Fix a pull request with AI Fix

AI Fix turns review findings, open threads, failing checks and merge conflicts into a change you can push. An agent (Claude Code or Codex) works in a sandbox on your computer, with a copy of the code, and can run the project's own tests. You watch it work, read what it changed, and push the result to the pull request's branch. Nothing leaves your computer until you press Push.

#Before you start

• An AI review. A fix starts from a finished AI review of the pull request. With none, the tab says so and offers Open AI Review. The exception: a finished CI check can start a fix of its failures on its own. See Choose what to fix.
• A container runtime, such as Docker or Colima, on this computer. With none, the tab shows one line with How to set up. See Docker and microsandbox.
• AI signed in: Claude Code or Codex. See Settings → AI.

#Start a fix

1. Open the pull request and choose the AI Fix tab. You can also press Generate fix from this review on the AI Review tab, or Fix with AI beside a CI check.
2. Under What to fix, tick the items the agent should work on.
3. Pick the model from the Model menu.
4. Optionally tick Plan first to talk the fix through before any code changes. See Plan first.
5. Press Start fix. It starts at once.

#Watch it work

⏺ Read(refunds/policy.py)
⏺ Edit(refunds/policy.py)
⏺ Bash(pytest tests/refunds)
  ⎿  42 passed

refunds/policy.py-    return order.status == "paid"+    if now() - order.paid_at > REFUND_WINDOW:+        return False+    return order.status == "paid"

The tab first says Setting up the sandbox., then The agent is fixing and testing. The conversation shows each step: the files it reads and edits and the commands it runs. While it works you can:

• Press Stop to stop the current turn. The sandbox stays, with the work so far.
• Ask for more in the message box. Each message is another build turn in the same session; the agent may edit and test again, and rewrites its report for the whole change.
• Open the Network, Services and Spend panels to see what the sandbox reached, which services it started and what the session has used.

#Read the result

When it is done, the tab says Fixed. Review the changes and push them, or ask for more. Above the diff, the agent's report lists, for each changed file, what it changed and which items it covers (each item is a chip, such as a finding). Then:

• Not addressed: items the agent left alone, each with its reason.
• Left out: edits it made in other repositories of the workspace. Only the pull request's repository is ever pushed.
• No report, when the agent did not write one. The diff still counts.

#Push or discard

Push sends the change, as one commit with the agent's commit message, to the pull request's branch, or to a new branch with a new pull request. Discard removes the sandbox and pushes nothing. See Push the fix, and auto fix.

#The sandbox

The sandbox is a container on your computer, made for this fix and removed after it.

• It holds the pull request's repository, at the pull request's latest commit, and the repositories of the other pull requests of its ticket, read only. Other workspace repositories are not loaded.
• The agent has a shell there, so it can run the project's own tests, lint and type checks, and start the services you picked for integration tests (Settings → Sandbox).
• It has no GitHub credentials and no way to change your own files. The only way out to the network is Limn's allow-list. A blocked host shows in the Network panel, where you can allow it.
• A session with nothing to do closes after an hour. Its changes are kept, and Resume starts it again.

Sharing the computer with Implement story

AI Fix and Implement story use the same sandboxes: up to four open sessions on this computer, and up to three building at once. One fix runs per pull request. When every sandbox is busy, a fix waits in line and starts by itself when one frees; the AI Fix tab shows its place, with Cancel. To make room, a fix session of yours that has done nothing for a minute is closed; its changes are kept, and Resume goes on. See AI and the sandbox for the full safety story.