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"
@@ -32,14 +33,15 @@ type OnIterationHook func(Iteration)
 
 // Config holds the dependencies and settings for the agent loop.
 type Config struct {
-	LLM          llm.LLMClient
+	LLM                llm.LLMClient
-	Persona      persona.Persona
+	Persona            persona.Persona
-	Tools        tools.Registry
+	Tools              tools.Registry
-	Sandbox      Sandbox
+	Sandbox            Sandbox
-	MaxIters     int
+	MaxIters           int
-	Approver     Approver
+	Approver           Approver
-	OnIteration  OnIterationHook
+	OnIteration        OnIterationHook
-	ToolTimeout  time.Duration
+	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.
@@ -86,8 +88,9 @@ func (l *Loop) Run(ctx context.Context, input string) (Response, error) {
 		iterations++
 
 		resp, err := l.cfg.LLM.Generate(ctx, llm.CompletionRequest{
-			Messages: messages,
+			Messages:           messages,
-			Tools:    l.getToolSchemas(),
+			Tools:              l.getToolSchemas(),
+			ChatTemplateKwargs: l.cfg.ChatTemplateKwargs,
 		})
 		if err != nil {
 			return Response{}, fmt.Errorf("LLM generate failed: %w", err)
@@ -143,11 +146,13 @@ func (l *Loop) RunStream(ctx context.Context, input string) iter.Seq2[llm.Stream
 			iterations++
 
 			stream := l.cfg.LLM.Stream(ctx, llm.CompletionRequest{
-				Messages: messages,
+				Messages:           messages,
-				Tools:    l.getToolSchemas(),
+				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)
@@ -172,21 +177,20 @@ func (l *Loop) RunStream(ctx context.Context, input string) iter.Seq2[llm.Stream
 					}
 				}
 
-		if !hasToolCalls && chunk.Delta != "" {
+				if !hasToolCalls && chunk.Delta != "" {
-				if !yield(chunk, nil) {
+					responseBuilder.WriteString(chunk.Delta)
-					return
+					if !yield(chunk, nil) {
+						return
+					}
 				}
 			}
-			}
 
 			if !hasToolCalls {
 				return
 			}
 		}
 
-		if iterations >= l.cfg.MaxIters {
+		yield(llm.StreamChunk{}, fmt.Errorf("max iterations (%d) reached", l.cfg.MaxIters))
-			yield(llm.StreamChunk{}, fmt.Errorf("max iterations (%d) reached", l.cfg.MaxIters))
-		}
 	}
 }
 
@@ -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

@@ -173,9 +173,10 @@ func (c *Client) buildRequest(req llm.CompletionRequest, stream bool) (io.Reader
 	}
 
 	openReq := llamaChatRequest{
-		Model:    req.Model,
+		Model:              req.Model,
-		Messages: messages,
+		Messages:           messages,
-		Stream:   stream,
+		Stream:             stream,
+		ChatTemplateKwargs: req.ChatTemplateKwargs,
 	}
 	if len(tools) > 0 {
 		openReq.Tools = tools
@@ -232,16 +233,17 @@ func (c *Client) toResponse(resp llamaChatResponse) llm.CompletionResponse {
 // llama.cpp API types
 
 type llamaChatRequest struct {
-	Model       string              `json:"model"`
+	Model              string              `json:"model"`
-	Messages    []llamaMessage      `json:"messages"`
+	Messages           []llamaMessage      `json:"messages"`
-	Tools       []llamaTool         `json:"tools,omitempty"`
+	Tools              []llamaTool         `json:"tools,omitempty"`
-	ToolChoice interface{}         `json:"tool_choice,omitempty"`
+	ToolChoice        interface{}         `json:"tool_choice,omitempty"`
-	Temperature float32             `json:"temperature,omitempty"`
+	Temperature       float32             `json:"temperature,omitempty"`
-	MaxTokens   int                `json:"max_tokens,omitempty"`
+	MaxTokens         int                `json:"max_tokens,omitempty"`
-	TopK        int                `json:"top_k,omitempty"`
+	TopK              int                `json:"top_k,omitempty"`
-	TopP        float32            `json:"top_p,omitempty"`
+	TopP              float32            `json:"top_p,omitempty"`
-	Stop        []string           `json:"stop,omitempty"`
+	Stop              []string           `json:"stop,omitempty"`
-	Stream      bool               `json:"stream"`
+	Stream            bool               `json:"stream"`
+	ChatTemplateKwargs map[string]any     `json:"chat_template_kwargs,omitempty"`
 }
 
 type llamaMessage struct {

pkg/llm/types.go

@@ -66,14 +66,15 @@ func (t *ToolRef) MarshalJSON() ([]byte, error) {
 
 // CompletionRequest is sent to an LLM provider.
 type CompletionRequest struct {
-	Model       string              `json:"model"`
+	Model             string              `json:"model"`
-	Messages    []Message          `json:"messages"`
+	Messages          []Message          `json:"messages"`
-	Tools       []json.RawMessage  `json:"tools,omitempty"`
+	Tools             []json.RawMessage  `json:"tools,omitempty"`
-	ToolChoice interface{}        `json:"tool_choice,omitempty"` // ToolChoice, ToolRef, or null
+	ToolChoice       interface{}        `json:"tool_choice,omitempty"` // ToolChoice, ToolRef, or null
-	Temperature *float32           `json:"temperature,omitempty"`
+	Temperature      *float32           `json:"temperature,omitempty"`
-	MaxTokens   *int              `json:"max_tokens,omitempty"`
+	MaxTokens        *int              `json:"max_tokens,omitempty"`
-	Stop        []string          `json:"stop,omitempty"`
+	Stop             []string           `json:"stop,omitempty"`
-	Metadata    map[string]string `json:"metadata,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.