Add agent-model-healer: self-healing model fallback for scheduled agents

[marlandoj]

Summary

• Autonomous watchdog skill that monitors model health across all Zo scheduled agents
• When a model fails (402 credits, 429 rate limit, 503 unavailable, timeout), automatically switches affected agents to the next healthy fallback model
• When the original model recovers, restores agents to their preferred model
• Daily heartbeat email acts as a dead-man switch — if you stop receiving it, the healer or Zo platform is down
• Zero AI model cost for orchestration: all agent management uses direct MCP API calls, only health probes use /zo/ask (~5 tokens per model)

How it works

1. Runs as a scheduled agent on zo:fast (every 30 min)
2. Probes all configured models with a tiny health-check prompt
3. Switches agents on unhealthy models to fallback chains defined in assets/fallback-chain.json
4. Restores agents when models recover
5. Sends email notifications only when switches/restores happen
6. Sends a daily heartbeat email with model health summary

Safety

The Watchmen Independence Rule prevents the healer from switching its own model (which would create a circular failure). Enforced via config-driven agent ID + title pattern match.

Files

File	Purpose
SKILL.md	Skill definition with setup and usage instructions
DISPLAY.json	Registry metadata (icon, tags)
scripts/healer.ts	Main healer engine (Bun/TypeScript, ~400 lines)
assets/fallback-chain.json	Example config with placeholder model IDs

Test plan

• Install skill and verify bun healer.ts probe runs against configured models
• Verify bun healer.ts status outputs JSON with probe results
• Create a scheduled agent and confirm auto pipeline completes
• Confirm self-exclusion by setting healerAgentId in config

🤖 Generated with Claude Code