feat(models): 模型映射新增最大输出 Token 列（max_output_size 直写 config.toml，models.dev 参考值一键填入/空值自动补全），移除无效的声明支持 1M 复选框；思考等级顺序修正为 low/medium/high/xhigh/max 且标签全英文 — bump to v0.8.0

CHANGELOG.md

@@ -6,6 +6,12 @@
 
 ---
 
+## v0.8.0 (2026-10-03)
+
+- **模型映射新增「最大输出 Token」列**：逐模型设置 `max_output_size`（写入 config.toml，留空用上游默认），models.dev 有 output 上限时显示可点击参考值一键填入；获取模型列表时自动预填
+- **models.dev 快照新增 output 上限字段**：同步至 2026-10-03（8,385 模型 / 226 供应商，output 覆盖 97.4%）；应用内同步（Rust 侧）同步支持，避免同步后字段丢失
+- **移除无效的「声明支持 1M」复选框**：该控件读时由上下文长度派生、写时不持久化，勾选从不生效，由「最大输出 Token」列取代
+
 ## v0.7.23 (2026-09-30)
 
 - **思考等级五档 + 能力感知**：全局配置的思考等级扩展为低 / 中 / 高 / Max / XHigh，按当前默认模型声明的 `support_efforts` 自动置灰不支持档位并提示原因；models.dev 标记不支持思考的模型整体禁用思考区；未声明档位的模型保留可选并提示"上游可能拒绝"

README.md

@@ -79,7 +79,7 @@
 | **连通性测试** | 实测 `base_url` 延迟，绿/橙/红彩色气泡，6 秒自动消失 |
 | **重复供应商** | 一键深拷贝供应商 + 全部模型，key 改 `xxx-copy` |
 | **图标按钮操作** | 启用 / 编辑 / 复制 / 测试连通 / 删除 全图标化（lucide-react） |
-| **模型映射** | 别名（"provider/model" 形式）↔ 实际请求模型 ID，自定义显示名、上下文长度、1M 上下文声明、能力 |
+| **模型映射** | 别名（"provider/model" 形式）↔ 实际请求模型 ID，自定义显示名、上下文长度、最大输出 Token（`max_output_size`）、能力（仅 Kimi Code） |
 | **自动上下文** | 拉取模型时 **API 返回 > models.dev ref > 正则兜底** 三级优先级自动适配 |
 | **能力自动推导** | `image_in / video_in / tool_use` 全部由 models.dev 推得；UI 仅暴露 `thinking` 一个手动开关 |
 | **全局设置** | `[thinking]` 表完整支持（enabled / effort / keep），仅 Kimi Code 生效 |
@@ -111,7 +111,7 @@
 
 ![编辑供应商-模型映射](docs/screenshots/provider-model-mapping.png)
 
-一张表管理全部模型映射：显示名、实际请求模型、上下文长度、1M 上下文声明、能力（仅"思考"）、设为默认、删除。
+一张表管理全部模型映射：显示名、实际请求模型、上下文长度、最大输出 Token（`max_output_size`，可点参考值填 models.dev 上限）、能力（仅"思考"）、设为默认、删除（后两项仅 Kimi Code）。
 
 **用量仪表盘**
 

README_EN.md

@@ -73,7 +73,7 @@ Build instructions: [`docs/BUILD.md`](./docs/BUILD.md).
 | **Connectivity test** | Real `base_url` latency, coloured bubble (green / orange / red), auto-dismiss in 6 seconds |
 | **Duplicate provider** | Deep-copy a provider + all its models; key auto-suffixed to `xxx-copy` |
 | **Iconified actions** | Activate / Edit / Duplicate / Test Connectivity / Delete via lucide-react |
-| **Model mapping** | Alias (`"provider/model"`) ↔ real model ID, with display name, context size, 1M-context flag, and capabilities |
+| **Model mapping** | Alias (`"provider/model"`) ↔ real model ID, with display name, context size, max output tokens (`max_output_size`), and capabilities (Kimi Code only) |
 | **Auto context size** | On model fetch: **API response > models.dev ref > regex fallback** — three-tier priority |
 | **Auto capabilities** | `image_in / video_in / tool_use` all derived from models.dev; UI only exposes `thinking` as a manual toggle |
 | **Global settings** | Full `[thinking]` table (enabled / effort / keep); Kimi Code only |
@@ -105,7 +105,7 @@ Provider name, notes, official URL, managed-provider toggle, API format, API key
 
 ![Edit provider — model mapping](docs/screenshots/provider-model-mapping.png)
 
-A single table for all model mappings: display name, real model ID, context length, 1M-context flag, capability (thinking only), default toggle, delete.
+A single table for all model mappings: display name, real model ID, context length, max output tokens (`max_output_size`, with a click-to-fill models.dev reference), capability (thinking only), default toggle, delete (the last two are Kimi Code only).
 
