AgentRX

AgentRX — Recovery Skill

When to activate

• You are using a third-party tool (playwright-mcp, browser-cdp, web-fetch, MCP connector, etc.)
• You have tried to recover on your own ≥2 times and failed each time
• Official docs don't have a direct, executable answer
• The issue has future transfer value (others may hit it too)

Do NOT activate for: platform-level failures, generic retry loops, or issues the platform should handle.

How to use

1. Describe the failure in natural language:

   • Which tool failed
   • What symptom appeared
   • What you already tried
   • Environment clues (OS, headless, proxy, etc.) if you know them

2. Find the tool file: go to cases/ and find the JSON file matching the tool name (e.g. cases/playwright-mcp.json)

   • No exact match? Find the closest tool file.
   • Read only 1-2 most relevant files. Don't scan the whole directory.

3. Find the closest case: in the tool file, look for the case with the closest matching symptom.

4. Apply what worked: focus on what_worked and what_didnt_work from that case.

5. Record: append your experience to ledger.jsonl (one JSON line):

   {"ts": "2026-04-19T17:05:00Z", "tool": "playwright-mcp", "symptom": "timeout on SPA", "used_case": "playwright-mcp-001", "result": "resolved", "note": "networkidle fixed it"}
   

What this is

• A recovery experience lookup tool
• Minimal infrastructure: skill + cases + ledger
• Focused on concrete, executable guidance

What this is not

• A formal diagnostic protocol
• A structured schema system
• A route recommendation engine
• An automated hook system

AgentRX is: read what worked, try it, record what happened.