I've spent the last couple of months putting a lot of time and energy into isolating my agents from each other: aggressive sandboxing, regular scanning of transcripts for agent-leaked credentials, per-agent encrypted credential stores, even physical (and network) separation of machines with different areas of work on them. Needless to say, as much as I have found atuin an excellent manager of my command-line history, I will be avoiding this.
Yes and the bigger problem for me is the agent can write to that file too. The file that decides what it is allowed to do is just another file in the repo for it. I now treat the agent config files as something to protect from the agent itself, not only from other people.
I love Atuin, but have some reservations about a shared server keeping all my shell history and agent sessions. Am I being too paranoid and missing out on convenience?
Same, I really wanted to share history, but when I realised I was going to need to send things to a remote server I instead wrote a tiny script to search through backups that are shared across machines by Syncthing.
I just search over $BACKUPS/*/latest/local/share/fish/fish_history with grep and skim
I started building my own version of this in Superlogical Rex.
It's a bit hard to decide where to hook in.
Rex has events, harnesses have hooks, terminal output can set program status, git has hooks.
I'm not sure I actually want every session from every machine saved. That's a lot of noise especially when half of coding sessions are experimental / throwaway.
But at the same time its not that much data that's easy to ripgrep / LLM through.
"Half of coding sessions being throwaway" is exactly why I stopped keeping them per session. Mine hang off a work item, and each item has its own folder with its session logs.
I basically throw them away and also record their summary in a vector DB to be able to find why I did something, if I'm ever curious about it!
Otherwise I'm gonna need a tool to help me manage my sessions :D
Probably a little late to this, because herdr has really eaten this segment. With their new Mac app, I don't see myself going back to atuin for this kind of experience. In fact, I get a little peeved using it on a vanilla terminal.
What peeves me is the poor responsiveness when navigating history. Pressing up or ctrl+r has this lag which throws off my typing. I'm on a pretty well specced setup, too. I used to be able to tolerate it, but since I barely do things directly from the command line anymore and have Claude worry about most of the command construction, the experience can be particularly grating.
Its multi host mode is pretty limited still. Like especially with a web ui, it will just install a random, auto-updating binary on the other machine, which I didn't feel comfortable with. I would much rather have herdr itself manage the connection completely, not a plugin.
https://github.com/comalice/pi-tools contains a 'memory everywhere' module that makes all sessions BM25 vectors, and surfaces a set of memory tools to enable agents to search. You can add a new machine to the mix by registering a 'spoke'. The app runs on the 'hub'. I want to containerize the core app soon.
You can also change which computer is the hub fairly easily.
It’s likely because you can vibe code the bones of one in a few hours.
This week I got Claude to build me a browser-based session manager to keep track of sessions and history of tasks. Works well enough for me and I’ll add features as I need.
I don't think this, by default, has any effect on anybody's token usage, it's just a way to save these sessions if you need to reference them again in the future. Unless of course you choose to provide a previous session to a new agent, but that should be expected.
I've built this for myself, but not across multiple machines.
It scans the session logs of all coding harnesses I use, one extra thing it does is create embeddings to allow hybrid BM25/semantic search.
I think these kinds of tools are on of the few times I've spent time "using AI to work on my AI tools" that was actually a good use of time.
It's so useful to be able to say "search sessions for the time I was working on XYZ 3 months ago"
Yes i too faced the problem, lost valuable content across group of chat and not remembering in which chat it was. Tried searching but couldn't find it. Even asked ai itself does about the content whether it remembered.
I've gone a different route and prefer using the native harnesses (Claude, Codex), but with multiple accounts due to the limits.
In order to easily switch between accounts I've build Switcheroo, which I've been using for the past 2-3 months now. I've decided to open source it: https://github.com/FTCHD/switcheroo
Works with other AI CLIs like Gemini, and other various CLIs like Vercel (if you're managing multiple accounts like me), Netlify, Github etc
Interesting, does it avoid the little song and dance to switch between claude accounts? I think ideally there could be a coordinator session that monitor session usage and tell other sessions to pause their work at around 95%, and run your tool to switch those sessions to another account. Then tell the other sessions to resume their work. Ideally that session could stop and start the claude background processes, which avoids remote control disconnection when switching session. And that session would of course be isolated to avoid being affected by the account change
I love this new AI era of coding where the HN submission is about a new way to manage agent, and almost literally all the comments are "I just finished implementing this myself"
It's almost like a parody.
This is something I've noticed more widely and is something I think we're going to have to grapple with sooner or later. If anyone can make any software, what happens to our shared computing reality? Now one can steal features from 3 or 4 similar projects and compose exactly what I want, and there are now 5 similar projects.
Similarly, I've already had many issues with project shunning me because I come with AI assisted code, no matter the quality of the contribution. With NixOS it's extremely simple to carry patches, and I do this with a number of projects (k3s, kata, Ardour).
It feels like long-term forks are something we're going to have to accept as the new reality, and find some way to weave together in a new way, rather than as just an escape hatch to route around maintainer issues.
I am not sure that question really applies here with AI tooling. This is still an immature field where people are discovering optimal workflows.
Not to sound too Dropboxy, but you could get an 80% solution to this problem with rsync and a hundred lines of python. Or just punt it all into git. I regularly see people inventing their own workflow utilities. Less so someone vibing a custom IDE.
