Skip to main content
The Review tab runs an AI code review on demand. You choose exactly what to review — a lane’s uncommitted changes, a range of commits, a whole pull request, or one lane’s diff against another — and ADE runs a read-only inspection and returns findings you can act on, plus the evidence and transcript behind each one. This is separate from the review-comments thread on a PR. Review runs are local by default: nothing is posted to GitHub unless you choose to publish. Think of it as a second pair of eyes you can summon for any slice of work, at any point, without opening a PR first.
The review agent is read-only. It can read and inspect files but never edits, commits, or pushes. A run produces findings and saved artifacts — acting on them is up to you.

​
Pick a review scope

Every run starts by choosing a lane and a target mode. The mode decides which diff gets reviewed.

Lane diff

Review a lane’s branch changes against the default branch — or against another lane. The most common scope: everything the lane has done since it split off.

Commit range

Review only a slice of a lane’s history. Pick an earlier base commit and a later head commit from ordered dropdowns; the base is excluded, the head is included.

Uncommitted changes

Review the staged, unstaged, and untracked changes in a lane right now — the working tree against the checked-out HEAD. No commit required.

Pull request

Review a whole PR’s diff as one scope, with findings you can optionally publish back as a GitHub review.
For a lane diff, you choose what to compare against: the lane’s default branch, or another lane (lane-to-lane). The launch dialog draws a small before/after diagram of the scope you’ve picked — which branch or commit is on each side — so you can confirm the comparison before spending a run.
Comparisons run against your local refs. If you want the latest remote changes included, fetch or pull the base branch first, then launch the review.

​
Run a review

Open Launch new review from the toolbar, set the lane, target mode, and (for lane diffs) the comparison, then pick a model and reasoning effort. Start the run and it appears in the runs list on the left while it works. A run moves through queued → running → completed (or failed / cancelled). Several specialist reviewers look at the diff from different angles — diff risk, cross-file impact, checks and tests, security and data, UI and regression — and ADE adjudicates their candidates into one merged set of findings. You can cancel a running review at any time, and Rerun re-runs the same scope after you’ve pushed more changes.

​
Read the findings

Each finding is a card you can act on. Cards carry a severity (critical, high, medium, low, info), a confidence score, the file and line they anchor to, and an evidence trail — quotes, diff hunks, file snapshots, and tool signals (typecheck, test, lint, build, CI) that back the claim.

Open in Files

Jump straight to the file and line in the Files tab to see the change in context.

Open in editor

Send the finding’s location to your external editor.

Copy findings

Copy one finding — or all of a run’s findings — as a single message to paste into a chat or issue.
Filter the list by severity, and act on each card to teach the engine: acknowledge it, dismiss it with a reason, snooze it, or suppress it so similar findings are filtered out of future runs.

​
Review learnings

Findings you dismiss or suppress feed a learning loop. The Learnings panel shows quality over time and the suppressions you’ve built up.
Counters for total runs, total findings, how many were addressed, and a noise rate — dismissed plus suppressed over total findings — so you can see how much of the review output the team actually acts on. A breakdown by finding class shows how well each category is being addressed.
Every suppression you’ve created, scoped to the repo, a path pattern, or globally. Each shows how many times it has filtered a matching finding, and you can remove any of them to let those findings surface again. Suppressions are matched by title and scope and persist across runs.

​
Review notes and evidence

Beyond the findings, each completed run keeps the full audit trail so you can check the engine’s work:
  • Review scope — the before/after diagram of exactly what was compared, the model used, and whether the run was kept local or published.
  • Review process — the specialist reviewers that ran, the context they were given (changed-file manifest, risk map, rule overlays, validation signals), and per-reviewer candidate counts.
  • Reviewer outputs — candidate counts, adjudication results, and the merged final set.
  • Artifacts — the raw diff bundle, prompts, and payloads for audit.
  • Review agent transcript — open the saved read-only session in Work for the full turn-by-turn trace.

Pull requests

Open, review, and merge GitHub PRs — and request an AI review on one.

Conflicts

Predict, simulate, and resolve merge conflicts before they bite.