@qt

Now that Claude Code has native loops and there is the (pretty good!) telegram skill, you might be asking yourself "how do I orchestrate the sessions NOT running the TG loop?" And you'd be right to ask yourself that. Your goal is to only have 1 session running the TG listener and the others working diligently with the tools and privileges you defined +you don't want to bloat the context of worker sessions with listener spam slop messages. And the way you do it is to assign inbox folders in your workspace to each session, usually best to just index 1, 2, 3 etc but you can assign them per project folder if that's your thing. Give one to each session. Then (before you leave your machine) you set up loops in each of the non-TG sessions to check their assigned inbox folder on a cadence that is less frequent than the TG checks. And you add those paths to your TG clerk session. Make sure you give the other sessions the TG clerk inbox path (effectively their outbox) and make sure that the TG clerk has a loop to check its own inbox. Basically you set up a simple in/out box folder structure that the different sessions share information through, and they check on some cadence. Your TG bot listens for your messages and assigns work out, your other sessions do them and share back, and you have a mobile multi agent orchestrator in your native terminal and folders. Pro: you keep your restrictions and permissions structures and don't have to install OpenClaw Con: sand boxing can cause issues There are a lot more Pros and Cons, and there are other ways to multi-agent orchestrate, just wanted to share this one, free, easy, and native way to get after it when running Claude Code shells.

0 reply

0 recast

4 reactions