# Review workflow

Canonical: https://snapvisor.io/docs/review-workflow

> How Snapvisor baselines, approvals, monitoring mode, flaky-change detection, and summary checks work together.



Every Snapvisor build diffs against an approved baseline and waits for a human (or an AI agent over MCP) to approve or reject the changes. This page covers the concepts that shape that workflow.

## Baselines and approvals [#baselines-and-approvals]

When a change is detected, the pull request gets a status check linking to a side-by-side review. Approving a change promotes it to the new baseline for that branch — intended changes never nag you twice. Until a reviewer decides, the check stays pending ("N changes awaiting review"); rejecting keeps the previous baseline and fails the check. See [check names and states](/github#check-names-and-states).

## Require a person to approve [#require-a-person-to-approve]

By default an agent connected over MCP can approve a build on the user's behalf; the review records which agent did it. A project can refuse that: switch on **Settings → Approvals → Require a person to approve**. The same rule is available on the API as the project setting `requireHumanApproval`.

While it is on:

* An approval has to come from the Snapvisor web app or the Snapvisor iOS app. A request that uses an access token (a personal access token, an OAuth token, the MCP server or the CLI) is answered with `403`. This applies whether or not the client says it is an agent, because a token cannot tell a person from a script.
* Rejections are still accepted with a token, so an agent can hold a change back. So are comments, except one that would replace the rejection of the same user: while that rejection is in place, a comment, or a draft comment added to a review, sent with a token is refused. Comment or change it signed in to the web app.
* An access token cannot turn the rule off, cannot change the auto-approved branches, the default base branch or the ignore settings, and cannot switch on tokenless or GitHub Actions OIDC uploads. Changing them takes someone signed in to the web app. Ignoring a change and dismissing a rejection need the web app or the iOS app too; unignoring a change and dismissing an approval are still accepted with a token, because they only bring review back or hold a change back.
* A build that an access token creates on an auto-approved branch is refused with `400`, and so is an access-token update to a build that is on one, or a parallel shard that joins one. Use another branch name for it. Page captures follow the same rule.

This narrows the ways to a new baseline without a person; it does not close all of them. A CI upload with the project token (including tokenless and GitHub Actions OIDC uploads) to an auto-approved branch still becomes the baseline, because that is how a baseline is made on your main branch. GitHub Actions OIDC tokens are tied to the commit but not to a branch, so a workflow run on any branch of the repository can upload with your main branch's name; tokenless uploads are tied to the run's branch. The review history and each build's creator (CI, agent or person) remain the record of who did what.

## Monitoring mode [#monitoring-mode]

Not every branch needs a human to approve each change. In the default **CI** mode, a build compares against the approved baseline on your base branch and waits for review. In **monitoring** mode, a build with no prior approval doesn't block — it reports a success check and, once approved, becomes the baseline future builds compare against.

Monitoring mode is configured with an **auto-approved branches** glob pattern in project settings. Any branch matching the pattern is automatically approved and gets a success status check. By default only your default base branch is auto-approved.

```text
main
{main,production}
release/**
```

Use it for:

* Production or release branches where the merged state is the source of truth, not something to re-review.
* Continuously monitoring a live app for unexpected visual drift rather than gating pull requests.

## Flaky test detection [#flaky-test-detection]

Some screenshots wobble for reasons that aren't real regressions — an animation frame, a timestamp, a random avatar.

* **Ignore changes manually** — turn on ignore changes in a project's settings and reviewers can mark a recurring change as ignored. Ignored changes no longer require review on future builds.
* **Auto-ignore** — with ignore enabled, Snapvisor can auto-ignore changes that keep reappearing. A change is treated as flaky once it shows up at least a configured number of times across auto-approved builds within a rolling 7-day window. The default threshold is **3** occurrences; raise it to be stricter, lower it to catch noise sooner.

Auto-ignore is scoped to auto-approved builds, so it learns from branches you already trust rather than from unreviewed pull requests.

## Summary checks [#summary-checks]

A project can produce several builds on the same commit — different apps, shards, or Storybook and end-to-end runs side by side. A **summary check** rolls them into one pass/fail status you can add to your branch protection rules. Three modes:

* **Auto** (recommended) — post a summary check only when a commit has more than one build.
* **Always** — always post a summary check, even for a single build. Useful when builds run conditionally, so the required check is always present.
* **Never** — don't post a summary check.

<Callout>
  To require Snapvisor before merge, set the mode to **always** and add the summary check to your repository's required checks.
</Callout>

