/deepask: Multi-Step Reasoning mit iterativer RAG-Suche

Neuer Discord-Command für tiefe Recherche in 3 Phasen:
1. Initiale Qdrant-Suche mit der Originalfrage
2. LLM generiert Folgefragen, sucht erneut (max 2 Iterationen)
3. Synthese aller gesammelten Chunks zu umfassender Antwort

Nutzbar via /deepask oder @bot deepask.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

cmd/discord/main.go

@@ -64,6 +64,13 @@ var (
 				{Type: discordgo.ApplicationCommandOptionString, Name: "frage", Description: "Die Frage", Required: true},
 			},
 		},
+		{
+			Name:        "deepask",
+			Description: "Tiefe Recherche mit Multi-Step Reasoning (mehrere Suchdurchläufe)",
+			Options: []*discordgo.ApplicationCommandOption{
+				{Type: discordgo.ApplicationCommandOptionString, Name: "frage", Description: "Die Frage", Required: true},
+			},
+		},
 		{
 			Name:        "asknobrain",
 			Description: "Stelle eine Frage direkt ans LLM (ohne Wissensdatenbank)",
@@ -449,6 +456,22 @@ func onInteraction(s *discordgo.Session, i *discordgo.InteractionCreate) {
 			return resp
 		})
 
+	case "deepask":
+		question := data.Options[0].StringValue()
+		channelID := i.ChannelID
+		handleAgentResponse(s, i, func() agents.Response {
+			resp := researchAgent.Handle(agents.Request{
+				Action:  agents.ActionDeepAsk,
+				Args:    []string{question},
+				History: getHistory(channelID),
+			})
+			if resp.RawAnswer != "" {
+				addToHistory(channelID, "user", question)
+				addToHistory(channelID, "assistant", resp.RawAnswer)
+			}
+			return resp
+		})
+
 	case "asknobrain":
 		handleAskNoBrain(s, i, data.Options[0].StringValue())
 
@@ -990,6 +1013,18 @@ func routeMessage(text, author, channelID string) agents.Response {
 			Source: "discord/mention",
 		})
 
+	case "deepask":
+		resp := researchAgent.Handle(agents.Request{
+			Action:  agents.ActionDeepAsk,
+			Args:    args,
+			History: getHistory(channelID),
+		})
+		if resp.RawAnswer != "" {
+			addToHistory(channelID, "user", strings.Join(args, " "))
+			addToHistory(channelID, "assistant", resp.RawAnswer)
+		}
+		return resp
+
 	default:
 		resp := researchAgent.Handle(agents.Request{
 			Action:  agents.ActionQuery,