I have been spending time on a small open source tool called agent context doctor, and the short version is that it is for the weird maintenance problem that shows up once a repo has more than one file telling agents how to behave.
Those files start useful, then they slowly collect stale paths, old commands, copied notes, conflicting rules, and little bits of local machine lore that no one meant to ship, which means the agent can be following context that feels official while quietly being wrong.
The tool is meant to give that context a basic health check before it wastes a coding session, trips over a missing script, or repeats a rule that made sense three months ago but no longer matches the project.
I am not opening the whole thing up yet, but it is close enough that I wanted to name the problem out loud first: agent instructions need upkeep like any other project surface, and right now most of us only notice they are sick after the agent has already acted on bad context.
No comments yet — be the first.
Comments are moderated before they appear. Plain text — blank lines start a new paragraph.