Live studio: https://futurefork-studio-39330700.vercel.app/ (no install, no login). The original downloadable evidence bundle is unchanged.

Release reviewers can miss a keyboard regression when the screen and mouse journey look unchanged. FutureFork puts the baseline, broken candidate and repair at the same captured browser checkpoint. A clickable div left keyboard users unable to reach Continue; GitLab Duo restored the native button. The authoring session timed out after its push, and a separate read-only continuation verified the passing repair. The public judge bundle preserves the original comparison and evidence; the repair MR stays draft and unmerged.

FutureFork is a release comparison studio built around real user journeys at exact Git commits. Baseline, candidate and repair share the same browser environment, fictional cart and verification contract. Each lane exposes screenshots, focus observations, elapsed times, traces and JUnit results from actual CI jobs. Move to one checkpoint and compare what each version allowed the user to do. When a journey stops, the candidate holds its last captured failure frame. It never generates an image of a successful future that did not happen.

The recorded comparison has three distinct outcomes. Baseline pipeline 2972490221, at 8389c503, passes mouse and keyboard. Original candidate pipeline 2972508230, at ca71f6e4, passes mouse but fails keyboard with KEYBOARD_CONTINUE_UNREACHABLE: 14 Tab presses never reach the control. Repair pipeline 2972655818, at 361f9a63, passes the unit/contract, mouse and keyboard jobs. The original failing candidate remains preserved, so the comparison does not erase the regression it explains.

GitLab Duo Agent Platform supplies the agentic repair, not a decorative chat panel. Its registered Assisted flow reproduces the regression, proposes a one-file change restoring native button semantics, applies that change and verifies the result. The human reviews separate plan, repair, verification/push and final-readback gates. Agents hold no tools; bounded programs validate proposals and restrict changes to CheckoutCTA.tsx. The browser assertions come from the baseline contract, so the candidate cannot simply weaken the test that caught it.

The successful evidence chain spans two native sessions, and both are retained. Authoring session 9826076 received the owner's genuine first three decisions and authored/pushed repair 361f9a63 to draft MR !25. It then timed out while waiting for CI and remains Failed. Separate read-only continuation 9833538 received the owner's genuine fourth decision and finished its final readback. Its executor job 17356747820 exited successfully without another repair, push, CI retry, merge or deployment. This is a completed evidence chain, not an uninterrupted successful session.

The creative idea is to make a release review about a user's changing ability to finish a task. Identical-looking captures can conceal opposite outcomes; the synchronized journey makes the difference inspectable. A reviewer can switch between mouse and keyboard, find the stopping checkpoint, inspect the exact source change and follow its evidence. The publicly hosted refined studio sharpens that workflow with an immediate lane-outcome scan, accessible stage controls and lane-specific artifact navigation.

The downloadable judge bundle packages the original read-only studio, its MIT source and all 35 unchanged report artifacts from six original CI browser archives. It runs locally with Python’s static server, without an account, credentials or dependency installation. Its schema, artifact hashes, source build and browser interactions were verified; anonymous download matches SHA-256 e426a2821fb1207e37a98ae518edbad31b3c7f28c4c596adee76f8f2227f9e4d. A fresh public clone can use the bundle’s reports directory as documented in its README. Private approval payloads are excluded; genuine session and public CI links remain available, although native session pages can require GitLab sign-in. The refined studio is publicly hosted at https://futurefork-studio-39330700.vercel.app/ with no install or login. It presents the same recorded evidence; hosting does not imply a new native run or protected production release.

FutureFork is a Path A, Assisted entry with a Most Creative emphasis. Its checkout and regression are deliberately controlled, and MR !25 remains unmerged. Maintainers still control CI, and inherited main-push capability means this is not an independently isolated release authority. The payoff demonstrated here is narrower and useful: a preserved keyboard failure, a bounded native repair and a reproducible comparison of what changed for the user.

Judges can compare the same browser checkpoint at the baseline, regression and repair commits in the public studio or downloadable evidence bundle. Duo authored the button repair; Codex supported engineering and evidence assembly, and Claude contributed read-only review. The edited demonstration uses Deepgram synthetic narration and retains the distinction between the timed-out authoring session and its read-only continuation.

Lifecycle stages demonstrated: verify, secure, package.