MergeGate: the merge queue that never ships a bad agent PR

The problem

Coding agents raised draft output 30-98% while verification capacity stayed flat: review time is up 49-91% and senior engineers have become the bottleneck, re-reading plausible-but-wrong diffs instead of building. Existing tools review and comment; nobody owns whether the merge should happen at all.

The idea

Why now

Three unlocks landed within days of each other: Liquid AI d1-ultra (October 16) is a SOTA decision model with zero output tokens, making per-PR verdicts cost fractions of a cent instead of dollars; CopilotKit CLiF (October 16) shows agent trajectories can be distilled into reusable skills; and GitHub itself is seeing ~17M AI-involved PRs per month, so the review debt is compounding daily.

What it combines

Liquid AI d1-ultra (cheap zero-token decision verdicts) + CopilotKit CLiF (distill repo trajectories into quality skills) + CodeDiver (understand AI-written code) + Jev (typed judgments as the decision layer). Review tools comment on PRs; none of them owns the merge outcome, and none of them closes the loop by sending the fix back to the originating agent. The combination is a gate with teeth, not another reviewer.

MVP

A GitHub App that (1) labels agent-authored PRs, (2) returns a cheap decision-model verdict with inline failure reasons, and (3) posts a machine-readable fix report the originating agent can consume. Deliberately skip: auto-merge, multi-platform support, and the learned-skills layer — ship the verdict first in report-only mode.

How it makes money

The moat

GitHub will not ship an opinionated auto-merge gate: it would conflict with their neutral-platform posture and cannibalize the review-tool partners (CodeRabbit, Greptile) in their own ecosystem. And the training signal the gate needs — per-repo merged/reverted agent trajectories — is generated inside Cursor, Claude Code, and Codex harnesses, outside GitHub's walls; GitHub cannot learn per-org quality bars from data it does not own.

Distribution

GitHub App sold B2B to engineering teams at 20-200 person startups where agents write most of the code; land through the GitHub Marketplace with platform teams and scale to the whole org. Report-only mode for the first 30 days removes adoption risk.

Why it wins

CodeRabbit, Greptile, and Cursor BugBot review PRs and post comments — a human still decides and merges. MergeGate owns the merge outcome instead: it gates, regenerates the fix through the originating agent, and only then merges, learning per-repo quality bars from trajectory history rather than static rules.

Risks

False-positive blocks erode trust fast. De-risked by shipping in report-only mode for 30 days and keeping human override one click away, so the gate earns authority before it exercises it.

Build it with

• Liquid AI d1-ultra: SOTA decision model with zero output tokensSOTA decision model with zero output tokens: per-PR verdicts at fractions of a cent
• CopilotKit CLiF: continual learning that turns agent trajectories into skillsTurns agent trajectories into reusable skills: learns each repo's quality bar from its own merged/reverted history
• CodeDiver RC1: desktop app + MCP server for understanding AI-written codebasesMaps and explains AI-written codebases so the gate's verdicts reference real structure
• Jev 1.14.0 — TypeSafe's judgment layer updateTyped judgments as the decision layer: merge / fix / escalate verdicts with calibrated confidence

Repo to start from

mergegate — GitHub App plus worker that gates agent-authored PRs with cheap decision-model verdicts and machine-readable fix reports.

Evidence

• CodeRabbit pricing (Pro $24/dev/mo annual)
• Harvard field data: +30% code, +49% review time, flat output
• AI coding costs land in review, not on the invoice (Faros AI data)
• AI is flooding curl with low-quality PRs
• AI coding agents are creating a code review bottleneck