 **Usage dashboard**
 
@@ -175,7 +175,7 @@ Per-workspace Kimi Code session browsing, active / archived / all filters; strea
 - **`config.toml` is the authoritative source for Kimi Code**: all providers and models are always written in full; `default_model` selects the active one (matching the CLI's native `/provider` behaviour). Switching only changes `default_model`; newly added providers are auto-promoted to the top of the list and never get overwritten
 - **SQLite holds Kimi Switch-private metadata only**: notes, official URLs, per-agent remembered default model (the `settings` table), ordering. Theme / language / last-update-check live in frontend `localStorage` (WebView2), not under `~/.kimi-switch`. It acts as a fallback when `config.toml` is incomplete
 - **`raw_other` passes unknown fields through untouched**, including `[oauth]` blocks — round-trips never drop fields
-- **models.dev snapshot**: derived from `https://models.dev/api.json`, cached to a local JSON; `capabilitiesFromRef` derives `thinking / image_in / video_in / tool_use`, `getModelRef` derives `max_context_size / display_name`
+- **models.dev snapshot**: derived from `https://models.dev/api.json`, cached to a local JSON; `capabilitiesFromRef` derives `thinking / image_in / video_in / tool_use`, `getModelRef` derives `max_context_size / display_name / max output tokens`
 - **Override env vars**: `KIMI_CODE_HOME` / `PI_CODING_AGENT_DIR` override the Kimi Code / Pi dirs; Kimi Switch's own data dir is fixed at `~/.kimi-switch` (no env override yet). See [Data Storage Locations](#data-storage-locations)
 
 ## Feature Details

docs/release-notes/release-notes-v0.8.0.md

@@ -0,0 +1,14 @@
+# KimiSwitch v0.8.0
+
+## 新功能
+
+- **模型映射表新增「最大输出 Token」列**（Kimi Code）：为每个模型单独设置 `max_output_size`，写入 config.toml 的 `[models."<alias>"]`，留空则用上游默认值。models.dev 收录了该模型的输出上限时，输入框下方显示可点击的参考值（如 `参考 128000`），一键填入；「获取模型列表」添加模型时也会自动预填参考值
+- **models.dev 快照新增 output 上限字段**：内置快照同步至 2026-10-03（8,385 模型 / 226 供应商，97.4% 模型带 output 上限）；应用内「同步 models.dev」（Rust 侧）同步支持该字段，同步后参考值不丢失
+
+## 修复
+
+- **移除无效的「声明支持 1M」复选框**（Kimi Code）：该控件在读取时由上下文长度派生、写入时不持久化，勾选永远不会生效，已由「最大输出 Token」列取代
+
+## 备注
+
+- 前端 + 抓取脚本 + Rust 快照构建逻辑；新增 `max_output_size` 读写纯函数 9 个单测 + config.toml 全链路往返测试，全量测试通过（前端 126 / Rust 162）

package.json

@@ -1,7 +1,7 @@
 {
   "name": "kimiswitch",
   "private": true,
-  "version": "0.7.23",
+  "version": "0.8.0",
   "type": "module",
   "scripts": {
     "dev": "vite",

scripts/fetch-models-dev.mjs

@@ -133,6 +133,11 @@ async function main() {
       if (typeof m.limit?.context === "number" && m.limit.context > 0) {
         entry.context = m.limit.context;
       }
+      // Max output tokens — the reference value behind the max_output_size
+      // column (model mapping table).
+      if (typeof m.limit?.output === "number" && m.limit.output > 0) {
+        entry.output = m.limit.output;
+      }
       if (m.reasoning === true) entry.reasoning = true;
       if (m.tool_call === true) entry.tool_call = true;
       if (m.structured_output === true) entry.structured_output = true;

src-tauri/Cargo.lock

@@ -1978,7 +1978,7 @@ dependencies = [
 
 [[package]]
 name = "kimiswitch"
-version = "0.7.23"
+version = "0.8.0"
 dependencies = [
  "anyhow",
  "chrono",

src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "kimiswitch"
-version = "0.7.23"
+version = "0.8.0"
 description = "Kimi Switch - model config manager"
 authors = ["codingplan.site"]
 edition = "2021"

src-tauri/src/kimi_code_io.rs

@@ -648,6 +648,114 @@ api_key = ""
         assert!(caps.iter().any(|v| v.as_str() == Some("thinking")));
     }
 
