Add reasoning effort

internal/job/job.go

@@ -157,11 +157,12 @@ func (j *Job) Run(ctx context.Context) Result {
 	}
 
 	res, err := j.llm.Completions(
-		ctx,
+		ctx, llm.CompletionsParams{
-		j.model,
+			Model:         j.model,
-		j.SystemMessage(),
+			SystemMessage: j.SystemMessage(),
-		j.MaxIterations(),
+			MaxIterations: j.MaxIterations(),
-		messages,
+			Messages:      messages,
+		},
 	)
 
 	duration := time.Since(start)

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,
+	ctx context.Context, p CompletionsParams,
-	model string,
-	systemMessage string,
-	maxIterations int,
-	messages []Message,
 ) (*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,
+		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,
+	ctx context.Context, p CompletionsParams, events chan<- StreamEvent,
-	model string,
-	systemMessage string,
-	maxIterations int,
-	messages []Message,
-	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)
 		}
@@ -304,12 +316,18 @@ 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,
+		ThinkingBudgetTokens: thinkingBudgetTokens,
+		ChatTemplateKWARGS:   kwargs,
 		StreamOptions:        &StreamOptions{IncludeUsage: true},
 	})
 	if err != nil {

internal/llm/client_completions_test.go

@@ -35,18 +35,18 @@ func TestCompletionsStreamStreamsReasoningDeltas(t *testing.T) {
 	errCh := make(chan error, 1)
 	go func() {
 		errCh <- client.CompletionsStream(
-			context.Background(),
+			context.Background(), CompletionsParams{
-			"test-model",
+				Model:         "test-model",
-			"",
+				SystemMessage: "",
-			0,
+				MaxIterations: 0,
-			[]Message{{
+				Messages: []Message{{
 					Role: RoleUser,
 					ContentParts: []ContentPart{{
 						Type: ContentTypeText,
 						Text: "hello",
 					}},
 				}},
-			events,
+			}, events,
 		)
 	}()
 
@@ -131,17 +131,18 @@ func TestCompletionsReturnsUsageAndTimings(t *testing.T) {
 	}
 
 	res, err := client.Completions(
-		context.Background(),
+		context.Background(), CompletionsParams{
-		"test-model",
+			Model:         "test-model",
-		"",
+			SystemMessage: "",
-		0,
+			MaxIterations: 0,
-		[]Message{{
+			Messages: []Message{{
 				Role: RoleUser,
 				ContentParts: []ContentPart{{
 					Type: ContentTypeText,
 					Text: "hi",
 				}},
 			}},
+		},
 	)
 	if err != nil {
 		t.Fatalf("Completions: %v", err)
@@ -205,18 +206,18 @@ func TestCompletionsStreamEmitsUsageEvent(t *testing.T) {
 	errCh := make(chan error, 1)
 	go func() {
 		errCh <- client.CompletionsStream(
-			context.Background(),
+			context.Background(), CompletionsParams{
-			"test-model",
+				Model:         "test-model",
-			"",
+				SystemMessage: "",
-			0,
+				MaxIterations: 0,
-			[]Message{{
+				Messages: []Message{{
 					Role: RoleUser,
 					ContentParts: []ContentPart{{
 						Type: ContentTypeText,
 						Text: "hi",
 					}},
 				}},
-			events,
+			}, events,
 		)
 	}()
 

internal/llm/llm.go

@@ -55,7 +55,8 @@ type ChatRequest struct {
 	ToolChoice           any             `json:"tool_choice,omitempty"`
 	StreamOptions        *StreamOptions  `json:"stream_options,omitempty"`
 	ParallelToolCalls    *bool           `json:"parallel_tool_calls,omitempty"`
-	ReasoningEffort     string          `json:"reasoning_effort,omitempty"`
+	ChatTemplateKWARGS   map[string]any  `json:"chat_template_kwargs,omitempty"`
+	ThinkingBudgetTokens *int            `json:"thinking_budget_tokens,omitempty"`
 	ServiceTier          string          `json:"service_tier,omitempty"`
 }
 

internal/llm/llm_integration_test.go

@@ -77,17 +77,18 @@ func TestIntegration_Completions(t *testing.T) {
 	}
 
 	res, err := client.Completions(
-		context.Background(),
+		context.Background(), CompletionsParams{
-		testModel(t),
+			Model:         testModel(t),
-		"",
+			SystemMessage: "",
-		0,
+			MaxIterations: 0,
-		[]Message{{
+			Messages: []Message{{
 				Role: RoleUser,
 				ContentParts: []ContentPart{{
 					Type: "text",
 					Text: "Say hello in three words.",
 				}},
 			}},
+		},
 	)
 	if err != nil {
 		t.Fatalf("Completions: %v", err)
@@ -117,18 +118,18 @@ func TestIntegration_CompletionsStream(t *testing.T) {
 	errCh := make(chan error, 1)
 	go func() {
 		errCh <- client.CompletionsStream(
-			context.Background(),
+			context.Background(), CompletionsParams{
-			testModel(t),
+				Model:         testModel(t),
-			"",
+				SystemMessage: "",
-			0,
+				MaxIterations: 0,
-			[]Message{{
+				Messages: []Message{{
 					Role: RoleUser,
 					ContentParts: []ContentPart{{
 						Type: "text",
 						Text: "Say hello in three words.",
 					}},
 				}},
-			events,
+			}, events,
 		)
 	}()
 
