Markdown edition

# Run Claude Code Ultrareview on one branch and verify its findings

> A documentation-based guide to checking repository upload and credits before one Claude Code cloud review, then verifying its findings against the branch before a human merge decision.

By BIG CHANGE Editorial

Published: 2026-10-20T11:00:41.997Z
Updated: 2026-10-20T11:00:41.997Z
Canonical: https://bigchange.ai/blog/claude-code-ultrareview-branch-findings-guide

![Conceptual charcoal illustration of an orange network cable connecting a desktop computer tower to a wall port.](https://bigchange.ai/api/media/file/ultrareview-repository-upload-boundary-hero-v2.png)
AI-generated conceptual illustration by BIG CHANGE.

Claude Code's Ultrareview can examine an existing Git branch in Anthropic's cloud before a human decides whether to merge it. The useful question is broader than whether the review finds a bug: which repository material will leave your machine, what will the run cost, and can you verify each finding against the code? This guide is for someone authorized to review a substantial branch and upload the repository under their organization's rules. It follows [Anthropic's current research-preview manual](https://code.claude.com/docs/en/ultrareview); BIG CHANGE did not run the command or test its findings.

## The big change

- **What changed:** `/code-review ultra` offers a cloud review of the current branch against its default branch, or against a base you name, including staged and uncommitted changes.
- **Why it matters:** The review can return file-specific findings for an existing change, but the branch route also uploads repository state. The visible diff alone does not define what leaves the machine.
- **What to watch:** Availability, upload scope, one-time free runs and usage-credit estimates can vary. A finding still needs local verification, and a review with no findings does not prove the branch is free of bugs.

### 1. Establish the branch and its original checks

Start in the repository that holds the change. Record the current branch, intended base, acceptance criteria, and the project's existing test or build command. Inspect the full diff, including staged and uncommitted tracked edits. The command without a base compares the current branch with the default branch; an alternate base changes that comparison. Keep the original check result so you can run the same check again after fixing any finding.

Confirm that this is the branch you mean to review. An empty comparison is refused. A first commit or a branch without a merge base can turn the target into every tracked file, subject to the manual's confirmation and limits. If that is not the intended review, correct the branch or base before launching. The current documented default limit is 500 changed files and 8,000 changed lines; the actual refusal reports the limits in force.

### 2. Decide whether the repository may go to the cloud

[Anthropic's upload rules](https://code.claude.com/docs/en/claude-code-on-the-web#send-local-repositories-without-github) describe a full clone bundle as repository history across **all branches**, plus uncommitted changes to tracked files. Untracked files are excluded unless deliberately added. On native Windows, uncommitted tracked edits upload even when their filenames resemble credentials. On macOS, Linux and WSL, certain sensitive uncommitted edits are omitted, but committed material remains in the history. Inspect the repository history and tracked edits under your code policy; if either contains material you lack authority to upload, preserve it and stop. A small review diff does not settle this decision.

The documented bundle limit is under 100 MB. For a larger bundle, Claude Code can fall back to the current branch and then a squashed working-tree snapshot, or fail. Check the actual scope before accepting a fallback. Reviewing a `github.com` pull request is a separate route: the cloud sandbox clones the PR through the connected GitHub account rather than uploading the local working tree.

Check account eligibility as well. Ultrareview needs a `claude.ai` sign-in, not API-key-only authentication. The current manual excludes Bedrock, Google Cloud Agent Platform, Microsoft Foundry, Zero Data Retention and HIPAA configurations; customer-managed encryption keys also block it. In some excluded cases, the command performs a **local** review instead. Label that result accurately if cloud access is unavailable. The separate [local `/code-review` workflow](https://code.claude.com/docs/en/code-review-local) is available for an authorized local review.

### 3. Read the launch dialog, then authorize one run

In an interactive Claude Code session on the reviewed branch, enter:

```text
/code-review ultra
```

For a reviewed alternate base such as `develop`, enter:

```text
/code-review ultra develop
```

The command targets a branch comparison, not a path. Read the dialog's branch, base, changed-file and line counts, remaining free runs and estimated usage credits alongside the repository upload scope you checked in step 2. Stop if the comparison, upload authority or estimate is wrong. The command starts only after you invoke and confirm it.

Anthropic currently lists three **one-time** free runs for Pro and Max accounts. Later runs typically use $5–$25 in usage credits, according to the manual; the account's launch estimate is the decision point. Paid reviews require usage credits enabled. A run counts once its cloud session starts. Stopping a free run consumes it, while paid work may be billed for the portion run. The feature, price and availability are research-preview terms, so verify the current dialog rather than budgeting from a fixed figure here.

### 4. Follow the same run and examine each finding

The manual says a review typically takes 15–25 minutes and runs in the background. Use `/tasks` to inspect the running or completed review and the session notification for findings. A quiet terminal is not a reason to launch a second review. Stopping a review returns no partial findings. If the session appears missing, check whether it was stopped, archived or deleted, or whether the signed-in account changed. A timed-out noninteractive command can leave a cloud run active; establish the original run's state before considering another paid or free run.

For each finding, open its cited file and location. Compare the claim with the intended behavior and run a targeted check where the claim is reproducible. Record a fix, false positive, accepted trade-off or unresolved question with the code and check that justify it. If Claude proposes a repair, inspect that diff before accepting it. Anthropic says its reviewer agents verify reported findings, but that vendor description does not replace your own check. A completed review with no findings is a review result, not evidence that there are no bugs.

### 5. Make the merge decision from the final branch

After any edits, inspect the complete final diff against the intended base. Rerun the original acceptance command and the targeted checks used to assess findings; record actual results, including failures. Leave an unresolved or failing branch unmerged until its owner decides how to proceed. The human reviewer or owner makes the merge decision. Another Ultrareview is a separate run with a separate free-run or credit consequence, justified only if the changed code warrants it.

This is a cloud-review step for an already developed branch. For the broader design, implementation and branch-review sequence, see BIG CHANGE's [Superpowers guide](https://bigchange.ai/blog/superpowers-claude-code-plan-test-review-change-guide).

## Sources & further reading

- [Anthropic, “Find bugs with ultrareview”](https://code.claude.com/docs/en/ultrareview) documents the command, branch/base scope, account restrictions, dialog, price, task tracking, limits and findings. It is a research-preview product manual, not evidence that BIG CHANGE ran a review or that any specific branch is safe.
- [Anthropic, “Use Claude Code in the cloud”](https://code.claude.com/docs/en/claude-code-on-the-web#send-local-repositories-without-github) explains the bundle, history, platform-specific tracked-file treatment, untracked files and size fallbacks. The reader must check the actual repository and policy before an upload.
- [Anthropic, “Review locally with /code-review”](https://code.claude.com/docs/en/code-review-local) describes the distinct local review route when cloud review is unavailable or unwarranted. Its results should not be reported as Ultrareview findings.

## Sources

- [Find bugs with ultrareview — Claude Code Docs](https://code.claude.com/docs/en/ultrareview) — Primary manual for the cloud command, scope, account exclusions, dialog, limits, credits, task tracking and findings. It does not prove a particular review ran or caught any bug.
- [Use Claude Code in the cloud — Claude Code Docs](https://code.claude.com/docs/en/claude-code-on-the-web#send-local-repositories-without-github) — Primary manual for full-clone history, tracked and untracked edits, platform-specific sensitive-file behavior and the bundle size/fallback rules. It cannot establish a reader's upload authority.
- [Review locally with /code-review — Claude Code Docs](https://code.claude.com/docs/en/code-review-local) — Primary manual for the distinct local workflow, which must not be described as a completed cloud Ultrareview.