+    #[test]
+    fn kimi_code_export_writes_and_drops_max_output_size() {
+        // max_output_size is not a first-class Model field: it round-trips
+        // through raw_other and must land in config.toml verbatim, and must be
+        // removable (unset = upstream default) without disturbing its siblings.
+        let mut providers = IndexMap::new();
+        providers.insert(
+            "p".to_string(),
+            Provider {
+                name: "p".to_string(),
+                provider_type: ProviderType::Openai,
+                base_url: Some("https://a.example.com".to_string()),
+                api_key: Some("sk-a".to_string()),
+                api_key_env: None,
+                env: IndexMap::new(),
+                note: None,
+                official_url: None,
+                managed: false,
+                enabled: true,
+                active: true,
+                icon: None,
+                icon_color: None,
+                raw_other: Value::Null,
+                usage_kinds: None,
+                usage_config: None,
+            },
+        );
+        let make_model = |raw_other: Value| Model {
+            alias: "p/m".to_string(),
+            provider: "p".to_string(),
+            model: "m".to_string(),
+            max_context_size: 128_000,
+            display_name: None,
+            supports_1m: false,
+            capabilities: vec![],
+            raw_other,
+        };
+
+        let mut models = IndexMap::new();
+        models.insert(
+            "p/m".to_string(),
+            make_model(serde_json::json!({
+                "max_output_size": 32768,
+                "support_efforts": ["low", "high"],
+            })),
+        );
+        let config = Config {
+            default_model: None,
+            providers,
+            models,
+            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
+        };
+
+        let exported = config_to_kimi_code(&config, None);
+        let root = exported.as_table().unwrap();
+        let model = root
+            .get("models")
+            .unwrap()
+            .as_table()
+            .unwrap()
+            .get("p/m")
+            .unwrap()
+            .as_table()
+            .unwrap();
+        assert_eq!(model.get("max_output_size").and_then(|v| v.as_integer()), Some(32768));
+        assert!(model.get("support_efforts").is_some());
+
+        // Re-import through real TOML text (the same serialize → parse path
+        // save_kimi_code_config / load_kimi_code_config use): the key must
+        // come back in the model's raw_other as an integer.
+        let toml_text = toml::to_string_pretty(&exported).unwrap();
+        assert!(toml_text.contains("max_output_size = 32768"), "toml: {toml_text}");
+        let reparsed: TomlValue = toml_text.parse().unwrap();
+        let parsed = kimi_code_to_config(&reparsed);
+        let loaded = parsed.models.get("p/m").unwrap();
+        assert_eq!(
+            loaded.raw_other.get("max_output_size").and_then(|v| v.as_i64()),
+            Some(32768)
+        );
+
+        // Clearing the override drops the key entirely.
+        let mut models = IndexMap::new();
+        models.insert(
+            "p/m".to_string(),
+            make_model(serde_json::json!({ "support_efforts": ["low", "high"] })),
+        );
+        let config = Config {
+            default_model: None,
+            providers: IndexMap::new(),
+            models,
+            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
+        };
+        let exported = config_to_kimi_code(&config, None);
+        let root = exported.as_table().unwrap();
+        let model = root
+            .get("models")
+            .unwrap()
+            .as_table()
+            .unwrap()
+            .get("p/m")
+            .unwrap()
+            .as_table()
+            .unwrap();
+        assert!(model.get("max_output_size").is_none());
+    }
+
     #[test]
     fn kimi_code_export_adds_opencode_go_session_header() {
         let make_provider = |base_url: &str, raw_other: Value| Provider {

