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)
 	}()

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 }}