These are just some of the updates we've shipped at Kodus over the past few days.
There are a few more improvements we didn't cover in this thread. You can check them all out in the full changelog.
And if you're already using Kody, I'd love to hear what you think of these updates. 🧡
docs.kodus.io/changelog
📂 Review Every Repository
Another update that should save a lot of time, especially for teams managing multiple repositories.
You can now enable Review every repository, and Kody will automatically review all repositories accessible through your Git integration.
That includes repositories created in the future.
So there's no need to manually add every new repository to Kodus.
And if there's a repository you don't want Kody to review, you can simply disable automated reviews for that specific repo.
Works with GitHub, GitLab, Bitbucket, Azure DevOps, and Forgejo.
✅ Approval Policies
We also made auto-approval a lot more flexible.
Instead of relying on a single configuration, you can now create multiple approval policies, each with its own conditions.
You can define rules based on:
・PR labels and authors
・Teams and groups
・Included or excluded files
・Number of changed lines
・Custom instructions Kody needs to verify
For example, you could create a policy that automatically approves dependency updates, but only if they stay within the same major version.
And if Kody identifies any risk or the PR requires human approval, it won't be approved automatically.
🔀 Routing
Not every PR needs the same level of review.
You can now configure Kody to automatically choose a review mode based on the risk it identifies.
For example:
Low-risk PRs → Balanced
Risky PRs → Heavy
This lets you use faster, more cost-effective reviews for everyday changes while saving deeper reviews for PRs that actually need extra attention.
You can also configure different models for Balanced and Heavy at the organization level.
🚨 Risk Assessment
Kody can now assess the risk of every PR, even when auto-approval is disabled.
You can also define custom risk criteria based on what matters to your team.
For example, changes that affect user permissions, delete customer data, or touch critical parts of your application.
Each criterion can have its own instructions and apply only to specific files.
And before enabling anything, you can test your criteria against a real PR without posting comments or approving it.
We shipped some pretty interesting updates to Kodus this week. 🐕
The biggest change is how Kody assesses risk, chooses review modes, and automatically approves PRs.
Teams now have much more control over which PRs need a deeper review, which ones can be approved automatically, and how the whole process works.
We also made it easier to enable automated reviews across all repositories, including new ones.
Here's what's new 👇
These are just some of the updates we've shipped at Kodus over the past few days.
There are a few more improvements we didn't cover in this thread. You can check them all out in the full changelog.
And if you're already using Kody, I'd love to hear what you think of these updates. 🧡
docs.kodus.io/changelog
Review Every Repository
Another update that should save a lot of time, especially for teams managing multiple repositories.
You can now enable Review every repository, and Kody will automatically review all repositories accessible through your Git integration.
That includes repositories created in the future.
So there's no need to manually add every new repository to Kodus.
And if there's a repository you don't want Kody to review, you can simply disable automated reviews for that specific repo.
Works with GitHub, GitLab, Bitbucket, Azure DevOps, and Forgejo.
Approval Policies
We also made auto-approval a lot more flexible.
Instead of relying on a single configuration, you can now create multiple approval policies, each with its own conditions.
You can define rules based on:
・PR labels and authors
・Teams and groups
・Included or excluded files
・Number of changed lines
・Custom instructions Kody needs to verify
For example, you could create a policy that automatically approves dependency updates, but only if they stay within the same major version.
And if Kody identifies any risk or the PR requires human approval, it won't be approved automatically.
Routing
Not every PR needs the same level of review.
You can now configure Kody to automatically choose a review mode based on the risk it identifies.
For example:
・Everyday, low-risk PRs → Balanced
・Risky PRs → Heavy
This lets you use faster, more cost-effective reviews for everyday changes while saving deeper reviews for PRs that actually need extra attention.
You can also configure different models for Balanced and Heavy at the organization level.
Risk Assessment
Kody can now assess the risk of every PR, even when auto-approval is disabled.
You can also define custom risk criteria based on what matters to your team.
For example, changes that affect user permissions, delete customer data, or touch critical parts of your application.
Each criterion can have its own instructions and apply only to specific files.
And before enabling anything, you can test your criteria against a real PR without posting comments or approving it.
we just updated our LLM code review benchmark with new models.
we've now tested 13 LLMs on the same 30 real-world PRs, covering 121 confirmed bugs.
here's the top 5 in Balanced mode:
Kimi K3.1 Pro
GPT-6.1 Sol
Claude Opus 5.5
GLM-5.3
DeepSeek V4.1 Flash
Kimi took the top spot with 68.6% recall, well ahead of the other models.
another interesting finding: DeepSeek came close to Opus in recall (40.5% vs. 42.1%) while costing over 5x less per PR.
you can also compare precision, cost per bug found, and results with specialized agents.
full leaderboard: codereviewbench.com/leaderbo…
A nova funcionalidade de agentes personalizados da Kodus ficou boa demais!
Antes, tínhamos um fluxo que executava um security review local após a abertura de uma pr, para validar com a Kody se um achado era realmente um problema.
Agora, com os agentes personalizados, consegui transferir essas instruções para um agente e eliminar essa etapa da minha sessão.
E como esses agentes também aceitam skills, aproveitei para testar uma adaptação das skills de segurança da Cloudflare (aquela que viralizou esses dias) e incorporei ao agente.
Até o momento, está funcionando bem 🎉
context is a liability, same as code
give a coding agent an AGENTS.md and it costs 20% more to do the same task, and does it no better
this is why we built self-pausing rules. the ones your team ignores or downvotes get turned off