Troupe Docs
AI agents

Follow the agent's work

Read the agent's plan, steps, and diffs, review changed files, and queue, stop, edit, or retry your prompts.

While an agent works, the chat shows what it is doing step by step. You can read its plan, open each file it touched, and line up your next requests. You can also stop it at any time.

A finished chat. The plan sits at the top with all four tasks crossed out. Below it are your message, a collapsed "Worked for 1s" row, and the agent's reply.

On a phone, the chat works the same way, with touch gestures for the reply actions. See Use Troupe on a phone.

Read the plan

Many agents write a plan before they start. The plan stays at the top of the chat while you scroll. Its header shows Plan and how many tasks are done, such as 3/4. A spinner shows while a task is in progress.

Each task has a status icon and a priority: High, Medium, or Low. Finished tasks are crossed out. Click a task that has details to read them. Click the header to collapse or expand the plan. On a phone, the plan starts collapsed.

A chat has one plan. When the agent updates a task, the plan changes in place. When the agent writes a new plan, it takes the place of the old one and opens. When every task is done or cancelled, the plan collapses. Click the header to open it again. The plan does not show up a second time inside the agent's reply. When a turn ends, a task that was still in progress goes back to waiting, so no spinner keeps turning.

See each step

While the agent works, a Working… row sits above its reply. When the turn ends, the row changes to Worked for and the time it took, with the number of lines added and removed. Click the row to open it. If a step failed, the row is red and says "needs attention".

Inside, the steps are grouped, for example "Read 1 file, ran 1 search". Click a group to see its steps. Each step names what the agent did and what it did it to, such as Ran bun test. A file step shows the file name, such as cart.ts, and a read can add the lines it read. A failed step has a failure icon.

An opened step list. A Thought row and a group of reads come first. The opened "Edited 1 file, ran 1 command" group shows a short diff of cart.ts, followed by "Ran bun test".

  • Click a step to see its output, such as a command's output and exit code, or the file content the agent read.
  • Click Open file on a step to open that file in the editor.
  • An edit shows a short preview of its diff. Click the preview to see the whole diff. Click the file name above it to open the file at the first changed line.
  • A Thought row holds the agent's reasoning and shows how many lines it has. Click it to read it.

Review changed files

When the agent changed files, started subagents, or left commands running, a small status bar floats at the top right of the chat. It has one part for each kind of activity, in this order: subagents, background commands, and changed files. Each part shows an icon and a count, and shows only while it has something to list. Hover a part to see what it counts, for example "3 changed files".

A part can also show a mark:

  • A spinner while a subagent or a background command runs.
  • A red dot when a subagent failed, or when Troupe could not read the Git status of the folder.
  • A blue dot on the changed files part when files changed since you last looked. It goes away when you open the list, or when the Changes panel shows the folder.

Click a part to open its list to the left of the bar. Only one list is open at a time. Click the part again or press Esc to close it. On a phone the list opens as a sheet from the bottom.

Open the changed files part to see the files this chat changed. In a Git repository, the list shows the files that this chat's edit, delete, and move steps touched and that are still uncommitted. Its header says how many, for example "This chat · 3 files, not committed", with the lines added and removed. Files you commit or discard drop out of the list. A file that only you, a terminal, or another chat changed is not listed. For a file this chat did change, the row and its line counts include all its uncommitted changes, also ones made outside this chat. A file the agent changed only through a shell command, for example a formatter, is not listed either. The Changes panel still shows it. Outside a Git repository, the list shows every file this chat touched, for example "This chat · 3 files edited".

The opened Changed files list shows cart.ts in the src folder, marked M, with 3 lines added and 2 removed.

The picture above was taken before the status bar moved to the top right of the chat. The rows now show a file icon instead of a letter.

Each row shows the file name, its folder, and how many lines were added and removed. A renamed file shows Renamed. Click a file to open it in the editor. A deleted file does not open.

The list shows the first 4 files. Click Show N more, or Open in Git, to open the Changes panel on the right with all of them. On a phone this opens the Git sheet. To stage and commit the changes, use the Changes panel. See Git.

To throw away this chat's changes, click Discard all. Troupe asks first, because new files are deleted and unstaged edits in these files go too, also your own. Staged changes stay. Click Discard N files to confirm, or Cancel. You can't undo a discard. Discard all is off while the agent works. It is not shown outside a Git repository, or when the server does not accept changes from your device.

