add reasoning effort setting for providers

Add a reasoning effort field to the completion provider config, sent as-is to the model. Reasoning models such as GPT-5.x need it set to "none" to use tools. Also switch to max_completion_tokens.

frontend/apps/main/src/features/admin/ai/ProviderForm.vue

@@ -79,6 +79,17 @@
               </FormItem>
             </FormField>
 
+            <FormField v-slot="{ componentField }" name="reasoning_effort">
+              <FormItem>
+                <FormLabel>{{ t('admin.ai.reasoningEffort') }}</FormLabel>
+                <FormControl>
+                  <Input type="text" placeholder="none" v-bind="componentField" />
+                </FormControl>
+                <FormDescription>{{ t('admin.ai.reasoningEffortHint') }}</FormDescription>
+                <FormMessage />
+              </FormItem>
+            </FormField>
+
             <FormField v-slot="{ componentField }" name="instructions">
               <FormItem class="md:col-span-2">
                 <FormLabel>{{ t('admin.ai.instructions') }}</FormLabel>
@@ -264,6 +275,7 @@ const form = useForm({
       base_url: z.string().optional(),
       api_key: z.string().optional(),
       instructions: z.string().optional(),
+      reasoning_effort: z.string().optional(),
       temperature: numberField().refine((v) => isBlank(v) || inRange(v, 0, 2), {
         message: t('admin.ai.temperatureRange')
       }),
@@ -281,6 +293,7 @@ const form = useForm({
     base_url: '',
     api_key: '',
     instructions: '',
+    reasoning_effort: '',
     temperature: fieldDefaults.temperature,
     max_tokens: fieldDefaults.maxTokens,
     dimensions: fieldDefaults.dimensions,
@@ -298,6 +311,7 @@ onMounted(async () => {
       base_url: data.base_url || '',
       api_key: data.api_key || '',
       instructions: data.instructions || '',
+      reasoning_effort: data.reasoning_effort || '',
       // Fall back to sensible defaults when the provider has nothing saved (0/absent).
       temperature: data.temperature != null ? String(data.temperature) : fieldDefaults.temperature,
       max_tokens: data.max_tokens ? String(data.max_tokens) : fieldDefaults.maxTokens,
@@ -321,6 +335,7 @@ const onSubmit = form.handleSubmit(async (values) => {
   if (props.showCompletionFields) {
     payload.instructions = values.instructions || ''
     payload.vision = !!values.vision
+    payload.reasoning_effort = (values.reasoning_effort || '').trim()
     if (values.temperature !== '' && values.temperature != null)
       payload.temperature = Number(values.temperature)
     if (values.max_tokens !== '' && values.max_tokens != null)

i18n/en-US.json

@@ -130,6 +130,8 @@
   "admin.ai.temperatureRange": "Temperature must be a number between 0 and 2.",
   "admin.ai.positiveNumber": "Enter a whole number greater than 0.",
   "admin.ai.maxTokens": "Max tokens",
+  "admin.ai.reasoningEffort": "Reasoning effort",
+  "admin.ai.reasoningEffortHint": "Leave blank for most models. Reasoning models such as GPT-5.x require 'none' to use tools; other values are 'low', 'medium' or 'high'.",
   "admin.ai.instructions": "Custom instructions",
   "admin.ai.instructionsHint": "Added to every Copilot and reply-draft prompt. Use it to set tone and format, e.g. keep replies under 3 sentences, match the brand voice, always reply in the customer's language.",
   "admin.ai.vision": "Model supports image input",

internal/ai/ai.go

@@ -218,15 +218,16 @@ func (m *Manager) UpdateProviderConfig(providerType string, in models.ProviderCo
 	}
 
 	cfg := models.ProviderConfig{
-		Provider:     "openai",
-		BaseURL:      in.BaseURL,
-		APIKey:       apiKey,
-		Model:        in.Model,
-		Temperature:  in.Temperature,
-		MaxTokens:    in.MaxTokens,
-		Dimensions:   in.Dimensions,
-		Instructions: in.Instructions,
-		Vision:       in.Vision,
+		Provider:        "openai",
+		BaseURL:         in.BaseURL,
+		APIKey:          apiKey,
+		Model:           in.Model,
+		Temperature:     in.Temperature,
+		MaxTokens:       in.MaxTokens,
+		Dimensions:      in.Dimensions,
+		Instructions:    in.Instructions,
+		Vision:          in.Vision,
+		ReasoningEffort: in.ReasoningEffort,
 	}
 	b, err := json.Marshal(cfg)
 	if err != nil {

internal/ai/models/models.go

@@ -45,6 +45,9 @@ type ProviderConfig struct {
 	MaxTokens    int      `json:"max_tokens"`
 	Dimensions   int      `json:"dimensions"`
 	Instructions string   `json:"instructions"`
+	// ReasoningEffort, when set, is sent as-is. Reasoning models (e.g. GPT-5.x) need "none" to use
+	// tools on /chat/completions; leave blank for models that reject the param.
+	ReasoningEffort string `json:"reasoning_effort,omitempty"`
 	// Vision marks the completion model as accepting image input; when false the harness drops image parts.
 	Vision bool `json:"vision"`
 }

internal/ai/openai.go

@@ -84,13 +84,19 @@ func (o *OpenAIClient) SendChatCompletion(payload models.ChatCompletionPayload)
 	o.lo.Debug("chat completion request", "model", model, "messages", len(messages), "images", sentImages, "vision", o.cfg.Vision, "tools", len(payload.Tools))
 
 	body := map[string]any{
-		"model":      model,
-		"messages":   messages,
-		"max_tokens": maxTokens,
+		"model":                 model,
+		"messages":              messages,
+		"max_completion_tokens": maxTokens,
 	}
+	// Only send optional params the admin set. Reasoning models (e.g. GPT-5.x) reject a non-default
+	// temperature and require reasoning_effort "none" to use tools on /chat/completions; leave
+	// temperature blank and set reasoning_effort accordingly for those.
 	if o.cfg.Temperature != nil {
 		body["temperature"] = *o.cfg.Temperature
 	}
+	if o.cfg.ReasoningEffort != "" {
+		body["reasoning_effort"] = o.cfg.ReasoningEffort
+	}
 	if len(payload.Tools) > 0 {
 		body["tools"] = payload.Tools
 		body["tool_choice"] = "auto"