Approvals and modes

Each chat runs in your own permission mode, and anything that mode doesn't allow comes to you as a card in the bar that shows exactly what will happen.

Permission modes

The mode decides what the AI may do on its own and what it asks you about first. These are Claude Code's and Codex's own modes; Saiki just lets you pick and switch them from the bar.

Claude

ChatGPT

Modes that skip asking altogether, such as Claude Code's bypass mode or Codex's full access, are never offered.

Change a chat's mode

The mode shows as a chip in the chat's header. Press ⇧⇥ or click the chip for the next one, in the same order as Claude Code's own ⇧⇥: Manual, Accept Edits, Plan, Auto (or Read-only, Ask, Auto-review). It's also Change Mode in ⌘K. A Claude chat switches at once; a ChatGPT chat from your next message.

Your default mode

New chats start in Claude mode and ChatGPT (Codex) mode from Settings → AI & Agents. The first choice, "Same as Claude Code" (or Codex), follows the default you've set in the tool itself.

When you pick a mode there, a card asks whether to make it the tool's own default too: Also Save to Claude Code (or Codex) changes only that one setting in its config file, and Only in Saiki leaves the file alone.

Confirm cards

When the AI wants to do something its mode doesn't allow, the answer pauses and a card appears above the field. Its title says what it wants ("Claude wants to run a command", "Claude wants to edit notes.txt", "ChatGPT wants network access"), and the card shows exactly that:

• a command, with the folder it runs in (a long one folds behind Show command);
• file edits as a diff, each file with its lines added and removed;
• why it asks, when the tool says (a path outside the folder, a web host);
• the AI's own one-line description of the step, marked as written by the AI, above what will actually happen.

Keys

The Always button says what it covers: Always Allow Similar Commands, Always Allow This Location (edits in that file's folder) or Always Allow Reading This Location for Claude; Always Allow Edits in This Chat or Always Allow in This Chat for ChatGPT. It lasts only for this chat and is never written to a settings file. Other tools are allowed one step at a time.

Allow All and Deny All

When the AI asks several things at once (its subagents can ask too, marked Subagent), the card shows "1 of 3". ⌘↑ and ⌘↓ page through them, ⌘⇧↵ allows them all and ⌘⇧⌫ denies them all.

Your tools' own allow and deny rules still apply first: something Claude Code or Codex already allows never reaches a card.

When the bar is closed

An ask doesn't wait for you to look. With the bar closed, the needs-you card slides in at the top of the screen, or a macOS notification if you're away. A notification offers Open and Deny, never Allow, so you always see what you approve. See The needs-you card.

Stopping an answer with ⌃C denies whatever it was asking.

Plan mode

In Plan mode the AI looks around and proposes a plan without changing anything. For Claude, switch to Plan with ⇧⇥ or type /plan; /default goes back to your usual mode. For ChatGPT, use the Plan chip in the Agents view's Detailed level, or Plan First in ⌘K there.

The plan waits under the answer in a Plan card:

• ⌘↵ Implement: Claude leaves Plan for Accept Edits and gets going (ChatGPT keeps its access as it is).
• ⌥⌘↵ Implement, Approving Each Edit: Claude switches to Manual, so every edit comes to you as a card.
• Or type feedback and press ↵ to keep planning.

Questions

Sometimes the AI asks you to choose ("Which database should I use?"). Its options appear in one picker:

• ↑ ↓ move and ↵ picks, then moves on. When it allows several answers, ↵ ticks each one and ⌘↵ moves on.
• Or type your own answer in the field and press ↵.
• Several questions come one after another, and your answers go once the last is answered.
• ⌘⌫ skips the question.

The steps row

The AI's tool calls fold into one row in the answer, with a one-line summary such as "Ran 2 commands, edited a.txt and b.txt" (asks you denied are counted too). While it works the row shows the current step. Click it, or press → with the field empty, to see each call.

Undo the agent's changes

When an answer changed files in the chat's folder, a chip under it says by how much: "+42 −7 · 3 files". Click it to see the files (marked added, modified or deleted) and:

• Revert Changes puts the folder back the way it was before that answer.
• Undo Revert puts the changes back again.
• View Changes, or clicking a file, shows the diff in the Agents view.

⌘K has Revert Changes (or Undo Revert) for the newest answer that changed something.

Before each answer Saiki takes a snapshot of the chat's folder in its own storage, never in your folder or a Git repository it holds. A revert is refused rather than lose work: while an answer is being written, while another chat is working in the same folder, or when files it would touch have changed since. Your home folder, very large folders and nested repositories aren't snapshotted, and snapshots unused for 30 days are cleared away. It needs Git, which comes with Apple's command-line developer tools.

Folder trust

Before an AI runs in a folder for the first time, the chat asks Do you trust this folder?, showing the folder. The AI will be able to read, edit and run the files there, so only trust folders you know.

• ⌘↵ Trust. The button is dimmed for a moment after the card appears, so a key meant for something else can't trust a folder ("Take a look first").
• ⌘⌫ Don't Trust leaves the chat, and nothing runs there.

Trusting a folder covers the folders inside it. Saiki remembers your answer in its own data, never in Claude Code's or Codex's settings. Saiki's own Workspace folder, where new chats run, is always trusted, so you see this card only for a chat that works in one of your folders, such as one continued from the Agents view.