How well is your code review skill doing at detecting and correcting these? History in comments in git is so annoying! As is missing why A didn't work but B does. None of the models I've tried get this right
I like it because it's a terminal first workspace and has bindings to use it programatically. Tried Superset and Intent, but they seemed like way overkill / unintuitive to use agents like I prefer. I use WezTerm as a terminal, so being able to have herdr exist in that application (not having another app to keep track of) as a tab really helps me stay in flow. Over the last few weeks, Herdr has given me the vibe of "works out of the box, but is highly customizable", and I think that power-user angle really sold me.
Oh, and I didn't know how much I'd love session persistence. I already use tmux for every other terminal process, so it being built in to herdr and a focused feature (server start, detach, stop) is nice to see.
I also tried Herdr. But I'm already using Aerospace, which makes MacOS window management a breeze and I don't need the remote session management stuff so the benefit wasn't quite there for me. It is a very thoughtful project so I did borrow lots of its features (Agents working now get an animation in my sketchy bar, a soft chime to indicate a finishes status).
> Why? Because it has more/better features. Or you mean something else?
I mean something else. Zoom out, why do people need, or want, a tool like herdr?
What has led people to this point? Is it because they need terminal multiplexing to spawn agents, instead of using built-in subagents? Do they like cute sheep icons? What made having multiple terminals open unappealing? Is it the built-in alerting when attention is needed? Is it only applicable to people running their agents on remote machines?
When I look at these kinds of tools, that is the question I can't answer because I suspect I haven't run into the pain points they address yet, despite running multiple LLM sessions in multiple projects simultaneously.
1) Built-in sub agents assume you use a single AI agent. Some people use multiple, so they need separate terminals/processes by definition. Even if you use the same agent sometimes you have different security boundaries. Think also the needs for a consultant working with many customers at once.
2)Multiple-open terminals. When that number passes 5-7 for me, then yes I don't like clicking on each one of those manually to see where my attention is needed. Herdr (and similar tools) show exactly what is working and what stalled.
3) No I use herdr locally mostly right now. But since it is just ssh I like the fact that you can run it from anywhere. I haven't done it yet, but I imagine you could open a herdr session from a steamdeck to a remote server to debug somthing
Another thing I like on herdr is easy work trees. Sure you can do it manually with git commands, or ask the agent itself to use a new worktree, but just having workrees with a single click and a visual tree/children hierarchy was the killer feature for making me look at herdr at the first place.
Yes technically you can do the same thing with just many terminals and custom scripts. But it is the same question of why Dropbox sells when you could do the same thing with cvs/ftp :-)
Totally fair, maker of cmux here. We fixed many RAM usage issues (latest version is 0.64.22), and also added an activity monitor to see how much RAM/CPU coding agents are taking.
My cmux has been running with 136 workspaces and 218 terminals, and has been hovering around 5.7 GB. Granted, the theoretical RAM minimum for this number of terminals is around 1-1.5 GB before real scrollback, so we definitely have more work to do.
Thanks, good to know. I installed the update and will check it out again.
If it helps, the main reason I love cmux is how can I manage my projects into folders, I have fast access to all the projects which made my life a lot easier working in the terminal. I can also have coding, test, monitoring spilt in one tab per project.
It’s incomplete but it show cases some of what I mean. There’s just a lot more charm in these earlier graphics. I’m fairly certain they had to consider how the different colours of ink bleed on one another.
Like take 1978 for example, this might be some kind of highlighter ink, and when the colours over lap the grid they interact in ways most overlapping geometry don’t in charting libraries (but in CSS you can get a similar effect with mix-blend-mode: multiply)
If you go to page 159 of the most recent RBA report, you’ll see otherwise clean clear charts that do a good job of communicating the relevant information (perhaps that’s all there should do from such an institution), but it’s admittedly soulless compared to the charts from previous decades
not the parent, but one thing i find charming is that a lot of them have a really good feel for line weight and inking style, and vary it interestingly in various parts of the plot. computer generated charts tend to have a much more uniform line weight.
> Fable 5 is sick. [It] solves almost any task from start to finish, the way it should be done — no errors, perfect code. It's a miracle.
> As for GPT-5.6 Sol — it doesn't even come close. I honestly don't understand why people even try to compare them.
What kind of problems are you working on? I like Fable but when planning work on a complex C codebase it's making more mistakes than 5.6 Sol xhigh for me.
In what scenarios is Fable giving you "no errors, perfect code"?
I have a large monorepo that includes about 15 TypeScript services and many Rust services. Everything is well-documented and organized, with standardized and structured custom code.
When an issue arises, I often test the systems by providing a minimal prompt, like: "this user, this is their email, this isn't working, figure it out in production." I send this to both Opus and ChatGPT, but it doesn't help. I've set up Agents.md and Quote.md identically, with the same access and linkers, so the Harness is consistent.
ChatGPT rarely succeeds. If the task is complex and requires a multi-step process to identify the true cause, ChatGPT usually stops after a few initial ideas and wrongly claims it has found the solution.
- For simple tasks, like identifying a missing item in a to-do list, ChatGPT performs well.
- However, for issues like memory leaks or file system corruption, it struggles.
On the other hand, Opus 4.8 always finds the solution, albeit slowly. I can rely on it without worrying about whether it will succeed. It just gets the job done.
Recently, Fable 5 has emerged, which resolves issues without needing any prompts. It operates even faster than Opus.
When I ask ChatGPT or Opus to create a new feature:
- ChatGPT often produces superficial results, ignoring existing code and building unnecessary independent code.
- Interestingly, the outcome from ChatGPT appears functional, but it's usually incorrect, focusing on a superficial "aha!" moment.
Opus, however, plans thoroughly, executes, and cleans up, ensuring everything works correctly. If needed, I can provide more realistic examples, though it's challenging due to the monorepo's size and complexity, with hundreds of thousands of lines of code.
Claude is Extremely slow, especially fast few days. Codex is so much faster in in my opinion with comparable quality. Anthropic is going to fix their problems, but for me as a user that depends for the service to work, it's not acceptable.
ChatGPT 5.5 is already a better version than Opus 4.8, at lest in my experience. Firstly, thing get done today and I don't see quality to be worse. I has many situation where Opus was just going in circles, finding more and more issues with the code it wrote a few hours back, to me this is counterproductive.
For gpt-5.5 I build up that session context into a markdown file, and then I start a new session and give it the markdown file with the instructions for what to do.
I'm guessing this works better because it can always go back and re-analyze the saved context.
Yes, I've just come to the end of implementing all the planning I did while Fable was available. And nothing now comes close to creating plans that could be coded and just worked like it did.
On a large C codebase, Claude hallucinates constantly, and GPT 5.5 gets there are with a lot of help, but still gets things wrong.
reply