How to make sure that an agent will go to existing PS console instead of spawning new one?

[luchezarno]

Hello dear @yotsuda !
First of all, that you very much for you mind-blowing idea and tool! 😎 I was dreaming of something similar since all this agents move started 😍
And especially thank you for brand new splashshell tool! Which seems even more prominent because of nowadays a few young people respect old good Powershell 😢 (well, most of them are on Mac now)

Since I am asking, then another question besides that in Subj. How do you think - is it realistic to make an agent to act as declared REPL-base language interpreter (such as Python, or Groovy, or Clojure and so on) if I have started REPL session in MCP invokes Powershell window ? 🤔 That way it could be even more powerfull as programming teacher!

With best regards,
Dima

[yotsuda]

@luchezarno thanks for the kind words! Answering in order:

1. Making the agent use an existing PS console instead of spawning a new one

PowerShell.MCP is designed so that the AI owns multiple consoles and switches between them automatically. The key here is the reason parameter on start_console: when omitted, an existing standby console is reused; only when specified is a new console forcibly spawned. The LLM is also nudged via the tool description to "not pass a reason unless a separate console is actually needed," so extra consoles normally don't pile up. On top of that, consoles outlive their parent MCP process, so restarting Claude Code reconnects to existing consoles via console re-claim. If you're still seeing new consoles spawn anyway, it's likely one of: (a) the agent is calling start_console with a reason, (b) sub-agent isolation has given it a different agent_id, or (c) re-claim is failing to detect the existing console. If you can share a reproduction, I'm happy to help narrow it down.

2. Making the agent drive a REPL started inside the PS console

This isn't possible with PowerShell.MCP's current architecture — commands run via Invoke-Expression on a PowerShell runspace, so any interactive REPL started underneath it (python / groovy / clojure / …) blocks on stdin and can't receive commands from the runspace side. It's the same class of stdin-ownership problem as #42 / #43.

But you can do exactly this with ripple (GitHub: https://github.com/yotsuda/ripple , npm: @ytsuda/ripple). On top of the four shells (pwsh, bash, zsh, cmd), ripple ships eleven language REPLs as first-class adapters: Python, Node.js, Racket, Clozure Common Lisp (ccl), Armed Bear Common Lisp (abcl), F# Interactive (fsi), Java (jshell), groovysh, Lua, Deno, and SQLite3. You launch one with start_console shell=python (or groovy, racket, …), and from there you drive the REPL with normal execute_command calls. Session state persists, so the AI pays the heavy initialization cost once and then explores interactively — which is exactly the "programming teacher" use case you described. Clojure isn't in the 18 adapters yet, but adapters are declarative YAML files, so it can be added in a future release.

Your question changed splash's direction

I want to be explicit about this: your question reshaped splash in two ways. splashshell already supported multiple shells, but I hadn't planned on language REPLs at all. Your framing — "what if the agent could drive a Python / Groovy / Clojure REPL the same way it drives a shell?" — made me realize the exact same machinery could host any interactive process, which is how the current 18-adapter setup (4 shells + 11 REPLs + 3 debuggers) came to be. From the same realization came the second change: the scope clearly no longer fit under the word "shell," so a few days ago I renamed the project from splashshell to splash (the old splashshell package is now deprecated). So your question ended up influencing even the name of the project. Thank you, genuinely.

To install the new package:

claude mcp add-json ripple -s user '{"command":"npx","args":["-y","@ytsuda/ripple@latest"]}'

I'd love to hear what you think of it — especially how the Python or groovysh REPL sessions turn out for educational use!

[luchezarno]

Wow! I am just happy to be a drop which I hope will lead to a avalanche in AI education :) Sincerely wish your project to became next big thing! This is exactly what I was dreaming to be able to leverage with AI.
I am gonna try and show it to my colleagues for sure, this way to work is going to be a kind of sensation