Skip to content
track

Follow-Ups

Use a follow-up when you want the agent to continue existing work rather than start from a clean slate.

Common examples:

  • Address the PR review comments.
  • Rework this to avoid cloning the config.
  • Add regression tests for the failing edge case.

What a follow-up reuses

Section titled “What a follow-up reuses”

If the latest task run already has a pull request, the follow-up continues on that same PR and branch.

If there is no PR yet, the follow-up still reuses the existing remote branch and worktree.

Automatic review follow-ups

Section titled “Automatic review follow-ups”

You can also let track create follow-ups automatically after PR review activity.

To enable that flow, open Finish Runner Setup and set both:

  • Main GitHub user
  • Review follow-up

Once those are configured, track watches for review feedback from that specific GitHub account and dispatches the next follow-up for you automatically. That means you do not have to reopen the task and manually start a new follow-up every time you leave review comments on the PR.

Automatic follow-up only listens to reviews from the configured main account.

That restriction is intentional:

  • it avoids turning arbitrary reviewers into a control channel for the agent
  • it keeps the automation scoped to the person who understands that review comments may be implemented directly

This matters for both safety and workflow clarity. A malicious reviewer should not be able to steer the agent, and an external reviewer may leave comments that are meant for a human to interpret rather than for an autonomous agent to apply blindly.

If somebody else leaves useful feedback, you can still read it yourself and start a manual follow-up with the instructions you actually want to send. Approved reviews are ignored.

Tool choice stays pinned

Section titled “Tool choice stays pinned”

Follow-ups do not switch runners midway through a thread of work. If the original run used Claude, the follow-up stays on Claude. If it used Codex, it stays on Codex.

That keeps branch history, worktree state, and prompts consistent.

Start fresh when you really mean start fresh

Section titled “Start fresh when you really mean start fresh”

If you want to throw away the old remote context and begin again:

  • discard the task’s run history, then dispatch again
  • or create a brand-new task if the work has drifted into a different problem

Use follow-ups for continuation, not for resets.