Skip to main content
Version: 0.1.147

Repave learning

After Repave delivers a feature, your developers often change it before they merge it. Repave learning reads those changes and learns from them: what Repave got wrong, and how your team works. It then suggests changes to this project, so the next feature needs fewer of them.

Repave learning is off for every project until someone turns it on in Project settings.

What Repave learns from​

Once a day, while learning is on, Repave reads your repositories with the access it already has. It only reads. It finds the changes your developers made between its delivery and the merge, and sorts each one into the project's Learning categories. Changes of the same kind become a lesson, such as "Log through AuditLog in every DataLayer endpoint, never console.log." Nothing is sent anywhere your features' code does not already go.

What Repave suggests​

From each lesson Repave suggests one of these:

  • Edit skill: a new version of one of this project's skills, with the lesson added where an agent doing that work would read it. The Claude Code copy and the Codex copy are suggested separately.
  • New skill: a skill the project does not have yet.
  • New validation rule: an advisory validation rule for a mistake Repave made. Repave suggests one only after it has checked it: the rule fails on what Repave delivered and passes on your developers' correction.

Some lessons have nothing to adopt. A lesson about Repave's own skills and prompts is recorded but changes nothing in your project, because those ship with Repave. A lesson the skill already follows needs no change.

Manual review or Auto-adopt​

When you turn learning on, you choose what happens to each suggestion:

  • Manual review (the default): nothing changes until someone adopts the suggestion on the Learning page.
  • Auto-adopt: Repave adopts each suggestion as soon as it is made. If it cannot adopt one, for example because the new skill's name is already taken, the suggestion waits on the Learning page and says why.

Switching from Manual review to Auto-adopt adopts the suggestions still waiting, the next time Repave learns. Either way, every change can be undone.

The Learning page​

Open Learning in the project's Management menu. The number beside it is how many suggestions are waiting for review. Only people who can edit the project see the number.

The list on the left sorts lessons by what is left to do with them: To review, Adopted, Rejected, and Learned, nothing to adopt. Choosing one shows the lesson, why the agent got it wrong, and Evidence: each change your developers made, with Show change to see its commit and diff. Below that are its suggestions.

Reviewing a suggestion​

Each suggestion opens on Changes, which shows exactly what would change. To adjust it before adopting it, open Edit. For a skill this is its SKILL.md, one copy per tab. For a new skill it is the name, description and instructions. For a rule it is the name and the rule, and an edited rule is not checked again against the evidence.

  • Adopt makes the change as shown, your edits included. The suggestion then says who adopted it and when.
  • Reject sets it aside, and Repave does not suggest it again. Review again brings it back.
  • Undo, on an adopted suggestion, puts the version before it back in use, removes the skill it added, or removes the rule. Repave does not make that change again.

If someone changed the skill after Repave made the suggestion, the suggestion is marked Outdated and cannot be adopted. Repave suggests it again on top of their version, so nobody's edit is overwritten. People who can only view the project can read every suggestion but not decide them.

When learning is off​

Turning learning off stops Repave reading your repositories and making new suggestions. Everything it already learned stays on the Learning page, and you can still adopt, reject or undo it.