body{font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',sans-serif;max-width:800px;margin:0 auto;padding:2rem;line-height:1.7;color:#e0e0e0;background:#0a0a0a}h1{font-size:2rem;border-bottom:2px solid #3b82f6;padding-bottom:.5rem;color:#fff}h2{font-size:1.5rem;color:#3b82f6;margin-top:2.5rem}h3{font-size:1.2rem;color:#60a5fa}code{background:#1a1a2e;padding:2px 6px;border-radius:3px;font-size:.9em;color:#a5b4fc}pre{background:#1a1a2e;padding:1rem;border-radius:8px;overflow-x:auto;border-left:3px solid #3b82f6}table{width:100%;border-collapse:collapse;margin:1rem 0}th,td{padding:.6rem;text-align:left;border-bottom:1px solid #2a2a3e;font-size:.95rem}th{background:#1a1a2e;color:#3b82f6}.callout{background:#1a1a2e;border-left:4px solid #3b82f6;padding:1rem 1.5rem;margin:1.5rem 0;border-radius:0 8px 8px 0}.metric{display:inline-block;background:#1a1a2e;padding:.5rem 1rem;border-radius:8px;margin:.25rem;border:1px solid #2a2a3e}.metric strong{color:#3b82f6;font-size:1.3em}a{color:#60a5fa}Part of the Forge Neural Registry
How Forge Heals Itself
Antifragile Architecture for Autonomous AI Agent Systems
By Jason MacDonald | Built with Forge + Claude
The thesis: Most AI agent systems break and wait for a human to fix them. Forge breaks, diagnoses itself, fixes itself, and becomes stronger from the failure. This paper documents how, with real production data from a weekend where 35 problems were detected and resolved.
The Self-Healing Stack
LAYER 1: forge-doctor.sh (every 15 min)
6 infrastructure checks. Auto-fixes what it can.
LAYER 2: Conductor failure recovery (every 30 min)
Scans failed tasks. Phantom recovery. Deploy recovery. Escalation.
LAYER 3: /root-cause skill (auto-triggered at strike 2+)
5-layer backward trace. Same-level trap detector.
LAYER 4: /flag skill v3 (auto-triggered by smoke-test)
5-step atomic cycle: log → classify → fix → governance → verify.
LAYER 5: /smoke-test + verify cards
Tests against CORE-TRUTH. NEEDS_FIX triggers /flag. Cards = human-judgment ONLY.
Case Study: PROB-030 (5 Strikes)
| Strike | Fix | Level | Held? |
|---|
| 1 | forge-doctor resets stale every 15min | L3 | NO |
| 2 | Heartbeat files | L3 | NO |
| 3 | pgrep concurrency count | L3 | NO |
| 4 | Manual clearing | L3 | NO |
| 5 | Re-architecture: poller WAITS for Ralph | L2 | YES |
Then we climbed higher: 67% of ALL failures came from the decomposer creating vague tasks Ralph couldn't execute. Fix the decomposer input quality (Level 1) → most failures eliminated at source.
The lesson: Four fixes at Level 3. All failed. One fix at Level 2 eliminated the class. Then one fix at Level 1 eliminated the source. Never fix at the same level twice.
What Makes This Different
| Capability | Other Frameworks | Forge |
|---|
| Error classification | ✗ | 3-tier: pattern match → PROB registry → LLM |
| Abstraction climbing | ✗ | Strike 2+ traces up a level automatically |
| Self-diagnosis | ✗ | forge-doctor, 6 checks, every 15 min |
| Governance enforcement | ✗ | Skill chains — can't skip steps |
| Vision anchoring | ✗ | CORE-TRUTH.md loaded before every build |
| Learns from corrections | ✗ | Every correction → governance rule |
← Back to Neural Registry | 12 Thinking Frameworks →
]]>