Groundtruth – checks your AI coding agent's claims against the Git diff

github.com

3 points·erapin_game··4 comments

From https://github.com/akahkhanna/groundtruth :

> Groundtruth — a Claude Code plugin that audits whether an agent actually did what it was asked; catches the false 'Done' on Stop (missed subtasks, stubs, false 'tests pass', overridden rules).

Which other agent IDEs does or could this work with?

Is this also where to attach provenance metadata and sign the agent trace, and debug?

Re: MCPSnoop, proxies like Aegis and LiteLLM, and awesome-auditable-ai: https://news.ycombinator.com/item?id=48777144#48779413

There are web standards for signing metadata with interoperable schema as linked data:

W3C JSON-LD/YAML-LD + W3C PROV + W3C DID + W3C VC Verifiable Claims

"Follow up to verify that the work was actually satisfactorily completed"

Are there other sound management practices that aren't yet effectively implemented in current gen agents?

I've had some success with using a more expensive model to plan a decent work breakdown structure in a document and then a low-cost model to implement the plan and keep following-up, but wonder about subsequent need for code review and code quality.

Can a lower-cost model verify completion? Iff tests and test coverage and e2e tests?

The same oracle / model routing and partitioning problem

Reply on news.ycombinator.com