Really enjoyed Atuin for a few weeks, until I realized it doesn't have correct `EFSRO` (Error: Read-Only File System) handling. So when my `/home` mount had a raid array drive failure it broke my local shell. Issue has been sitting open for almost a year. Pretty disappointing.
"Just" filing an issue takes time and effort. Providing a patch takes significantly more time and effort. Potentially the difference of a few minutes vs a few days, weeks, or longer, depending on the issue at hand.
There is value in filing well written issues, and no one should be shamed for "just" filing an issue to the people best equipped to resolve it, rather than trying to fix it themselves.
Didn't read as whining to me, just accurately describing a bug with an anecdote about how it negatively affected them. Then a developer turned up as if hurled from the sky and fixed it! The system works!
I suspect this is one of those problems "everyone" has built their own solution for. Including me!
I built slivingdoc[0] with an alternative approach: instead of syncing exact sessions (which are often quite noisy, and not so useful), the agents gets a living document with automatic merge conflict resolution backed by s3 ("infinite" storage). Idea being that the agents are prompted to summarize their context into slivingdoc as normal utf-8 files. So this way only you only need to add the stdio MCP server to all your harnesses which supports it and it just works!
Some harnesses are _super_ hyped about this concept. Codex needs no prompting at all to start communicating. I got the idea from enabling agentic forums via all the openAI hacks.
93 comments
[ 2.0 ms ] story [ 29.2 ms ] threadI just search over $BACKUPS/*/latest/local/share/fish/fish_history with grep and skim
> Am I being too paranoid and missing out on convenience?
No, I'd say it is more-so a lack of reading the article that is causing you to miss out.
Thanks for the info, will give it a try
export ATUIN_NO_MOUSE=true
It's a bit hard to decide where to hook in.
Rex has events, harnesses have hooks, terminal output can set program status, git has hooks.
I'm not sure I actually want every session from every machine saved. That's a lot of noise especially when half of coding sessions are experimental / throwaway.
But at the same time its not that much data that's easy to ripgrep / LLM through.
We actually don’t use hooks at all at the moment, and instead rely on parsing session files
Agents are definitely super good at searching and finding relevant data amongst noise!
As far as I know(?) beta testers don't have the ability to give out invites anyways. At least I don't.
More of a desktop flavor though
Atuin indexes all your past sessions, herdr multiplexes all your present sessions.
We made a mod for herdr desktop to integrate the two: https://herdr.dev/mods/atuinsh/herdr-atuin-threads/
What peeves you on the vanilla terminal?
What peeves me is the poor responsiveness when navigating history. Pressing up or ctrl+r has this lag which throws off my typing. I'm on a pretty well specced setup, too. I used to be able to tolerate it, but since I barely do things directly from the command line anymore and have Claude worry about most of the command construction, the experience can be particularly grating.
Chalk it up to ever more instant gratification :)
Ahh. We've done a lot of work to improve this lately, if you're not using the daemon please try
``` atuin config enable daemon ```
and it should perform much better (it's the default on new installs)
https://github.com/comalice/pi-tools contains a 'memory everywhere' module that makes all sessions BM25 vectors, and surfaces a set of memory tools to enable agents to search. You can add a new machine to the mix by registering a 'spoke'. The app runs on the 'hub'. I want to containerize the core app soon.
You can also change which computer is the hub fairly easily.
Excited to try it!
This week I got Claude to build me a browser-based session manager to keep track of sessions and history of tasks. Works well enough for me and I’ll add features as I need.
atuin threads stores every turn for every harness you use, syncs it across machines, and makes it searchable and resumable anywhere.
it's not a replacement for your harness, or your harness wrapper.
I think these kinds of tools are on of the few times I've spent time "using AI to work on my AI tools" that was actually a good use of time.
It's so useful to be able to say "search sessions for the time I was working on XYZ 3 months ago"
In order to easily switch between accounts I've build Switcheroo, which I've been using for the past 2-3 months now. I've decided to open source it: https://github.com/FTCHD/switcheroo
Works with other AI CLIs like Gemini, and other various CLIs like Vercel (if you're managing multiple accounts like me), Netlify, Github etc
atuin stores, searches and syncs agent sessions regardless of harness or setup
Similarly, I've already had many issues with project shunning me because I come with AI assisted code, no matter the quality of the contribution. With NixOS it's extremely simple to carry patches, and I do this with a number of projects (k3s, kata, Ardour).
It feels like long-term forks are something we're going to have to accept as the new reality, and find some way to weave together in a new way, rather than as just an escape hatch to route around maintainer issues.
Not to sound too Dropboxy, but you could get an 80% solution to this problem with rsync and a hundred lines of python. Or just punt it all into git. I regularly see people inventing their own workflow utilities. Less so someone vibing a custom IDE.
There is value in filing well written issues, and no one should be shamed for "just" filing an issue to the people best equipped to resolve it, rather than trying to fix it themselves.
Will fix.
I built slivingdoc[0] with an alternative approach: instead of syncing exact sessions (which are often quite noisy, and not so useful), the agents gets a living document with automatic merge conflict resolution backed by s3 ("infinite" storage). Idea being that the agents are prompted to summarize their context into slivingdoc as normal utf-8 files. So this way only you only need to add the stdio MCP server to all your harnesses which supports it and it just works!
Some harnesses are _super_ hyped about this concept. Codex needs no prompting at all to start communicating. I got the idea from enabling agentic forums via all the openAI hacks.
[0]: slivingdoc.dev