Azure provider improvements

.github/instructions/azure-provider-settings.instructions.md

@@ -0,0 +1,273 @@
+---
+applyTo: '**/byok/**'
+description: Azure provider settings format and configuration guidelines
+---
+
+# Azure Provider Settings Configuration
+
+Guidelines for configuring Azure OpenAI models in the BYOK (Bring Your Own Key) system using the `chat.azureModels` VS Code setting.
+
+## Settings Format
+
+The `chat.azureModels` setting accepts a record of model configurations where each key is a unique model identifier and the value is a model configuration object.
+
+### Model Configuration Schema
+
+```typescript
+{
+  "chat.azureModels": {
+    "[modelId]": {
+      "name": string,                    // Display name for the model
+      "url": string,                     // Azure endpoint URL (base URL without API path)
+      "deploymentType": "completions" | "responses",  // API type (default: "completions")
+      "deploymentName": string,          // Azure deployment name (optional, defaults to modelId)
+      "apiVersion": string,              // Azure API version (optional, default: "2025-01-01-preview")
+      "temperature": number,             // Temperature setting (optional, 0.0-2.0)
+      "toolCalling": boolean,            // Whether model supports function/tool calling
+      "vision": boolean,                 // Whether model supports vision/image inputs
+      "maxInputTokens": number,          // Maximum input context window size
+      "maxOutputTokens": number,         // Maximum output tokens per response
+      "requiresAPIKey": boolean,         // Whether model requires API key (optional, default: true)
+      "thinking": boolean                // Whether model supports thinking/reasoning tokens (optional)
+    }
+  }
+}
+```
+
+## URL Resolution Behavior
+
+### Base URL Format
+The `url` field should contain the **base Azure endpoint URL** without the API path:
+
+```json
+// ✅ Correct - base URL only
+"url": "https://my-resource.openai.azure.com"
+"url": "https://my-resource.models.ai.azure.com"
+
+// ❌ Wrong - includes API path
+"url": "https://my-resource.openai.azure.com/openai/deployments/gpt-4/chat/completions"
+```
+
+### Deployment Types
+
+#### 1. Chat Completions API (`deploymentType: "completions"`)
+**Default deployment type.** Uses the standard Azure OpenAI Chat Completions endpoint.
+
+- **URL Pattern**: Deployment name is included in the URL path
+- **Azure OpenAI endpoints** (`openai.azure.com`):
+  - Resolved URL: `{url}/openai/deployments/{deploymentName}/chat/completions?api-version={apiVersion}`
+- **Azure AI endpoints** (`models.ai.azure.com`, `inference.ml.azure.com`):
+  - Resolved URL: `{url}/v1/chat/completions`
+
+**Example Configuration:**
+```json
+{
+  "chat.azureModels": {
+    "gpt-4-turbo": {
+      "name": "GPT-4 Turbo",
+      "url": "https://my-resource.openai.azure.com",
+      "deploymentType": "completions",
+      "deploymentName": "gpt-4-turbo-deployment",
+      "apiVersion": "2024-08-01-preview",
+      "toolCalling": true,
+      "vision": false,
+      "maxInputTokens": 128000,
+      "maxOutputTokens": 4096
+    }
+  }
+}
+```
+
+#### 2. Responses API (`deploymentType: "responses"`)
+Uses the Azure OpenAI Responses API for streaming responses with structured outputs.
+
+- **URL Pattern**: Deployment name is sent in the request body, NOT in URL