src-tauri/src/models_dev.rs

@@ -138,6 +138,14 @@ fn build_snapshot(raw: &Value) -> Result<Value, String> {
             {
                 entry.insert("context".into(), num(ctx));
             }
+            // Max output tokens — reference for the max_output_size column.
+            if let Some(out) = m
+                .pointer("/limit/output")
+                .and_then(|x| x.as_f64())
+                .filter(|o| *o > 0.0)
+            {
+                entry.insert("output".into(), num(out));
+            }
             for flag in ["reasoning", "tool_call", "structured_output"] {
                 if m.get(flag).and_then(|x| x.as_bool()) == Some(true) {
                     entry.insert(flag.into(), Value::Bool(true));
@@ -286,7 +294,7 @@ mod tests {
                     },
                     "kimi-image": {
                         "name": "Kimi Image",
-                        "limit": { "context": 0 },
+                        "limit": { "context": 0, "output": 0 },
                         "modalities": { "input": ["text", "image", "video"] }
                     }
                 }
@@ -309,6 +317,7 @@ mod tests {
         let k3 = &obj["moonshotai/kimi-k3"];
         assert_eq!(k3["name"], "Kimi K3");
         assert_eq!(k3["context"], 262144);
+        assert_eq!(k3["output"], 16384);
         assert_eq!(k3["reasoning"], true);
         assert_eq!(k3["tool_call"], true);
         assert_eq!(k3["structured_output"], true);
@@ -318,9 +327,10 @@ mod tests {
         assert_eq!(k3["cost"]["cache_read"], 0.3);
         assert_eq!(k3["cost"]["cache_write"], 0.0);
 
-        // Context 0 → dropped; video modality → flagged.
+        // Context 0 / output 0 → dropped; video modality → flagged.
         let img = &obj["moonshotai/kimi-image"];
         assert!(img.get("context").is_none());
+        assert!(img.get("output").is_none());
         assert_eq!(img["video"], true);
         assert!(img.get("cost").is_none());
 

src-tauri/tauri.conf.json

@@ -1,6 +1,6 @@
 {
   "productName": "Kimi Switch",
-  "version": "0.7.23",
+  "version": "0.8.0",
   "identifier": "com.kimiswitch.app",
   "build": {
     "beforeDevCommand": "npm run dev",

src/components/ProviderEdit.tsx

@@ -1,11 +1,17 @@
-import { useEffect, useId, useMemo, useState } from "react";
+import { useEffect, useId, useMemo, useReducer, useRef, useState } from "react";
 import { createPortal } from "react-dom";
 import { invoke } from "@tauri-apps/api/core";
 import { useTranslation } from "../i18n";
 import type { TranslationKey } from "../i18n/zh";
 import { findPresetForProvider } from "../config/providerPresets";
 import { getDefaultMaxContextSize } from "../lib/model-defaults";
-import { capabilitiesFromRef, getModelRef } from "../lib/models-dev";
+import {
+  parseMaxOutputInput,
+  readMaxOutputSize,
+  setMaxOutputSize,
+  withMaxOutputSize,
+} from "../lib/model-max-output";
+import { capabilitiesFromRef, getModelRef, modelsDevReady } from "../lib/models-dev";
 import { getIconMetadata } from "../icons/extracted/metadata";
 import { AgentSettingsPanel } from "./AgentSettingsPanel";
 import { KimiOAuthDialog } from "./KimiOAuthDialog";
@@ -567,6 +573,33 @@ function ModelMapping({
   const [discoverError, setDiscoverError] = useState<string | null>(null);
   const [fetchThinking, setFetchThinking] = useState(true);
 
