder is inside a git project.
- While it is on, the agent cannot use git in that project: not to see changes, not to save versions. If it tries, the app tells it the reason so it can tell you. You keep using git as always.
- It does not cover other copies of the project on the disk (another working copy, a backup), nor what the agent already read before you closed the folder.
If the project uses git and git already knows that folder (it has files of it now, or once saved something of it), the page says so under the folder. With the switch on it says: «This folder is in the project's git history, closed to the agent too while «With a closed folder, the agent does not read that project's git history either» is on.» With the switch off it says: «This folder is already in this project's git history (the record where git keeps every version): closing it does not hide it from agents.» Closing a folder does not erase what git already saved. If git never saw the folder, if the project does not use git, or if git does not answer, that notice does not appear. The app asks git when you close the folder and when you open that page, not on every change.
A closed folder stays closed however its name is written, and in every copy the app makes of the project:
- The app compares folders by the name your Mac gives them, not by the letters you typed. A folder closed as «Secret» is also closed when it is reached as «secret» on a Mac that does not tell capitals apart, or through a link or another path that leads to the same place. An entry the app cannot name safely stays closed, and the page says which entries it did not understand.
- When the app makes a separate copy of the project to work in (for a helper that works apart, several attempts at one task or the workers of a long job), the copy leaves out what your profile closes, files with a secret's name (.env, keys) included, and it stays closed there too. Your chat says it once per job, with what was left out. If the work needs it, open it for that project in «Settings» › «Agents & helpers» › «Advanced permissions».
- A protected file (a shell start-up file, the SSH folder, a launch agent) can no longer be written by naming it with look-alike letters or through another path to the same place: the file tools judge a place by the name the disk gives it.
Two limits of the code tools worth knowing:
- Code files in a closed folder do not show in the code search or the project map, and an agent cannot ask for their contents. But the Swift language server (the program that understands Swift code) does not run sandboxed: it can read the whole project. The app only filters what that program answers, so nothing from a closed folder reaches the agent.
- Code files whose name looks like a secret's (keys, credentials, passwords) no longer go into anyone's code index, not with «Full access» either.
When a Nobreinier asks for permission
A Nobreinier of your business does not ask you first: it follows the order of your organization.
- Its department's head decides. A head can decide only on its department's tools, accounts, files or messages, as you set when you created it; what is not its call it passes up.
- If the head cannot, your business's Breini decides.
- Only what they cannot decide, or what only you decide (paying, deleting data from a database, destroying files outside your projects), reaches you. The card says who saw it: «Told first: …» or «… passed it up to you.»
A Nobreinier's questions take the same path: its head answers, then Breini, and only then you. What your team decided shows in «Steps» marked «Your team». A Nobreinier that is in a department is already active: nothing else needs turning on.
When an agent wants to change your code
A Nobreinier never changes code that already exists (your files, your git history, a server over ssh) without your direct word. A change goes ahead when your own message is the work, when it is the run of a task card you ordered, or after your yes on one card in that chat. The card says who asks, what it wants to change and why: «Sales wants to change code: …». «Always allow» on that card keeps your yes for that agent in that project.
A file the agent itself made during the task is not yours yet, so editing it does not ask again, as long as it is still as the agent last wrote it. If you changed it, or anything else changed it, the next edit asks as for any other file of yours.
Answer an approval
- When an agent needs permission, the card takes the message box's place and says in one sentence what it wants to do and who asks. For example: «Sales wants to open the Simulator app on your Mac.» You never see code or a tool's internal name.
- When the agent said why it needs it and what it is for, you read it right under, in its own words, as «Why» and «What for».
- A quiet line then says why the app asks you: one of your three things («Paying always asks you first.»), your rule by its words («Your rule «…» asks you first.») or the chat's level («It asks because this chat is set to «Project only».»).
- The action as it would be sent is behind «See details», closed at first.
- Choose with one of the three buttons:
| Button | What it does |
|---|---|
| «Skip» | It is the no: that action is not done, the agent goes on with the rest and says at the end what was left undone. The card for using one of your paid models says «No». |
| «Always allow» | From now on this goes without asking in every chat of this project, or of this business if the chat is not in a project. The button says what it keeps, for example «Always allow this command in this project», «Always allow commands like this in this business» or «Always allow this site». It is kept as a rule in «Settings» › «Rules», where you see it in plain words and remove it. |
| «Allow once» | Lets this one action go; next time it asks again. |
| «Allow» | Lets it go for more than one action; resting the pointer on it says for how long (for example, for this conversation or this turn). |
Each button says what it does when you rest the pointer on it. Paying and deleting a database's data never have «Always allow»: they ask every time. Cards answered only once (a program brought from another tool, an app of your Mac) have a single yes, «Always allow», because it is remembered.
- If the card asks only because of the chat's level, it also carries «Switch this chat to Full access»: the chat goes to Full access and the card is decided again at that level.
- The card has one box, «Ask a question, or say why not…»:
- To ask before you decide, write your question and press the arrow (or Return). The request keeps waiting; the agent answers right under your question, on the card. Where the agent cannot answer now (an automatic task or a task at rest, paused work), the box only takes a reason.
