fix: make reasoning effort opt-in, disable by default

CHANGELOG.md

@@ -10,6 +10,8 @@
 
 ### Fixed
 
+- 默认不再发送 `reasoning.effort`：移除 `modelInfo.defaultReasoningEffort` 自动兜底，`default_reasoning_effort` 默认改为 `null`，彻底消除简单对话触发 medium 推理导致的 token 暴涨；Dashboard 新增 "Disabled (no reasoning)" 选项，用户可按需开启
+
 - 上游 401 时立即触发 RT→AT 刷新，而非等待定时器（修复 token 被提前作废后账号一直显示 expired 的问题）
 - Dashboard session 滑动窗口续期：每次有效请求自动延长过期时间，不再固定 TTL 后断连
 - Dashboard 前端全局 401 拦截：session 过期后自动跳回登录页，不再卡死在空白页

config/default.yaml

@@ -10,7 +10,7 @@ client:
   chromium_version: "144"
 model:
   default: gpt-5.2-codex
-  default_reasoning_effort: medium
+  default_reasoning_effort: null
   default_service_tier: null
   inject_desktop_context: false
   suppress_desktop_directives: false

shared/hooks/use-general-settings.ts

@@ -8,7 +8,7 @@ export interface GeneralSettingsData {
   inject_desktop_context: boolean;
   suppress_desktop_directives: boolean;
   default_model: string;
-  default_reasoning_effort: string;
+  default_reasoning_effort: string | null;
   refresh_enabled: boolean;
   refresh_margin_seconds: number;
   refresh_concurrency: number;

src/__tests__/config-local-override.test.ts

@@ -44,7 +44,7 @@ client:
   chromium_version: "136"
 model:
   default: test-model
-  default_reasoning_effort: medium
+  default_reasoning_effort: null
   default_service_tier: null
   inject_desktop_context: false
   suppress_desktop_directives: false

src/__tests__/config-schema.test.ts

@@ -37,7 +37,7 @@ describe("ConfigSchema", () => {
     expect(result.auth.max_concurrent_per_account).toBe(3);
     expect(result.auth.request_interval_ms).toBe(50);
     expect(result.model.default).toBe("gpt-5.2-codex");
-    expect(result.model.default_reasoning_effort).toBe("medium");
+    expect(result.model.default_reasoning_effort).toBeNull();
     expect(result.tls.force_http11).toBe(false);
     expect(result.quota.refresh_interval_minutes).toBe(5);
     expect(result.quota.warning_thresholds.primary).toEqual([80, 90]);

src/config-schema.ts

@@ -17,7 +17,7 @@ export const ConfigSchema = z.object({
   }),
   model: z.object({
     default: z.string().default("gpt-5.2-codex"),
-    default_reasoning_effort: z.string().default("medium"),
+    default_reasoning_effort: z.string().nullable().default(null),
     default_service_tier: z.string().nullable().default(null),
     inject_desktop_context: z.boolean().default(false),
     suppress_desktop_directives: z.boolean().default(true),

src/routes/__tests__/general-settings.test.ts

@@ -11,7 +11,7 @@ import { describe, it, expect, vi, beforeEach } from "vitest";
 const mockConfig = {
   server: { port: 8080, proxy_api_key: null as string | null },
   tls: { proxy_url: null as string | null, force_http11: false },
-  model: { default: "gpt-5.2-codex", default_reasoning_effort: "medium", inject_desktop_context: false, suppress_desktop_directives: true },
+  model: { default: "gpt-5.2-codex", default_reasoning_effort: null as string | null, inject_desktop_context: false, suppress_desktop_directives: true },
   quota: {
     refresh_interval_minutes: 5,
     warning_thresholds: { primary: [80, 90], secondary: [80, 90] },
@@ -114,7 +114,7 @@ describe("GET /admin/general-settings", () => {
       inject_desktop_context: false,
       suppress_desktop_directives: true,
       default_model: "gpt-5.2-codex",
-      default_reasoning_effort: "medium",
+      default_reasoning_effort: null,
       refresh_enabled: true,
       refresh_margin_seconds: 300,
       refresh_concurrency: 2,
@@ -325,6 +325,30 @@ describe("POST /admin/general-settings", () => {
     expect(res.status).toBe(400);
   });
 
+  it("accepts null default_reasoning_effort to disable reasoning", async () => {
+    mockConfig.model.default_reasoning_effort = "medium";
+    const app = makeApp();
+    const res = await app.request("/admin/general-settings", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ default_reasoning_effort: null }),
+    });
+    expect(res.status).toBe(200);
+    const data = await res.json();
+    expect(data.success).toBe(true);
+    expect(mutateYaml).toHaveBeenCalledOnce();
+  });
+
+  it("rejects invalid default_reasoning_effort", async () => {
+    const app = makeApp();
+    const res = await app.request("/admin/general-settings", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({ default_reasoning_effort: "ultra" }),
+    });
+    expect(res.status).toBe(400);
+  });
+
   it("requires auth when proxy_api_key is set", async () => {
     mockConfig.server.proxy_api_key = "my-secret";
     const app = makeApp();