A quick chat has no changed files part, because it has no project.

Follow subagents

Some agents hand part of a task to a helper agent, called a subagent. Subagents that the agent starts together show as one group in the turn. The group header shows up to three agent icons, the number of subagents, a summary such as "1 working · 1 done · 1 failed", and how long they have run. The group opens while a subagent runs and folds when all of them finish. You can fold it earlier by clicking the header. A single subagent shows as one card without the header. The group stays in the chat after the turn ends, and when you reopen the chat later.

Each row shows the subagent's title and type, its latest step while it runs or the first line of its result when it ends, its status (Working, Needs input when it waits for you, Done, Failed or Stopped) and its time.

Click a row to open the subagent. On a wide window it opens beside the chat; on a narrow one it slides in from the right, and on a phone it opens as a full-height sheet. The panel reads like the main chat:

  • The task, the prompt the agent gave the subagent, sits at the top as a message bubble. Click Show full prompt for the rest of a long prompt, and Show less to fold it again.
  • Below it, the subagent's messages and steps show in the order they happened, the same way as in the chat: a message, a folded group of steps such as "Read 1 file", then the next message. Click a group to see its steps. A subagent that started its own subagents shows them among its steps; click one to open it, and click Back to return.
  • When the subagent ends, a Result section shows its result only when the result adds something its last message doesn't already say. The section also shows an error if the subagent failed.

A subagent that the chat's agent runs itself, with its built-in subagent tool, shows the chat agent's icon. When its panel explains a missing section, it names that agent instead of "The agent". A subagent that the agent started from another of your agents, such as Codex, shows that agent's icon. Its row starts the detail line with the agent and the model, for example "Codex · gpt-5.5", and shows where it works, for example "worktree · troupe/sub/sub-9f2c1a3b4d5e". Its panel adds a Workspace section, and when it ends, a summary of what it changed. See Let an agent start other agents.

When the focus is on a control in the panel header, such as Back, use the up and down arrow keys to move between subagents. The keys don't switch subagents while the focus is in the panel body. From a group row, the keys stay within that group. When the agent did not report a section, the panel says why.

Stop one subagent

To stop one subagent and leave the others running, click Stop in its panel. On a phone, tap the stop button on its row in the Subagents sheet and confirm. The work it did so far is kept. Stop shows only while the subagent runs and when the agent supports it. To stop the whole turn instead, use the Stop button of the chat.

Subagents that run in the background

An agent can end its turn while subagents it started keep working in the background. The chat then keeps those subagents as running. When you hover the subagents part of the status bar, its text ends with "waiting on N background subagents". Its opened list shows "Waiting on N background subagents" above the subagent rows. While they run, Troupe does not switch the agent or the Claude account of the chat, because that would stop them.

Commands that run in the background

An agent can end its turn while a command it started keeps running. For example, Claude Code can run a command in the background, and Codex can leave a shell command running when it ends its turn. The step of that command reads Running in background with a timer that counts up until the command ends.

The background commands part of the status bar lists these commands. It shows how many commands there are, with a spinner while one runs. Click it to see each command, its state, and how long it ran. The list header says how many are running and how many have ended, for example "1 running, 2 ended". Click a row to go to the step of that command in the chat and open its output.

The agent gets the result of each command when the command ends. An ended command stays in the list until you send your next message. You can keep chatting while a command runs. Troupe does not switch the agent or the Claude account of the chat while a command runs, because that would stop the command. The one exception is a chat that hit a usage limit and continues on another Claude account: it continues at once and stops its commands. Reload this chat also stops it, so Troupe asks you to confirm the reload first. When you close the chat, its commands leave the list.

When you stop a reply, the list keeps only a command that had already started. A command that was waiting for your approval does not run, so it is not listed. When a reply fails or times out, Troupe lists no command for it.

Troupe shows background commands of agents that report when the command ends, such as Claude Code, Codex, and Factory Droid. The list has no stop button for a command.

See earlier steps of a long turn

A long turn shows only its latest 40 steps. Click Show N earlier steps at the top of the turn to load 50 more.

See images, videos, and audio

When the agent's reply includes an image, a video, or an audio clip, the chat shows it in the reply. This works in the desktop app and in a browser, for a file in your project, a file on the web, or one the agent sends inside its reply.

