Review finished work

When an agent hands a task back, it sends a report: what it changed, its answer to each item of the task's Done when list, screenshots, and how to check the result. The session is then finished, waiting for you and the task is In review, as they are when a session is marked finished. The agent never marks its own task done: that's yours to do after looking at the result.

Tasks that wait for this one (dependencies) don't wait for your review: a task in review counts as finished for them, unless the agent handed it back partly done or blocked (below).

Where finished sessions show up

A session that is still running can need you too, when its agent waits for you in its terminal or asks you a question: it shows in the task's details, the task list and Sessions with a navy dot, and the desktop app notifies. See When the agent waits for you.

The report

The report shows in the task's details, at the top of the panel, and in Sessions when the session is selected. From top to bottom, with the parts the agent left empty left out:

An item added to the Done when list, or reworded, after the report shows as Not answered.

Select an image to open it in the viewer. ← and → move between the images, a click on the image shows it at its actual size, Open opens it on its own in the browser, and Esc closes the viewer. Images the agent attaches while it's still working show under Images so far, and join its report when it hands the task back. PacedMind keeps its own copy of each image in its data folder. See Where your data lives.

Every hand-back adds a report. With more than one, the arrows next to Report 2 of 2 page back to the earlier ones.

Files changed

The report says what the agent did; Files changed shows what actually changed, so you can compare the two. When a session starts on this computer in a folder that's in a git repository, PacedMind notes the commit it's on, whether it has changes that aren't committed, and which files git doesn't track. When the agent hands the task back, PacedMind compares the folder with that and shows:

• a count, such as since the session started · 3 files · +120 −14 · 2 commits;
• the commits made since, newest first;
• each changed file with its lines added and removed, marked A (added), M (changed), D (deleted), R (renamed, with its old path) or N (new, not tracked by git); binary for a file without lines. Up to 8 show, with Show all … files;
• Show the diff: the full diff, changes that aren't committed included, and new files git doesn't track. Added lines are raised, removed ones dimmed.

Nothing changed in git since the session started. means just that, whatever the report says. When the folder already had changes that weren't committed at the start, the report says so: they count too. Files git didn't track that were already there, unchanged, are left out. Every report of a session compares with the session's start, also after Request changes.

Only the summary goes with the report, to your other computers and the web app. The full diff stays in PacedMind's data folder on the computer the session ran on, up to 1 MB (The diff was longer than PacedMind keeps (1 MB), so it's cut short.); elsewhere, Show the diff says The diff itself stays on the computer the session ran on.

There's no Files changed for sessions in the cloud or started outside PacedMind, for a folder that isn't in a git repository, or for sessions that started before this version. It says why when PacedMind couldn't look:

• No diff: PacedMind doesn't read folders in Desktop, Documents, Downloads or a cloud drive, where macOS asks first. Move the project to another folder, such as one in your home folder, to get diffs.
• No diff: git wasn't there to compare the folder with. Install git: on macOS, Apple's command line tools (xcode-select --install) or Homebrew's git.

A session without a report shows the agent's note instead, if it left one. That's the case for sessions marked finished with Mark finished, for Codex cloud tasks, and for hand-backs from before reports existed.

Review a session

1. Open the session: from the notification, from Sessions, or from the task's details.
2. Read the report: the summary, the Done when answers, the images and the questions.
3. Check the work, starting with the report's How to check it. The agent's terminal or app stays open, and the changed files are in the session's folder. For a cloud session, open its pull request.
4. Decide:
   • Mark done when the result is right.
   • Request changes when it isn't, for a session that ran in a terminal on this computer (below). You can also continue the conversation in the agent's terminal or app yourself: the agent hands the task back again, with a new report, when it's done. If it picks the task up again with start_task first, the session shows as Running in the meantime.
   • To start over with a fresh agent: for a running session, select Close session, then New session. For a finished session, select New session with Claude Code (or Codex) in the task's details; the finished one keeps waiting for review until the task is marked done.