- To say why you decline, write it and press «No». The agent reads your words, changes its plan and does not repeat the action as it was. Your words also reach the agent that delegated the work, so it does not ask a helper for the same thing again.
- From Telegram, Slack or your phone, answer «yes» or «no», or «always», which does the same as «Always allow» where the card offers it. The Mac's notification and the phone's say the same plain sentence. Deleting a database's data is the exception: only a click on the app's card or on your phone answers it.
A mail or message before the page sends it
When an agent writing a mail or a message in the app's browser reaches «Send» (or presses Enter) and the app asks for your yes, the card shows the draft as the page holds it, under «The message that goes out»: who it goes to (with the address when the page shows it; if not, «The card found no recipient on this page.»), the subject and the other fields, «Conversation», the text in «Message text» and the «Attachments» by name.
- You can change the text in its box before you press «Allow»: «Edited by you: your text is what goes out.» «Back to the agent’s text» puts the agent's back.
- On «Allow», your text is typed in and the app reads the page again: it presses «Send» only if the page holds what the card showed. Otherwise nothing is sent.
- An edit counts for that «Allow» only; it never rides on «Always allow».
- A key or password the session typed shows hidden, and it goes out as the page has it: hidden words cannot be changed here.
- If the agent ran its own code in that page, the message cannot go out from it: «The agent ran its own code in this page, so this message cannot go out from it. Reload the page and have the agent write the message again.»
Change the permission while it works
You can change a chat's permission whenever you like, also while the agent works: from the «+» in the message box (or the permission button), by asking Breini, or with «Switch this chat to Full access» on a card. The cards waiting are decided again at the new level: those the level no longer asks go by themselves.
When you answer with a message
On your phone, in Telegram or Slack, or from the command line, if you write a message in the chat while cards of an agent wait, your message answers them: those cards close without running, and the agent reads your words and decides what to do (if you agreed, it asks again and a new card comes). Payments, deletions for good, merges, network connections and a database's two cards still wait for your own click.
What you asked for does not ask again
With «Project only», what you ask for in your message counts as your permission. If you write «open the simulator», it opens it with no card, and in «Steps» that action is marked «You asked for it». Only what protects you still asks: paying, deleting a database's data, deleting for good or destroying files outside the project, your Mac's secrets and your rules.
On the simulators, turning on, installing and opening the app that was built, looking and touching ask nothing. Building does (it runs the project's own programs on your Mac), and so does what deletes an app's data.
Learning or using an app of your Mac
When you ask an agent to learn or use an app of your Mac, and your message names that app, the agent may use it for that task without asking you at each click or key: «… may use … for this task, without asking you at each step. The Stop button takes it back.» It holds only for that app and that task, and «Stop» ends it. macOS asks for the screen and control permissions once, in the first-run setup, not in the middle of a task. Paying still asks: a button that pays or places an order in a Mac app gets its own card in every level. Your rules, a database's two cards and deleting for good still ask.
Your orders in a network's conversation
When you write yourself in a conversation with a network's channel or a mailbox, your order has every tool of your main chat and is answered as an order of yours; what someone else writes there keeps the reduced set.
A deletion for good asks once, and says what goes
When you ask Breini to delete something that cannot be restored («delete the old projects», «delete the archived cards»), it does not ask in the chat. One card appears, «Confirm deletion», naming exactly what goes and what stays. For example: «Delete 2 projects: «old-app», «tests». Their chats stay, without a project; no file on the Mac is touched.» The line below says: «You asked for this deletion: this is exactly what goes. Nothing is deleted until you press Delete.»
- «Delete» removes exactly what is listed, right then.
- «Cancel» leaves everything as it is, and Breini is told you cancelled.
This card shows with «Project only» and has no «Always allow»: every deletion for good is confirmed. With «Full access» your order is enough and it deletes with no card; deleting a database's data and destroying files outside your projects still ask.
Deleting a database's data asks twice
When Breini or a Nobreinier would delete a database, a table, a collection or any data of a database (rows, documents, the keys of a key store, a database file, a framework's reset), it must first say why, and you say yes twice. It holds in every mode, Full access too, and however the agent does it: a command, a script or a connection's call. Without a reason in words, the action does not run.
- The first card says in one sentence who asks and what it wants to delete, with the agent's reason («Why, in their words: «…»»). «See details» lists what would be deleted and ends with «Nothing is deleted unless you say yes twice.»
- Press «Yes, I authorize» to go on, or «No» to stop. Nothing is deleted yet.
- A second card comes right away, in the agent's voice: «I'm going to go ahead: … Is that all right?», with «Yes» and «No».
- Press «Yes» and it is deleted right then. If you say «No» on either card, nothing is deleted.
Only your own click answers these cards, on the app or by approving on your phone. «Always allow», a reply from Telegram or Slack, «Let your rules decide», a rule or a hook cannot. If you try to answer from a channel, the app replies «Deleting a database's data is answered only on the app's own card or on your phone.» Each deletion asks anew.
