codeSplitters: dependency doctor — scheduled dependency updates that build, test and fix themselves #194

On a schedule the user can tune, find outdated packages, try each update on a branch, build and test it in a throwaway Rusty Buns sandbox, and have an agent fix whatever breaks. People only see updates that are ready to merge.

Flow

1. Cron fires (a Workers Cron Trigger on the edge; Rusty Buns has no cron yet, so the desktop twin needs one).
2. Read the lockfile and check the registry for updates.
3. Open one line branch per update, or per group.
4. Build and test in the sandbox container, then tear it down.
5. If it breaks: an agent reads the changelog and the repo's line history (who calls this API, since when) and patches the affected code as line ops on the branch, blamed on agent-deps.
6. Green → a merge that's ready for review. Still red after N tries → one short report on why.

Settings to keep updates from piling up

• How often it runs, and whether to group updates (all patches in one branch, say).
• Policy by level: auto-merge patches, review minors, have the agent attempt majors.
• Minimum release age (wait N days after a version is published). This also guards against compromised packages.
• Ignore or snooze packages that release too often.

How it differs from Dependabot and Renovate: they open a PR and leave you to fix what breaks. Here the build, the test environment and the fix all happen before you're asked to look, and every changed line says which agent wrote it and why.

Depends on running tests in the sandbox (shared with Cut).

🤖 Generated with Claude Code