The conversation view

A thread's conversation view is a single scrolling stream: your messages, the agent's replies, and everything the agent did to produce them. This page covers what you'll see in that stream. For what a thread is and how its lifecycle works, see threads.

A conversation: a user message, a collapsed Explored activity group, an expanded edit call with an inline diff, and the end-of-turn changed-files rollup

Screenshot coming soon

/images/docs/interface-conversation.png

Writing a message

The composer at the bottom of a thread takes more than plain text:

• @ references. Type @ to insert a reference to a file in the workspace or to the output of a running terminal, so the agent gets the content without you pasting it.
• Images. Attach JPEG, PNG, GIF, or WEBP images — up to 8 MB each and 12 MB per message. Every agent except Grok Build takes them, and one that can't says so rather than dropping them silently.
• / commands. Start a message with / to run one of the agent's own commands. LinkCode lists what your setup actually advertises. Every agent except Grok Build supports this.
• $ shell. Start a message with $ to run a shell command through Codex or OpenCode. The output comes back as a normal turn.

A command or shell directive has to be the first thing in the message and there can only be one, so LinkCode offers to move, convert, or remove a stray one rather than sending something ambiguous.

Each message you send renders as a bubble on the right. Long pastes (over 20 lines) collapse automatically with a Show more toggle. Hovering a message reveals the send time and a Copy button.

Agent replies

The agent's turn can contain several kinds of content, in the order the agent produced them:

• Text — rendered as markdown. Links carry the site's favicon and open in the Browser panel or your system browser.
• Thought — a collapsible reasoning block. It streams as the agent thinks and collapses to a one-line summary ("Thought for 12 seconds") once the agent moves on.
• Tool activity — see below.
• Images, audio, and linked resources the agent's underlying provider returns as part of its output.

Tool activity

Every action an agent takes — reading a file, running a command, editing code — is a tool call. LinkCode renders a run of consecutive same-kind calls as a single collapsed row instead of one row per call, so a burst of file reads doesn't flood the transcript. The groupings you'll see most often:

• Explored — file reads and searches
• Ran commands — shell/terminal execution
• Changed files — edits, deletes, and moves

(Web fetches, integrations, and explicit "thinking" tool calls group the same way, under their own labels.)

Click a grouped row to expand it into its individual calls. Each call shows a kind badge (Read, Edit, Delete, Move, Search, Run, Think, Fetch, or Tool) and, where relevant, an expandable body: raw input, a live terminal block, or an inline diff. A call that's gated behind a permission request is never folded into a group — it always renders as its own row so you don't miss it.

Diffs. With Claude Code and Codex, an edit call's expanded body shows a per-file diff with +/- counts. The other agents report edits as plain text — see the agent comparison. When a turn finishes, LinkCode adds a rollup card below it summarizing every file the turn touched, with Review and Undo actions and a per-file breakdown (the first three files, with the rest behind a Show more toggle).

Subagents

When an agent spawns a subagent, its work appears as its own Subagent entry with a View full transcript action, rather than being flattened into the parent turn.

Long threads

Two things help once a thread gets long:

• The minimap — a slim rail beside the conversation with a mark per turn; click to jump.
• Compaction — when the agent's context fills up, it summarizes the conversation so far. LinkCode shows that as a Context compacted marker in the stream, with before/after token counts where the agent reports them, rather than silently dropping the earlier conversation.

Plan progress

When the active agent reports a step-by-step plan, LinkCode pins a collapsible Plan card above the composer showing the current step (for example "Step 2/5") and, expanded, the full list with each step's status. It stays pinned until the plan completes or the agent stops reporting one.

Permission prompts and questions

When an agent needs your approval to proceed, LinkCode pins a permission request above the composer instead of burying it in the transcript. Some agents can also ask a multiple-choice question, which appears the same way. You can answer either with the keyboard — number keys, arrows, and Enter all work. See approvals and keyboard shortcuts.

Artifact cards

When a tool call produces a file — for example an edit or a move — LinkCode shows a small card for it under the call, up to four per call. Clicking a card opens the file in the Files panel. This is one of two ways content produced by an agent can surface in the conversation; see artifacts for the other (content that renders inline, like diagrams).

Copying a reply

Every agent turn ends with a small action bar. Copy copies the turn's text and is the one action there that works today (the others are covered in the callout above).