[Feature] Real-time correction, session memory & in-terminal feedback for long agentic sessions

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Feedback from a Developer — Real Daily Usage of Claude Code

From a conversation on claude.ai, May 2026. The developer uses Claude Code in the terminal throughout the entire workday on real production tasks.

---

1. Quality Degradation in Long Sessions (Task Horizon)

Problem: After several hours of work — and especially after context compaction — there is noticeable quality degradation:

• Previously made fixes are repeated
• The same errors occur again
• The agent seems to forget decisions that were already made

Context: The developer actively maintains CLAUDE.md — logging errors, decisions, and patterns like "don't do X again". Claude is asked to append entries itself after incidents. This helps, but does not fully solve the problem.

What would help: A mechanism that preserves not just what was done, but why — especially across the compaction boundary. (Dreaming from the recent announcement seems to address exactly this — looking forward to it.)

---

2. No Way to Correct the Agent Mid-Task

Problem: When Claude works on a complex task (5+ minutes), it writes to the terminal describing what it's doing. The developer can see in real time that the agent is going in the wrong direction — but there's no way to course-correct without interrupting it.

• Typing in the terminal = injecting text into the active context, disrupting the task
• The only option is Escape (full stop)

What would help: A soft correction mechanism — something like a "side note" that the agent takes into account after the current step, without dropping the work.

---

3. No Feedback Button in the Terminal

Problem: claude.ai has 👍/👎. The terminal Claude Code has nothing. There's no way to quickly send feedback on a specific response during a working session.

What would help: A simple command like /feedback or a keyboard shortcut.

---

Overall Context

All three problems are connected to the task horizon concept (mentioned at Code with Claude 2026): the longer the agent works autonomously, the more critical the mechanisms for correction, memory, and feedback become.

These are not theoretical observations — they come from real, daily production use.

Proposed Solution

1. Degradation / memory

Persist decision rationale across compaction, not just actions. A lightweight "why we did X" log that survives context resets — similar to the announced Dreaming feature.

1. Mid-task correction

Allow a non-blocking input channel during agentic runs. The agent processes the note after completing the current step, without interrupting the active task.

1. Feedback in terminal

Add a /feedback command or keyboard shortcut (e.g. Ctrl+F) to submit a thumbs down on the last response without leaving the session.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_