Review
Human-in-the-loop quality assurance. Review the AI's work feature by feature, verify done whens and flag issues, then complete the cycle to create a pull request.
The review phase is where people take back control. After the AI builds your features, you review every task against its done whens. You can accept work that meets your standards, flag defects, note requirements that were missed, and request small changes.
This is the quality gate that ensures AI-generated code meets your team's standards before it reaches your codebase.
AI can write code fast, but speed without quality creates technical debt. Review provides a structured way to evaluate AI output: not just "does it look right" but "does it meet the criteria we agreed during planning".
- •Criteria-based review — Every task has verifiable 'done when' conditions from planning
- •Batch efficiency — Accept passing work in one click and focus your time on issues
- •Issue tracking — Issues can be fixed now or deferred to the roadmap for a future cycle
The review page with feature cards, task status, and overview metrics.
Overview Assessment
Feature-by-Feature Review
Mark Task Status
Handle Issues
Complete the Cycle
Key Features
Cycle Quality Card
Completed cycles show the AI validated pass rate, fix rounds, human-caught defects and actual versus estimated cost, plus a one-line summary to copy for stand-ups.
Review Chat
Ask the AI about what was built. It has read-only access to the code and, once a preview is deployed, can click through the live app and record a video of its investigation.
Iterate
Request small cosmetic changes without a full rebuild. A narrowly scoped session edits only allowed files, such as styles and components, and reverts if it strays outside them.
Preview Deployments
When enabled, reviewers can click around the running app at a preview URL.
GitHub PR Creation
Completing a cycle creates a GitHub pull request with a detailed description of what was built and reviewed.
Defer to Roadmap
Defects and missed items can be sent to the Later lane instead of blocking the cycle.
Manual Steps Checklist
If your plan includes manual deployment steps, such as DNS changes or API keys, they appear as a checklist.
Each task receives one of three verdicts:
The task meets all its done whens. The implementation is correct and complete.
The task was attempted but has issues, such as bugs, incomplete implementation or quality concerns.
The task was not implemented at all, or the implementation is fundamentally wrong.
Can I accept all tasks at once?
Yes. Features can be bulk-accepted, and you can filter to focus on unreviewed tasks.
What happens to defects?
They appear in the Issues Panel, where you can trigger a fix build, defer them to the roadmap's Later lane, or resolve them if they aren't real issues.
Can I view the code changes?
Yes. You can view diffs and build logs inside the review page, and open the branch on GitHub from the overview header.
What if I want to go back to an earlier phase?
Use the Rewind button on the read-only banner. A confirmation dialog explains what data will be cleared, such as review statuses and generated documentation.
Who can review?
Users with the right project permissions. All reviewers see the same task statuses, and each reviewer has their own chat thread.
We've thought carefully about these areas. Some are intentional design choices, others are on our roadmap.
Iterate is scoped to one feature and an allowlist of files. Anything larger should be a fix build or a new cycle.
Feedback is captured at task level, not as inline comments. For detailed feedback, use the task notes or the review chat.
The review chat can only drive the running app once a preview has deployed.