Google Cloud Rapid Agent Hackathon · GitLab Track
Minari実成
Autonomous flaky-test intelligence. The agent that doesn't just flag a flaky test, it fixes one.
Detect → diagnose → fix → verify → deliver. Minari proves every fix in real GitLab CI and opens a reviewed merge request. It never weakens an assertion, and never merges on its own.
- ASYNCT-2847test_async_result_ready_after_sleep94
- RACET-1823test_event_order_dependent_result78
- REST-0042test_pool_acquire_under_contention71
- NETT-8807test_service_responds_within_latency_budget55
- DATAT-1142test_cart_total_isolation28
Built with Gemini 2.5 · LangGraph · GitLab MCP · Google Cloud Run
Problem Statement
Flaky tests don't fail because the code is wrong.
They fail because timing, ordering, and shared state are unreliable, and that unreliability quietly erodes trust in the whole suite. Most tools can only point at them; fixing by hand is slow, and the lazy fix, weakening the test, is worse than the flake itself.
Flakiness rate from Google's published testing research; other figures are representative of industry experience.
How It Works
One trigger.
The entire repair, end to end.
Detect → Diagnose → Fix → Verify → Deliver. Five specialist agents, one LangGraph pipeline, no human in the loop until the merge request is ready for review.
Continuous data flow from a flaky run to a verified, reviewed merge request.
Architecture
A pipeline of specialists,
not one monolithic prompt.
Five agents, each with one job and one well-defined hand-off, orchestrated by LangGraph and grounded in real GitLab and Supabase data.
FastAPI · LangGraph agent
LLM Integration
Gemini is the
reasoning layer.
Diagnosis, repair, and recall all flow through Gemini 2.5. Deterministic gates (syntax, assertions, real CI) own the safety. The model reasons; the pipeline verifies.
Deep root-cause reasoning on the uncertain cases (PFS 40–75).
Fast generation of three ranked fix candidates (~8s).
768-dim embeddings power similar-failure recall.
Features
Built to be trusted.
An AST walk proves every original assertion still exists in the fix. The rule is even enforced in the database, so a fix that edits an assertion can't be stored.
+ await until(ready, timeout=2s)
assert result == expected
The fix runs in a fresh pipeline five times; the gate demands 5/5 green plus a real drop in run-to-run variance before it ships.
Minari opens a merge request with a senior-engineer write-up and a reviewer, and never merges on its own. The final call always belongs to a person.
Every MCP call, decision, and fix candidate streams to the dashboard in real time.
Tech Stack
Every layer, in production.
The full stack behind Minari: agent, backend, data, integration, frontend, and the security boundaries that make autonomous test repair safe to run.
Why Minari
Others detect.
Minari completes the loop.
These are the tools teams reach for today. Most stop at detecting or quarantining a flaky test. Minari is the only one that carries it all the way to a verified, reviewed merge request.
| Capability | Datadog | Katalon | Mergify | BuildPulse | Trunk | TestDino | Launchable | Copilot | Minari |
|---|---|---|---|---|---|---|---|---|---|
| Detect flaky tests | |||||||||
| Diagnose root cause | |||||||||
| Generate fix | |||||||||
| Verify fix | |||||||||
| Create merge request | |||||||||
| Free | |||||||||
| Open source | |||||||||
| GitLab-native | |||||||||
| Multi-language |
Positioning reflects each product's publicly documented scope as of 2026 and is meant to show category boundaries, not to disparage any tool.
Target Market
Every team with CI
has flaky tests.
Minari enters through GitLab-native teams who feel flakiness every day, expands across Git-based CI, and scales into the broader quality-tooling market. Figures are top-down estimates from public market reports.
Business Model
Three revenue streams.
Priced on value, not promises.
Open-core adoption seeds teams for free, a per-developer plan monetises private repos, and an outcome-based add-on ties price directly to verified fixes shipped.
The agent is open-core: public repos and self-hosting are free. Teams pay a flat per-developer fee for hosted runs, the live dashboard, forecasting, and the MCP server on private repos.
Regulated and large orgs license a self-hosted deployment with SSO/SAML, audit logging, private model routing, and a support SLA. Annual contracts, seat- or org-based.
An optional usage add-on prices Minari on value delivered: a fee per verified fix merged. Because the safety gate ships only 5/5-green fixes, customers pay strictly for repairs that worked.
Roadmap
Live today.
Built to become a platform.
The five-stage pipeline ships verified fixes today. From here, Minari widens across languages and platforms, then shifts from repairing flakiness to preventing it.
- Five-stage LangGraph pipeline live on Cloud Run + Vercel
- Python / pytest with real GitLab CI verification (5× runs)
- Assertion-safety gate + confidence cascade, never auto-merges (reviewer via blame or config)
- Live dashboard, predictive forecast, sustainability tracker
- Minari exposed as its own MCP server
- JavaScript / TypeScript (Jest, Vitest) and Go support
- GitHub Actions alongside GitLab CI
- Persistent MCP session to shave seconds per run
- Org-wide flakiness analytics across repos
- Slack + merge-request annotations on every fix
- Forecast-driven pre-merge flakiness gating
- Auto-quarantine with batched, scheduled fix MRs
- Self-hosted enterprise: SSO, audit, private routing
- Fine-tuned diagnosis on accumulated repair history
- Shared library of verified fix strategies
GitLab Track · Rapid Agent Hackathon · 2026
Stop re-running.
Start trusting the suite.
Watch Minari diagnose a flaky test, prove the fix in real CI, and open a reviewed merge request, end to end, in under two minutes.