Review and recover task work

When a coding agent submits a work run, the task enters Reviewing. Open Review from the task to read the change summary, inspect any linked code review, and verify the local Git changes against the task details. The AI hand-off describes the completed work and where the agent stopped so a future coding agent can continue if you request more work. The resolved coding-agent command is also available for reference.

The review header uses a subtle task-type icon, including when both types share Reviewing or Done. Hover over or focus the icon to read Coding Agent Task or Person Task.

For a Person task in In progress, open task details and choose Submit for review. Use the larger Latest changes editor to describe what changed, how to verify it manually, and any remaining issues or limitations. You can use Markdown headings and lists. Optionally add an absolute HTTP or HTTPS Code review URL, then submit. Empty changes and invalid links cannot be submitted. The task enters Reviewing only after the submission succeeds.

The Person review workspace shows Latest changes, any code-review link, exact lifecycle timestamps, and review comments. It has no AI hand-off, coding-agent command, listener approval, or agent recovery controls. Current review comments follow the same author and project-moderator editing permissions as Coding-agent reviews.

• Choose Mark done to accept the submitted work.
• Choose Return to work to start a linked continuation for the task’s current assignee, shown in the dialog. To change who does the work, edit the task assignment first. An unassigned task or an assignee who is no longer a project member must be assigned to a current member before returning it to work. Incomplete dependencies require acknowledgement. The previous submission, comments, and review resolution remain in history.

You can also choose Mark done directly from an In progress Person task. This records completion without creating a submission or a review interval.

If the task or active run changed while a dialog was open, close it, refresh, and review the current state before retrying. If the action succeeded but another view could not refresh, refresh the page before taking another action.

Open a Reviewing Coding-agent task to enter its visually distinct Review workspace and inspect the change summary, AI hand-off, code-review link, command, and review comments. A prominent purple review header carries the Review workspace label and review icon, distinguishing this decision point from ordinary task details before you begin reading. You can then:

• Mark done to accept the current result.
• Request follow-up work after adding at least one review comment. The dialog optionally offers Check and update review branch, initially unchecked. Leave it unchecked to send your existing feedback as usual. Selecting it shows editable guidance that will be saved as another review comment when you confirm.
• In Review actions, open the three-dot More actions menu and choose Check and update review branch to request a branch check even without existing review comments. It opens the same follow-up dialog with the branch-check option selected and required. Review or edit the guidance and choose a coding agent before confirming.
• Start fresh to choose a coding agent and queue a clean, unlinked run. The reviewed run stays in review history, but its feedback is not continuation context for the new run. The new work-run context explicitly tells the coding agent to restart from the current task details and general comments instead of prior review or agent output.

Current review comments can be edited or deleted by their author and by people with project moderation permission. Follow-up work requires a completed review run with a change summary and AI hand-off, plus either existing review feedback or the branch-check guidance you confirm. The guidance and linked follow-up are saved together; canceling or a failed confirmation adds no comment or queued work. Existing feedback remains intact, and the selected coding agent receives it alongside the guidance, prior summary, AI hand-off, and linked code-review URL. Both follow-up work and Start fresh ask you to choose a coding agent and move the task to Queued.

Both branch-check entry points begin with the project's effective default comment: its saved Markdown when configured, or the platform default otherwise. Organization owners, organization admins, and project leads can configure review branch guidance. Review or edit the suggested comment for this request before confirming. A project default change leaves an already-open dialog's draft and your edits intact; newly opened dialogs use the updated default. It does not rewrite saved review comments, earlier work-run context, or review history.

The platform default branch-check guidance asks the agent to verify the existing task worktree and branch, then merge the latest available main into that task branch. With a configured remote, it fetches and uses the current remote main according to repository guidance; without a remote, it uses local main. The agent resolves conflicts while preserving intended task changes, runs affected validation, and commits the result. It pushes the existing task branch only when a remote is configured and the repository workflow calls for it. A linked code-review URL is not required for this local update. Applicable review feedback still accompanies the merge instruction. If main is already included, or the worktree, task branch, or main cannot be verified, the agent reports that outcome. The reviewer retains the final merge into main; this action does not authorize merging the task branch into main or pushing main.

If a task was moved to Reviewing without a completed work run, Start fresh still queues a clean run.

To stop pursuing a task without accepting its result, open its Edit page and choose Close beside Delete. Closing preserves work runs, summaries, hand-offs, comments, and review evidence as read-only history. Person tasks also offer Close on Edit; use Mark done or Return to work when that work should continue through its normal review lifecycle.

A Closed task's review-history header offers Reopen. Confirming it returns the task to Backlog and then opens task details; it does not restart the old review or queue new work.

On a Done Coding-agent task, View code changes appears as a right-aligned button on a separate row below View review history and Edit when a work run recorded a code-review URL. It opens the most recent recorded URL in a new tab, even when a newer run omitted a URL, and leaves the task Done. Earlier links remain available in review history.

Open the task details while it is in Agent working and find Recover agent work beneath the task header. Use this section when a coding-agent process stopped, when it finished without submitting its work, or when you moved a task directly to Agent working and now want alterac.ai to queue it. The section remains available whether or not a listener or active work-run record exists:

• Start fresh: choose a coding agent and queue a clean run without prior-run linkage or a continuation comment. If an incomplete run exists, alterac.ai removes it from visible history. The replacement receives an explicit fresh-start instruction and should rely only on the task's current details and general comments.
• Requeue and continue: choose whether the agent completed work without submitting it, its terminal closed before completion, or something else happened. Add a custom description for the third choice, choose a coding agent, and queue a run that receives the selected recovery details. If an incomplete run exists, alterac.ai retains it as continued after a problem and links the replacement to it. Use this when the next agent needs to know what happened or what useful work to preserve.

The selected recovery details are saved as a continuation comment with the task and included when the replacement run starts. A retained run marked Continued after a problem shows who recovered it and when. It is excluded from duration, concurrency, and other work-run metrics. If the task has no work-run record because its state was changed directly, there is no old run to remove, retain, link, or exclude; the difference is whether the replacement receives a continuation comment.

Late output from the replaced run is rejected, so it cannot move the task into review or overwrite the replacement run.

Both recovery paths require an explicit coding-agent selection. Review the command preview and every selector before confirming. alterac.ai does not automatically reuse the previous run's coding agent or selector choices.

If the task or active-run presence changed after you opened the dialog, refresh the task and repeat the action against the current state.

Use complete your first task for the full first-run path, including the listener and queueing steps.