  • When the reply has two or more images in a row, they show as one row of square tiles. Scroll the row sideways to see them all.
  • Click an image or a tile to open it full screen. Images that a message sends together, and the tiles in one row, open in the same viewer.
  • In the viewer, swipe, use the left and right arrow buttons at the sides, or press the left and right arrow keys to move between images. The bottom of the viewer shows where you are, for example "2 of 5". A single image shows no count.
  • Click the download button at the top right of the viewer to save the image. In the desktop app, a save dialog opens. In a browser, the browser downloads the file. For an image from the web that Troupe cannot save, Troupe asks to open the image link in your browser instead. If the download fails, Troupe shows "Failed to download image".
  • To close the viewer, click the X at the top left, click outside the image, or press Esc. On a phone, the back button also closes it.
  • A video from your project first shows its file name. Click it to load the video, and it starts playing. A video from the web shows a player right away.
  • An audio clip shows a player.
  • When the agent links to an image or video file in the middle of a sentence, the link stays in the sentence, and the preview shows below that paragraph.

A video can be up to 100 MB. If a video is too large or cannot be read, Troupe says why in its place, for example "clip.mp4 is too large to play inline (128.0 MB, limit 100.0 MB)". If an image cannot be shown, its name shows instead.

Queue prompts

You do not have to wait for the agent to finish. Type your next message and press Enter. While the agent works, the send button becomes Queue message, and your message goes into a queue at the top of the message box.

Troupe sends the queued messages in order, from top to bottom, one each time the agent finishes a turn. The queue header shows how many messages wait, for example 2 queued. Click it to collapse or expand the list. On a phone the list starts collapsed. Click the × button in the header (Clear queue) to remove all queued messages.

Hover a queued message to see its buttons. On a touch screen, the buttons are always visible. Each queued message has these buttons:

  • Steer stops the agent's current turn and sends this message right away.
  • Edit takes the message out of the queue and puts its text in the message box, after anything you already typed there. A message with files or images has no Edit button.
  • Remove from queue, the trash button, deletes the message without sending it.

To change the order, drag a message by its handle on the left. You can also focus the handle and press the up and down arrow keys. Troupe announces the new place, for example "Moved to place 1 of 3". The handle shows when two or more messages wait.

When the agent works after its turn

An agent can keep working after its turn ends. For example, Claude Code starts a turn of its own when a background command or a subagent finishes. While that happens, the chat shows the same working cue and the same stop button as during a turn. A message you send goes into the queue and goes out when the work ends. Troupe counts the work as ended when no step is still running and the agent sends nothing for 30 seconds.

Stop a turn

To stop the agent, do one of these:

  • Press Esc in the message box.
  • Click Cancel turn, the square button. It takes the place of the send button while the agent works and the message box is empty.

Troupe tells the agent to stop. Any approval request or question that was waiting is cancelled. If you have queued messages, Troupe sends the next one.

This also works while the agent works after its turn. Troupe tells the agent to stop that work. If no messages are queued, the chat shows it as idle right away. If messages are queued, Troupe sends the next one, and the chat shows as working again.

Edit or retry a message

Hover a message to see its buttons. The last message in the chat always shows them.

  • Edit, on your message, puts its text back in the message box. Change it and send it as a new message. The earlier messages and replies stay in the chat.
  • Retry, on a finished reply, sends your last message again as a new turn. The old reply stays above the new one. Retry is hidden while the agent works.
  • Copy copies the message. On the agent's reply, it copies everything the agent wrote in that turn.

Your message, "Try a discount parameter on cartTotal", with the Edit and Copy buttons below it.

If the chat stopped, Retry starts the agent again before it sends your message. When a turn ends with an error, a notice at the top of the message box says "The last turn did not finish." and shows the error. It can have its own Retry or Resume chat button. When the error has a code or data, click Details to see them. Click the × button to dismiss the notice.

Code blocks in a reply have their own Copy and Download buttons. Tables have copy, download, and View fullscreen buttons.

When you click a link in a reply or in a Thought, Troupe asks first. The Open external link? dialog shows the full address. Click Open to open it, or Cancel to stay in the chat. Pressing Esc also cancels. Clicking outside the dialog does not close it. A link to a heading or footnote in the same reply, one that starts with #, scrolls to it without asking.

On this page