feat(llm): add ChatTemplateKwargs to CompletionRequest for provider-specific template params

- Add ChatTemplateKwargs field to llm.CompletionRequest
- Propagate kwargs through agent loop in both Run() and RunStream()
- Pass kwargs to llama.cpp client chat request
- Fix tool schema marshaling to include type/function wrapper
- Fix stream indentation logic in RunStream with responseBuilder
- Remove indirect marker from uuid dependency

go.mod

@@ -4,4 +4,4 @@ go 1.26
 
 require gopkg.in/yaml.v3 v3.0.1
 
-require github.com/google/uuid v1.6.0 // indirect
+require github.com/google/uuid v1.6.0

pkg/agent/loop.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"iter"
+	"strings"
 	"time"
 
 	"github.com/VictorVargas/rony-llm-agent/pkg/llm"
@@ -40,6 +41,7 @@ type Config struct {
 	Approver           Approver
 	OnIteration        OnIterationHook
 	ToolTimeout        time.Duration
+	ChatTemplateKwargs map[string]any // passed to the LLM provider (e.g. Qwen enable_thinking)
 }
 
 // Iteration represents a single cycle of the agent loop.
@@ -88,6 +90,7 @@ func (l *Loop) Run(ctx context.Context, input string) (Response, error) {
 		resp, err := l.cfg.LLM.Generate(ctx, llm.CompletionRequest{
 			Messages:           messages,
 			Tools:              l.getToolSchemas(),
+			ChatTemplateKwargs: l.cfg.ChatTemplateKwargs,
 		})
 		if err != nil {
 			return Response{}, fmt.Errorf("LLM generate failed: %w", err)
@@ -145,9 +148,11 @@ func (l *Loop) RunStream(ctx context.Context, input string) iter.Seq2[llm.Stream
 			stream := l.cfg.LLM.Stream(ctx, llm.CompletionRequest{
 				Messages:           messages,
 				Tools:              l.getToolSchemas(),
+				ChatTemplateKwargs: l.cfg.ChatTemplateKwargs,
 			})
 
 			var hasToolCalls bool
+			var responseBuilder strings.Builder
 			for chunk, err := range stream {
 				if err != nil {
 					yield(llm.StreamChunk{}, err)
@@ -173,6 +178,7 @@ func (l *Loop) RunStream(ctx context.Context, input string) iter.Seq2[llm.Stream
 				}
 
 				if !hasToolCalls && chunk.Delta != "" {
+					responseBuilder.WriteString(chunk.Delta)
 					if !yield(chunk, nil) {
 						return
 					}
@@ -184,10 +190,8 @@ func (l *Loop) RunStream(ctx context.Context, input string) iter.Seq2[llm.Stream
 			}
 		}
 
-		if iterations >= l.cfg.MaxIters {
 		yield(llm.StreamChunk{}, fmt.Errorf("max iterations (%d) reached", l.cfg.MaxIters))
 	}
-	}
 }
 
 func (l *Loop) buildInitialMessages(input string) []llm.Message {
@@ -201,7 +205,16 @@ func (l *Loop) buildInitialMessages(input string) []llm.Message {
 func (l *Loop) getToolSchemas() []json.RawMessage {
 	var schemas []json.RawMessage
 	for _, tool := range l.cfg.Tools.List() {
-		schemas = append(schemas, tool.InputSchema)
+		def := map[string]interface{}{
+			"type": "function",
+			"function": map[string]interface{}{
+				"name":        tool.Name,
+				"description": tool.Description,
+				"parameters":  json.RawMessage(tool.InputSchema),
+			},
+		}
+		data, _ := json.Marshal(def)
+		schemas = append(schemas, data)
 	}
 	return schemas
 }

pkg/llm/providers/llamacpp/client.go

@@ -176,6 +176,7 @@ func (c *Client) buildRequest(req llm.CompletionRequest, stream bool) (io.Reader
 		Model:              req.Model,
 		Messages:           messages,
 		Stream:             stream,
+		ChatTemplateKwargs: req.ChatTemplateKwargs,
 	}
 	if len(tools) > 0 {
 		openReq.Tools = tools
@@ -242,6 +243,7 @@ type llamaChatRequest struct {
 	TopP              float32            `json:"top_p,omitempty"`
 	Stop              []string           `json:"stop,omitempty"`
 	Stream            bool               `json:"stream"`
+	ChatTemplateKwargs map[string]any     `json:"chat_template_kwargs,omitempty"`
 }
 
 type llamaMessage struct {

pkg/llm/types.go

@@ -74,6 +74,7 @@ type CompletionRequest struct {
 	MaxTokens        *int              `json:"max_tokens,omitempty"`
 	Stop             []string           `json:"stop,omitempty"`
 	Metadata         map[string]string `json:"metadata,omitempty"`
+	ChatTemplateKwargs map[string]any    `json:"chat_template_kwargs,omitempty"` // model-specific chat template params, e.g. Qwen enable_thinking
 }
 
 // CompletionResponse is returned from an LLM provider.