Chat

The Chat Interface

How the conversation model works - streaming responses, inline actions, and the controls that shape every message.

How the conversation works

PandaOS chat is not a simple question-and-answer box. It is a live working session. When you send a message, Panda streams its response as it works - and that work can include editing files, running terminal commands, querying databases, or calling external services. Each action appears inline as a card in the conversation thread, so you see exactly what happened and what it produced.

This means a single message can kick off a chain of real actions: "Set up a new API route for user profiles, add the Supabase migration, and test it" might produce five or six action cards before Panda finishes its response. You can stop a response mid-stream at any point and continue from where it left off - nothing is lost.

Shaping each message

Every message you send carries several decisions beyond the text itself:

Which model answers it - you can switch models between messages without starting a new conversation. See Models & Harnesses.
How much autonomy Panda gets - the interaction mode (Default, Auto-Edit, Agent, Plan) controls whether Panda asks before acting or works independently. See Permissions & Approvals.
What context it has - files, images, and references you attach travel with the message. See Attachments & Context.
Which agent handles it - you can hand a message to a specialized agent instead of the general assistant.
Skills and mentions - type / to invoke a reusable skill, or @ to reference a specific file, app, or agent by name.

The Tools menu in the chat bar switches capabilities on for the conversation without going near Settings:

Image generation - lets Panda produce pictures, which render inline in the thread. Configure the provider under Models.
Generated interfaces - lets Panda answer with a real component instead of prose where that reads better: a comparison table, a chart, a status board, or a short form to fill in. Answers you give in one of those forms come back into the conversation as your reply.
Spotlight - whether an app's tab opens when Panda uses it. See Spotlight.
Agent models - whether an agent may switch model when it activates. Off means everything runs on the model you picked.

What is available depends on the harness you are on, so the menu only ever offers what the current engine can actually do.

Each chat remembers its engine

The engine and model belong to the conversation, not to the app. Switching a chat to Codex does not move your other chats, and coming back to a chat later resumes it on what it was running. Switching mid-conversation carries the thread across, so you can start on one harness and finish on another without losing the context you built up.

Context window awareness

Every AI model has a finite context window - the amount of information it can hold in working memory at once. The context donut in the input area shows how full that window is. As a conversation grows with messages, file contents, and tool results, the window fills up.

When the context window is nearly full, the model may start losing track of earlier details or slow down. That is a good signal to start a fresh conversation for unrelated work. Think of it like a desk: a cluttered desk makes it harder to focus on the task in front of you.

Session usage

The second ring, to the right of the context donut, tracks something different. The context donut measures one conversation. The session ring measures how much of your provider plan you have used across all of them.

It appears only when you are on a plan the provider reports limits for, which today means a Claude subscription or OpenAI Codex. On PandaOS Cloud your usage is measured in Credits instead, and on your own API keys you are billed directly by the provider, so neither shows a ring.

The ring fills with your session window alone, the rolling few hours a provider allows before a short cooldown. It changes color as that window fills, amber past 70 percent and red past 90 percent.

Click it for the full breakdown. The popover lists every limit on your plan, not just the session one: weekly totals, per-model weekly limits where the provider sets them, and any extra usage. Each row shows how much is used and when it resets, both as a countdown and as a real time you can plan around.

Check the popover rather than the ring before a long stretch of work. The two can disagree, and the ring only ever answers the short question. You can sit at a nearly empty session ring while your weekly limit is almost gone, and the popover is the only place that shows it.