src/routes/admin/settings.ts

@@ -132,7 +132,7 @@ export function createSettingsRoutes(): Hono {
       inject_desktop_context?: boolean;
       suppress_desktop_directives?: boolean;
       default_model?: string;
-      default_reasoning_effort?: string;
+      default_reasoning_effort?: string | null;
       refresh_enabled?: boolean;
       refresh_margin_seconds?: number;
       refresh_concurrency?: number;
@@ -161,9 +161,12 @@ export function createSettingsRoutes(): Hono {
 
     if (body.default_reasoning_effort !== undefined) {
       const validEfforts = ["low", "medium", "high", "xhigh"];
-      if (!validEfforts.includes(body.default_reasoning_effort)) {
+      if (
+        body.default_reasoning_effort !== null &&
+        !validEfforts.includes(body.default_reasoning_effort)
+      ) {
         c.status(400);
-        return c.json({ error: `default_reasoning_effort must be one of: ${validEfforts.join(", ")}` });
+        return c.json({ error: `default_reasoning_effort must be one of: ${validEfforts.join(", ")} or null` });
       }
     }
 

src/routes/responses.ts

@@ -486,19 +486,21 @@ export function createResponsesRoutes(
       codexRequest.previous_response_id = body.previous_response_id;
     }
 
-    // Reasoning effort: explicit body > suffix > model default > config default
+    // Reasoning effort: explicit body > suffix > config default
     const effort =
       (isRecord(body.reasoning) && typeof body.reasoning.effort === "string"
         ? body.reasoning.effort
         : null) ??
       parsed.reasoningEffort ??
-      modelInfo?.defaultReasoningEffort ??
       config.model.default_reasoning_effort;
-    const summary =
-      isRecord(body.reasoning) && typeof body.reasoning.summary === "string"
-        ? body.reasoning.summary
-        : "auto";
-    codexRequest.reasoning = { summary, ...(effort ? { effort } : {}) };
+    const clientReasoningRecord = isRecord(body.reasoning) ? body.reasoning : null;
+    if (effort || clientReasoningRecord) {
+      const summary =
+        clientReasoningRecord && typeof clientReasoningRecord.summary === "string"
+          ? clientReasoningRecord.summary
+          : "auto";
+      codexRequest.reasoning = { summary, ...(effort ? { effort } : {}) };
+    }
 
     // Service tier
     const serviceTier =

src/routes/shared/proxy-handler.ts

@@ -136,13 +136,24 @@ export async function handleProxyRequest(
     const inputItems = req.codexRequest.input?.length ?? 0;
     const instrLen = req.codexRequest.instructions?.length ?? 0;
     const affinityHit = preferredEntryId && entryId === preferredEntryId;
+    const reasoningField = req.codexRequest.reasoning
+      ? `effort=${req.codexRequest.reasoning.effort ?? "none"} summary=${req.codexRequest.reasoning.summary ?? "none"}`
+      : "off";
     console.log(
-      `[${fmt.tag}] Account ${entryId} | model=${req.model} | input_items=${inputItems} instr=${instrLen}B payload=${reqJson.length}B` +
+      `[${fmt.tag}] Account ${entryId} | model=${req.model} | input_items=${inputItems} instr=${instrLen}B payload=${reqJson.length}B reasoning=[${reasoningField}]` +
       (prevRespId ? ` | affinity=${affinityHit ? "hit" : "miss"}` : ""),
     );
     if (reqJson.length > 50_000) {
+      // Log per-item size breakdown to diagnose large payload origin
+      const itemSizes = (req.codexRequest.input ?? []).map((item, i) => {
+        const sz = JSON.stringify(item).length;
+        const role = typeof item === "object" && item !== null && "role" in item ? (item as Record<string, unknown>).role : (item as Record<string, unknown>).type;
+        return `  [${i}] ${role} ${sz}B`;
+      });
       console.warn(
-        `[${fmt.tag}] ⚠ Large payload (${(reqJson.length / 1024).toFixed(1)}KB) — input_items=${inputItems} instr=${instrLen}B`,
+        `[${fmt.tag}] ⚠ Large payload (${(reqJson.length / 1024).toFixed(1)}KB) — input_items=${inputItems} instr=${instrLen}B\n` +
+        `  instructions: ${instrLen}B\n` +
+        itemSizes.join("\n"),
       );
     }
   }
@@ -208,11 +219,21 @@ export async function handleProxyRequest(
               affinityMap.record(capturedResponseId, capturedEntryId, conversationId, upstreamTurnState);
             }
             if (usageInfo) {
+              const uncached = usageInfo.cached_tokens
+                ? usageInfo.input_tokens - usageInfo.cached_tokens
+                : usageInfo.input_tokens;
               console.log(
-                `[${fmt.tag}] Account ${capturedEntryId} | Usage: in=${usageInfo.input_tokens} out=${usageInfo.output_tokens}` +
-                (usageInfo.cached_tokens ? ` cached=${usageInfo.cached_tokens}` : "") +
+                `[${fmt.tag}] Account ${capturedEntryId} | Usage: in=${usageInfo.input_tokens}` +
+                (usageInfo.cached_tokens ? ` (cached=${usageInfo.cached_tokens} uncached=${uncached})` : "") +
+                ` out=${usageInfo.output_tokens}` +
                 (usageInfo.reasoning_tokens ? ` reasoning=${usageInfo.reasoning_tokens}` : ""),
               );
+              if (usageInfo.input_tokens > 10_000) {
+                console.warn(
+                  `[${fmt.tag}] ⚠ High input token count: ${usageInfo.input_tokens} tokens` +
+                  (usageInfo.reasoning_tokens ? ` (reasoning=${usageInfo.reasoning_tokens})` : ""),
+                );
+              }
             }
             releaseAccount(accountPool, capturedEntryId, usageInfo, released);
           }
@@ -295,6 +316,19 @@ async function handleNonStreaming(
       if (result.responseId && affinityMap && conversationId) {
         affinityMap.record(result.responseId, currentEntryId, conversationId, turnState);
       }
+      if (result.usage) {
+        const u = result.usage;
+        const uncached = u.cached_tokens ? u.input_tokens - u.cached_tokens : u.input_tokens;
+        console.log(
+          `[${fmt.tag}] Account ${currentEntryId} | Usage: in=${u.input_tokens}` +
+          (u.cached_tokens ? ` (cached=${u.cached_tokens} uncached=${uncached})` : "") +
+          ` out=${u.output_tokens}` +
+          (u.reasoning_tokens ? ` reasoning=${u.reasoning_tokens}` : ""),
+        );
+        if (u.input_tokens > 10_000) {
+          console.warn(`[${fmt.tag}] ⚠ High input token count: ${u.input_tokens} tokens`);
+        }
+      }
       releaseAccount(accountPool, currentEntryId, result.usage, released);
       return c.json(result.response);
     } catch (collectErr) {

