fix(azure): use max_completion_tokens and omit temperature for reasoning deployments

src/chrome/src/providers/azure-openai.js

@@ -1,5 +1,6 @@
 import { BaseLLMProvider } from './base.js';
 import { fetchWithFallback } from './fetch-with-fallback.js';
+import { configuredMaxTokensField } from './provider-compatibility.js';
 
 /**
  * Azure OpenAI provider (deployment-based OpenAI-compatible API).
@@ -62,11 +63,24 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
   }
 
   _addMaxTokens(body, options) {
-    // Azure OpenAI follows the legacy OpenAI contract here.
+    // The Compatibility panel's "max tokens field" is the explicit
+    // per-deployment switch. `_addConfiguredMaxTokens` already honors
+    // `compat.maxTokensField` (and `config.maxTokensField`) through
+    // `configuredMaxTokensField`, so a reasoning deployment set to
+    // `max_completion_tokens` in Settings sends the new field with no
+    // name-guessing here.
     this._addConfiguredMaxTokens(body, options, 'max_tokens');
   }
 
   _addTemperature(body, options) {
+    // Azure deployment names do not reveal the model behind them (prod-chat
+    // may back o3-mini; o365-assistant may back gpt-35-turbo), so never guess
+    // the contract from the name. A reasoning deployment opts in by setting
+    // the Compatibility "max tokens field" to `max_completion_tokens` (which
+    // also drops temperature — reasoning models reject any non-default), or
+    // by enabling `omitTemperature` directly.
+    if (this.config.omitTemperature) return;
+    if (configuredMaxTokensField(this.config, 'max_tokens') === 'max_completion_tokens') return;
     body.temperature = options.temperature ?? 0.7;
   }
 

src/firefox/src/providers/azure-openai.js

@@ -1,5 +1,6 @@
 import { BaseLLMProvider } from './base.js';
 import { fetchWithTimeout } from './fetch-timeout.js';
+import { configuredMaxTokensField } from './provider-compatibility.js';
 
 /**
  * Azure OpenAI provider (deployment-based OpenAI-compatible API).
@@ -60,10 +61,24 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
   }
 
   _addMaxTokens(body, options) {
+    // The Compatibility panel's "max tokens field" is the explicit
+    // per-deployment switch. `_addConfiguredMaxTokens` already honors
+    // `compat.maxTokensField` (and `config.maxTokensField`) through
+    // `configuredMaxTokensField`, so a reasoning deployment set to
+    // `max_completion_tokens` in Settings sends the new field with no
+    // name-guessing here.
     this._addConfiguredMaxTokens(body, options, 'max_tokens');
   }
 
   _addTemperature(body, options) {
+    // Azure deployment names do not reveal the model behind them (prod-chat
+    // may back o3-mini; o365-assistant may back gpt-35-turbo), so never guess
+    // the contract from the name. A reasoning deployment opts in by setting
+    // the Compatibility "max tokens field" to `max_completion_tokens` (which
+    // also drops temperature — reasoning models reject any non-default), or
+    // by enabling `omitTemperature` directly.
+    if (this.config.omitTemperature) return;
+    if (configuredMaxTokensField(this.config, 'max_tokens') === 'max_completion_tokens') return;
     body.temperature = options.temperature ?? 0.7;
   }
 

test/run.js

@@ -49153,6 +49153,65 @@ test('Azure OpenAI Ask streams require DONE and distinguish terminal API errors'
   }
 });
 
+test('Azure OpenAI does not guess the wire contract from the deployment name', () => {
+  for (const Provider of [AzureOpenAIProviderCh, AzureOpenAIProviderFx]) {
+    // Deployment names don't reveal the model behind them: o1-named
+    // deployments stay legacy, and a name that merely starts with o3/o4
+    // must not flip a gpt-35-turbo deployment to the new contract.
+    for (const deployment of ['o1', 'o1-mini', 'o3-mini', 'o4-mini', 'gpt-5', 'gpt-4.1', 'o365-assistant', 'prod-chat']) {
+      const provider = new Provider({
+        providerName: 'azure-openai',
+        baseUrl: 'https://example.openai.azure.com',
+        model: deployment,
+        apiVersion: '2025-09-01',
+        apiKey: 'test-key',
+      });
+      const body = provider._buildRequestBody(
+        [{ role: 'user', content: 'hello' }],
+        { maxTokens: 2048 },
+      );
+      assert.equal(body.max_tokens, 2048, `${Provider.name}/${deployment}: default contract keeps max_tokens`);
+      assert.equal(body.max_completion_tokens, undefined, `${Provider.name}/${deployment}: no name-guessing into the new contract`);
+      assert.equal(body.temperature, 0.7, `${Provider.name}/${deployment}: default contract keeps temperature`);
+    }
+  }
+});
+
+test('Azure OpenAI switches to max_completion_tokens and omits temperature via the explicit compat field', () => {
+  for (const Provider of [AzureOpenAIProviderCh, AzureOpenAIProviderFx]) {
+    const provider = new Provider({
+      providerName: 'azure-openai',
+      baseUrl: 'https://example.openai.azure.com',
+      model: 'prod-chat',
+      apiVersion: '2025-09-01',
+      apiKey: 'test-key',
+      compat: { maxTokensField: 'max_completion_tokens' },
+    });
+    const body = provider._buildRequestBody(
+      [{ role: 'user', content: 'hello' }],
+      { maxTokens: 2048 },
+    );
+    assert.equal(body.max_completion_tokens, 2048, `${Provider.name}: explicit field must use max_completion_tokens`);
+    assert.equal(body.max_tokens, undefined, `${Provider.name}: legacy field must be absent`);
+    assert.equal(body.temperature, undefined, `${Provider.name}: reasoning contract omits temperature`);
+  }
+});
+
+test('Azure OpenAI omitTemperature config suppresses temperature on any deployment', () => {
+  for (const Provider of [AzureOpenAIProviderCh, AzureOpenAIProviderFx]) {
+    const provider = new Provider({
+      providerName: 'azure-openai',
+      baseUrl: 'https://example.openai.azure.com',
+      model: 'custom-reasoning-deployment',
+      apiVersion: '2025-09-01',
+      apiKey: 'test-key',
+      omitTemperature: true,
+    });
+    const body = provider._buildRequestBody([{ role: 'user', content: 'hello' }]);
+    assert.equal(body.temperature, undefined, `${Provider.name}: omitTemperature must be respected`);
+  }
+});
+
 test('Ask stream aggregation exposes text live but withholds tool calls until response.completed', async () => {
   for (const [label, AgentClass] of [['chrome', AgentCh], ['firefox', AgentFx]]) {
     let releaseCompleted;