+  // The models.dev index loads in the background; re-render once it lands so
+  // the max-output "参考" hints appear even when this panel mounted first.
+  const [, forceModelsDevReady] = useReducer((x: number) => x + 1, 0);
+  // One-shot backfill: models with no max_output_size yet get the models.dev
+  // `output` cap filled in automatically. onModelChange only mutates the
+  // in-memory config — the user still reviews and presses 保存配置.
+  const backfillDone = useRef(false);
+  const modelsRef = useRef(models);
+  modelsRef.current = models;
+  useEffect(() => {
+    let alive = true;
+    modelsDevReady().then(() => {
+      if (!alive) return;
+      forceModelsDevReady();
+      if (backfillDone.current || agent !== "kimi_code") return;
+      backfillDone.current = true;
+      for (const m of modelsRef.current) {
+        if (readMaxOutputSize(m.raw_other) !== undefined) continue;
+        const output = getModelRef(m.model)?.output;
+        if (output !== undefined) onModelChange(withMaxOutputSize(m, output));
+      }
+    });
+    return () => {
+      alive = false;
+    };
+  }, []);
+
   const handleDiscover = async () => {
     setDiscovering(true);
     setDiscoverError(null);
@@ -636,6 +669,12 @@ function ModelMapping({
           : fetchThinking
             ? ["thinking"]
             : [],
+        // Seed the max_output_size override only when models.dev knows the
+        // cap — an absent key means the upstream default applies. kimi_code
+        // only: Pi's equivalent is `maxTokens` (not written by this UI).
+        ...(agent === "kimi_code" && ref?.output
+          ? { raw_other: setMaxOutputSize(undefined, ref.output) }
+          : {}),
       });
     }
     onBulkAdd(toAdd);
@@ -648,7 +687,10 @@ function ModelMapping({
       <div className="flex items-center justify-between">
         <div>
           <h3 className="font-medium text-content-primary">{t("modelMapping")}</h3>
-          <p className="text-xs text-content-muted mt-1">{t("modelMappingDesc")}</p>
+          <p className="text-xs text-content-muted mt-1">
+            {t("modelMappingDesc")}
+            {agent === "kimi_code" ? ` ${t("maxOutputSizeDesc")}` : ""}
+          </p>
         </div>
         <div className="flex items-center gap-2">
           <button
@@ -756,7 +798,9 @@ function ModelMapping({
               <th className="text-left px-4 py-3 font-medium">{t("displayName")}</th>
               <th className="text-left px-4 py-3 font-medium">{t("actualModel")}</th>
               <th className="text-left px-4 py-3 font-medium w-28">{t("contextSize")}</th>
-              <th className="text-center px-4 py-3 font-medium w-28">{t("supports1M")}</th>
+              {agent === "kimi_code" && (
+                <th className="text-left px-4 py-3 font-medium w-32">{t("maxOutputSize")}</th>
+              )}
               {agent === "kimi_code" && (
                 <th className="text-left px-4 py-3 font-medium">{t("capabilities")}</th>
               )}
@@ -797,19 +841,14 @@ function ModelMapping({
                     }}
                   />
                 </td>
-                <td className="px-4 py-2 text-center">
-                  <input
-                    type="checkbox"
-                    checked={m.supports_1m || false}
-                    onChange={(e) =>
-                      onModelChange({
-                        ...m,
-                        supports_1m: e.target.checked,
-                      })
-                    }
-                    className="w-4 h-4 rounded border-border bg-input text-blue-600 focus:ring-blue-500"
-                  />
-                </td>
+                {agent === "kimi_code" && (
+                  <td className="px-4 py-2">
+                    <MaxOutputCell
+                      model={m}
+                      onChange={(next) => onModelChange(next)}
+                    />
+                  </td>
+                )}
                 {agent === "kimi_code" && (
                   <td className="px-4 py-2">
                     <CapabilitiesCell
@@ -985,3 +1024,47 @@ function CapabilitiesCell({
     </div>
   );
 }
+
+/**
+ * Max output tokens override (`[models."<alias>"] max_output_size`).
+ *
+ * Not a first-class Model field, so the value lives in `raw_other` (see
+ * lib/model-max-output.ts). Blank input removes the key — "unset" means the
+ * upstream default applies. When models.dev knows the model's cap it is shown
+ * as a click-to-fill reference underneath.
+ */
+function MaxOutputCell({
+  model,
+  onChange,
+}: {
+  model: Model;
+  onChange: (next: Model) => void;
+}) {
+  const { t } = useTranslation();
+  const current = readMaxOutputSize(model.raw_other);
+  const refOutput = getModelRef(model.model)?.output;
+
+  return (
+    <div>
+      <input
+        type="number"
+        min={0}
+        step={1024}
+        className="w-full bg-transparent border border-border rounded px-2 py-1.5 text-sm focus:ring-2 focus:ring-blue-500 focus:outline-none"
+        value={current ?? ""}
+        onChange={(e) =>
+          onChange(withMaxOutputSize(model, parseMaxOutputInput(e.target.value)))
+        }
+      />
+      {refOutput !== undefined && refOutput !== current && (
+        <button
+          type="button"
+          onClick={() => onChange(withMaxOutputSize(model, refOutput))}
+          className="mt-1 text-[11px] text-content-muted hover:text-blue-500"
+        >
+          {t("maxOutputRef", { value: refOutput })}
+        </button>
+      )}
+    </div>
+  );
+}

