Userz

Review loop

The agent opens a draft PR. You read the diff. If it isn't right, you tell the agent — inline — and it tries again on the same branch.

The shape of an iteration

  1. Feedback arrives, the agent opens a draft PR with a summary and a warning banner.
  2. You open the PR in Userz (or in GitHub / GitLab). Leave per-line diff comments wherever something needs changing.
  3. Click Request revision. Your comments are bundled with the original feedback as the agent's next input.
  4. The agent re-runs, commits on top of the same branch, and updates the PR.
  5. Repeat until you're happy — then merge. Or close, if you've changed your mind.
Each iteration is one full agent run with its own audit row, tokens-in/tokens-out, and cost line. Revisions count toward your plan's monthly PR limit only on the first opened PR; iterations on the same PR are free.

Diff comments

Comments are anchored to a file + line and survive the agent's edits to that file. Each comment has:

  • path: the file the comment is attached to.
  • line: the line number in the post-diff view.
  • body: your prose to the agent. Markdown is fine.
  • severity (optional): nit, suggestion, or must_change. The agent treats must_change as non-negotiable; nit as advisory.

Comments you mark as resolved in the dashboard are excluded from the next iteration's context. Unresolved threads stay visible to the agent until they're either resolved or deleted.

Revision requests

A revision request is a single batched message that the agent reads alongside its diff and the original feedback. Pattern:

What works

  • "Fix the failing test cases — they're failing because of X."
  • "Reuse the existing formatCurrency helper in lib/money.ts instead of re-implementing the rounding logic."
  • "Pull the loading skeleton into a shared component."
  • "Don't touch the wire format — just fix the rendering bug."

What doesn't work as well

  • "Make this better." (Be specific about the property you want.)
  • "Refactor everything." (One PR, one feedback report — refactors that cross the feedback's scope tend to get reverted in the next iteration.)

Plan caps

Each plan caps how many revisions you can request per PR before the loop closes:

  • Free / Trial: 2 revisions per PR.
  • Starter: 5 revisions per PR.
  • Pro: 10 revisions per PR.
  • Enterprise: unlimited.

The cap exists to keep agent costs bounded — if you're hitting it regularly, the underlying feedback probably needs to be split into smaller items, or the model you've configured isn't strong enough for the change you're asking for. You can switch the agent model independently from the helper model under Settings → AI.

Auditing iterations

Every PR's Runs tab shows each iteration's agent run: which provider + model ran, tokens consumed, cost, the comments that drove it, and the resulting commit SHA. You can open the full agent transcript for any past run from the same view.

If you're calling the REST API, GET /v1/agent-runs/:id returns one of these rows.