Skip to contentSaltar al contenido
Kensa documentation

Run your first PR review

Open a pull request, start a Quick or Deep Review, and read your first Kensa findings.

With a repository connected, you can review any pull request from the Pull Requests view. A Quick Review analyzes the diff alone; a Deep Review lets agents read the repository on disk for deeper context.

Quick path

  1. Open the Pull Requests view and pick a PR — search, filters, and column sorting help you find it.
  2. On the PR page, choose Start Quick Review — or Start Deep Review when a local clone is attached.
  3. Watch the agents work in the Review tab; each specialist reports progress live. You can Stop at any time.
  4. Read the findings when the review completes.

Expected result: findings appear in the Review tab grouped by file with severity badges. The review is cached locally and restored the next time you open the PR.

Quick vs. Deep Review

Quick ReviewDeep Review
What agents readThe PR diffThe diff plus your local clone
Needs a local clone pathNoYes
Speed and costFast and cheapThorough, more turns

Without a clone path, Deep Review stays disabled behind a Set up local access action. On the PR page, Deep Review for a local-type repository asks for confirmation before reading your working copy — including uncommitted changes.

What happens during a review

  • Kensa pre-selects the specialist agents that match the PR's files. Under Review Settings you can adjust the agent set, add a one-off Focus instruction, or override the provider and model for that run.
  • Agents run on your machine through your configured provider — such as Claude Code, Codex, or Gemini CLI — with limited concurrency.
  • When the agents finish, a consolidator deduplicates overlapping findings, and a Learn analysis adds background context in the Assist tab.
  • Every finding carries a severity, the producing agent and provider/model, and the file and line it refers to.

Severity levels: critical (security, data loss, correctness), major (logic flaws, missing error handling), minor (quality and conventions), and info (observations and notes).

If something goes wrong

  • Provider errors mid-review (quota, authentication, network) — a banner names the failing provider and hints at the cause.
  • Some agents fail — the review ends as incomplete; use Retry pending for the failed agents or Accept as-is for a partial review.
  • New commits land after a review — the PR list marks the review stale; run it again. The previous result stays available in the review history.

Next step

Interpret findings and navigate the diff — or review local changes that are not a pull request yet.