Orange Codens is an AI code review product in the Codens suite. Its distinguishing feature, as described in “When AI reviews AI’s code, you’ve built an infinite loop. Here’s how we stopped it.”, is that it does not stop at review comments: it hands the fix for a finding to another Codens agent — Purple or Red — which opens a fix pull request, which Orange then reviews in turn. Because that arrangement puts AI on both sides of the loop, the product’s design centres on guaranteeing that the cycle terminates and that its cost is bounded.

Capabilities

Orange reviews pull requests through GitHub webhooks on open, synchronize and reopen events. Before a PR is merged it only posts GitHub suggestions; handoff of fix tasks is triggered on the PR’s closed event when it was merged, and queued handoffs are discarded if the PR is closed unmerged. Handoff policy has modes — the post shows a full-automatic mode and a threshold mode that dispatches findings at or above a configured severity and within configured categories — and a human can queue an individual finding for fixing, which dispatches even where automatic handoff would not.

Several guards keep the review-and-fix cycle from looping. PRs authored by the Codens bots are reviewed but never trigger automatic handoff; findings from reviews of the fixing agent’s own verify runs are not handed off; and a finding already handed off is never dispatched again. Findings in the same file bound for the same target agent are coalesced into one fix task, after per-finding fix PRs were found to block one another. For bot-authored fix PRs, the review blocks only on unresolved high-severity carry-over findings or new blockers, recording other new findings while approving; human PRs receive comments only and are never automatically blocked. The post also mentions a cap on review iterations, reached at five rounds in one incident.

Adoption & Ecosystem

The only account of the product here is its builder’s own post, which ends by promoting it. It works alongside the other Codens agents it hands fixes to, and is an instance of the configuration described on Closed-Loop AI Review — AI authoring and AI reviewing the same changes — with the termination problem discussed on Review Loop Non-Convergence.