Session: input field, modes, questions

A session is one conversation with the agent in a project folder. You write a task, the agent reads the code, edits files and runs commands, and under every answer you can see what it did.

The feed

• “Used N — …” — which tools the agent called between replies: “Read”, “Edit”, “Shell”, “To-do list”, “Knowledge base: entry” and others. A click expands the calls with their results.
• The agent’s answer — plain text with markup. File paths in it are links.
• The summary of the turn’s changes — under the answer, if the turn changed files: “Changed 2 files +23 −1” and “Rewind the turn”. In detail — Changes and undo.
• Service marks — as a collapsed line: an agent switch, session compaction, marks of the goal.

The input field

Under the field there is a control line:

• / — commands: built-in (/goal, /side, /compact, /undo…), the project’s own commands and skills. A command without arguments runs on Enter; for a command with optional text (/side, /goal) the first Enter inserts it into the field — add the text or press Enter again.
• @ — a project file or folder as part of the message; the agent gets their contents.
• ! — a shell command from you, without the agent: !bun test. It runs in the project folder, and the output shows up in the feed right away and goes into the agent’s context.

The request map

A long session has request map ticks to the left of the feed, one for each of your messages. Hovering shows the request and the start of the answer; a click takes you to it.

Attachments and quotes

• Images and files — with the button, by dragging or by pasting from the clipboard. Each has a thumbnail; a click opens a preview. If the model does not read images itself, the attachment goes as a file, and the hint on the badge explains it: “The model does not read images itself — it opens the file with its tools”. There is no limit on the number of attachments.
• A large paste — three lines or more, or over 150 characters — becomes a chip “Pasted text · N lines”. A click shows the whole text; you can “Expand into text” or delete it. The model gets the whole paste.
• A quote from an answer — select a fragment in the agent’s answer, and an “Add to message” button appears next to it. A “From the answer · N lines” chip goes into the field, and the agent gets the fragment as a quote.

Permission modes

A mode is the rules of the current session: what the agent does without asking and what it asks about. You choose the mode — in the input field or with the /mode command; the agent does not change it. The organisation does not fix the mode: it is a developer setting. A new session starts with the mode chosen last.

Permission request

When the agent needs permission, a request appears above the input field: the whole command, and for a file edit — the edit itself, as in a diff — and three buttons.

• Allow once — only this action.
• Allow always — remember it for the project. Under the command it says exactly what will be remembered: “«Always» remembers for this project: commands cat *, echo *”.
• Deny — with a “What to do instead” field. The text goes to the agent and the turn continues: the agent does it your way rather than stopping.

A request from a subagent is marked “from a subagent”.

The agent’s questions

If the task lacks a decision, the agent asks — with a card above the input field. A question has answer options with explanations, and you can write your own answer in the last line. There may be several questions: “1 of 2 questions”, “Next”, “Back”, “Submit”.

Background work

The agent can start a long command in the background and keep working. While such work runs, the line under the input field shows “1 background task running”; a click opens the list.

• Each task has its running time, “Output” (the whole output, appended while the command runs) and “Halt”.
• “Halt” does not wake the agent. A “Shell: cancelled” mark goes into the feed; the agent sees it on the next turn.
• Your own command (!) or the agent’s command, while running, is moved to the background with the “Move to background” button or Ctrl+B (on macOS — also Ctrl+B).

Balance and usage

The organisation’s balance is at the end of the line under the input field. A click opens “Balance and spending”: how many days it will last at the organisation’s spending over 30 days, the plan and prices per million tokens, the estimate for this session in roubles, your spending over 30 days — as bars and by day, “Open cabinet” and “Refresh”. Requires company connector v0.18.0 or newer; with an older one the window says “the company connector does not know this request — update it”.

• When funds are low, the balance turns yellow; when they have run out — red, and a bar “Open cabinet / More details / Hide” appears above the input field.
• A turn refused for lack of funds is explained in words, with a “Balance and cabinet” button.
• The session cost is an estimate by the plan, not an invoice. Charges are visible in the workspace, see Models and usage.

Session compaction

When a conversation gets long, you can compact it: /compact. The recent part of the conversation is kept verbatim, and the agent retells the earlier part as a summary for itself. In the feed compaction shows as a “Session compacted” divider with the number of tokens; the summary opens with the “Summary for the model” button.

In the console