Large Language Model (LLM)-based Multi-Agent Systems (MASs) are vulnerable to stochastic reasoning and coordination failures, yielding verbose, unstructured trajectories that are difficult to debug. Existing debuggers typically rely on either post-hoc static analysis, which suffers from ambiguous causal attribution, or dynamic verification, which incurs prohibitive computational costs and execution instability. To bridge this gap, we propose SynDebug, a static–dynamic synergistic debugging framework. Statically, SynDebug reconstructs execution topologies from flat logs via a Hierarchical Causal Graph and prioritizes root-cause candidates using synthesized Virtual Oracles.Dynamically, it performs targeted interventions on prioritized candidates to progressively recalibrate suspiciousness rankings based on execution impact. Evaluated across three MAS frameworks and four benchmarks, SynDebug outperforms six baselines, improving the repair success rate by 5.59%–81.81% while maintaining competitive efficiency.