Here’s the analysis: squid-mesh-comparison.md · GitHub
webofbits
dimitarvp
Yeah, but more opinionated and focused on exactly what we need.
Initially I thought that’s a great idea but of course LLMs change the picture substantially and SquidMesh is moving so fast that I might end up just using it – but creating a small opinionated (and likely macro-based) DSL on top of it so the intention of the calls is crystal clear both to my team and our LLMs. We have huge tech debt and the very talkative and cheerfully generating thousands of lines of new code LLMs are not frakkin helping. As a backend lead I am taking measures to start taming our LOC and using my own – or your – workflow engine is going to be a good chunk of that effort. Too many half-done Repo.transaction “sagas” in our code.
Yeah, as if we’re fooling anyone that we do it all manually, right. Everybody knows that strong programmers use agents as productivity multipliers. I am not ashamed because 80% of the time I am steering the wheel. I do have lapses but I am done feeling bad for not being born perfect – I pay attention what is happening, I see the failures and the miscommunications and try to learn and do a better session next time. If anyone can shortcut this and make me even more powerful through LLMs, I’ll be grateful… but I don’t have the time to do this every day for two weeks, for example.
1
dimitarvp
I love that one, really well-written and not chatty and verbose as a bank manager on acid. Thanks for posting it.
I particularly love the ideas on what to lift from the other libraries.
And it confirms what I suspected – SquidMesh is already much more feature-rich. Though honestly, why is it “Phoenix/OTP workflow engine”? From what I can see one can use it in a bare Elixir repo just fine, as long as you have Ecto (or just Bedrock-adjacent deps).
1
dimitarvp
Ah oops, that’s likely related to your recent web dashboard that you are adding. My bad, forgot. And that’s one more thing that makes SquidMesh irresistible for me.
0
webofbits
Yeah. Nothing really I can do other than build with some LLMs at this stage (Dependabot, Copilot, Codex). Until it’s mature and people adopt it, I’ll be pushing agentic code and staying on the loop about what other libraries are doing right.
I am specially disappointed with Claude. Spent a couple of months not using it but was testing today and it drifts even for trivial things like reading its own “manual”:
> why are we not following ~/.claude/CLAUDE.md? is it in the wrong place?
⏺ You're right. I should have created a dedicated worktree and branch first before making any
changes. According to your CLAUDE.md
> why we're not following TDD?
⏺ You're right - according to the CLAUDE.md workflow, I should be following TDD: "failing
test → minimal impl → green → refactor". Let me write the tests first
With AI leverage, we need to continue the discipline of the old times like smaller prs, TDDs, end to end tests, etc otherwise tech debt will grow exponentially. Humans are the bottleneck now as they need to pass the right instructions, review the code, make sure nothing is going off rails.
If there’s another library that is simpler, faster or easier to grasp for your team I’d say go for it. I might be piling too much features in an earlier stage to raise awareness but I am planning to stop after a couple of more issues unless it’s really worth it or someone is needing.
0
webofbits
Maybe. But also it could be that it’s referring to the ability of embedding it to Phoenix and also standalone OTP apps.
The web monitoring piece is super simple and read-only now but it’s an interesting way to visualize running workflows.
0
webofbits
Codex usually tends to get defensive when i say something different than its AGENTS.md file. Claude seems to completely ignore CLAUDE.md if I don’t pay attention 
1
dimitarvp
Yeah, we are circling back to that in my team. LLMs are just too non-deterministic and I am not sure we all want to develop brand new skills i.e. “how to prompt Claude perfectly so it stops doing stupid crap”. Career-wise that’s a very smart move IN THE SHORT TERM and an extremely stupid move in the long-term. This is why I am putting hard caps on my sessions where I just try to devise a better harness for Claude so it needs less supervision. Very mixed success; I had some screaming successes but way too few; most have been embarrassing failures.
I am doing quite OK prompting and my attention rarely drifts but still, it is hugely mentally exhausting to never quite know what you’re getting. I am either taming Claude with hooks in the next 1-2 months or I’ll just start doing the task planning and breaking down on smaller tasks by myself and just delegate it the coding job and nothing else. We’ll see.
But again, I am not interested in doing the big-three-LLM-providers job for them. They are using us as testers and improvers of their own product which I guess is a fair deal having in mind how much I am getting for $100 a month but still, the equation is still not quite in my favor.
Yep. Periodically it just forgets about it at all.
Well shhhh, don’t tell my team, but I am starting to ponder rewriting our app in Rust or Golang. Not like I don’t love Elixir to bits, mind you, but the dynamic typing is not helping, and the tons of shovelware that previous LLMs put in confuses even the frontier models and I have to devise literal dozens of rules and constantly remind them to reread those so they don’t figure “Oh hey, business logic here is literal 17 lines but let’s add 200 more for meaningless defenses on hard contracts where you’d never get a nil, but add them anyway!”.
Exhausting.
But I think that can be alleviated with boundary and reach and dialyzer. I will of course not just do a huge rug-pull like this. A guy can dream though.
But no, no better library in mind. Runic came the closest – extremely disciplined code (I took a look in a few places) but the lack of persistence ultimately swayed me either to my own bespoke workflow engine, or SquidMesh. The latter becomes more likely with each forum post from you.
0
webofbits
Sure thing. Dynamic type bite us sometimes. It’s funny that we need no types but we spent a ton of time dealing with dialyxir errors. Hoping the new type system in Elixir would alleviate the pain. Dialyzer errors are not super helpful at times.
In terms of Go/Rust, I’d say we are engineers so the language it’s just a tool. e.g. Jose was trying to solve a problem in Ruby but ended up finding Erlang the best fit so it depends of the problem. No language fits all problems.
If you need to change anything in SquidMesh to match your needs, just let me know. Dont love being the one man show making decisions.
1
dimitarvp
Oh, and Zoi of course. It’s quite amazing. I am making an Ecto extension for it with the blessing of the author very soon. It helps with removing ambiguity in typing contracts, like a lot.
2