Feature: Separate task pane and feedback pane for agentic work

Problem

During agentic tasks, the work channel and the feedback channel are the same. This creates a fundamental UX problem:

• To correct Claude mid-task, you interrupt the work stream
• If you wait, you watch it go wrong and correct after the fact
• The moment a correction is most valuable (at the point of violation) is also the hardest moment to deliver it cleanly

Proposed Feature

Two panes:

1. Task pane — Claude executes the task; the user watches
2. Feedback pane — the user provides real-time corrections, context, or "stop/redirect" signals asynchronously

The feedback pane would let corrections land at the moment of violation without requiring the user to interrupt the task flow or wait for a natural break.

Why This Matters

Expert users working on complex, multi-step tasks accumulate corrections as Claude works. Right now those corrections either interrupt the task or evaporate. A side channel would make real-time course correction low-friction — which is the highest-leverage moment for steering agentic behavior.

This is especially relevant for users whose domain expertise runs against Claude's trained priors: the correction at violation time is the signal, and the current UX makes that signal expensive to deliver.

Conversation Context

This feature request emerged from a conversation about how to efficiently transfer deep SWE expertise (30 years) to an LLM across sessions. The core problem we were discussing:

Rules aren't pre-articulable — they crystallize at violation. Expert knowledge is largely a library of "I've seen that fail", and that library is tacit until triggered. The moment Claude does something wrong is the moment the rule becomes speakable.

The conversation surfaced a related problem: LLM training corpora are dominated by mediocre-to-bad code. Popular patterns ≠ correct patterns. Expert users spend significant effort doing "manual gradient descent against bad priors" — reprompting to get Claude past the statistically common (but wrong) solution to the fundamental one. This correction work resets every conversation unless captured.

We discussed techniques for making corrections durable (CLAUDE.md, memory, capture-at-violation). The two-pane idea emerged as the missing UX primitive: corrections need to flow at the moment of violation, and the current single-channel model makes that expensive.

The feedback pane is the UX complement to persistent memory — memory makes corrections durable across sessions; the feedback pane makes corrections deliverable within a session without disrupting the task flow.