+- **Azure OpenAI endpoints** (`openai.azure.com`):
+  - Resolved URL: `{url}/openai/responses?api-version={apiVersion}`
+- **Azure AI endpoints** (`models.ai.azure.com`, `inference.ml.azure.com`):
+  - Resolved URL: `{url}/v1/responses?api-version={apiVersion}`
+- **Model Field**: The `deploymentName` is passed as the `model` field in the request body
+
+**Example Configuration:**
+```json
+{
+  "chat.azureModels": {
+    "o1-preview": {
+      "name": "OpenAI o1-preview",
+      "url": "https://my-resource.openai.azure.com",
+      "deploymentType": "responses",
+      "deploymentName": "o1-preview-deployment",
+      "apiVersion": "2025-01-01-preview",
+      "toolCalling": false,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 128000,
+      "maxOutputTokens": 32768
+    }
+  }
+}
+```
+
+### Deployment Name Behavior
+
+- **If `deploymentName` is provided**: Uses the specified deployment name
+- **If `deploymentName` is omitted**: Falls back to using `modelId` as the deployment name
+- **For Completions API**: Deployment name is included in the URL path
+- **For Responses API**: Deployment name is sent as the `model` field in the request body
+
+## Complete Configuration Examples
+
+### Example 1: Azure OpenAI - Chat Completions
+```json
+{
+  "chat.azureModels": {
+    "gpt-4": {
+      "name": "GPT-4",
+      "url": "https://my-eastus-resource.openai.azure.com",
+      "deploymentType": "completions",
+      "deploymentName": "gpt-4-prod",
+      "apiVersion": "2024-08-01-preview",
+      "temperature": 0.7,
+      "toolCalling": true,
+      "vision": false,
+      "maxInputTokens": 8192,
+      "maxOutputTokens": 4096,
+      "requiresAPIKey": true
+    }
+  }
+}
+```
+
+### Example 2: Azure AI - Responses API with Reasoning
+```json
+{
+  "chat.azureModels": {
+    "o1": {
+      "name": "OpenAI o1",
+      "url": "https://my-resource.models.ai.azure.com",
+      "deploymentType": "responses",
+      "deploymentName": "o1-deployment",
+      "apiVersion": "2025-01-01-preview",
+      "toolCalling": false,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 200000,
+      "maxOutputTokens": 100000,
+      "requiresAPIKey": true
+    }
+  }
+}
+```
+
+### Example 3: Multiple Models
+```json
+{
+  "chat.azureModels": {
+    "gpt-35-turbo": {
+      "name": "GPT-3.5 Turbo",
+      "url": "https://my-resource.openai.azure.com",
+      "deploymentType": "completions",
+      "deploymentName": "gpt-35-turbo-16k",
+      "toolCalling": true,
+      "vision": false,
+      "maxInputTokens": 16384,
+      "maxOutputTokens": 4096
+    },
+    "gpt-4-vision": {
+      "name": "GPT-4 Vision",
+      "url": "https://my-resource.openai.azure.com",
+      "deploymentType": "completions",
+      "deploymentName": "gpt-4-vision-preview",
+      "toolCalling": true,
+      "vision": true,
+      "maxInputTokens": 128000,
+      "maxOutputTokens": 4096
+    },
+    "o1-preview": {
+      "name": "OpenAI o1 Preview",
+      "url": "https://my-resource.openai.azure.com",
+      "deploymentType": "responses",
+      "deploymentName": "o1-preview",
+      "toolCalling": false,
+      "vision": true,
+      "thinking": true,
+      "maxInputTokens": 128000,
+      "maxOutputTokens": 32768
+    }
+  }
+}
+```
+
+## Implementation Details
+
+### URL Resolution Logic (`resolveAzureUrl`)
+
+The `resolveAzureUrl` function in `azureProvider.ts` handles URL construction:
