Add reasoning effort

internal/service/service.go

@@ -450,6 +450,8 @@ type Request struct {
 	SystemMessage string `json:"system_message,omitempty"`
 	// MaxIterations caps the number of agent loop turns (0 = unlimited).
 	MaxIterations int `json:"max_iterations,omitempty"`
+	// ThinkingBudgetTokens is the llama.cpp reasoning budget (-1 = unlimited, 0 = off).
+	ThinkingBudgetTokens *int `json:"thinking_budget_tokens,omitempty"`
 }
 
 // Response is the response format for chat and voice endpoints.
@@ -545,7 +547,13 @@ func (svc *Service) chat(w http.ResponseWriter, r *http.Request) {
 	}
 
 	res, err := llmc.Completions(
-		ctx, model, req.SystemMessage, req.MaxIterations, req.Messages,
+		ctx, llm.CompletionsParams{
+			Model:                model,
+			SystemMessage:        req.SystemMessage,
+			MaxIterations:        req.MaxIterations,
+			Messages:             req.Messages,
+			ThinkingBudgetTokens: req.ThinkingBudgetTokens,
+		},
 	)
 	if err != nil {
 		log.ErrorContext(
@@ -717,7 +725,13 @@ func (svc *Service) chatStream(w http.ResponseWriter, r *http.Request) {
 	)
 	go func() {
 		errs <- llmc.CompletionsStream(
-			ctx, model, req.SystemMessage, req.MaxIterations, req.Messages, events,
+			ctx, llm.CompletionsParams{
+				Model:                model,
+				SystemMessage:        req.SystemMessage,
+				MaxIterations:        req.MaxIterations,
+				Messages:             req.Messages,
+				ThinkingBudgetTokens: req.ThinkingBudgetTokens,
+			}, events,
 		)
 		close(errs)
 	}()