How about expect tool and ability to start console by user side?

[luchezarno]

Hello dear @yotsuda !
Well, I am not sure how useful this could be, but just thought that could be nice to be able to start console/s by human or some pipeline (CI/CD or kind of) and then somehow pass it's #number to AI to attach it with peek_console for example. What do you think?

And second idea - what if to add kind of expect/pexpect tool for long-polling MCP calls from AI? Though I am not sure about cases except for monitoring something (like errors in log or long running commands or file appearance)

[yotsuda]

Hey @luchezarno — thanks so much for taking the time to write these up. Both ideas touched on core design tradeoffs, and I ended up thinking through them more carefully than I expected. Sharing where I landed.

1. Human-/CI-started consoles AI can attach to

Passing on this. execute_command routes by shell family, not by console identity — AI never names a specific console. A switch_console would technically work but composes badly with auto-route-to-standby: if the attached console is busy, AI would silently land in a fresh sibling that shares only cwd, losing whatever state the user was trying to hand over.

Most concrete cases (SSH+MFA, az login/SSO, license activation, debugger attach, vcvarsall, REPL preload) are already covered by "AI calls start_console → user prepares in the visible console → user says go".

The one exception is elevation (UAC) — it has to spawn a new process with an elevated token, you can't escalate an existing one. A narrow start_console(elevated: true) feels like the right shape there; I'll consider it separately.

For observing a console you're driving, peek_console already accepts PID or display-name substring — that's already there.

2. expect/pexpect-style long-poll

Also passing:

• MCP tool calls have a ceiling (few minutes), so true long-poll isn't possible at the protocol level.
• ripple already piggybacks finished-command status onto every MCP response, so completion surfaces naturally.
• Pattern-match-in-stream works today by making the match the exit condition:

  tail -f app.log | grep -m 1 ERROR
  until test -f /tmp/ready; do sleep 0.5; done

If you hit a case where the exit condition isn't shell-expressible, happy to revisit with a concrete example.

Thanks again for pushing on these.