Feature Request: /feedback command in Amp CLI for user preference signals

[IgorGanapolsky]

Feature Request: Feedback Command/Hook for Amp CLI

Filed: 2026-03-10
Target: Sourcegraph Amp CLI (ampcode.com)
Contact: amp-devs@ampcode.com / @ampcode on X
Related: anthropics/claude-code#4569 (closed as "not planned")

Summary

Amp CLI currently has no built-in way for users to signal satisfaction or dissatisfaction with an agent response. Request a /feedback command (or similar mechanism) that fires a hook event, allowing local feedback systems and MCP servers to capture user preference signals.

Problem

Amp is a terminal-based CLI tool. There is no mechanism to:

• Express approval/disapproval of an agent response
• Fire a hook event that local systems can intercept
• Call a registered MCP tool with a feedback signal

The only workaround is typing phrases like "thumbs up" or "that was wrong" as a regular prompt, which the UserPromptSubmit hook can regex-match. This is indirect, unreliable, and wastes an agent turn.

Proposed Solutions

Option A: /feedback Command (Preferred)

Add a command palette entry (like existing /thread, /model, etc.):

/feedback up      — signal positive feedback on last response
/feedback down    — signal negative feedback on last response

This fires a new FeedbackSignal hook event with environment variables:

• AMP_FEEDBACK_SIGNAL — "positive" or "negative"
• AMP_FEEDBACK_THREAD_ID — the current thread ID
• AMP_FEEDBACK_TURN_INDEX — which agent turn was rated

Option B: Keyboard Shortcut

A quick key combo (e.g., Ctrl+Y / Ctrl+N) that emits the same hook event without requiring a command.

Option C: MCP Tool Callback

When a feedback signal is given, Amp calls a designated MCP tool (e.g., capture_feedback) if one is registered, passing { signal, threadId, turnIndex }.

Use Cases

1. Local RLHF feedback loops — Track approval rates per skill, action type, and session
2. DPO training pair export — Pair positive/negative signals with agent outputs for preference optimization
3. Autonomy calibration — Adjust agent confidence thresholds based on user satisfaction trends
4. Team quality dashboards — Aggregate feedback across workspace members

Context

The rlhf-feedback-loop npm package already has full receiving infrastructure:

• captureFeedback() engine in scripts/feedback-loop.js
• capture_feedback MCP tool in adapters/mcp/server-stdio.js
• JSONL logging, DPO export, self-audit, prevention rules
• Multi-platform support (Claude Code, Amp, Codex, Gemini CLI, Cursor)

All that's missing is a first-class way to emit feedback signals from the Amp CLI itself.