Skip to main content
After every implementation run pushes a branch, AI-Implement runs a second Claude pass that reviews the PR diff for merge readiness — bugs, missing requirements, unsafe behavior, test gaps, and other follow-up that should be fixed before merge. It posts its findings as PR comments and, when it finds blocking issues, has Claude fix them and re-reviews automatically before handing the PR to a human.

​
How the review runs

The review posts a starting comment, then reviews the full diff against the ticket. If it finds no blocking issues, it approves the PR and posts an ”✅ Approved” comment along with a native GitHub PR review. If it finds blocking issues, it posts them as a “⚠️ Reviewer found issues” comment, has Claude fix every listed item, and reviews the updated diff again. This repeats until the review approves the diff or the iteration cap is reached — 3 passes by default for the Anthropic provider, 2 for Bedrock. If the cap is reached without approval, the review posts a “⚠️ Reached review cap” comment summarizing what’s still blocking, and the run ends without approval (see Reference: error codes for what happens to the PR in that case). Each new pass’s comment carries a distinct marker, so a long-running review shows up as a sequence of comments on the PR rather than one comment edited in place.
If you’ve configured an external review provider (see reviewProviders in Customize pipeline behavior), the review waits for it to finish and folds its blocking findings in alongside its own before deciding whether to approve.
The review only runs when the implementation step successfully pushes a branch and opens (or already has) a PR. If Claude doesn’t get that far, no review comment is posted.

​
Model

The review uses the same model as implementation — there is no separate, cheaper model for it on either provider.

​
Triggering a gap-fill run

After reviewing the PR, you can ask Claude to address remaining feedback without starting a new implementation from scratch.
1

Open the PR on GitHub

Navigate to the pull request that AI-Implement opened.
2

Post the trigger comment

Add a comment that starts with /ai-implement on the PR. You must have write access to the repository.
Anything after /ai-implement is passed to Claude as an instruction for that run, as in /ai-implement also add tests for the error paths. Leave it bare to let Claude work from the existing review comments alone.
3

Wait for the run

AI-Implement checks out the existing branch and pushes gap-fill commits directly to the PR — no new branch or PR is created. The post-push review then runs again against the updated diff.
This requires the orchestrator’s GitHub webhook to be configured — see Configure webhook delivery. Without it, the comment has no effect.
The issue link, identifier, and description are all preserved from the original run — Claude sees the same ticket context it had during the first pass.
You can run /ai-implement multiple times on the same PR. Each pass reviews the latest diff and posts its own comment, so you can track progress across the comment thread.