Code Reviews
AI-assisted code review on every pull request, with a visual report in the app
Rock8Cloud can automatically review pull request diffs with AI. Findings are posted on the PR as a GitHub comment, and every review also produces a visual report in the app.
Enable
- Open the service Workflows tab
- Enable the AI Code Review workflow
- Configure which PR events trigger it (defaults to opened and synchronize)
Reviews run on every matching PR and post results as a comment on the pull request.
Trigger Behavior
- Opened - review runs when the PR is created
- Synchronize - review runs again when new commits are pushed
- Branch patterns control which PRs are reviewed (defaults to
*- all branches)
The Visual Review
Open the service Code Reviews tab to see the full report. The GitHub PR comment, check run, and channel notifications all link straight to it.
Each review is a document built from the actual changes:
- Summary - what the PR accomplishes, in plain language
- Changed files - a tree of every touched file with change flags. Click any file to open its diff in a side-by-side browser
- Data model changes - shown only when the PR touches database schemas
- API changes - shown only when the PR adds or modifies endpoints
- Architecture - a diagram of how the changed components relate, shown only when the PR changes structure
- Diff walkthrough - the important hunks grouped into narrative tabs, with findings pinned to the exact lines they refer to
Every section collapses, and reviews from older versions without a visual report fall back to a plain findings list.
Fixing Findings with the Agent
You can hand findings straight to a coding agent that works on the PR branch:
- Check the findings you want fixed, in the diff or in the findings list
- Optionally click any line number in a diff to leave the agent an instruction on that line
- Click Send to agent
The agent starts a session on the PR branch, applies the batch, and pushes a commit. The push triggers a fresh review automatically. The review shows the agent's status with a link to its session in the Agents tab, where you can keep chatting with it about the same PR.
Sending again reuses the same session, so follow-up batches become new instructions to the agent you already briefed.
This requires the coding agent feature to be enabled for your organization. When it is off, the review stays read-only.
Learnings
The reviewer can be taught what not to flag. Code Reviews > Learnings holds per-service instructions that are appended to every review run on that service.
Two ways to add one:
- From a finding - dismiss it, check Don't flag this again, and edit the instruction it proposes
- By hand - open the Learnings sub-tab and write the rule yourself
Active learnings go into the reviewer's prompt on every run. Toggle one off to pause it without losing the text, or delete it outright.
Good learnings are narrow and about your codebase, not about style in general. "Don't flag missing JSDoc on internal helpers, we only document exported APIs" works. "Be less picky" does not.
Coding agents have their own separate list, see Instructions.
Reading Results on GitHub
Review findings also appear as a GitHub PR comment with:
- Issues grouped by severity
- File and line references
- A link to the full visual review in the app
Treat automated findings as guidance - always validate before merging.
Notifications
If the workflow has Notifications enabled, your project's notification channels receive alerts when reviews complete or fail, including the link to the visual review. See Notifications.