Agents > Xedant Agent > Docs > Prompts

Prompts

Prompts are a queue of saved requests that the agent sends on its own, one by one, into new chats. Kanban boards with cards and tags used to live here — they are gone now, replaced by the prompts dialog with server-side AutoSend. Write your requests down in advance, and the agent will run them one after another as soon as a slot frees up.

The most common uses: a batch of small subtasks (“rename the files”, “bump the versions”, “check the build”), a “morning run” — a list of prompts that must complete without you, and handing out prepared requests one by one while you do other things.

The Prompts dialog

Click the “Prompts” button on the input panel — a dialog opens with a list of rows. A prompt row is the request text plus the model and skill selectors for the chat it will be sent to.

  • Text — an auto-sizing field: it grows with the content. Enter inserts a new line, and sending is done with the button or the Ctrl/Cmd+Enter shortcut;
  • Model and skill — the selectors in the row: exactly these go into the new chat. Empty values mean the globally selected model/skill;
  • Microphone — record the request by voice right into the row (with speech recognition enabled);
  • Paste from clipboard — paste text, an image, or a file from the clipboard;
  • Move — the up/down arrows change the order of the rows in the queue;
  • Delete — remove the row from the queue;
  • Send — launch the prompt right away (see below).

The “New Prompt” button in the dialog header adds a new row at the top. Next to it is a paste-from-clipboard button: it adds a new prompt from the clipboard contents — text, an image, or a file. At the bottom there is a row counter (for example, “3 prompts”) and the AutoSend panel. When there are no rows, the dialog shows one empty draft row. Changes save automatically — the save status indicator in the header tells you so.

Launching a prompt

Sending a row creates a new chat with the model and skill from the row’s selectors, sends the text into it, and takes you to that chat. When a fresh empty conversation is still there, the prompt goes into it instead of a new chat. AutoSend behaves the same way — the only difference is who pressed the button.

  1. A new chat is created; the title is taken from the first line of the prompt;
  2. The prompt text is sent into the chat;
  3. The row is consumed — use-once: after launch it disappears from the queue (if the send fails, the row is not restored; you need to add the prompt again);
  4. The interface switches to the new chat (the “Auto redirect” setting), and a successful send shows a “Prompt sent to a new chat” confirmation.

AutoSend: automatic sending

The AutoSend switch at the bottom of the dialog turns on auto-sending: the server itself launches prompts from the queue into new chats, keeping up to Max active chats chats busy at the same time. When one chat finishes its work, the next prompt is sent into its place.

  • It runs on the server, not in the browser: the queue keeps sending even when the tab is minimized or closed. (The scheduler used to live in the browser, and a hidden tab “froze” its timers — prompts waited for minutes.)
  • A 1-second metronome: the server checks the conditions every second, while the launches themselves happen no more often than once per poll interval (Check every);
  • Instant pickup: as soon as a chat finishes, the next prompt is sent almost immediately, without waiting out the interval;
  • FIFO order: prompts launch in queue order — from the oldest to the newest;
  • When a fresh empty conversation is still there, the agent reuses it instead of spawning new chats;
  • While auto-send is running, auto-fix does not start — the queue works without interference from background repairs (that is how the “AutoFix disabled until done” switch, on by default, works);
  • Active chats are counted from the server’s list of busy chats plus the launcher’s own in-flight launches — one chat is never counted twice.

When AutoSend is turned on from one device, the queue and its settings synchronize across all open tabs and devices.

Queue settings

The “AutoSend settings” gear next to the switch opens all the queue settings. Inside: the maximum number of active chats, the poll interval, the peak-hours mode, auto-redirect, auto-commit before sending, and the auto-fix pause for the duration of a send.

  • promptsItems — the prompt list: rows with text, a model, and a skill (JSON);
  • promptsAutoSend — turns auto-sending on and off (off by default);
  • promptsMaxActiveChats — the maximum number of simultaneously active chats: 1–10 (default 1);
  • promptsPollSeconds — the poll interval in seconds: 1–3600 (default 20);
  • promptsAutoRedirect — automatic switch to the chat launched by a prompt (on by default);
  • promptsIgnorePeakHours — whether auto-send should respect peak hours. While it is on (the default), auto-send works as usual, ignoring peaks; when it is off, prompts whose model is currently in peak wait in the queue, while prompts with off-peak models keep sending;
  • promptsAutoCommit — commit the project repository before each prompt is sent (on by default). The commit message becomes the title of the most recently finished chat;
  • promptsAutoFixDisabledUntilDone — while auto-send is working through a non-empty queue, auto-fix waits for it to finish and does not inject build errors into the middle of the batch (on by default).

About models’ peak hours and the MODEL_PEAK_HOURS/MODEL_PEAK_DAYS variables — on the Models page.

The chat message queue

Prompts are not the only queue: every chat has its own message queue. When a chat is busy (the agent is processing a request), new messages are not lost — they join a persistent queue and are sent as the agent frees up.

  • FIFO order — messages are delivered by creation time, one at a time: the pause between “one message finished” and “the next one started” is minimal;
  • Pause — a queued message can be paused: dispatch skips it, and when the pause is lifted it returns to its place by creation time;
  • Editing — the text of a waiting message can be changed right in the queue; edits synchronize across devices;
  • Clearing — the queue can be cleared entirely, or single messages removed;
  • The queue in the interface — with a chat open, its current queue is visible and manageable.

Delivery is “at most once”: a message is deleted from the queue right before it is sent, and if the send fails, an error appears in the chat. Every message source puts things into the queue: the central Telegram bot and your own bots, agent buttons, the task scheduler, auto-fix, the Remote API, Yandex Disk sync — and the web client itself, when the chat is busy.

Links

← Models

Analytics →