Using the agent chat
What your users see when they chat with the AI Agent.
When the agent is enabled with the native placement, users see Agent Chat in the app sidebar. The page includes a conversation history rail and a full-screen chat view. On smaller screens, the history opens as a sheet.
With the floating placement, a launcher button appears in the bottom-right corner of the app. It opens a chat panel with a history dropdown and a New conversation button.
Starting a conversation
A conversation is created when the first message is sent, not when the chat opens. Users can type a message and press Enter, or use Shift+Enter for a new line.
Conversations are private to the user who created them. If an organization is active, the conversation is scoped to that organization, but it is still private to the user.
Following the response
Answers stream in as Markdown. Code blocks include copy and download buttons. When the agent calls a tool, an inline chip shows whether the action is running, succeeded, failed, or was interrupted.
Each reply groups its reasoning, tool calls, intermediate text, and final answer in one container. The compact header has one expandable reasoning trace and a tool-call summary; expand either to inspect the full details. Generated images and their loading placeholders remain visible within the reply, separate from the ordinary tool count.
If the model supports reasoning, an expandable Thinking section appears above the answer.
Each answer has its own container with a Copy answer button below its text. Copying preserves the source Markdown, including code blocks.
Copy, edit, and rerun messages
Sent messages have Copy message, Edit message, and Rerun message actions in both chat layouts. Edit opens an inline field with Send and Cancel; Enter sends, Shift+Enter adds a line, and Escape cancels. Rerun immediately submits the original text again. Stop the current response before editing or rerunning.
Editing or rerunning starts a new active branch from that message, replacing it and its later answers in the visible conversation. Pi retains the original branch in saved history. The request reuses the original image and file attachments and image-generation intent, with the currently selected model and thinking level. Any draft and pending attachments in the main composer stay intact.
Rerunning may execute tools again; it does not undo earlier application actions. If a request fails, retry the same action. Network retries reuse the admission ID so an already accepted request does not start another run.
Attachments and tools
From the plus menu, users can attach up to four images to a message when the model supports image input. The menu also lists the available tools as a reference.
Choose Files to attach supported UTF-8 text, Markdown, data, configuration, or source files, including names such as Dockerfile, Makefile, README, and LICENSE. The default is four text files per message in addition to images. Add a text request describing what the agent should do, then send. Pending chips show filenames and sizes and can be removed; file contents are not pasted into the message box.
The agent receives the filename and full text within the configured limits. Saved user turns show downloadable attachments; downloads return the original bytes. Files use the same private storage as images. PDF and Word conversion are not supported. Unsupported names, invalid UTF-8, and exceeded limits produce specific errors without silently truncating files.
Pending files remain with their scope's draft across navigation and floating-panel toggles. A failed send retains them for retry; refresh or choosing a conversation clears unsent attachments. Saved files remain available from history. Compaction may remove older file content from model context; reattach a file when its full detail is needed again.
When image tools are enabled, the Image generation button sits after the text-size control. Turn it on to reveal the configured aspect ratios, choose a ratio, and send a message. A successful saved generation or edit ends the image-mode run and displays the image with a download button. Click the image button again to return to normal chat. See image tools for provider configuration and editing behavior.
The Stop button interrupts the current run. It does not undo completed actions.
When a conversation reaches its limit
Pi automatically summarizes older model context when a conversation approaches the model's limit and can retry recoverable model errors. Original conversation history remains saved. If recovery fails, the run fails. If a growing conversation cannot be saved within the storage limits, the chat shows an unsaved warning; the latest output may not be durable even when an action already completed.
Start a New conversation and include the relevant context. The old saved conversation remains in history. See session limits and compaction for the separate model, storage, and image limits.
See setup for changing the attachment limit or enabling the thinking selector.