Reading a response

A reply is more than text. This page decodes everything that can appear in one.

Parts of a reply

Thinking

Some models reason before they answer. You’ll see a collapsible row with a brain icon that reads Thinking… while it works and Thought for 12 seconds (or similar) when it’s done. Expand it to read the reasoning.

Status and tool cards

While the assistant works, a status row shows what it’s doing right now — Searching…, Executing… — and expands into the full list of steps, including the web pages it found.

Each tool the assistant used gets a collapsible card. Expand it to see the Request (what the assistant asked for) and the Response (what came back). If the work was handed to a sub-agent, the card says so.

Sub-agents

For bigger jobs the assistant can delegate to several sub-agents at once. A row of agent cards shows each one’s task and live progress, with an overall progress bar.

Sources

When a reply draws on web pages, papers or documents, a Sources button appears under it with the sites’ icons. Expand it to see every source; click one for the snippet that was used and where it came from.

If the chat belongs to a project, each source has a + to add it to the project’s knowledge base.

Citations

A claim drawn from a research paper ends with a chip naming the paper’s first author and year — Chen 2019. Several papers can share one claim; past three, a +N more chip holds the rest. Hover a chip (or tap it on a phone) to see the paper: its title, authors, journal, citation count and whether its details were verified, with a link to the paper’s public page.

A chip in a muted warning style names a paper that is not in the chat’s citations, so its details were not checked — treat that claim with care. Numbered chips, like [1], point at the reply’s sources.

Documents cite the same papers as linked author-year references — (Chen et al., 2019) — with the same card on hover. See Markdown.

Verification

For factual answers the assistant can run a Verification pass that checks its own claims against the sources. Expand Details to see the claim-by-claim result. On free plans you’ll see that a verification ran, but not the detail.

A confidence badge at the right of the message’s action row summarises how well the reply is supported. On the latest reply, you can turn a flagged claim into a follow-up question with one click.

Citations and execution plan

A toggle in the chat header opens a side panel showing the reply’s citations, the task list the assistant is working through, or the transcripts of any sub-agents — whichever the conversation has produced.

Code

Every code block has a toolbar: Preview / Code, Edit, Run, Download and Copy code. Run executes JavaScript in your browser so you can try a snippet in place.

Diagrams and charts

Diagrams and charts render inline. Hover for Expand and Copy source.

Papers

When the assistant finds academic papers, each appears as a card with actions to add it to the chat’s citations or to your Library.

Artifacts

When the assistant builds a document, board, deck or chart, a card appears in the reply. Click it to open the artifact in the side panel, or use the card’s menu to download it.

Suggested tools

If the assistant thinks a tool you haven’t enabled would help, it suggests it as a chip you can switch on with one click.

Actions on a message

Your own messages: Edit, Copy, Delete.

Assistant replies: Copy, 👍 Good response, 👎 Bad response, Regenerate, and a More menu with Edit, Read aloud, Continue, Generate image and Delete.

When a reply fails

If the model’s provider could not answer, the reply shows a short explanation with one or two actions:

• Try again re-runs the same request on the same model. It is offered when a retry can plausibly help (a timeout, an overloaded provider); a request the provider rejected outright will not offer it, because sending the identical request again would fail the same way.
• Retry with re-runs the same request on your default model instead (the one you set in Settings → Models, or the workspace default if you have not chosen one). It appears when the failed reply ran on a different model. Only that reply changes model — the chat’s own model selection stays as it was, so your next message still goes to the model you picked.

A model whose provider keeps failing is taken out of the model picker for a while. Chats already using it keep working, and it comes back as soon as a request to it succeeds.

A reply that could not start because your Free plan’s daily allowance is spent shows a different card, with the reset countdown, Upgrade, and Notify me when it resets — an email with a one-click Continue where you left off button. See Plans and limits.

Editing and branching

Editing a message opens it in place. Save keeps your change without re-running anything; Send re-runs the conversation from that point. Either way the original is kept — a ‹ 1/2 › control appears under the message so you can switch between versions. Regenerating a reply works the same way.

Deleting a message that has replies below it offers Delete branch, which removes everything that followed.

Rating

Thumbs up or down help improve the assistant. A thumbs-down opens an optional box for what went wrong. If a reply failed or came back blank, a bug icon lets you report the problem with the technical details attached.

Ask about a selection

Select any text in a finished reply and floating Ask and Explain buttons appear. The answer streams in just below the message.

Who said what

Hover a message to see when it was sent. In a shared chat, each person’s messages carry their name and avatar; each assistant reply shows the model that wrote it.