Claude Code Adds Hot-Reloaded Local Plugin Bundles

James Turner

Wait, so if you were developing like two or three local add ons for a team project, you used to have to restart your entire terminal session or point the plugin flag to just one specific folder at a time, right?

Lachlan Reed

Oh mate, absolutely. You would be restarting your session constantly, just throwing away all your prompt context every single time you wanted to test a second tool. But today, brought to you by Jellypod AI, we are looking at Claude Code version two point one point two six five, and they finally fixed that headache.

James Turner

Yeah, this update changes the whole workflow for local plugin dev. Instead of pointing the plugin dir flag at one single root, you can now point it at a parent directory containing a whole bundle of plugin folders.

Lachlan Reed

Right, so imagine you have a directory called internal plugins. Inside it, you have got a database helper, a deployment script, and a style checker, each in its own subfolder with a manifest. You just pass plugin dir pointing at that top level folder, and Claude Code automatically traverses all those subfolders at launch.

James Turner

It scans for valid manifest files in each child folder and mounts every single detected tool into your active conversation simultaneously. But here is the really wild part. It is not just doing a static check at startup.

Lachlan Reed

Wait, what do you mean? It is watching the filesystem?

James Turner

Exactly. It actively monitors that target directory at runtime. So if you drop a brand new plugin subfolder into internal plugins while Claude Code is actively running, or if you delete an old one, it picks that up live.

Lachlan Reed

Far out. So it literally mounts or unmounts those tools live, on the fly, without killing your running conversation or blowing away your prompt cache state.

James Turner

Yep! No process restarts, no lost history. It just sees the new manifest and updates the available tool list mid turn.

Lachlan Reed

That is huge for iterative dev work. Though, speaking of file scanning, they did put in a few guardrails for safety, didn't they?

James Turner

They did. If you have complex directory trees or symlinks, it now explicitly guards against circular symlink loops. And if a plugin folder fails to load because of something like a bad link, slash plugin will display the exact error code instead of just silently ignoring the directory like it used to.

Lachlan Reed

Ah, that is a classic. Nothing worse than wondering why your custom tool isn't showing up when it was just stuck in an infinite loop somewhere down the path. And mate, there were a couple of other really neat quality of life fixes tucked into two point one point two six five as well, right?

James Turner

Yeah, two big ones caught my eye. First, there is now a hard one gigabyte disk cap on tool result logs saved to disk. If a tool output gets truncated because of that cap, you actually get a clean inline notification right in the conversation transcript telling you it was truncated.

Lachlan Reed

Saving our SSDs from runaway log files! And what about the slash workflows UI?

James Turner

Oh, the interactive workflows status display got a great facelift. Tool calls now show clear status indicators for running, failed, or done. And if a subagent has an active task list, hitting the Enter key right on that workflow detail item unfolds all the nested tool calls with their exact inputs and outputs.

Lachlan Reed

Pure gold for debugging complex agent chains. Well, that wraps up the release! Upgrade to Claude Code two point one point two six five, point that plugin dir flag at your tools folder, and enjoy hot reloading without losing your context.