Feature: Split-pane session chaining (two-way mirror observation model)

Summary

A way to split a Claude Code session into multiple linked panes, where each pane can observe everything happening in panes to its left, but panes to the left are unaware of those to the right. Think of it like a one-way mirror chain.

The concept

Right now, if I want to step back and think about what Claude is doing mid-session — or get a second AI opinion on its approach — I have to break my flow, copy context, or start a separate session from scratch.

Instead, imagine:

• You're in a Claude Code session (Pane A)
• You hit a command to open an observer pane (Pane B) to the right
• Pane B automatically sees everything Pane A is doing: messages, tool calls, file edits, decisions
• Pane A has no idea Pane B exists
• You can talk to a separate Claude in Pane B: "Is what it just did to auth.ts safe?" / "Review the approach so far"

Session chaining

The model extends naturally to more than two panes. The rule is simple: every pane sees everything to its left, nothing to its right.

• A → B → C → D
• D sees A, B, and C
• C sees A and B
• A sees nothing

You can insert a new pane at any position in the chain. If you're at B and spawn a new observer between B and C, the chain becomes A → B → (new) → C and visibility recalculates.

Use cases

• Code review in real-time — code with Claude in A, have a reviewer watching in B
• Agent debugging — test an agent in A, observe its decisions in B
• Side quests — mid-conversation, spawn a new pane for a tangential idea without losing your thread
• Coaching — junior dev works with Claude in A, lead observes and coaches from B
• Compliance/audit — B generates a log of everything A did

What I built

I prototyped a VS Code extension that does the two-pane version of this: https://github.com/jimmyjshang/two-way-mirror

It works as a standalone extension with a plugin interface for wiring up any AI backend. But it can't wrap or observe an actual Claude Code session because the extension is closed.

Native support in Claude Code would make this dramatically more useful — the observer pane would have real visibility into tool calls, file edits, reasoning, etc., not just terminal output.

Why this matters

Every AI coding tool today is a single-thread experience. This introduces a second (or third, or fourth) axis of interaction without breaking the primary one. It's the difference between working alone and having a team in the room — except the team sees everything you see, in real time, without you having to explain anything.