[FEATURE] Allow users to provide a reason/instruction when denying a tool call

[mizunashi-mana]

Preflight Checklist

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

Problem Statement

When a user denies a tool call at the permission prompt, the model receives no context about why it was denied. This causes the model to repeatedly attempt workarounds or similar tool calls, essentially "going rogue" trying to bypass the denial without understanding the intent. The user then has to separately explain the reason or desired alternative, adding extra steps and breaking the autonomous workflow.

Proposed Solution

Allow users to attach a short message/instruction when denying a tool call, similar to how permission rules can include instructions (e.g. ["Bash(pnpm:*)", "Use npm instead"]). For example, when the permission prompt appears and the user selects "No", they could type a reason like "Use npm command instead" or "Don't modify this file, change X instead". This message would be passed to the model as context for the denial, so it can adjust its approach immediately.

Before: Claude Code is going rogue...

{"permissions": {"deny": ["Bash(pnpm:*)"] } }

After: Claude Code works well!

{"permissions": {"deny": [["Bash(pnpm:*)", { "reason": "Use npm instead" }] ] } }

Alternative Solutions

Currently, after denying, the user must type a follow-up message explaining the reason, or preemptively set up hooks/rules. This adds friction and delays autonomous operation.

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Use Case Example

1. Claude attempts Bash(pnpm install ...)
2. User denies with reason: "Use npm, not pnpm"
3. Claude immediately understands and runs npm install ... instead
4. No extra back-and-forth needed

Additional Context

No response

[github-actions]

Found 3 possible duplicate issues:

1. [BUG] Additional instructions ignored during action rejection #25646
2. Feature request: PostToolUseRejection hook event #29118
3. PermissionRequest hooks should support additionalContext #28035

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[mizunashi-mana]

These are related but not duplicates. The linked issues are either bug reports or proposals for specific mechanisms (hooks, additionalContext) that would partially address the problem. This issue is more comprehensive — it describes the general problem of no context being passed to the model on denial and proposes a first-class solution at the UI/permission config level. They are complementary, not overlapping.

[xificurC]

Hit TAB and you can add a message. This works today

[ssalka]

Despite the fact that Tab to amend the denial message is good for one-off cases, I still see a need for the reason metadata in settings.json:

I notice claude frequently tries to run git -C /path/to/repo <command> rather than just <command>, and I want it to avoid doing so. I have "Bash(git -C *)" in the permissions.deny list, but Claude is given no reason as to why those commands are denied, and it will sometimes try to run it several times before moving on (sometimes not even completing the thing it was trying to do originally).

With this feature, I could add a reason, like "You are already in the root directory of the repository", which would automatically be sent to Claude when these tool calls come up, effectively course-correcting it rather than leaving it in the dark about what to do.

[rcalrissian80]

Despite the fact that Tab to amend the denial message is good for one-off cases, I still see a need for the reason metadata in settings.json:

I notice claude frequently tries to run git -C /path/to/repo <command> rather than just <command>, and I want it to avoid doing so. I have "Bash(git -C *)" in the permissions.deny list, but Claude is given no reason as to why those commands are denied, and it will sometimes try to run it several times before moving on (sometimes not even completing the thing it was trying to do originally).

With this feature, I could add a reason, like "You are already in the root directory of the repository", which would automatically be sent to Claude when these tool calls come up, effectively course-correcting it rather than leaving it in the dark about what to do.

I was trying to figure out how to write the exact same hook