mijoharas

Maybe I need to try it again, but I have never liked the opus models.

I keep having problems with them not doing the thing I ask, (and also failing to call tools correctly in pi. I think they've very much trained to the harness, which is frustrating.).

brent crude futures are trading above 109$/barrel right now. America didn't want that to happen, Iran did. They won that one.

So... Anthropic reinvents SVN?

Interested to try this out.

I played around with Freecad mcp, and then also with having an agent generate some openscad before.

For the Freecad mcp I didn't like it at all. The agent didn't work in a way that would let me easily modify things. I hadn't given it instructions, but it didn't really create sketches and then extrude them (after checking they're all fully parameterized.)

For openscad the problem was just that while the model could do it, the lack of "visual reasoning" (for want of a better word) really held it back. There was a nice workflow there though by giving it a little "generate_image" of part tool and it could make a nice iterative loop.

Overall for both approaches it was slower than if I'd done it myself (and I'm not particularly good at 3d modelling). Now, that's not that surprising given that I was trying something new, but without being able to edit things generated by the AI myself I wouldn't want to try it again.

Now, maybe this solution can work in the way I'd want to. I may give this a go next time I have something I want to make.

so, how are people running these locally? is there a llama.cpp/ollama solution for systemone/jev style apis?

UK too.

not at all, it's nice! solves the composability problem while keeping a trust separation between harness/ai bash calls.

I agree with most of what you said, but wanted to add that another point is there is a different level of trust in code executing from the harness, and code executing from the bash shell by the agent.

I block all network access from bash calls, but want to provide certain tools that can call certain services in a specific way. I can have a tool (or mcp) that does just that, and runs outside of the sandbox.

> There is no "initialize a new thing on push" and there never has been

Well, there hasn't ever been a real reason for it, and now there is?

Ooi, can I ask for the reasoning for the new default of fullscreen?

The codemode stuff/mcp has been explained to death at this point but I haven't seen anything around fullscreen.

(I don't have strong feelings either way, just curious)

I was thinking the same in terms of similarity to org-babel (which can make some very nice runnable notebooks).

To be clear, is the issue with org-babel that it's not accessible to other people on the team? or something else that this solves?

I find it pretty handy. I run any bash tools in a restricted sandbox, and so that it cannot hit the network (usually). I don't want any credentials to be available to the agent, and my trust boundary for that is precisely the harness vs. the agent. Harness can touch secrets, agent can't (filesystem is restricted from it being able to read any secrets as well).

that's my one problem with the cli vs mcp debate. I prefer cli, because they're tools I'm familiar with, and they're composable. My problem with it is some cli tools need to use secrets to access things. By making sure they only go to the harness, then I've got my problem solved.