Context

When coordination metrics collection is enabled (see #135), the system can optionally classify coordination errors into structured categories. This enables targeted diagnosis of multi-agent coordination problems — e.g., if coordination failures spike, the topology may be too complex; if context omissions spike, the orchestrator's synthesis is insufficient.

This is an M5 feature because it requires semantic analysis infrastructure (embedding computation, contradiction detection) and baseline comparison data from M4 coordination metrics.

Acceptance Criteria

Error Taxonomy Categories

• Logical contradiction — agent asserts both "X is true" and "X is false"; detected via semantic contradiction analysis on agent outputs
• Numerical drift — accumulated computational errors from cascading rounding/unit conversion (>5% deviation); detected via numerical comparison against ground truth or cross-agent verification
• Context omission — failure to reference previously established entities/state required for current reasoning; detected via missing-reference analysis across agent conversation history
• Coordination failure — MAS-specific: message misinterpretation, task allocation conflicts, state synchronization errors; detected at the protocol/orchestration layer

Classification Pipeline

• Opt-in via coordination_metrics.error_taxonomy.enabled: true config
• Configurable category selection (can enable/disable individual categories)
• Pipeline runs post-execution — never blocks agent work
• Classification results logged as structured events to observability layer
• Pipeline gracefully degrades if semantic analysis is unavailable

Testing

• Unit tests for each error category detection logic
• Integration test: multi-agent task with known error patterns → verify classification
• Verify pipeline never blocks execution (async/post-execution enforcement)

Dependencies

• Implement LLM call categorization, coordination metrics suite, and orchestration tracking (DESIGN_SPEC §10.5 M4) #135 — Coordination metrics suite (M4) — must exist before error taxonomy can classify errors
• M5 memory/persistence infrastructure for storing classification results
• Embedding computation capability for context_omission and logical_contradiction detection

Design Spec Reference

• §10.5 — M4/M5: Coordination Error Taxonomy
• §16.3 — Agent Scaling Research (Kim et al., 2025 + Cemri et al., 2025 MAST)