Claude Code Adds Agentic Bash Command Loops
This episode breaks down Claude Code’s new bash-command behavior, where failed terminal commands can automatically trigger a debugging loop, plus how to disable it when you want more control.
We also cover workflow filtering, the new Skills section in /plugin, and a handy setting to stop terminal mouse clicks from hijacking your cursor.
Is this your podcast and want to remove this banner? Click here.
Chapter 1
Bash Commands Get Agentic
Lachlan Reed
If you've been running terminal commands inside Claude Code... like, you know, just spinning up a quick test with an exclamation mark... things just got a whole lot more agentic, brought to you by Jellypod AI. I- I- I actually got caught out by this in my backyard shed yesterday, James. I ran a basic build command, and instead of just dumping the error and letting me think, Claude literally went... wait, hold on, that broke, let me dive in and fix that for you. It just started looping on its own.
James Turner
Oh, wait, so it's not just a passive terminal wrapper anymore. It's actually parsing the stderr from your, say, !npm run build and automatically initiating a debugging loop?
Lachlan Reed
Exactly. It's the new 2.1.186 update. If you prepend a command with that exclamation mark, Claude now treats the output as a prompt. If a test fails, it just... rolls up its sleeves. Which is cool, but, er, honestly, sometimes you just want to run the test and think for yourself, right? You don't want the AI running off on a tangent, spending your API tokens trying to fix a typo you already spotted.
James Turner
Right, you don't want to get stuck in a token-burning loop because of a simple syntax error. So, how do we shut that off? Is there an opt-out config?
Lachlan Reed
Yeah, thank goodness there is. You can run /config respondToBashCommands=false right in the session. Or, if you want it permanent, just throw "respondToBashCommands": false into your global settings file. It stops the automatic loop dead in its tracks.
James Turner
That's a huge relief. I mean, agentic workflows are great, but predictability is king when you're in the flow. What about other quality-of-life updates in this release? I saw something about filtering workflows.
Lachlan Reed
Oh, yeah. If you run /workflows to see what Claude is actively chewing on, you can now just press the 'f' key. It lets you filter by status, so you're not scrolling through a massive wall of completed steps just to find the one active run that's currently hung up.
James Turner
Ah, so 'f' for filter, right there in the interactive CLI. Nice. And didn't they do something with custom tools as well? Like, how they're displayed?
Lachlan Reed
Spot on. They've overhauled the /plugin list. There is now a dedicated 'Skills' section specifically for your custom user-defined tools. It makes it way easier to see exactly what extra capabilities you've given Claude versus the built-in system stuff. And, oh! One more thing that was driving me absolutely mental... terminal mouse clicks. If you hate Claude hijacking your terminal cursor when you try to highlight text, you can now set CLAUDE_CODE_DISABLE_MOUSE_CLICKS to true in your env. No more accidental cursor jumps when you're just trying to copy-paste.
James Turner
Oh, man, that mouse click hijacking was the bane of my existence. Glad they sorted that. Alright, I'm going to go update my global config right now.