Danny Tuppeny

@luke_pighetti I don't know the details, but maybe allowing they agent to trigger it coud be beneficial? (it wouldn't happen unexpectedly while it's making changes, and it can trigger just once after many edits?) Guessing though, haven't tried using AI for Flutter apps yet!

@DanTup only a google engineer could overthink just adding a global dart environment setting that enables reload on save

refusal to trigger flutter hot reload on file save looks pretty dumb now that hot reload is totally broken for agentic coding workflows

@luke_pighetti Ah, I see. I don't know the details, but maybe github.com/flutter/flutte… is something in this area?

@DanTup Claude CLI

@anaisbetts Although, executing commands falls back to non-sandbox automatically, but browser use does not, so you need to explicitly disable sandbox for that to work

@anaisbetts It wsn't, that gave an error about it being invalid. With v1 it just complained about no compatible models.. But see follow-up, there is an option for local LLMs in settings (it's just not in the onboarding). All up and running now :)

Has anyone successfully set up an OpenClaw'alike with a local LLM model? Even with a 27B model it seems too Dumb, and it's incredibly slow too (~7tok/sec on Strix Halo)

@anaisbetts FWIW, shell commands worked fine without me configuring anything (ofc. they're running in the main container, I'd rather that than any part of this own the host!)

@anaisbetts "If you cannot mount the Docker socket, Moltis will run in “no sandbox” mode" sounds like it just always does this though. I filed an issue that the docs are confusing anyway, and I'll just test it out.

@anaisbetts FWIW, with Qwen3-Coder-Next it says ~28t/s with all defaults, just a few prompts asking for some code and poems

@anaisbetts Turns out there is this option once you've gone through onboarding (by skipping the LLM) and then go into settings. Not sure why this isn't in onboarding 🙃

@anaisbetts hmm, wait - these seem to contradict? One says they'll fail, other says they will run. I guess I will have to try it out...

@anaisbetts I'd seen the name but not looked at it, will have a look. Simpler = good IMO, as long as I can extend it if needed. It'd just be nice to have a base to start from and not have to implement everything myself (I've started doing so a few times, and keep aborting 😄)

@anaisbetts This looked good, but fails for me because of this. Giving it the Docker socket is basically giving it root access to the host? #docker-socket-sandbox-execution" target="_blank" rel="nofollow noopener">github.com/moltis-org/mol…

@DanTup So far Moltis has been relatively sane though a bit bare bones

@luke_pighetti It was a joke 😄 Some people are putting a little too much trust in AI!

@DanTup i haven’t changed my CI strat so yes tests and linters. you?

CI is too slow in the age of agentic coding

@mraleph "If helpful, I can try to provide a minimal repro" Like the one already posted in the original report it was replying to? 🤔

you ain't fooling me mr. I-use-slop-generator-to-make-what-I-think-are-useful-and-thoughtful-comments

@anaisbetts Yeah, I was surprised/disappointed at this response. If your bot username is guessable (or searchable?) or you post a screenshot, and one can extract all your info, use your LLM, execute terminal commands on your box? (and I suspect not everyone runs these in containers 😞)

@DanTup "Regarding access control: you're not missing anything. CoPaw is currently designed primarily as a personal assistant, so we don't yet have built-in user permission features like allowlists to restrict who can DM the bot." That is *nuts*. Every bot has this

@anaisbetts I want something I can run entirely in a container (which rules out some that want to launch their own containers for tasks), in a language I'm familiar with, and that doesn't have huge security issues (CoPaw seemed decent until I tried telegram... -> #issuecomment-3987841004" target="_blank" rel="nofollow noopener">github.com/agentscope-ai/…)

@anaisbetts Yeah, I haven't tried OpenClaw because of the boasting of unreviewed AI code, but I do like the idea of delegating some tedious life tasks to something so searching for something similar but smaller/simpler (that works well w/ local LLM).

@anaisbetts For Qwen3.5? Haven't tried it yet and hadn't thought about that. I wish there were more benchmarks for the quants that fit on 128GB machines (both speed and intelligence - it doesn't seem reliable to compare the unquantized scores)

@DanTup For whatever reason, none of the documented ways to disable it with llama.cpp seem to work