Chatting
Chat is the heart of Personal Agent. This page covers the conversation itself — the reading pane, the composer, and what you can do with each message. The per-chat controls (model, security, memory, modes) have their own page: Chat controls & modes.
The reading pane
Your conversation reads top-to-bottom: your messages and the assistant’s answers, each turn rendered as clean Markdown — including Mermaid diagrams, drawn inline with a small toolbar to view the source, download the SVG or maximize them. Older history loads as you scroll up.
Under each answer you’ll find a compact footer with (depending on your preferences):
- Cost - input ↑ and output ↓ tokens, a ⚡ cache indicator (read + write), this single answer’s cost, and throughput in tokens per second, all under one toggle.
- Model name - which model produced it.
- Timestamp - when the message was sent.
You choose which of these appear under Settings → Appearance → Message details (the three toggles are Show timestamp, Show model name and Show cost), so the chat can be as clean or as instrumented as you like.
Reasoning, tool calls and plans
An answer is more than text. Inline, in the order they happened, you may see:
- Reasoning — a collapsible block showing the model’s thinking, on models that support it. It’s collapsed by default so it never gets in the way.
- Tool-call cards — one per tool the agent used, with friendly names (“Read file”, “Searched the web”, “Drafted reply”, …). Each card’s icon is colour-coded by status — running, succeeded or failed — so you can scan a long turn at a glance. Expand a card to see its arguments and result. File edits render as diffs; a plan renders as a prominent plan card; a to-do list renders with live status icons. When the agent fires many tools in a row, they collapse into a single “Ran N actions” header you can expand.
The context gauge
A thin bar at the top edge of the composer shows how full the model’s context window is, measured against the auto-compaction threshold. Click it for a breakdown - messages, system prompt, skills, project rules, memory, history summary, system tools and MCP tools, plus the free space left - with a detail list of memory files, MCP tools and the biggest system tools. When a conversation gets long, older messages are automatically compacted into a summary so the chat keeps going; the gauge turns amber past 80% and red once it is full or compaction has happened, and the breakdown notes how many tokens were folded into the summary. (Compaction is applied on both the inline and the durable run paths, and an overflow self-heals even when no window is known.)
Alongside the live window, the breakdown shows the session token total - everything this chat has spent across all its turns, with the running cost - split by kind (input, output, cache write, cache read) and with the share attributed to sub-agents broken out (count, their tokens, their cost and the average cost per agent), so you can see how much of the total a delegated worker accounted for.
The composer
Type your message in the box at the bottom and press Enter to send (you can switch this to Cmd/Ctrl+Enter under Settings → Appearance → Composer, so plain Enter inserts a newline; Shift+Enter always inserts a newline). While the agent is working, the send button becomes a stop button, and a new message you type is queued — sent automatically once the current answer finishes, and shown muted in the thread until then.
From the composer you can:
- Attach images or documents (PDF, text, Markdown, CSV). Vision models read images directly, so you can paste a screenshot and ask about it.
- Speak — tap the microphone to dictate; your speech is transcribed into the box. See Voice below.
- Run a slash command — type
/to open the command palette. See slash commands. - Open the per-chat controls — model, security, mode, memory and more, covered in Chat controls & modes.
Suggested follow-ups
After an answer, the assistant may offer a few follow-up suggestions as clickable chips. Tap one to send it, or just keep typing — they disappear as soon as you reply.
What you can do with a message
Hover over any message (or tap it on mobile) for its actions:
| Action | What it does |
|---|---|
| Copy | Copy the answer’s text to the clipboard |
| Regenerate | Re-run the last answer — e.g. after switching model — discarding the old one |
| Edit | Edit one of your messages and re-run the conversation from there |
| Rewind to here | Delete this message and everything after it (in coding chats, the workspace is rewound too) |
| Branch from here | Fork a new, independent chat from this point, leaving the original untouched |
| Revert | Undo the file changes a coding answer made, and trim history back to before it |
!!! tip “Rewind vs. branch vs. revert” Rewind rewrites the current chat and is destructive; Branch keeps the original intact and continues a copy; Revert is the coding-specific “undo this answer’s edits”. Each asks for confirmation before doing anything irreversible.
You can also export the whole conversation to Markdown from the chat’s header menu.
When the agent needs you
Some turns pause for your input — these appear as cards right in the conversation, and clear themselves once you respond:
- A question — the agent asks you to choose between options (single- or multi-select, sometimes with a recommended choice and a free-text field). Answer it and the same turn continues from where it paused.
- An approval — when a chat requires approval for tool calls, each call waits for Approve, Reject, or Always allow (which whitelists that command on that device so you’re not asked again). See security modes.
- A draft reply — for incoming messages, the agent proposes a reply you can edit and then Approve & send. More in Inbox & contacts.
When something goes wrong
Runs are resilient. A transient provider hiccup (rate-limit, a 5xx, a dropped connection) is retried automatically, and the fallback chain can cascade to another model rather than failing the turn. If an answer does end in an error, any text the model had already streamed is kept — it isn’t thrown away — and an explicit Retry button appears so you can re-run the turn without retyping. Work that was already done can resume rather than starting from scratch.
Sharing a chat
From the chat header menu, Share creates a read-only, password-protected public
link to the conversation. Set an expiry (never, or 7, 30 or 90 days) and revoke
it anytime. Whoever has the link and the password can read the chat at a public /s/…
address; the shared view is read-only and updates live as the chat continues. Send the
password separately — and note that the whole transcript, including tool calls, becomes
readable, so don’t share a chat that contains secrets like API keys or passwords.
Voice
If your admin has configured speech models, chat becomes hands-free:
- Speak — the microphone button records you and transcribes speech to text.
- Read aloud — play any answer as speech.
- Read responses aloud — toggle auto-read so every answer is spoken as it streams in.
The main chat
Your main chat is special: it’s pinned at the top of the sidebar, can never be archived or deleted, and is where proactive messages land — due-date nudges, briefings, and anything the assistant reaches out about on its own. When it has something for you, the sidebar shows a gentle “reached out” badge. The main chat always runs in standard mode.
Sessions and folders
Every other conversation is a session. The sidebar shows recents; the full, searchable list lives on the Sessions page. You can rename, archive or delete any session.
Folders group related chats and carry folder-specific instructions that apply to every chat inside them (for example, “Always answer with sources and in German.”). A folder can be private or shared with your group; deleting a folder keeps its chats, which are simply detached.