.png&w=256&q=75)
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Jointly - Evidence Before Merge
Two changes. Two passing test suites. One broken checkout. AI coding agents can build features independently while making incompatible assumptions about shared behavior. Git can merge their code cleanly, and existing tests can still pass. The missing test is often the interaction between the changes. Jointly is an intent-aware pre-merge verification tool for developers, engineering teams using parallel coding agents, and reviewers integrating independently developed features. It combines IBM Bob’s reasoning with executable tests and a deterministic verification core. A developer opens Jointly’s custom mode in IBM Bob IDE and supplies two changes, their original task prompts, and a common Git base. Jointly freezes the commits and prepares four separate workspaces: the base, each change independently, and their combination. Bob analyzes both intents, identifies a risky interaction, and generates a focused test through Jointly’s MCP workflow. Our checkout demonstration combines percentage coupons with payment retry validation. Both branches and the existing combined test suite pass. But retrying a discounted payment returns HTTP 400 instead of the original successful response. Bob’s interaction test exposes the incompatible accounting assumptions. When requirements conflict, an authorized decision is recorded before repair. Changes stay in the isolated combined candidate; original branches remain untouched. Jointly reruns the unchanged interaction test, checks existing suites, and requires complete stability and verification evidence before issuing a scoped verdict. The browser dashboard makes the resulting Merge Safety Passport inspectable: exact commits, intent, failure evidence, requirement decisions, repair, and verification results. Reviewers can follow the evidence from failure to fix. Jointly’s distinguishing feature is connecting the intent behind two changes to an executable interaction test and a reviewable repair. Model confidence cannot substitute for passing checks, and missing evidence blocks a safe verdict. The current delivery supports the local Bob IDE workflow and browser evidence review on a trusted TypeScript checkout fixture. A watsonx.ai adapter is implemented and tested offline; hosted GitHub investigations and approved integration-PR publication remain the next milestone. Jointly complements code review and CI without claiming universal correctness or automatically merging changes.
27 Sep 2026