src/translation/anthropic-to-codex.ts

@@ -228,14 +228,15 @@ export function translateAnthropicToCodexRequest(
     request.tool_choice = codexToolChoice;
   }
 
-  // Reasoning effort: thinking config > suffix > model default > config default
+  // Reasoning effort: thinking config > suffix > config default
   const thinkingEffort = mapThinkingToEffort(req.thinking);
   const effort =
     thinkingEffort ??
     parsed.reasoningEffort ??
-    modelInfo?.defaultReasoningEffort ??
     cfg.default_reasoning_effort;
-  request.reasoning = { summary: "auto", ...(effort ? { effort } : {}) };
+  if (effort) {
+    request.reasoning = { effort, summary: "auto" };
+  }
 
   // Service tier: suffix > config default
   const serviceTier =

src/translation/gemini-to-codex.ts

@@ -219,16 +219,17 @@ export function translateGeminiToCodexRequest(
     request.tool_choice = codexToolChoice;
   }
 
-  // Reasoning effort: thinking config > suffix > model default > config default
+  // Reasoning effort: thinking config > suffix > config default
   const thinkingEffort = budgetToEffort(
     req.generationConfig?.thinkingConfig?.thinkingBudget,
   );
   const effort =
     thinkingEffort ??
     parsed.reasoningEffort ??
-    modelInfo?.defaultReasoningEffort ??
     cfg.default_reasoning_effort;
-  request.reasoning = { summary: "auto", ...(effort ? { effort } : {}) };
+  if (effort) {
+    request.reasoning = { effort, summary: "auto" };
+  }
 
   // Service tier: suffix > config default
   const serviceTier =

src/translation/openai-to-codex.ts

@@ -182,13 +182,14 @@ export function translateToCodexRequest(
     request.tool_choice = codexToolChoice;
   }
 
-  // Reasoning effort: explicit API field > suffix > model default > config default
+  // Reasoning effort: explicit API field > suffix > config default
   const effort =
     req.reasoning_effort ??
     parsed.reasoningEffort ??
-    modelInfo?.defaultReasoningEffort ??
     cfg.default_reasoning_effort;
-  request.reasoning = { summary: "auto", ...(effort ? { effort } : {}) };
+  if (effort) {
+    request.reasoning = { effort, summary: "auto" };
+  }
 
   // Service tier: explicit API field > suffix > config default
   const serviceTier =

tests/integration/account-routing.test.ts

@@ -32,7 +32,7 @@ vi.mock("@src/config.js", () => ({
     },
     model: {
       default: "gpt-5.2-codex",
-      default_reasoning_effort: "medium",
+      default_reasoning_effort: null,
       default_service_tier: null,
     },
     server: { proxy_api_key: null },
@@ -73,7 +73,7 @@ describe("account-routing integration", () => {
       },
       model: {
         default: "gpt-5.2-codex",
-        default_reasoning_effort: "medium",
+        default_reasoning_effort: null,
         default_service_tier: null,
       },
       server: { proxy_api_key: null },
@@ -166,7 +166,7 @@ describe("account-routing integration", () => {
       },
       model: {
         default: "gpt-5.2-codex",
-        default_reasoning_effort: "medium",
+        default_reasoning_effort: null,
         default_service_tier: null,
       },
       server: { proxy_api_key: null },