Repository intelligence
Understand Repository Learnings, Codebase Index, and Deep Code Scan without confusing their roles.
Kensa keeps three kinds of repository knowledge separate: rules your team wants reviews to honor, a descriptive map of the codebase, and an optional way to propose new rules. Keeping them separate makes it clear what can affect a review and what is only context.
Know which tool answers which question
| Term | What it is | What it changes |
|---|---|---|
| Repository Learning | A repo-specific rule or decision that Kensa should remember for future reviews. | Active, relevant learnings guide review judgment. |
| Learning Signal | Evidence that might become a learning: an instruction, review outcome, or feedback. | Nothing until it is turned into a learning. |
| Codebase Index | An AI-generated map of modules, technologies, conventions, and paths. | Background context for review agents. |
| Deep Code Scan | An optional analysis of written rules and representative code. | Proposes new learnings for people to review. |
A Repository Learning changes what Kensa looks for; it is not a writing-tone preference or a replacement for a skill. Suggested Learnings do not influence reviews until you approve them. A Stale Learning is quarantined when its supporting evidence changed or disappeared, so it stops affecting reviews until someone confirms, edits, or archives it. Supersession keeps the older learning as history while its replacement becomes the rule to use.
Use the Codebase Index for orientation
Generate or refresh the Codebase Index when Kensa needs an architectural map of a local clone. It describes the repository's structure and conventions so agents can interpret a diff in context. A Codebase Index describes the repository; it is not a source of Repository Learnings. It is kept separately from the Learnings view and can be refreshed when the repository context changes.
Use Deep Code Scan to propose rules
Run Deep Code Scan when you want to mine durable review knowledge from a repository's own documentation, configuration, and representative source files. It is not part of a normal review. Deep Code Scan is opt-in and shows a cost estimate before it runs. Its output is proposed as Suggested Learnings, so you can inspect the evidence and decide which rules become active.
Practical sequence
- Add a clear Repository Learning when you already know a rule the reviewers should honor.
- Keep the Codebase Index current when the project structure changes materially.
- Run a Deep Code Scan only when you want candidates to review, then approve or archive them.