Request changes

Request changes, under the report in the task's details or in Sessions, sends the work back to the agent with what you want changed. It's offered on the task's latest report, for a session that ran in a terminal on this computer, and still after the task was marked done.

1. Select Request changes.
2. Write what should change in the field (What should change? Be as specific as you'd be with a colleague.). The agent reads it as written, so answer its questions there too.
3. Select Send to Claude Code (or Send to Codex), or press Ctrl+Enter. Cancel or Esc drops the text.

PacedMind adds your changes to the report, under You asked for changes, and opens a new terminal in the task's folder: a new tab in Windows Terminal, a new window otherwise. Sent to Claude Code in a new terminal confirms it. Then:

• The agent starts a new conversation, with its last report and your changes, as the line under the box says (Claude Code starts again in a new terminal, with its last report and your changes.). Claude Code doesn't go back into its earlier conversation, so that PacedMind's tools are connected when it reads your changes. The old terminal can stay open or be closed.
• The task goes back to In progress, from Done too, and the session to Running: Claude Code is working on your changes since 14:05. The earlier report folds away under The report from 25 Sep 12:22.
• The agent reads your changes first, makes them without starting the task over, and hands the task back with a new report whose summary starts with what it changed. You're notified again. The earlier report keeps your changes.

For a session in the Claude or Codex app, or in the cloud, write the changes to the agent where it runs. A session that ran on another computer takes its changes from PacedMind on that computer. An agent you're talking to can send changes the same way, with the MCP tool request_changes, when you ask it to.

Answer the agent's questions

When the agent's latest report has questions for you, Answer its question (or Answer its questions) appears next to Request changes, wherever that is offered. It sends your answers back to the agent the same way as changes.

1. Select Answer its question.
2. Write an answer under each question, in the fields marked Your answer. A question you leave empty is left out. The last field, Anything else it should know (optional), takes anything else.
3. Select Send to Claude Code (or Send to Codex), or press Ctrl+Enter, once at least one answer is written. Cancel or Esc drops the answers.

The line under the fields says Claude Code starts again in a new terminal, with its last report and your answers. PacedMind sends your answers as one message, which starts with My answers to your questions: and lists each question you answered with your answer. From there it goes as with Request changes: the report shows your answers under You answered its questions, the session's history records You answered its questions: and the start of your answers, and the session runs again with Claude Code is working with your answers since 14:05. The agent reads your answers first when it calls start_task, goes on with the task using them, and hands it back with a new report.

While a session runs, the agent can also ask you something in its terminal, or with report_progress: answer it there. A question it asks with ask_user waits for your answer in PacedMind, on the task and in Sessions: see Answer while it works. See When the agent waits for you.

Partly done or blocked

An agent that can't finish everything still hands the task back, and says so in its report:

The notification reads DEV-21 is partly done or DEV-21 is blocked, and the row in Sessions partly done 12:22 or stuck 12:22. The session stays in Finished, waiting for you.

A partly done or blocked hand-back holds up what comes after it: the tasks that wait for this one (dependencies) keep waiting until you mark the task done, the agent hands it back again as done (for example after Request changes), or a newer session of the task is marked finished. Meanwhile the Roadmap says When you mark DEV-21 done for them, the Timeline Waits for you, and get_next_task doesn't offer them.

What Mark done does

Mark done, in the task's details or in Sessions, marks the task Done. Setting the status to Done any other way, including by the status icon in a list, has the same effect. Then:

• the task's finished sessions become Done (You marked it done) and leave the waiting count;
• the check block leaves the week plan;
• after a partly done or blocked hand-back, the tasks that wait for this one stop waiting for it.

Nothing starts by itself: start the next task when you're ready, for example with Start session next to Waits for it: in Sessions. Request changes stays available afterwards, and moves the task back to In progress.

Follow-up work

Agents record work they found outside the task as new tasks in the same project, and list them in their report, where each one opens its task. They appear in the project's task list.