diff --git a/apps/ai-bot/router.go b/apps/ai-bot/router.go
index 59cc41c0..8058e5bc 100644
--- a/apps/ai-bot/router.go
+++ b/apps/ai-bot/router.go
@@ -118,6 +118,14 @@ func (b *Bot) classify(ctx context.Context, body, rcx string, cost *CostBreakdow
 	if !b.cfg.RouterClassifierEnabled || b.gemini == nil {
 		return d
 	}
+	// Emit the exact conversation window handed to the classifier — its INPUT, which the
+	// truncated llm-exchange body log cuts off (the static prompt fills the ~4 KB cap before
+	// rcx). Gated by the LOG_BODIES_USERS verbose flag exactly like logLLMExchange, since rcx
+	// carries user content. This is the only way to tell a real misclassification apart from
+	// a context-starvation problem (empty/insufficient rcx) when about_project misfires.
+	if ri, ok := reqInfoFromContext(ctx); ok && ri.verbose {
+		b.log.DebugContext(ctx, "router context", "rcx", rcx)
+	}
 	// 4s router sub-deadline derived from genCtx (a budget cancel still propagates).
 	rctx, cancel := context.WithTimeout(ctx, routerStageTimeout)
 	defer cancel()