Project Secretary
Stops maintainers from repeating work already discussed: a Python service finds related issues and pull requests across GitHub repositories. Local embeddings and a graph-aware reranker separate true duplicates from overlap and useful context.
The problem Link to heading
In a busy backlog the same idea gets filed twice, the issue describing a feature lives in one repo while the code that implements it lives in another, and nobody has time to wire those connections by hand.
What I built Link to heading
Built a memory backbone over SurrealDB that syncs issues, PRs, comments, and Project v2 items, with a local-embeddings semantic layer and a reranker that classifies items as duplicate, overlap, prior context, or noise.
Implementation Link to heading
- Indexed issues, pull requests, comments and Project items in SurrealDB using repository-scoped identifiers.
- Combined local embeddings with labels, milestones and graph relationships to classify related work.
- Added review queues, explained label suggestions and a read-only MCP interface for agent access.
- Made enrichment writes idempotent and backlog-maintenance commands dry-run by default; documented remaining recovery and deployment gaps.