You're responsible for code you don't fully understand.

Most of it, an agent wrote. mergecomplete turns the diff into a review: what's risky, what's not, and where to look first. Understand it in seconds, and sign off knowing what's actually there.

mergecomplete
feature/ENS-343-cdss-polish context layers CdssDetailsRow.tsx does not compile seed flag left on in CdssOrderRequest useRowState() is unreachable host screen context details page shared write path shared & server PatientChart.tsx hosts the row useCdssGate() feature flag CdssSummaryLane.tsx the other reader CdssDetailsRow.tsx the hub does not compile CdssRowResolution collects; never interprets useRowState() unreferenced dead CdssOrderSubmission shapes the payload CdssOrderRequest one rule set, two readers seed flag on PUT /orders/:id the write analytics.track outside the feature cdssPolicy.ts one rule set, two readers
This is a review: everything that changed, and whether you can trust it.

You didn't write it. You still sign off on it.

Eight tickets today, each one a diff an agent wrote in minutes.

Read every line and the time it saved you evaporates. Skim it and you're guessing. Either way, your name is on the merge.

How to read it

The review.

mergecomplete replaces the diff with a review: every file and function that moved, grouped by the part of the system it touches. Read it in seconds, and know what's risky before you open a single file.

  • The verdict, before you scroll

    Three ranked lines in plain language:

    • CdssDetailsRow.tsx does not compile
    • seed flag left on in CdssOrderRequest
    • useRowState() is unreachable

    Five seconds, and you already know where to look.

  • The note, not the diff

    Between the diagram and the source sits a note: the code restated as pseudocode, with the reasoning inline. It is not a summary of the diff. When a file is mid-refactor and does not compile, the note renders both halves separately, because intent only reads in the new code and current behavior only reads in the old. No diff view does that.

  • Risk on an unchanged file

    Change state and risk are independent. A file nobody touched can be the riskiest node in the diagram, because of what changed around it.

Separation of duties

The reviewer isn't the author.

mergecomplete keeps the model that reviews a change independent of the model that wrote it. A team can require that code from one agent is checked by a different one; the trace shows which model signed off. Separation of duties, for agents.

Your existing workflow

Fits the workflow you already have.

Your coding agent already talks to git. mergecomplete gives it two more tools: latest, to work off the current base, and submit, to send the change for review.

That works from inside Cursor, Claude Code, Codex, or a plain terminal. GitHub stays the remote: where the result shows up, not where review happens.

mcp · tools
latest
work off the current base
submit
send the change for review

For whoever signs off on the change.

Solo on a repo, running review for a team, or setting policy for an org: it's the same job. Decide whether an AI-written change is safe to ship, without reading every line.

No procurement, no team rollout: just one repo, and nobody else's permission. That's what getting in early gets you.

Get early access.

The review surface is in build now. Leave your email and we'll bring you in as it opens.