+
+1. **Strips trailing slashes and `/v1` suffixes** from base URLs
+2. **Detects endpoint type** based on URL patterns:
+   - Azure AI: `models.ai.azure.com`, `inference.ml.azure.com`
+   - Azure OpenAI: `openai.azure.com`, `cognitiveservices.azure.com`
+3. **Applies deployment-specific URL patterns** based on `deploymentType`
+4. **Uses defaults** if optional fields are omitted:
+   - `deploymentType`: `"completions"`
+   - `deploymentName`: Falls back to `modelId`
+   - `apiVersion`: `"2025-01-01-preview"`
+
+### Model Info Resolution (`getModelInfo`)
+
+The `getModelInfo` method in `AzureBYOKModelProvider`:
+
+1. **Retrieves deployment configuration** from settings
+2. **For Responses API**: Overrides the model name to use `deploymentName` (sent in request body)
+3. **For Completions API**: Uses `deploymentName` in URL path
+4. **Sets `modelInfo.id`** to the deployment name for proper API routing
+5. **Configures `supported_endpoints`**:
+   - Responses API: `[ModelSupportedEndpoint.Responses]`
+   - Completions API: `[ModelSupportedEndpoint.ChatCompletions]`
+6. **Applies temperature** from configuration if specified
+
+### API Key Management
+
+- API keys are stored securely using VS Code's secrets storage
+- Keys are stored per-model using the pattern: `copilot-byok-Azure-{modelId}-api-key`
+- The `requiresAPIKey` field controls whether authentication is required (default: `true`)
+- Users configure API keys through the UI or via commands
+
+## Best Practices
+
+1. **Use explicit deployment names**: Always specify `deploymentName` to avoid confusion with model IDs
+2. **Match API versions**: Use the API version that matches your Azure deployment capabilities
+3. **Set accurate token limits**: Configure `maxInputTokens` and `maxOutputTokens` based on your deployment
+4. **Enable appropriate capabilities**: Set `toolCalling`, `vision`, and `thinking` flags based on model support
+5. **Test endpoint URLs**: Verify base URLs are correct and accessible before adding models
+6. **Group related models**: Use descriptive model IDs for easy identification in the UI
+7. **Document custom configurations**: Add comments in settings.json to explain non-standard configurations
+
+## Troubleshooting
+
+### Common Issues
+
+**Issue**: Model not appearing in selection
+- **Check**: Verify `requiresAPIKey` is set correctly and API key is configured
+- **Check**: Ensure all required fields are present in configuration
+
+**Issue**: 404 errors when using model
+- **Check**: Verify `deploymentName` matches your Azure deployment
+- **Check**: Ensure `url` points to the correct Azure resource
+- **Check**: Confirm `apiVersion` is supported by your deployment
+
+**Issue**: Unsupported endpoint errors
+- **Check**: Verify `deploymentType` matches your deployment's API type
+- **Check**: For reasoning models (o1), use `deploymentType: "responses"`
+
+**Issue**: Model name mismatches in API calls
+- **For Responses API**: Ensure `deploymentName` is set (it's sent as the model name)
+- **For Completions API**: Deployment name is in URL, not request body
+
+## Related Files
+
+- Implementation: `src/extension/byok/vscode-node/azureProvider.ts`
+- Base provider: `src/extension/byok/vscode-node/customOAIProvider.ts`
+- Configuration: `src/platform/configuration/common/configurationService.ts`
+- Storage: `src/extension/byok/vscode-node/byokStorageService.ts`