Entering «Settings» asks, and your yes lasts for one request
A Nobreinier never enters «Settings» or its pages: if something there should change, it tells Breini, and Breini enters only with your permission. If your message for that request orders the change in your own words, that is enough. If not, a single card appears that says, in a sentence like «Breini wants to enter Settings», what it would change there and what for. This happens at every permission level, also with «Full access».
- «Allow» lets Breini in for as long as that request lasts.
- Your yes counts only for the request in which you gave it. When you write something new, even in the same chat, and it wants to enter «Settings» again, the card appears again. Before, the yes lasted the whole chat until the app closed; it no longer does.
- With «No», nothing changes and Breini knows.
The price, before you say yes
When the assistant asks permission to make a picture, a video or audio with a model paid by use, the card says what it costs before you answer, on a line of its own under the question: «This costs about …», or «This costs up to about …» when it depends on the result. With a plan that already includes it, no price is shown.
When a command wants to connect to the internet
If a command tries to reach a site that commands do not reach without asking, a card appears: «… wants to connect to …», with the site; the command that asked is under «See details».
| Answer | What it does |
|---|---|
| «No» | Refuses it. |
| «Always allow» | From now on every command may reach that site without asking. It is kept in «Settings» › «Rules», where you can remove it. |
| «Allow» | Lets this chat's commands reach that site until the chat ends. |
On your phone and in the channels the card also keeps «Only this command», which lets a single command through.
An auto-accepting project
To let one project accept on its own, put one of its chats on screen and press ⌥⌘A («Auto-accept in this project, or back»). The project goes to Full access: its chats ask only about paying, deleting a database's data and destroying files outside your projects, and a quiet shield mark appears in the title strip («… accepts on its own: its chats ask only before paying, deleting data from a database, and deleting files outside your projects. Click or ⌥⌘A to turn it off»). Click the mark or press ⌥⌘A again to go back. If the chat is not in a project, the app says so.
To choose what the new chats of a project start with, open the project's row dialog and set «New chats start with». «Your default (Settings)» follows the permission of new chats.
Your rules
- Open «Settings» (⌘,) › «Rules». Above the list, «Find a rule: words, a command, a site, a project» hides the rules that do not hold what you type.
- In «What the rule covers», type for example «buying something» or «posting on my networks».
- In «What the agents do», choose:
- «Do it without asking»
- «Do it if you asked»
- «Ask you first»
- «Hand it to you»
- In «For whom», choose «The whole business» or one Nobreinier.
- Optional, under «Make it more precise»: «If this text appears (optional)», a text that fires the rule when an agent writes it (for example «delete everything»). Under «Advanced: patterns» in the same place, «How the text is read»:
- «Exactly as I wrote it»: looks for those words, capitals or not.
- «As a pattern»: for people who describe many texts with one line of symbols.
- «As a pattern, capitals or not»: the same, capitals or not.
- Optional, also under «Make it more precise», «Write this instead (optional)», only for «Do it without asking» rules: what is written in place of that text.
- Press «Add rule».
Below, «Kinds of action» lets you set what agents do with «Buying», «Sending», «Deleting», «Accounts», «Commands», «Files» and «Anything else». Reorder rules by dragging them or with ⌥↑ ⌥↓.
A rule reaches every action it names
A rule you write in words applies to every action that changes something and fits it, edits included. A rule such as «Changing files in /infra: ask me first» (with «Ask you first») makes a card appear on each edit of those files with «Project only», and the card names your rule. With «Full access» a rule that asks shows no card (except for paying); a rule that stops does stop. The agents' own notes (to-do list, memory, board) are not actions on your Mac, so they never ask.
A pattern that reads too slowly
A text or an expression has half a second to read an agent's call. If it does not answer in time, a rule that stops or asks acts as if it had matched (the call is stopped, or you are asked, and the card says why); a rule that changes the text or only warns does nothing that time. In «Settings» › «Rules» the rule's row says how many times that happened and ends with «simplify it». The other automatic rules that were slow (the ones in files, yours, a project's or a skill's, and the built-in ones) are listed under it.
If the app cannot read your rules
If the rules were written by a newer version of the app, or cannot be read, «Settings» › «Rules» says «The app could not read your rules», and the card is titled «Your rules (the app could not read them)». Until it can, the app asks you before every action one of them could cover, and agents start nothing that changes something outside on their own. Open the newer version again and the rules are there as they were. A rule another version wrote is read as far as it means something; a behavior this version does not know reads as «Ask you first».
A refusal always says why
When a rule of yours stops something, the agent is not left guessing: it reads which rule it was, why it exists and what it can do instead. And it knows you can change the rule or do the step yourself.
- A rule can carry its reason. Tell Breini when you create it, for example «install nothing with sudo: this Mac belongs to the office». In the list you see it as «Why: …».
- The files that control a project (git hooks, editor settings, the terminal's startup files) are guarded. If an agent needs to change one, the refusal tells it which file it is and how to ask you for permission in one click.
- When a tool fails, the agent is told what kind of failure it was (refused, not found, no network, wrong arguments and so on) and what to do instead, so it repeats the same wrong call less. You do not see that line in the chat.
