I started using Pi with Astra on a whim after really enjoying Astra and reading somewhere that you get close to identical results as with the codex harness but for a significant hunk less token usage.

Coming from mostly using Claude models, the terse factual statements coming from Astra via the Pi harness are a breath of fresh air over having to wade through the flowery verbose nonsense that Claude constantly outputs

I recently had Astra review a fairly detailed design doc I have for an audio VST fork, that I originally wrote with Opus and/or Fable a few months ago. The doc reaches deep into signal flow and module topology while lifting most of the DSP code from other open-source projects. I had it review for feasibility and architectural soundness.

Astra found a number of flaws that would have come up during implementation and we worked through them. But then I had Fable 5.1 review that document and it found a number of issues with Astra's changes, the least of which had was that Astra duplicated a lot of technical notions that it added rather than using references to an authoritative section. It also flagged some of Astra's designs as technically impossible, pointing out why and I'm actually in the process of digesting its feedback and updating the design spec. (I hand-review each point and we work through a solution together -- I don't trust either model to come up with something that follows my vision on their own)

I'm not promoting one or the other, I just found it interesting how this sort of adversarial review found pretty significant flaws in the other model's work. I am curious as to whether this process will eventually converge on a document that both agree on or if the models are going to perpetually nitpick each other.

I haven't actually started implementation yet, so maybe one or the other is full of shit. Just trying to come up with an architecturally sound design for something I want to write, when I lack the DSP knowledge to be able to write it myself. But the intent is to pass an agent the design doc and list of milestones and let it handle implementation.

In my experience, rather than converging, you end up with a minced up concept. You have to know when to stop the loop. I filter the feedback too, and need to challenge some of the challenges as these models tend to be very conservative. I believe this is intentional, to control AI psychosis, which is indeed quite easy to get. My 2c. If you don’t have good control of what you’re working with you are either searching blind or end up with something basic.

Agreed, I believe strongly that human-in-the-loop is the way to go. LLMs are fantastic thought partners but ask it to critique something and it will go absolutely nuts. For Claude Code's /code-review command the highest I will set it is 'medium', otherwise it will produce so much feedback that you'd never get anything done.

Ignorant questions for you and anyone else using DS:

1. who hosts the inference

2. which harness are you using with it, still CC?

So everyone has their preferred way of doing it.

1. I go direct to source, i.e. DS platform, I find it cheaper than paying the openrouter tax -- I also switch it up a bit

2. I built a local LLM router, that I update with new profiles that have my preferred provider of the week (lowest token costs/speed) with fallbacks, like mimo --> DS4 etc.. if there is overloading,

3. I use 3 diff harnesses, CC/Codex + Opencode -- they all talk to each other through a custom rig system that routes messages between llms using a Rust backed structured JSON system

Not saying this is the best, it's just what I like and works for me^.

I can flow quite naturally between Opus/Astra/K3/GLM/MiMo/DS/etc.. this way and often do...more so these days with subs no longer great as they used to be.

doesn't DS train on your prompts when you go through their platform?

Whatever openrouter puts me on, running in pi (though I do all comms over my xmpp wrapper).

I'm using DeepSeek's own API with opencode. The pricing is absurdly good.

Deepseek harness is great!