Feature Request: Agent Delegation Quality Assurance & Verification Protocols #7247

Activity

1. github-actions

   added

   enhancementNew feature or requestNew feature or request

   area:tools

   on Sep 6, 2025
2. coygeek commented on Sep 6, 2025

   

   coygeek

   on Sep 6, 2025

   More actions

   Hey, this is a great write-up and a super important issue for enterprise use. I've run into similar consistency problems with agents on large-scale refactors, where a small, systematic error gets amplified across dozens of files. It's a critical point for making agents truly reliable.

   While built-in VERIFY steps would be a fantastic addition, I think you can get pretty close to what you're asking for today by combining a few existing features: Subagent Prompts, Memory Imports, and especially Hooks.

   1. For your Verification Protocol & Quality Gates (PostToolUse Hooks)

   This is a perfect use case for Claude Code Hooks. Specifically, you can use a PostToolUse hook that triggers after an agent uses the Edit or Write tool. This hook can run a script to perform your verification checks.

   The best part is that if the script returns a specific exit code (2), it blocks the agent and feeds the script's error output back to Claude as feedback. This acts as your "Quality Gate."

   Here’s how you could set it up:

   a. Create a verification script in your repo, for example .claude/hooks/verify-component.sh:

   #!/bin/bash
   # .claude/hooks/verify-component.sh
   
   # Read the file path from stdin (passed by the hook)
   # Note: The hook provides a JSON object to stdin
   FILE_PATH=$(jq -r '.tool_input.file_path' < /dev/stdin)
   
   # Only run checks on component files
   if [[ "$FILE_PATH" != *"components"* || "$FILE_PATH" != *".tsx"* ]]; then
     exit 0
   fi
   
   ERRORS=""
   
   # --- Your Verification Logic ---
   # Check for default exports
   if grep -q "export default" "$FILE_PATH"; then
     ERRORS+="ERROR: Component uses a default export. Please use named exports.\n"
   fi
   
   # Check for bad hook import paths
   if grep -q "from 'hooks'" "$FILE_PATH"; then
     ERRORS+="ERROR: Invalid import path. Hooks must be imported from '@enterprise/hooks'.\n"
   fi
   
   # Run TSC for type errors
   TSC_OUTPUT=$(npx tsc --noEmit "$FILE_PATH" 2>&1)
   if [ $? -ne 0 ]; then
     ERRORS+="\nTypeScript Errors:\n$TSC_OUTPUT\n"
   fi
   
   # --- Quality Gate ---
   if [ -n "$ERRORS" ]; then
     echo -e "Verification failed for $FILE_PATH:\n$ERRORS" >&2
     exit 2 # Exit code 2 feeds this stderr output back to Claude
   fi
   
   exit 0 # Success

   Make sure to chmod +x .claude/hooks/verify-component.sh

   b. Configure the hook in .claude/settings.json:

   {
     "hooks": {
       "PostToolUse": [
         {
           "matcher": "Edit|Write",
           "hooks": [
             {
               "type": "command",
               "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/verify-component.sh"
             }
           ]
         }
       ]
     }
   }

   Now, every time the agent modifies a component, this script will run. If it fails, the agent will be stopped, shown the errors, and will likely try to fix them on its own.

   2. For Enterprise Context Inheritance

   You're spot on that agents need to reliably follow project standards. The Memory documentation shows you can use a root CLAUDE.md file to import other standards files using the @ syntax. This is more robust than just hoping the agent finds them.

   Your project's CLAUDE.md could look like this:

   # Project Standards
   This project follows strict enterprise development standards.
   
   - Core Contribution Guidelines: @CONTRIBUTING.md
   - Enterprise File Structure Rules: @ENTERPRISE-FILE-STRUCTURE.md
   - Agent Delegation Checklist: @AGENT-DELEGATION-CHECKLIST.md
   
   **IMPORTANT**: All component modifications are automatically verified by a hook. If you receive a 'Verification failed' error, you MUST correct all listed issues before proceeding.

   3. For Batch Size Limits

   This is a tough one. I don't see a direct way to configure this in the docs, and it's a great suggestion. The best workaround I've found is to be explicit in the prompt:

   "Process the first 5 components from the list. After you are done, stop and wait for my confirmation before proceeding with the next batch."

   It's manual, but it prevents the agent from running away with all 52 components before you can spot a systematic error.

   ---

   Tying It All Together

   So, a potential workflow could be:

   1. Beef up your sub-agent's system prompt to explicitly mention the verification step and the standards documents.
   2. Use CLAUDE.md with @ imports to ensure all standards are in context.
   3. Implement the PostToolUse hook to act as your automated quality gate.
   4. Prompt the agent to work in smaller batches for large tasks.

   Anyway, hope this helps bridge the gap until the team can build a more native solution. It's definitely a +1 from me on the core feature request!

   P.S. I saw in the latest changelog (v1.0.94) they added a /todos command. Could be useful for having the agent explicitly track and check off components as it processes them in batches.
3. github-actions commented on Dec 5, 2025

   

   github-actionsBot

   on Dec 5, 2025 – with GitHub Actions

   More actions

   This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
4. github-actions

   added

   autocloseIssue will be closed automaticallyIssue will be closed automatically

   on Dec 5, 2025
5. github-actions commented on Jan 5, 2026

   

   github-actionsBot

   on Jan 5, 2026 – with GitHub Actions

   More actions

   This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.
6. github-actions

   closed this as not plannedon Jan 5, 2026
7. github-actions commented on Jan 14, 2026

   

   github-actionsBot

   on Jan 14, 2026 – with GitHub Actions

   More actions

   This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.
8. github-actions

   locked as resolved and limited conversation to collaborators on Jan 14, 2026