Plan & code review
Approve a plan before an agent starts, review the diff before you merge, and pull in a subagent for a second opinion on either one.
Proliferate gives you two natural checkpoints on agent work: the plan, before any files change, and the diff, before it merges. Both show up right where you're already working, and you can loop in a subagent for a second opinion on either one.
A note on this page. An earlier build of Proliferate had a dedicated review workflow with named reviewer personalities, multi-round auto-review, and its own settings pane. That workflow isn't in the app today. What follows is what you actually have right now: plan approval, diff review, and subagent-assisted critique.
Approve the plan before work starts
When an agent that supports plan mode (Claude Code, Codex, and others) works out an approach before touching any files, it posts the plan as a card in the transcript instead of jumping straight into changes. The card shows an Awaiting approval chip and two buttons: Reject, if the direction is wrong, and Approve, if it's ready to run.
Proposed plan card, awaiting approval
A plan card in the transcript rendering the plan's markdown body, with an 'Awaiting approval' status chip and Reject / Approve buttons in the footer.
This is the moment to actually read the plan: does it touch the right files, does it skip a migration or a test, is it solving the problem you asked for. Rejecting sends the agent back to rethink it; approving unlocks the next step.
Once a plan is approved, the card's status chip flips to Approved and the footer swaps to a primary Run here button, with New session tucked into the overflow menu if you'd rather hand the approved plan to a fresh session instead of continuing in the current one.
Approved plan card with follow-up actions
The same plan card after approval, showing an 'Approved' status chip, a primary 'Run here' button, and a '...' overflow menu containing 'New session'.
Review the diff before you merge
Every workspace has a Changes panel: a diff view of everything the agent has touched, compared against a base branch you can change. It's where you actually look at the code before it goes anywhere, file by file, with the usual collapse/expand and word-wrap controls for scanning a big diff quickly.
Changes panel
The Changes panel in the workspace, showing a file tree of the current diff against a selectable base branch, with per-file status badges and panel options for collapsing all diffs and toggling word wrap.
When you're happy with the diff, publish the branch (and open a pull request in the same step) from the workspace. See Publish and review for that flow, including how pull request status shows up in the workspace afterward.
Get a second opinion from a subagent
There's no built-in reviewer-personality workflow right now, but you don't need one to get a second set of eyes. Since subagents can read the same workspace, just ask your agent to spawn one with a review job instead of an implementation job:
The subagent shows up as a normal delegated agent: a collapsed row in the transcript you can expand to see who it is and what it was asked to check.
Subagent review request in the transcript
A collapsed 'Created subagent' row in the transcript, expanded to show the child agent's name and the review prompt it was given.
The same pattern works for a plan: ask your agent to spawn a subagent to critique the plan for gaps before you approve it, or ask a running subagent to check a specific file or pattern across the codebase. While it works, you can check on it from the composer's Agents popover, which lists every active subagent with its current status and a Wake action if you want to be pinged the moment it finishes.
Agents popover with a running reviewer subagent
The composer's Agents popover open, listing a subagent by name with a status label and a Wake button to be notified when it completes.
Because it's a real subagent, you control the lens completely: ask for a security pass, a test-coverage pass, a "does this match our existing patterns" pass, or anything else, just by changing the prompt.