[Bug] `/feedback` command should not be consumed by LLM mid-stream

/feedback command should not be consumed by LLM mid-stream

Description

When /feedback appears in text content being processed by the LLM (e.g., as part of the feedback message itself), the command is consumed and executed immediately during the LLM's message stream. Unlike other slash commands, /feedback is a side-effecting, user-facing command that should be deferred until the LLM finishes processing the current message queue.

Other commands (e.g., /skills, /otherplugins) are designed to work inline during LLM streaming and are fine being consumed mid-stream. /feedback is the exception — it triggers an external submission flow and should wait until the stream completes.

Environment

• Platform: macOS (darwin)
• Terminal: Cursor
• Claude Code Version: 2.1.80
• Feedback ID: 5b92b1e2-6568-4bfe-afd0-8d7d969ff67e

Steps to Reproduce

1. Run a long chat
2. Queue chat message: /feedback inline text
3. Observe the /feedback being consumed by the LLM mid-stream and executed as a command. Dispite LLM dont have the tool to handle that.

Working case (no /feedback in content):

❯ /feedback refuse when shouldn't
  ⎿  Feedback / bug report submitted ✅

Broken case (/feedback appears in content):

When LLM still generating:
❯ /feedback inline /feedback command was interpreted by chat
LLM: got it, I will store it as memory.

The /feedback is consumed during LLM message processing, triggering cascading errors.

Expected Behavior

/feedback should not be consumed by the LLM during stream processing. It should be deferred until the current message queue is fully drained, then executed. Any /feedback appearing as part of message content mid-stream should be passed through as literal text.

Root Cause

/feedback is missing a deferral guard that prevents it from being consumed by the LLM mid-stream. Other slash commands are designed to execute inline during streaming and work correctly in this context. /feedback, however, is a side-effecting command (external submission, UI flow) that needs special handling to defer execution until after the LLM completes its current response.

Suggested Fix

Add /feedback to whatever deferral or exclusion mechanism prevents side-effecting commands from being consumed by the LLM during stream processing. The command should only execute after the message queue is fully drained.

Workaround

Avoid including /feedback (or other slash command patterns) in the text content when submitting feedback.