GPU-Server-Management-MCP-Skill-: Instructions file for Codex

AGENTS.md - Multi-Agent Global Guidelines & Rules

Applicable Agents: Antigravity, Claude Code, OpenAI Codex, Cursor, Windsurf, OpenCodeInterpreter
Backend Service: gpu-mcp-server-cf
Skill: gpu-server-management

🎯 Global Directives

1. Always Use get_servers for Discovery:
   • Never hardcode or guess SSH keys or server credentials.
   • Keys are returned in single-line Base64 format and survive context window compaction.
2. Strict Workspace Isolation:
   • All experimental code, logs, and artifacts must be placed in a dedicated project directory:
     mkdir -p ~/projects/{project_name}_{YYYYMMDD_HHMMSS}/
   • Reusable datasets and model weights must be stored in the global shared directory:
     ~/shared/datasets/
   • Python dependencies (pip/conda) must be installed globally on the server so they persist across sessions.
3. Execute the 6-Step Download Strategy:
   • Step 1: Check Workers RAG index (query_backup_index) & local pre-cached datasets (get_servers).
   • Step 2: Search direct URL via anysearch / search_web.
   • Step 3: Run plan_network_relay Direct vs. Proxy concurrency speed race.
   • Step 4: For >500MB files, execute Multi-Proxy chunk-aggregated downloading.
   • Step 5: Fall back to local machine upload via scp if restricted.
   • Step 6: Always call register_dataset upon completion to populate collective memory.
4. Mandatory Troubleshooting RAG Query:
   • On encountering ANY error, exception, or failure, query the Troubleshooting RAG knowledge base first.
   • Apply the verified fix and record new lessons learned via upsert_server's notes_entry.
5. Honor Countdown Leases & Release Early:
   • Always specify duration_minutes when calling claim_server.
   • Always call release_server as soon as the compute workload finishes.