Context
AI-assisted codebases can accumulate unclear contracts, duplicated work, and changes that are difficult to trust. RepoDiet is designed around the maintenance workflow, not merely the act of generating a patch.
What was built
RepoDiet supports repository triage, contract-scoped cleanup, isolated branches, independent verification, review-ready pull requests, and signed receipts. It frames maintenance as evidence that a specific change was checked against an agreed boundary.
Architecture and workflow
The workflow uses GitHub-oriented automation, A2MCP and A2A patterns, and an evidence trail for the work that happened. The product is structured so a reviewer can see the scope, verification, and resulting artifact rather than accept an untraceable automation claim.
Selected technical decisions
- Scope changes to an explicit contract before cleanup begins.
- Separate implementation from independent verification.
- Preserve a review-ready trail rather than reporting an uninspectable success state.
What this demonstrates
This work demonstrates the ability to take a complex domain, turn it into a legible workflow, and connect interface decisions to data, systems, and real execution paths. For a client, that means a build can begin with the actual problem instead of a generic feature checklist.