package.json

@@ -3035,13 +3035,17 @@
 									"type": "string",
 									"description": "URL endpoint for the Azure model"
 								},
-								"toolCalling": {
-									"type": "boolean",
-									"description": "Whether the model supports tool calling"
+								"deploymentName": {
+									"type": "string",
+									"description": "Azure deployment name for the model"
 								},
-								"vision": {
-									"type": "boolean",
-									"description": "Whether the model supports vision capabilities"
+								"deploymentType": {
+									"type": "string",
+									"description": "Type of Azure deployment"
+								},
+								"apiVersion": {
+									"type": "string",
+									"description": "Azure OpenAI API version to use"
 								},
 								"maxInputTokens": {
 									"type": "number",
@@ -3051,11 +3055,25 @@
 									"type": "number",
 									"description": "Maximum number of output tokens supported by the model"
 								},
+								"toolCalling": {
+									"type": "boolean",
+									"description": "Whether the model supports tool calling"
+								},
+								"vision": {
+									"type": "boolean",
+									"description": "Whether the model supports vision capabilities"
+								},
 								"thinking": {
 									"type": "boolean",
 									"default": false,
 									"description": "Whether the model supports thinking capabilities"
 								},
+								"temperature": {
+									"type": "number",
+									"description": "Temperature parameter for model responses (typically 0.0 to 2.0)",
+									"minimum": 0,
+									"maximum": 2
+								},
 								"requestHeaders": {
 									"type": "object",
 									"description": "Additional HTTP headers to include with requests to this model. These reserved headers are not allowed and ignored if present: forbidden request headers (https://developer.mozilla.org/en-US/docs/Glossary/Forbidden_request_header), forwarding headers ('forwarded', 'x-forwarded-for', 'x-forwarded-host', 'x-forwarded-proto'), and others ('api-key', 'authorization', 'content-type', 'openai-intent', 'x-github-api-version', 'x-initiator', 'x-interaction-id', 'x-interaction-type', 'x-onbehalf-extension-id', 'x-request-id', 'x-vscode-user-agent-library-version'). Pattern-based forbidden headers ('proxy-*', 'sec-*', 'x-http-method*' with forbidden methods) are also blocked.",

src/extension/byok/common/byokProvider.ts

@@ -55,6 +55,7 @@ export interface BYOKModelCapabilities {
 	toolCalling: boolean;
 	vision: boolean;
 	thinking?: boolean;
+	temperature?: number;
 	editTools?: EndpointEditToolName[];
 	requestHeaders?: Record<string, string>;
 }
@@ -153,6 +154,9 @@ export function resolveModelInfo(modelId: string, providerName: string, knownMod
 	if (knownModelInfo?.requestHeaders && Object.keys(knownModelInfo.requestHeaders).length > 0) {
 		modelInfo.requestHeaders = { ...knownModelInfo.requestHeaders };
 	}
+	if (knownModelInfo?.temperature !== undefined) {
+		modelInfo.temperature = knownModelInfo.temperature;
+	}
 	return modelInfo;
 }
 

src/extension/byok/node/openAIEndpoint.ts

@@ -113,6 +113,7 @@ export class OpenAIEndpoint extends ChatEndpoint {
 	private static readonly _maxCustomHeaderCount = 20;
 
 	private readonly _customHeaders: Record<string, string>;
+	private readonly _temperature: number | undefined;
 	constructor(
 		_modelMetadata: IChatModelInformation,
 		protected readonly _apiKey: string,
@@ -144,6 +145,7 @@ export class OpenAIEndpoint extends ChatEndpoint {
 			logService
 		);
 		this._customHeaders = this._sanitizeCustomHeaders(_modelMetadata.requestHeaders);
+		this._temperature = _modelMetadata.temperature;
 	}
 
 	private _sanitizeCustomHeaders(headers: Readonly<Record<string, string>> | undefined): Record<string, string> {
@@ -289,13 +291,22 @@ export class OpenAIEndpoint extends ChatEndpoint {
 		}
 
 		if (body) {
+			// Apply temperature from config if specified (before thinking model checks)
+			if (this._temperature !== undefined) {
+				body.temperature = this._temperature;
+			}
+
 			if (this.modelMetadata.capabilities.supports.thinking) {
+				// Thinking models don't support temperature parameter
 				delete body.temperature;
+				// Thinking models use max_completion_tokens instead of max_tokens
 				body['max_completion_tokens'] = body.max_tokens;
 				delete body.max_tokens;
+			} else {
+				// For non-thinking models, removing max_tokens defaults to the maximum which is what we want for BYOK
+				delete body.max_tokens;
 			}
-			// Removing max tokens defaults to the maximum which is what we want for BYOK
-			delete body.max_tokens;
+
 			if (!this.useResponsesApi && body.stream) {
 				body['stream_options'] = { 'include_usage': true };
 			}