Multiverse context split — allow user input and tool permission dialogs to run concurrently
Description
When Claude Code requests tool permission mid-conversation, user input is modally blocked — but critically, it is preserved: pressing Escape restores exactly what the user was typing. This means the input buffer survives the interruption.
As Claude Code becomes more conversational and agentic ("go mode"), these interruptions become increasingly disruptive. The user is mid-thought while Claude is mid-execution — two parallel streams that the current UI serializes unnecessarily.
Proposed: Multiverse Context Split
When a permission prompt fires while the user is actively typing, instead of blocking, spawn a parallel agent context:
- Path A (user): Continue typing uninterrupted. Your conversational thread stays live.
- Path B (agent): The permission inquiry + execution continues in a parallel agent, operating on the current plan/execution context.
Context sharing — open question worth discussing
- Shared context via diffs — both paths operate on the same context, synchronized via lightweight diffing. Most correct, most complex.
- Full context copy — each path gets a snapshot. Simple, potentially expensive.
- Reduced plan context — the go-mode agent gets only the executable plan (analogous to the "clear context and implement" handoff that already exists in Claude Code). Cheapest, and arguably sufficient for pure execution tasks.
Option 3 is particularly elegant: it mirrors the existing pattern where Claude surfaces a plan, asks for approval, and can hand off to a fresh implementation context. The multiverse split would just make that handoff non-blocking.
Why this matters
Conversations with Claude Code are getting longer, richer, and more iterative. The permission dialog is a seam where the conversational and agentic modes collide. A multiverse split turns that seam into a feature.
---
Appendix: Also-Ran Names (a moment of silence)
- Multiverse Agent Spawn — technically accurate, sounds like a SyFy original series. Rejected on vibes alone.
- Parallel Universe Execution — too many syllables, not enough drama.
- Schrödinger's Prompt — the user's message is simultaneously sent and not sent until the permission box collapses. Tempting. Too cute.
- The Permission Singularity — what happens when every tool call requires approval. A cautionary tale, not a feature name.
This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