[codex] Fix OpenAI reasoning defaults

deno.json

@@ -1,6 +1,6 @@
 {
   "name": "veryfront",
-  "version": "0.1.987",
+  "version": "0.1.988",
   "license": "Apache-2.0",
   "nodeModulesDir": "auto",
   "minimumDependencyAge": {

extensions/ext-llm-openai/README.md

@@ -57,7 +57,7 @@ Any model accessible through the OpenAI Chat Completions, Responses, or Embeddin
 
 - **Flagship:** `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-4o`, `gpt-4o-mini`
 - **Frontier:** `gpt-5`, `gpt-5-mini`, `gpt-5-nano`
-- **Reasoning:** `o3`, `o4-mini`, `o1`, `o1-mini`, `o3-mini` (sampling parameters are automatically dropped with warnings)
+- **Reasoning:** `gpt-5.4-nano`, current `gpt-5`/`gpt-5.x` reasoning models, `o3`, `o4-mini`, `o1`, `o3-mini` (sampling parameters are automatically dropped with warnings)
 - **Embeddings:** `text-embedding-3-small`, `text-embedding-3-large`
 - **OpenAI-compatible:** Any third-party model reachable via an OpenAI-compatible endpoint (set `OPENAI_BASE_URL`)
 
@@ -74,7 +74,11 @@ The extension accepts configuration through `LLMProviderConfig` when creating ru
 
 ## Model-Specific Behavior
 
-### Reasoning Models (o3, o4-mini, o1)
+### Reasoning Models (GPT-5.x, o3, o4-mini, o1)
+
+Default reasoning params are applied only for native `openai` and `veryfront-cloud` providers.
+OpenAI-compatible providers require explicit `reasoning` options. `gpt-5-chat-latest`,
+`gpt-5.1`, `o1-mini`, and `o1-preview` are left unmodified by default.
 
 Reasoning models automatically:
 

extensions/ext-llm-openai/src/openai-chat-request-builder.test.ts

@@ -27,6 +27,102 @@ function createWarningCollector() {
 }
 
 describe("ext-llm-openai/openai-chat-request-builder", () => {
+  it("sets default reasoning effort for GPT-5.5 chat requests", () => {
+    const warnings = createWarningCollector();
+
+    const body = buildOpenAIChatRequest(
+      "gpt-5.5",
+      "openai",
+      {
+        prompt: [{ role: "user", content: [{ type: "text", text: "Think carefully." }] }],
+        temperature: 0.2,
+      },
+      true,
+      warnings,
+    );
+
+    assertEquals(body.reasoning_effort, "medium");
+    assertEquals(body.temperature, undefined);
+    assertEquals(warnings.drain().map((warning) => warning.setting), ["temperature"]);
+  });
+
+  it("does not set default reasoning effort for GPT-5 chat snapshots", () => {
+    const warnings = createWarningCollector();
+
+    const body = buildOpenAIChatRequest(
+      "gpt-5-chat-latest",
+      "openai",
+      {
+        prompt: [{ role: "user", content: [{ type: "text", text: "Be concise." }] }],
+        temperature: 0.2,
+      },
+      true,
+      warnings,
+    );
+
+    assertEquals(body.reasoning_effort, undefined);
+    assertEquals(body.temperature, 0.2);
+    assertEquals(warnings.drain(), []);
+  });
+
+  it("does not set default reasoning effort for legacy o1 chat variants", () => {
+    const warnings = createWarningCollector();
+
+    const body = buildOpenAIChatRequest(
+      "o1-mini",
+      "openai",
+      {
+        prompt: [{ role: "user", content: [{ type: "text", text: "Be concise." }] }],
+        temperature: 0.2,
+      },
+      true,
+      warnings,
+    );
+
+    assertEquals(body.reasoning_effort, undefined);
+    assertEquals(body.temperature, undefined);
+    assertEquals(warnings.drain().map((warning) => warning.setting), ["temperature"]);
+  });
+
+  it("does not set default reasoning effort for OpenAI-compatible providers but still drops rejected sampling params", () => {
+    const warnings = createWarningCollector();
+
+    const body = buildOpenAIChatRequest(
+      "gpt-5.5",
+      "azure",
+      {
+        prompt: [{ role: "user", content: [{ type: "text", text: "Be concise." }] }],
+        temperature: 0.2,
+      },
+      true,
+      warnings,
+    );
+
+    assertEquals(body.reasoning_effort, undefined);
+    assertEquals(body.temperature, undefined);
+    assertEquals(warnings.drain().map((warning) => warning.setting), ["temperature"]);
+  });
+
+  it("drops rejected sampling params when explicit reasoning is disabled", () => {
+    const warnings = createWarningCollector();
+
+    const body = buildOpenAIChatRequest(
+      "o3-mini",
+      "openai",
+      {
+        prompt: [{ role: "user", content: [{ type: "text", text: "Be concise." }] }],
+        reasoning: { enabled: false },
+        temperature: 0.2,
+      },
+      true,
+      warnings,
+    );
+
+    assertEquals(body.reasoning_effort, undefined);
+    assertEquals(body.temperature, undefined);
+    assertEquals(warnings.drain().map((warning) => warning.setting), ["temperature"]);
+  });
+
   it("preserves chat request shaping, provider option merge order, and warnings", () => {
     const prompt: RuntimePromptMessage[] = [
       { role: "system", content: "You are concise." },

