---
title: "Review and approve work"
description: "Follow agent activity, inspect changes, and resolve approvals within the thread."
canonical_url: "https://hoplite.sh/docs/threads/review"
markdown_url: "https://hoplite.sh/docs/threads/review.md"
---

# Review and approve work
URL: /docs/threads/review
LLM index: /llms.txt
Description: Follow agent activity, inspect changes, and resolve approvals within the thread.
Related: /docs/threads/run, /docs/agent/tools, /docs/sandboxes/previews, /docs/github

# Review and approve work

Follow agent activity, inspect code changes, and direct the next step from the thread. Tool activity streams into the timeline, and actions that require approval wait for a decision before proceeding.

## The activity timeline

While the agent works, the working header shows its latest thought and a scrolling live feed of narration, file reads, and commands. When the run ends, the header folds into **Worked for**. On mobile, live progress and elapsed time stay in one activity dropdown, collapsed until you open it.

Open the activity record to inspect the run's thoughts and steps. File reads appear as chips, and commands show their targets. Select a command to expand its input and output, or a file chip to inspect what was read. This keeps the overview compact without hiding the detail needed to review the work.

When you return to a thread or a previously hidden tab, a **New** divider marks the first transcript activity that arrived while you were away.

## Approvals

Cloud runs automatically approve tool calls and do not pause for a tool-approval response. Command policy, protected-path checks, repository scope, integration write settings, and workspace policy still apply; auto-approval cannot turn a blocked operation into an allowed one. Public media sharing also requires workspace policy permission.

Local CLI and desktop runs retain their approval flow. When an action requires approval, approve or deny it in the inline bar. Approvals left by ended runs are cleared, and a response to a stale approval is rejected rather than resuming finished work. See [Tools](/docs/agent/tools#permissions-and-approvals) for the current boundaries.

## Questions

When an agent needs structured input, its question appears in the same pending
action bar. A request with several questions shows one question at a time behind
tabs labelled with each question's header, and answered tabs are marked with a
check. Picking a single-choice option moves on to the next unanswered question,
and **Enter** submits once every question is answered. You can type your own
answer in the **Other…** field, or select **Dismiss** to decline the whole request.
Answers are persisted before execution resumes, so the same response reaches
both cloud and local-agent runs without being lost to a refresh or worker retry.

On iOS, a notification for a single pending question can offer **Answer**, so you can reply with free text without opening the app. Multiple questions open the thread instead. If AI data-sharing consent is missing, Answer opens the app to request it first. Responses are bound to the pending question, so stale notifications cannot answer a different question.

## Agent tasks

When Agent tasks are enabled for your workspace, a progress indicator and a collapsible **Agent tasks** dock appear above the composer. They summarize the agent's planned work and show whether individual tasks are pending, in progress, completed, or cancelled. The overall task list can also be marked interrupted.



## Subagents

When a run delegates work, the **Subagents** control opens a durable panel for its child agents. The panel separates active work from recent results and reports whether each child is running, needs approval, completed, failed, or was interrupted. Selecting a child shows its tool activity in the same collapsible format as the main timeline, refreshed while it runs, followed by its summary once available; the delegation prompt the parent wrote is not shown.

A child that needs approval exposes the same review path in the panel, so you can resolve its blocked action without hunting through the timeline.

## References in messages

When an agent cites a GitHub pull request, commit, Linear issue, source location, or task, Hoplite renders it as an inline reference link. References that Hoplite can resolve show live metadata on hover; unresolved links remain ordinary links so the message stays readable.

## Review the diff

In a cloud thread, open the **Diff** tab (or add it from the **+** tab menu). The cloud diff compares the thread workspace against its base branch. Its folder tree lets you filter and jump between changed files, while the patches appear in one continuous scroll. Collapse or mark files as viewed, and use **[** and **]** to step between files. Comment on a line or range, then send your comments to the agent together. Review the patches before creating or merging a PR; use the activity timeline to inspect test results.

If a file's patch cannot be displayed, that file shows an unavailable-patch notice; the other files remain reviewable.

Send a follow-up in the same thread when you need revisions, then reopen the diff to review the updated result.

The desktop diff view includes a changed-files tree with file, addition, and deletion counts. Select a file to jump to its patch, drag the divider to resize the tree, or use the button beside **Changed files** to hide and restore it. Hoplite remembers the layout for that workspace; smaller viewports show patches at full width without the tree.

Cloud threads open Chat, Diff, pull requests, previews, Terminal, and Subagents as tabs. On desktop, split panes right or down, drag tabs between panes, and resize the dividers; phones use one pane with a scrolling tab strip. The layout is remembered in your browser. **Cmd+B** on macOS or **Ctrl+B** elsewhere switches between Chat and the secondary view without changing the main application navigation.

## Rate an agent response

Use the thumbs-up or thumbs-down control on an assistant reply to save a private rating. You can add reason chips and an optional comment, change the rating later, or remove it; your choice persists across reloads and is not shared with other workspace members. Hoplite attributes the response to its producing model server-side so the team can triage the right behavior.

## Previews and terminal

When the agent starts your app, a **Preview panel** shows it live inside the thread, and a **Terminal panel** gives you a shell into the sandbox. For UI changes, the agent can attach an ordered checklist to guide your review. See [Previews](/docs/sandboxes/previews).
