Gate conversational context with a direct-tool text ceiling

Before adding conversational fusion to an audio router, test whether the preceding agent turn improves tool selection when the caller transcript is already available. Use a call-disjoint split, predict the tool directly, constrain candidates by the active node, and select any agent-context weight only inside training data. Long agent prompts can overwhelm short replies in additive models; bounded interactions may help a few cases without explaining the broader audio gap. Preserve the turn for train/serve parity even when measured incremental value is small.

Applies when: a voice router appears to miss context from the agent’s immediately preceding question. Keywords: audio routing tool classification conversational context transcript legal mask