Context

Deep dive on Hive (aden-hive) revealed that their self-healing mechanism starts with structured failure diagnosis -- not just "task failed" but which specific node failed, which criteria it fell short on, and the full decision log.

SynthOrg's RecoveryResult currently lacks this structure, making checkpoint recovery and task reassignment routing less informed.

Action Items

• Add failure_category enum to RecoveryResult (e.g., TOOL_FAILURE, STAGNATION, BUDGET_EXCEEDED, QUALITY_GATE_FAILED, TIMEOUT, DELEGATION_FAILED)
• Add criteria_failed: list[str] -- which specific acceptance criteria were not met
• Add stagnation_evidence: StagnationEvidence | None -- link stagnation detection data when applicable
• Add failure_context: dict[str, Any] -- structured bag for domain-specific failure data (tool error messages, provider errors, etc.)
• Use enriched diagnosis in checkpoint reconciliation messages
• Use enriched diagnosis for smarter task reassignment routing (e.g., tool failure -> reassign to agent with different tool access)

Design Notes

This is a model extension, not a new module. Extends existing RecoveryResult in engine/recovery.py. Low effort, immediate improvement to recovery quality.

References

• Hive -- structured failure diagnosis in evolution loop
• Related: research: audit engine against multi-agent failure patterns (swarm drift, microservices anti-patterns) #690 (multi-agent failure patterns -- diagnosis data feeds the audit)
• Related: feat: step-level quality signals + accuracy-effort observability metric #697 (step-level quality signals -- complementary granularity)