Following the work
Read what an agent is doing while it works, see what it changed, and undo a turn.
Each turn's work sits between your message and the agent's reply. While the agent works, you see the step it's on. When it finishes, the work folds into one row that says what the turn came to, and a card lists the files it changed.
While the agent works
Each turn has a head with the agent's name and model. While the turn runs, the head says working with a clock, such as working 12s, and Writing once the reply starts. Below it, the agent's work reads as steps in one card. Click the row under the head to fold the steps away. The step in progress then stays visible as a single line. To end the turn early, click Stop in the composer.
A step starts when the agent says what it's about to do, and takes its title from those words. With Claude Code, that's the sentence it writes before a batch of tool calls. With Codex, it's usually the title of its reasoning. When the agent says nothing first, the step is named after what it did, such as Read 3 files, ran 2 commands.
The current step shows a spinner and what it's doing now. A finished step gets a check, and a failed one a cross. A finished step folds to its title and what it came to, such as Opened #218 or 1 failed, or a hint of what's inside, such as 6 files · 2 commands.
Inside a step, each thing the agent worked on gets a row, with a verb for what it did:
| Verb | What the row shows |
|---|---|
| Read or Listed | The files the agent read or the folders it listed, each with its folder. Past six, the rest wait behind a count. |
| Edited or Wrote | The file with the lines added and removed, and under it the lines the edit took out and put in. Wrote marks a new file. |
| Searched | What the agent looked for, then the first lines it found. Each line opens its file at that line. A web search lists the pages it found. |
| Ran | The command, by the agent's own words for it or its first line. While it runs, or once it fails, its newest output lines show under it. |
| Tested | A test run as one cell per test, with the count passed and failed, what ran, and how long it took. A failing test is named under it. OpenOrc reads the output of Vitest, Jest, pytest, Go, and Cargo. |
| Committed, Pushed, or Opened | Git work as the branch it happened on: the commits as dots on a line from the base, the push to origin filling the line as it sends, and the pull request it opened. |
Other rows appear when the agent does other things, such as Read for a web page it fetched, or the name of an MCP server whose tools it used. Subagents get a card of their own. See Subagents.
When the agent keeps a to-do list, the card's head reads Plan with the step it's on, such as Step 2 of 5, over a progress bar, and the items check off as the agent works through them. Before the first step, the live line reads Thinking while the agent reasons.
Open a step
Click a step's title to open or fold it. Inside, each row opens what's behind it:
- A file the agent read opens beside the conversation.
- A search result opens its file at the line it matched.
- A verb, such as Ran or Edited, lists the calls behind the row. A command shows the folder it ran in and its output.
In an open call, Raw call shows the exact input and output. When a tool's output was very large, OpenOrc keeps a short digest of it, so the row can still say what the call came to after the full output is gone from the page.
When a turn finishes
The head says how long the turn took, or Failed after or Stopped after and the time. Under it, one row says what the turn came to, and the steps fold away. Click the row to open them again.
| Item | What it says |
|---|---|
| Thought 8s | How long the agent spent reasoning, when it shared its reasoning. |
| #218 open | The pull request the turn opened, or how its last commit or push went. |
| +21 −8 in 3 files | The lines added and removed, and how many files changed. |
| 12 passed | The last test run, counted when the test runner printed the counts. Another check, such as typecheck failed, shows when there was no counted test run or the check failed. |
| 2 agents or 3 sources | For a turn that changed nothing, the subagents it ran and the web pages it read. |
An item with nothing to report is left out. A turn with nothing else to report says how many steps it took. The sidebar gives you the same news without opening the thread. A thread's row shows the step its agent is on, Needs approval with the command it's waiting on, or Has a question for you. A finished turn you haven't read shows what it came to, such as Done · 2 files changed · tests passed, until OpenOrc closes the idle agent.
A reply can show images, such as a screenshot the agent took, and they open full size when you click them. It can also hold a visualization: a diagram, chart, or small interactive explanation that the agent draws with OpenOrc's visualize tool. A visualization runs in a sandbox of its own with no network access, and stays in the conversation's history.
Review or undo a turn
A turn that changed files ends with a change card. It reads Changed 3 files, with the lines added and removed. Click the chevron to list the files.
- Click a file, or Review for all of them, to read the turn's diff in the Changes tab.
- Click Undo, then Confirm undo within four seconds, to put the files back as they were before the turn.
Undo is unavailable while the agent is working, and for older turns that have no starting checkpoint. In a folder that isn't a Git repository, turns get no change card at all.
When files change between turns, a line between the turns says so. It reads You changed 2 files when you made the edits, or 2 files changed between turns when OpenOrc can't tell who did. It has its own Undo, which puts the files back as they were after the last turn.
Undo rolls back the whole folder
Undo returns the thread's folder to how it was before that turn. Every change made since is undone and new files are deleted, including later turns, your own edits, and work by other threads in the same folder. Files Git ignores are left alone. OpenOrc first saves the files as a checkpoint in the Checkpoints tab, so you can bring them back with Restore.Questions and approvals
An agent can stop to ask you something. The question card shows its options, or a text field when there are none.
- Pick an option. When the card has a Write a different answer… field, you can type your own answer there instead.
- Click Answer. Decline closes the question without an answer.
A question waits for you in every mode, Autonomous included. When it asks for something secret, the field hides what you type, and the card shows Answer hidden afterwards.
An approval card looks similar. It asks whether to Allow a tool and offers Allow, Allow for this run and Deny, or only Allow and Deny for a request that can't be allowed for longer, such as a password. Which requests ask first depends on the thread's mode. See Modes and permissions.
Both cards appear under the turn's work, and the turn waits until you answer.
The Inbox
The Inbox collects what's waiting on you across every project. Open it with Inbox on the sidebar's rail, under your projects, or choose Inbox in the command palette (⌘K). The number on the button counts what's waiting.
| Item | What the row says |
|---|---|
| An approval or a question | The request itself, such as the command or the question. |
| A task an agent proposed | Proposed by the agent, waiting for your go-ahead |
| A task whose run finished | Finished run waiting for your review |
Click a row to open its thread or task. When nothing is waiting, the Inbox says Inbox zero.
Next steps
Checked against OpenOrc's source on November 9, 2026.