@@ -182,17 +183,18 @@ func TestIntegration_ToolCalling(t *testing.T) {
 	}
 
 	res, err := client.Completions(
-		context.Background(),
+		context.Background(), CompletionsParams{
-		testModel(t),
+			Model:         testModel(t),
-		"",
+			SystemMessage: "",
-		10,
+			MaxIterations: 10,
-		[]Message{{
+			Messages: []Message{{
 				Role: RoleUser,
 				ContentParts: []ContentPart{{
 					Type: "text",
 					Text: "Use the echo tool to say 'hello'",
 				}},
 			}},
+		},
 	)
 	if err != nil {
 		t.Fatalf("Completions: %v", err)

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)
 	}()

internal/service/static/main.css

@@ -730,7 +730,7 @@ body {
 
 .settings-label__sub {
   display: block;
-  font-size: var(--s-2);
+  font-size: var(--s-1);
   font-weight: 400;
   color: var(--color-text-muted);
   margin-top: 0.125rem;
@@ -756,10 +756,6 @@ body {
   outline-offset: 2px;
 }
 
-.settings-field {
-  margin-bottom: var(--s1);
-}
-
 .settings-textarea {
   display: block;
   width: 100%;

internal/service/static/main.js

@@ -15,6 +15,7 @@ import { WakeLock } from './wakelock.js';
 const STREAM_ENDPOINT = '/v1/chat/voice/stream';
 const SYSTEM_MESSAGE_KEY = 'odidere_system_message';
 const MAX_TURNS_KEY = 'odidere_max_turns';
+const REASONING_BUDGET_KEY = 'odidere_reasoning_budget';
 const MUTE_KEY = 'odidere_mute';
 const ICONS_URL = '/static/icons.svg';
 
@@ -111,6 +112,9 @@ class Odidere {
       'compose-settings-close',
     );
     this.$maxTurnsInput = document.getElementById('max-turns-input');
+    this.$reasoningBudgetInput = document.getElementById(
+      'reasoning-budget-input',
+    );
     this.$compose = document.querySelector('.compose');
 
     this.init();
@@ -268,9 +272,6 @@ class Odidere {
       if (e.target === this.$composeSettingsOverlay)
         this._closeComposeSettings();
     });
-    this.$maxTurnsInput.addEventListener('input', () => {
-      this._setMaxTurns(this.$maxTurnsInput.value);
-    });
 
     // --- Model/voice change persistence ---
     this.settings.$model.addEventListener('change', () =>
@@ -640,6 +641,11 @@ class Odidere {
           payload.max_iterations = maxTurns;
         }
 
+        const reasoningBudget = this._getReasoningBudget();
+        if (reasoningBudget !== null) {
+          payload.thinking_budget_tokens = reasoningBudget;
+        }
+
         // Clear attachments and location after building payload.
         this.clearAttachments();
         this.clearLocation();
@@ -1170,11 +1176,14 @@ class Odidere {
 
   _openComposeSettings() {
     this.$maxTurnsInput.value = this._getMaxTurns();
+    this.$reasoningBudgetInput.value = this._getReasoningBudget() ?? -1;
     this.$composeSettingsOverlay.classList.add('open');
     this.$composeSettingsClose.focus();
   }
 
   _closeComposeSettings() {
+    this._setMaxTurns(this.$maxTurnsInput.value);
+    this._setReasoningBudget(this.$reasoningBudgetInput.value);
     this.$composeSettingsOverlay.classList.remove('open');
     this.$composeSettings.focus();
   }
@@ -1188,6 +1197,16 @@ class Odidere {
     localStorage.setItem(MAX_TURNS_KEY, String(value));
   }
 
+  _getReasoningBudget() {
+    const stored = localStorage.getItem(REASONING_BUDGET_KEY);
+    if (stored === null) return null;
+    return Number(stored);
+  }
+
+  _setReasoningBudget(value) {
+    localStorage.setItem(REASONING_BUDGET_KEY, String(value));
+  }
+
   // ====================
   // TWO-PRESS SAFETY PATTERNS
   // ====================

internal/service/templates/static/modal/compose-settings.gohtml

@@ -29,6 +29,24 @@
           aria-label="Max iterations"
         />
       </div>
+
+      <div class="settings-field">
+        <label class="settings-label" for="reasoning-budget-input">
+          Reasoning Budget
+          <span class="settings-label__sub">
+            Only supported by some models and inference engines.
+            <br>
+            -1 = unlimited, 0 = off · 512 ≈ low · 2048 ≈ med · 8192 ≈ high
+          </span>
+        </label>
+        <input
+          type="number"
+          class="settings-select"
+          id="reasoning-budget-input"
+          value="-1"
+          aria-label="Reasoning budget"
+        />
+      </div>
     </div>
   </div>
 {{ end }}