I've built an orchestrator that solves some of the issues you ran into (although it doesn't do anything about cheating): https://navels.dev/blog/neal/. Features:
- lets you configure different models for planner, coder, and reviewer roles. (e.g., using Claude as an adversarial reviewer against Codex)
- breaks your plan up into reasonable-sized chunks of work with clearly defined success criteria
- runs each chunk of work through a coder / read-only reviewer loop. Once both agents are satisfied, neal moves on to the next chunk. Once everything is complete there is a final pass through the coder / reviewer loop to ensure the implementation satisfies the entire plan.
- resets the coder's context with each chunk of work to prevent context drift, leaving the reviewer's context long-running.
Following up on @killix's comments, which were helpful but he was flagged (presumably for sounding too much like AI).
Thanks for the feedback. I've made a couple of updates:
- Starting with 0.4.0, the reviewer gets the diff of the earlier chunk for any file the current chunk touches again. Also, if a new chunk weakens or removes a test or assertion from an earlier chunk, the reviewer will block it unless the plan says to do so.
- About read-only: The reviewer's tools were already limited by the SDK (no shell or write tools). However, it was still finding MCP servers from my Claude config. I have now blocked those. The docs now explain what is enforced by the system and what is just a prompt instruction.
I have accomplished code base (small size) migrations with it as well. Currently I do review each PR. For a large code base migration I think the core skills would still work but need a different way of driving it as you have come up with.
I don't think even the frontier models recognize something was produced by the same model in order to maliciously review it positively.
They may share some blind spots with the producer but generally I think they will review the other agent's output as harshly as they can if that is their task.
It says it can use any model/provider (but recommends Gemma?)
I must investigate. This is probably where 80% of my cognitive load comes from these days: the "language barrier". (How ironic!)
EDIT:
> You rewrite the assistant's message into much simpler, plain English. Keep every fact, name, number, and file path. Use short sentences and everyday words. Leave fenced code blocks unchanged. Output ONLY the rewritten message with no preamble, labels, or commentary.
> For context, the user asked the assistant: "$userq". Use this only to understand the message. Do NOT rewrite, answer, or repeat the user's question — rewrite only the assistant's message that follows.
The migration was not a single, long run of neal (because I was developing/tweaking neal along the way). However I have used it for large-ish implementations and haven't noticed any issues with the reviewer drifting. It is re-prompted each review turn with its adversarial review instructions plus a summary of the run state (completed scopes, open findings, etc). Also, the final review is executed with a fresh context.
same parent had said "It's crazy that a language that prioritizes security so highly in it's design itself is only compiled through such insecure methods."
I'm not an expert by any means, but everything I've seen of LLMs / machine learning looks like mathematical computation no different than what computers have always been doing at a fundamental level. If computers weren't AI before than I don't think they are now just because the maths they are doing has changed.
Maybe something like the game of life is more in the right direction. Where you set up a system with just the right set of rules with input and output and then just turn it on and let it go and the AI is an emergent property of the system over time.
Why do you have a preconception of what an implementation of AGI should look like? LLMs are composed of the same operations that computers have always done. But they're organized in novel ways that have produced novel capabilities.
Even as someone who is skeptical about LLMs, I’m not sure how anyone can look at what was achieved in AlphaGo and not at least consider the possibility that NNs could be superhuman in basically every domain at some point
> This website is not affiliated with the Proxmox VE Helper Scripts repository. This website is simply a redesign of the original website, with a focus on readability and security.
I'm assuming there's no malicious intent here, but tteck explicitly warns against these kinds of copycat sites.
I also use NginxProxyManager (8 hosts) and I'm not seeing any replies to your post that would explain why caddyserver or traefik provide any benefit over NPM.
- lets you configure different models for planner, coder, and reviewer roles. (e.g., using Claude as an adversarial reviewer against Codex)
- breaks your plan up into reasonable-sized chunks of work with clearly defined success criteria
- runs each chunk of work through a coder / read-only reviewer loop. Once both agents are satisfied, neal moves on to the next chunk. Once everything is complete there is a final pass through the coder / reviewer loop to ensure the implementation satisfies the entire plan.
- resets the coder's context with each chunk of work to prevent context drift, leaving the reviewer's context long-running.