extensions/ext-llm-openai/src/openai-chat-request-builder.ts

@@ -5,14 +5,11 @@ import {
   unwrapToolInputSchema,
 } from "veryfront/provider/shared";
 import type { OpenAICompatibleChatRequest, RuntimePromptMessage } from "veryfront/provider/shared";
-
-type ProviderReasoningEffort = "low" | "medium" | "high" | "max";
-
-type ProviderReasoningOption = {
-  enabled?: boolean;
-  effort?: ProviderReasoningEffort;
-  budgetTokens?: number;
-};
+import {
+  type OpenAIProviderReasoningOption,
+  rejectsOpenAISamplingParams,
+  resolveOpenAIReasoningConfig,
+} from "./openai-reasoning-models.ts";
 
 export type RuntimeToolDefinition =
   | {
@@ -44,7 +41,7 @@ export type OpenAICompatibleLanguageOptions = {
   providerOptions?: Record<string, unknown>;
   includeRawChunks?: boolean;
   abortSignal?: AbortSignal;
-  reasoning?: ProviderReasoningOption;
+  reasoning?: OpenAIProviderReasoningOption;
   userId?: string;
   serviceTier?: "auto" | "default" | "flex" | "scale";
   parallelToolCalls?: boolean;
@@ -75,10 +72,6 @@ type WarningCollector = {
   }>;
 };
 
-function isOpenAIReasoningModel(modelId: string): boolean {
-  return /^o[134](-|$)/.test(modelId);
-}
-
 function isNativeOpenAIModel(modelId: string): boolean {
   return /^(gpt-|o[134](-|$)|chatgpt-)/.test(modelId);
 }
@@ -87,36 +80,18 @@ function isFixedSamplingModel(modelId: string): boolean {
   return /^kimi-k2\.5/.test(modelId);
 }
 
-function resolveOpenAIReasoningEffort(
-  option: ProviderReasoningOption | undefined,
-): "low" | "medium" | "high" | undefined {
-  if (!option || option.enabled !== true) {
-    return undefined;
-  }
-  switch (option.effort) {
-    case "low":
-      return "low";
-    case "high":
-    case "max":
-      return "high";
-    case "medium":
-    default:
-      return "medium";
-  }
-}
-
 export function buildOpenAIChatRequest(
   modelId: string,
   providerName: string,
   options: OpenAICompatibleLanguageOptions,
   stream: boolean,
   warnings: WarningCollector,
 ): OpenAICompatibleChatRequest {
-  const isReasoningModel = isOpenAIReasoningModel(modelId);
-  const reasoningEffort = resolveOpenAIReasoningEffort(options.reasoning);
-  const reasoningEnabled = isReasoningModel || reasoningEffort !== undefined;
+  const reasoning = resolveOpenAIReasoningConfig(modelId, providerName, options.reasoning);
+  const reasoningEnabled = reasoning !== undefined;
+  const samplingRejected = rejectsOpenAISamplingParams(modelId);
   const fixedSampling = isFixedSamplingModel(modelId);
-  const dropSamplingParams = reasoningEnabled || fixedSampling;
+  const dropSamplingParams = reasoningEnabled || samplingRejected || fixedSampling;
 
