Replay2PR - bug replays into verified patches.
Replay2PR is a Gemini 3 + Playwright pipeline that turns a short bug replay into a reproducible Playwright test, an automated patch attempt, and a shareable Evidence Pack. It ships with a built-in demo target carrying an intentional bug and a demo video, so the full extract-reproduce-patch-verify-report loop runs locally with no paid services. Built for the Gemini 3 Hackathon.
A bug report is usually a screenshot and a sentence. Someone still has to reproduce it, write a failing test, fix it, and prove the fix worked - the slowest, least glamorous part of shipping.
Replay2PR compresses that whole loop into one automated run. A short replay becomes reproduction steps, a failing Playwright test, an applied patch, and a verified, shareable Evidence Pack - so the proof travels with the fix instead of living in someone's head.
From a replay to a verified PR.
Extracts repro steps from a replay
A short screen recording goes in and Gemini reads it as evidence, extracting an ordered set of reproduction steps and a plain-language statement of the failure ("submitting the replay form never shows the success banner") - the raw material every later stage builds on.
Generates a Playwright test that fails
From the extracted steps, Replay2PR writes a real Playwright test that drives the actual UI and asserts the missing behaviour. The bug has to reproduce red before anything is allowed to claim it green - no test, no proof.
Attempts the patch, then verifies it
The patch step edits the target source, flips the broken state, and re-runs the same Playwright test. A fix only counts when the previously-failing test now passes - verification is the gate, not an afterthought.
Ships a shareable Evidence Pack
Every run persists artifacts and renders a report card: status, repro steps, the generated test, the patch diff, verification logs, and a downloadable Evidence JSON - turning a bug report into portable, reviewable proof instead of a screenshot in a chat.
A pipeline you can watch run.
The run isn't a black box. Each of the five stages - extract, reproduce, patch, verify, report - reports its own status and detail live on a Mission Control timeline, so you can see exactly where a job is and read the failure the model extracted before it ever writes a line of test code.
Extract, reproduce, patch, verify, report.
Each stage feeds the next: Gemini reads the replay, the steps become a real Playwright test, the patch edits the target and re-runs that test, and only a genuine red-to-green transition earns a passing Evidence Pack.
- 01Ingest a short bug-replay video (built-in demo video included)
- 02Extract reproduction steps and the failure statement with Gemini
- 03Generate a Playwright test that reproduces the bug
- 04Apply a patch and re-run the test to verify the fix
- 05Assemble a shareable Evidence Pack with downloadable JSON
Proof you can hand to anyone.
The Evidence Pack is the deliverable: a share page that bundles the reproduction steps, the generated test, the patch diff, the verification logs, and the run artifacts - with a downloadable Evidence JSON so the whole run is portable, auditable proof rather than a claim.
Three layers, one deterministic demo.
Job runner
A Next.js 14 app with an in-process job runner drives the five-stage mission, streams a Mission Control timeline to the page, and caps concurrency so parallel Playwright runs never collide.
Gemini + Playwright
Gemini 3 (Flash and Pro) handles extraction and code generation; Playwright executes the generated tests against a real target - with a mock-Gemini path so the whole loop runs without an API key.
Evidence layer
Artifacts are written to disk and rendered as an Evidence Pack share page, complete with a deterministic built-in demo target whose intentional bug the patch step actually fixes.
The work behind Replay2PR
A pipeline that turns a short bug replay into a reproducible Playwright test, an automated patch attempt, a verification run, and a shareable evidence pack.
Built by Team Kanban, the studio's competition and experimental build team, at Gemini 3 Hackathon. See the full competition record.
Services this build draws on
- Custom SoftwareWhen off-the-shelf software almost fits but never quite does, we build the system that does. Web apps, internal tools, portals, and APIs on dependable, maintainable foundations - software you own and can grow into.Read the service page
- Monthly Software SupportSoftware doesn't stop needing attention at launch. We act as your ongoing technical partner - fixing issues, improving workflows, adding integrations, and guiding decisions - so your systems keep working, improving, and adapting as you grow.Read the service page
- Artificial IntelligenceFrom machine learning and natural language to computer vision and predictive analytics, we build a full suite of AI capabilities - tailored to how your business actually works, with a person reviewing the decisions that matter. No black boxes, no hype.Read the service page
