"Tab to amend" instructions delivered post-execution — agent fires original tool call unmodified

Resolved 💬 3 comments Opened Feb 27, 2026 by loganmoon Closed Mar 27, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report
  • [x] I am using the latest version of Claude Code

What's Wrong?

When a user presses Tab to amend a pending tool approval and types amendment instructions, the original tool call executes unchanged, and the amendment text arrives in the agent's context only afterward — as a new user turn. The agent then treats it as a follow-up instruction rather than a pre-execution modification.

The intended contract of "Tab to amend" is that the amendment modifies the action before it runs. The current behavior inverts this: the irreversible action fires first, then the agent attempts to reconcile the amendment with the already-completed result.

Steps to Reproduce

  1. Ask the agent to perform any tool call that requires approval (e.g. a gh issue create or a destructive Bash command)
  2. At the approval prompt, press Tab to amend
  3. Type instructions that would change what the agent does (e.g. "update the body to include X before posting")
  4. Confirm

Result: The original tool call fires as-is. The amendment text appears in context afterward as a user message. For irreversible actions (network requests, file writes, etc.), the damage is already done.

Expected Behavior

The amendment text is injected into the agent's context before the tool executes, allowing the agent to revise the pending call — or cancel and re-plan — before taking action. The tool approval gate is the natural point to enforce this: the tool should not fire until the agent has processed the amendment.

Why This Matters

For any irreversible or public action (posting to GitHub, pushing code, deleting files), "Tab to amend" is the user's last opportunity to correct the agent's intent. If amendments arrive post-execution, the feature provides false assurance — users believe they are modifying the action, but they are only queuing a correction after the fact.

Claude Code Version

v2.1.59

Platform

Anthropic API

Operating System

Linux

Terminal/Shell

zsh

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