Feedback Agent

The Feedback Agent is an autonomous Claude Code instance that processes user feedback (bug reports and feature requests) submitted through the Rondo Club UI. It runs on a Mac Mini every 30 minutes, creating pull requests for resolved items and posting follow-up questions when it needs more information.

Architecture

User submits feedback → WordPress REST API → Feedback queue (status: approved)
                                                    ↓
Mac Mini launchd (every 30 min) → bin/get-feedback.sh --loop --optimize
                                                    ↓
                                        Fetch oldest approved item
                                        Set status to in_progress
                                        Pipe to Claude Code --print
                                                    ↓
                            ┌────────────┬──────────┴──────────┐
                         RESOLVED     NEEDS_INFO            DECLINED
                            ↓            ↓                     ↓
                     Create PR     Post comment          Set declined
                     Set resolved  Set needs_info
                     Store PR URL
                                                    ↓
                                        When no feedback left:
                                        Run optimization review

Feedback Statuses

Status	Meaning
new	Just submitted by a non-admin user, awaiting approval
approved	Ready for agent pickup
in_progress	Currently being processed by the agent
needs_info	Agent posted a question, waiting for user reply
resolved	Agent created a PR or the issue was fixed
declined	Not actionable

REST API

Feedback Comments

Comments enable a conversation thread between the agent and users on feedback items.

List comments: GET /rondo/v1/feedback/{id}/comments

Returns an array of comments ordered by date ascending:

[
  {
    "id": 1,
    "content": "Can you clarify which page this happens on?",
    "author_id": 1,
    "author_name": "Admin",
    "author_type": "agent",
    "created": "2026-02-14 12:00:00"
  }
]

Create comment: POST /rondo/v1/feedback/{id}/comments

{
  "content": "It happens on the People list page",
  "author_type": "user"
}

When a user replies to a needs_info feedback item, the status automatically transitions back to approved so the agent picks it up again.

Agent Meta Fields

The feedback API includes two agent-specific fields in the meta response:

• pr_url — URL of the GitHub PR created by the agent
• agent_branch — Git branch name used by the agent

These are stored as post meta (_feedback_pr_url, _feedback_agent_branch) and can be set via the update endpoint.

Script: bin/get-feedback.sh

Flags

Flag	Description
--run	Process one feedback item with Claude Code
--loop	Process all approved items until none remain
--optimize	When no feedback, review one file for optimization
--status=X	Filter by status (default: approved)
--type=X	Filter by type: bug or feature_request
--id=X	Process a specific feedback item
--json	Output raw JSON (no Claude processing)

Safety

• Lock file at /tmp/rondo-feedback-claude.lock prevents concurrent runs
• Crash cleanup resets feedback status to approved if the script exits unexpectedly
• ensure_clean_main() aborts if the working directory is dirty
• Branch cleanup deletes merged feedback/* and optimize/* branches after each run
• Status set to in_progress before Claude runs, preventing other instances from picking up the same item

Agent Prompt

The agent’s instructions live at .claude/agent-prompt.md. Key rules:

• Create branch feedback/{id}-{slugified-title}
• Make changes, build, commit, push, create PR via gh pr create
• Do NOT deploy — only create PRs
• Output STATUS: RESOLVED + PR_URL: or STATUS: NEEDS_INFO + QUESTION:

Optimization Mode

When --optimize is set and no feedback items are found, the script reviews one file per run:

1. Build a queue of PHP includes and React source files
2. Track reviewed files in logs/optimization-tracker.json
3. Claude reviews one file using .claude/optimize-prompt.md
4. If improvements found, creates a PR on optimize/{module-name} branch
5. Max 1 optimization PR per run

Mac Mini Setup

launchd Configuration

The plist template is at bin/com.rondo.feedback-agent.plist. To install on the Mac Mini:

# Copy to LaunchAgents
cp bin/com.rondo.feedback-agent.plist ~/Library/LaunchAgents/

# Load the job
launchctl load ~/Library/LaunchAgents/com.rondo.feedback-agent.plist

# Check status
launchctl list | grep rondo

# Unload if needed
launchctl unload ~/Library/LaunchAgents/com.rondo.feedback-agent.plist

Prerequisites on Mac Mini

• Claude Code CLI installed and authenticated (Max OAuth)
• gh CLI authenticated with GitHub
• SSH keys configured for git push
• .env configured with API credentials
• jq installed (brew install jq)

Logs

• logs/feedback-processor.log — Main script log
• logs/launchd-stdout.log — launchd stdout capture
• logs/launchd-stderr.log — launchd stderr capture
• logs/optimization-tracker.json — Tracks which files have been reviewed

Frontend

FeedbackDetail Page

• Shows PR link when meta.pr_url is set
• Shows “Waiting for your response” banner when status is needs_info
• Shows conversation thread with agent/user messages
• Shows reply form when status is needs_info

FeedbackManagement Admin Page

• Includes needs_info (“Info nodig”) in status dropdown
• Shows PR column with link to GitHub PR