Phase 7: Error Handling & Recovery

Design error handling architecturally, not through prompts alone.

Core components

1. LoopGuard

Monitor for repetitive behavior across three dimensions:

• Action signature fingerprints (deduplicate tool+args)
• Normalized error pattern matching (cosine similarity on stripped messages)
• Semantic similarity via embeddings
• Graduated intervention: inject reflection prompt → force different approach → escalate/terminate
• Reduces iteration counts from 30+ to ~8

2. Checkpoint strategy by deployment type

3. Retry vs. pivot decision

• Retry when: transient error, actionable error info, quality trending up, within budget (max 3 for infrastructure)
• Pivot when: quality flat/declining across 3+ iterations, same semantic error repeats, token consumption escalating without quality gain
• Escalate to human when: irreversible side effects, 2-3 failed pivots, ambiguous error, security-sensitive operations

4. Context poisoning prevention

• Failed code accumulates in context, causing probability drift toward the failing pattern
• On doom loop detection: revert code, flush failed-attempt context, restart with original task + lessons learned only
• Preserve error traces during compaction (they serve as implicit negative examples)

5. Multi-agent error isolation

• Schema validation at every agent boundary (typed schemas, not natural language)
• Circuit breakers between agent clusters using 95th percentile response times
• SagaLLM compensation agents for rollback in multi-step workflows

6. Premature completion prevention

• Never trust agent self-assessment for completion
• Maintain structured feature list with items initially marked failing
• QA agent validates against the list before accepting completion
• Pre-completion checklist middleware intercepts exit and forces verification