   // OpenAI Chat Completions has no top_k surface.
   if (options.topK !== undefined) {
@@ -128,7 +103,7 @@ export function buildOpenAIChatRequest(
     });
   }
 
-  // Reasoning models (o1 / o3 / o4) and models with fixed sampling params
+  // Reasoning models and models with fixed sampling params
   // reject sampling params outright. Emit warnings.
   if (dropSamplingParams) {
     const dropped: Array<[keyof typeof options, string]> = [
@@ -145,7 +120,9 @@ export function buildOpenAIChatRequest(
           setting: key,
           details: fixedSampling
             ? `Dropped because this model uses fixed sampling parameters.`
-            : `Dropped because OpenAI reasoning models reject ${openaiName}. Reasoning was active for this request.`,
+            : samplingRejected
+            ? `Dropped because this model rejects ${openaiName}.`
+            : `Dropped because reasoning was active for this request and OpenAI rejects ${openaiName} with reasoning.`,
         });
       }
     }
@@ -178,7 +155,7 @@ export function buildOpenAIChatRequest(
     ...(!dropSamplingParams && options.frequencyPenalty !== undefined
       ? { frequency_penalty: options.frequencyPenalty }
       : {}),
-    ...(reasoningEffort !== undefined ? { reasoning_effort: reasoningEffort } : {}),
+    ...(reasoning !== undefined ? { reasoning_effort: reasoning.effort } : {}),
     ...(typeof options.userId === "string" && options.userId.length > 0
       ? { user: options.userId }
       : {}),

extensions/ext-llm-openai/src/openai-reasoning-models.test.ts

@@ -0,0 +1,59 @@
+import { assertEquals } from "#veryfront/testing/assert.ts";
+import { describe, it } from "#veryfront/testing/bdd.ts";
+import {
+  getDefaultOpenAIReasoningEffort,
+  rejectsOpenAISamplingParams,
+} from "./openai-reasoning-models.ts";
+
+describe("ext-llm-openai/openai-reasoning-models", () => {
+  it("defaults known reasoning models while excluding chat snapshots and legacy o1 variants", () => {
+    const cases: Array<[string, "medium" | undefined]> = [
+      ["gpt-5", "medium"],
+      ["gpt-5-mini", "medium"],
+      ["gpt-5.4-nano", "medium"],
+      ["gpt-5.5", "medium"],
+      ["gpt-5.1", undefined],
+      ["gpt-5-chat-latest", undefined],
+      ["o1", "medium"],
+      ["o1-2024-12-17", "medium"],
+      ["o1-mini", undefined],
+      ["o1-preview", undefined],
+      ["o3-mini", "medium"],
+      ["o4-mini", "medium"],
+    ];
+
+    for (const [modelId, expected] of cases) {
+      assertEquals(getDefaultOpenAIReasoningEffort(modelId), expected, modelId);
+    }
+  });
+
+  it("only enables default reasoning params for native OpenAI providers", () => {
+    assertEquals(getDefaultOpenAIReasoningEffort("gpt-5.4-nano", "openai"), "medium");
+    assertEquals(getDefaultOpenAIReasoningEffort("gpt-5.4-nano", "veryfront-cloud"), "medium");
+    assertEquals(getDefaultOpenAIReasoningEffort("gpt-5.4-nano", "azure"), undefined);
+    assertEquals(getDefaultOpenAIReasoningEffort("gpt-5.4-nano", "moonshot"), undefined);
+  });
+
+  it("detects models that reject sampling params separately from default reasoning params", () => {
+    const cases: Array<[string, boolean]> = [
+      ["gpt-5", true],
+      ["gpt-5-mini", true],
+      ["gpt-5.4-nano", true],
+      ["gpt-5.5", true],
+      ["gpt-5.1", false],
+      ["gpt-5-chat-latest", false],
+      ["o1", true],
+      ["o1-2024-12-17", true],
+      ["o1-mini", true],
+      ["o1-preview", true],
+      ["o1-pro", true],
+      ["o3-mini", true],
+      ["o4-mini", true],
+      ["gpt-4o-mini", false],
+    ];
+
+    for (const [modelId, expected] of cases) {
+      assertEquals(rejectsOpenAISamplingParams(modelId), expected, modelId);
+    }
+  });
+});