Add reasoning effort

internal/llm/client_completions.go

@@ -25,6 +25,15 @@ type CompletionsResult struct {
 	Timings  *Timings
 }
 
+// CompletionsParams holds parameters for Completions and CompletionsStream.
+type CompletionsParams struct {
+	Model                string
+	SystemMessage        string
+	MaxIterations        int
+	Messages             []Message
+	ThinkingBudgetTokens *int
+}
+
 // Completions sends messages to the LLM using the specified model.
 // If systemMessage is non-empty, it overrides the configured system message.
 // maxIterations caps the number of agent loop iterations.
@@ -32,15 +41,14 @@ type CompletionsResult struct {
 // Returns all messages generated during the completions request, including
 // tool calls and tool results, along with usage and timings from the final response.
 func (c *Client) Completions(
-	ctx context.Context,
-	model string,
-	systemMessage string,
-	maxIterations int,
-	messages []Message,
+	ctx context.Context, p CompletionsParams,
 ) (*CompletionsResult, error) {
 	ctx, cancel := context.WithTimeout(ctx, c.timeout)
 	defer cancel()
 
+	model, systemMessage, maxIterations, messages, thinkingBudgetTokens :=
+		p.Model, p.SystemMessage, p.MaxIterations, p.Messages, p.ThinkingBudgetTokens
+
 	// Use request system message if provided, config default otherwise.
 	smsg := c.systemMessage
 	if systemMessage != "" {
@@ -68,7 +76,7 @@ func (c *Client) Completions(
 	var finalUsage *Usage
 	var finalTimings *Timings
 	for i := 0; maxIterations == 0 || i < maxIterations; i++ {
-		cr, err := c.completions(ctx, model, messages)
+		cr, err := c.completions(ctx, model, thinkingBudgetTokens, messages)
 		if err != nil {
 			return nil, fmt.Errorf("completions: %w", err)
 		}
@@ -106,12 +114,18 @@ func (c *Client) Completions(
 // completions sends a single non-streaming request and returns the
 // full API response.
 func (c *Client) completions(
-	ctx context.Context, model string, messages []Message,
+	ctx context.Context, model string, thinkingBudgetTokens *int, messages []Message,
 ) (*ChatResponse, error) {
+	kwargs := map[string]any{}
+	if thinkingBudgetTokens != nil {
+		kwargs["enable_thinking"] = *thinkingBudgetTokens != 0
+	}
 	body, err := json.Marshal(ChatRequest{
-		Model:    model,
-		Messages: messages,
-		Tools:    c.tools,
+		Model:                model,
+		Messages:             messages,
+		Tools:                c.tools,
+		ThinkingBudgetTokens: thinkingBudgetTokens,
+		ChatTemplateKWARGS:   kwargs,
 	})
 	if err != nil {
 		return nil, fmt.Errorf(
@@ -232,17 +246,15 @@ const (
 // means unlimited (no cap).
 // Returns ErrMaxIterations if the iteration limit is exceeded.
 func (c *Client) CompletionsStream(
-	ctx context.Context,
-	model string,
-	systemMessage string,
-	maxIterations int,
-	messages []Message,
-	events chan<- StreamEvent,
+	ctx context.Context, p CompletionsParams, events chan<- StreamEvent,
 ) error {
 	defer close(events)
 	ctx, cancel := context.WithTimeout(ctx, c.timeout)
 	defer cancel()
 
+	model, systemMessage, maxIterations, messages, thinkingBudgetTokens :=
+		p.Model, p.SystemMessage, p.MaxIterations, p.Messages, p.ThinkingBudgetTokens
+
 	// Use request system message if provided, config default otherwise.
 	smsg := c.systemMessage
 	if systemMessage != "" {
@@ -265,7 +277,7 @@ func (c *Client) CompletionsStream(
 
 	// Loop for tool calls. maxIterations == 0 means unlimited.
 	for i := 0; maxIterations == 0 || i < maxIterations; i++ {
-		choices, err := c.completionsStream(ctx, messages, model, events)
+		choices, err := c.completionsStream(ctx, messages, model, thinkingBudgetTokens, events)
 		if err != nil {
 			return fmt.Errorf("completions stream: %w", err)
 		}
@@ -304,13 +316,19 @@ type streamAccum struct {
 // response, and returns all assistant messages from the response choices.
 // Sends each message to the events channel. Usage and timings are included
 // on the done event for the last message.
-func (c *Client) completionsStream(ctx context.Context, messages []Message, model string, events chan<- StreamEvent) ([]Message, error) {
+func (c *Client) completionsStream(ctx context.Context, messages []Message, model string, thinkingBudgetTokens *int, events chan<- StreamEvent) ([]Message, error) {
+	kwargs := map[string]any{}
+	if thinkingBudgetTokens != nil {
+		kwargs["enable_thinking"] = *thinkingBudgetTokens != 0
+	}
 	body, err := json.Marshal(ChatRequest{
-		Model:         model,
-		Messages:      messages,
-		Stream:        true,
-		Tools:         c.tools,
-		StreamOptions: &StreamOptions{IncludeUsage: true},
+		Model:                model,
+		Messages:             messages,
+		Stream:               true,
+		Tools:                c.tools,
+		ThinkingBudgetTokens: thinkingBudgetTokens,
+		ChatTemplateKWARGS:   kwargs,
+		StreamOptions:        &StreamOptions{IncludeUsage: true},
 	})
 	if err != nil {
 		return nil, fmt.Errorf(