Kintsugi - Autonomous QA for IBM Bob

The problem: Teams ship features faster than they can test them, so bugs slip through. End-to-end tests are the safety net for real user flows, and they are the first thing teams skip, because writing and maintaining them is slow, brittle, and manual. A red CI line is not a root cause, and finding a bug and fixing it are two disconnected jobs done by two different people. The solution: Kintsugi turns IBM Bob into an autonomous QA engineer. You do not write a single test. Bob reads your codebase, writes Playwright end-to-end tests, runs them in a real browser through the Kintsugi MCP server, triages every failure with watsonx.ai, and heals the application code until the suite is green. How it works: It is a closed loop. First Bob analyzes the app, crawling it into a route map and reading the source. Then it plans and writes, designing lane-based test plans (ui, api, auth, edge, visual) whose subagents author the Playwright specs. Then it runs and triages: the tests run in the MCP sandbox, and every failure becomes a screenshot, a video, a Playwright trace, a watsonx.ai triage, and a Kanban ticket. Finally it heals, fixing the application code and re-running each test until the board is green. Target users: development teams that want real end-to-end coverage without owning a brittle test suite, and startup or hackathon teams that need QA to keep pace with feature velocity. What makes it unique: Kintsugi ships as two Bob modes over one loop. kintsugi-qa plans, writes, and runs the tests and can edit only the test files. kintsugi-healer fixes the application code and can never touch a test. Neither mode has a shell, so Playwright runs only inside the MCP sandbox behind a host allowlist and a test-directory jail. The healer physically cannot weaken a test to make it pass, and after three failed attempts a ticket is set to NeedsHuman, honestly. Every failure is triaged by watsonx.ai into a severity, a root cause, and the suspected files, and every fix ships with tamper-proof evidence in IBM Cloud Object Storage. The result: in the demo, ShopLite is a small React storefront with five genuine bugs across five lanes. Bob writes 22 tests, finds all five bugs, triages them, and heals every one, taking the board from red to green. No human wrote a line of test code.