src/components/SubagentSettingsPage.tsx

@@ -45,7 +45,7 @@ interface SubagentSettingsPageProps {
 
 // Effort tiers the CLI accepts; the value is forwarded verbatim upstream, and
 // per-model support comes from `[models."<alias>"] support_efforts`.
-const EFFORTS = ["low", "medium", "high", "max", "xhigh"] as const;
+const EFFORTS = ["low", "medium", "high", "xhigh", "max"] as const;
 const EFFORT_LABELS: Record<(typeof EFFORTS)[number], TranslationKey> = {
   low: "thinkingLow",
   medium: "thinkingMedium",

src/i18n/en.ts

@@ -111,7 +111,8 @@ export const enTranslations: Record<TranslationKey, string> = {
   baseUrlHint: "Enter a {type} API-compatible endpoint without a trailing slash.",
   envPairs: "Env pairs",
   addEnv: "+ Add",
-  modelMappingDesc: "Display name only affects the /model menu; 1M declares context capability for Kimi Code.",
+  modelMappingDesc: "Display name only affects the /model menu.",
+  maxOutputSizeDesc: "Max output tokens writes max_output_size — leave blank to use the upstream default.",
   oneClickSetup: "One-click setup",
   fetchModels: "Fetch models",
   fetchingModels: "Fetching...",
@@ -121,7 +122,8 @@ export const enTranslations: Record<TranslationKey, string> = {
   displayName: "Display name",
   actualModel: "Actual model",
   contextSize: "Context size",
-  supports1M: "Supports 1M",
+  maxOutputSize: "Max output tokens",
+  maxOutputRef: "Ref {value}",
   default: "Default",
   operation: "Operation",
   isDefault: "Default",

src/i18n/zh.ts

@@ -108,7 +108,8 @@ export const zhTranslations = {
   baseUrlHint: "填写兼容 {type} API 的服务端点地址，不要以斜杠结尾",
   envPairs: "Env 键值对",
   addEnv: "+ 添加",
-  modelMappingDesc: "显示名称只影响 /model 菜单；1M 只是给 Kimi Code 的上下文能力声明。",
+  modelMappingDesc: "显示名称只影响 /model 菜单。",
+  maxOutputSizeDesc: "「最大输出 Token」写入 max_output_size，留空则用上游默认值。",
   oneClickSetup: "一键设置",
   fetchModels: "获取模型列表",
   fetchingModels: "获取中...",
@@ -118,7 +119,8 @@ export const zhTranslations = {
   displayName: "显示名称",
   actualModel: "实际请求模型",
   contextSize: "上下文长度",
-  supports1M: "声明支持 1M",
+  maxOutputSize: "最大输出 Token",
+  maxOutputRef: "参考 {value}",
   default: "默认",
   operation: "操作",
   isDefault: "已默认",
@@ -158,9 +160,9 @@ export const zhTranslations = {
   enableThinking: "启用思考",
   thinkingLevel: "思考等级",
   thinkingKeep: "保留思考内容",
-  thinkingLow: "低",
-  thinkingMedium: "中",
-  thinkingHigh: "高",
+  thinkingLow: "Low",
+  thinkingMedium: "Medium",
+  thinkingHigh: "High",
   thinkingMax: "Max",
   thinkingXHigh: "XHigh",
   thinkingEffortUnsupported: "当前默认模型不支持思考。",

src/lib/model-max-output.test.ts

@@ -0,0 +1,83 @@
+import { describe, expect, it } from "vitest";
+import {
+  parseMaxOutputInput,
+  readMaxOutputSize,
+  setMaxOutputSize,
+  withMaxOutputSize,
+} from "./model-max-output";
+
+const model = (raw_other: unknown) =>
+  ({
+    alias: "a",
+    provider: "p",
+    model: "m",
+    max_context_size: 0,
+    display_name: null,
+    raw_other,
+  }) as const;
+
+describe("readMaxOutputSize", () => {
+  it("reads the key from a model entry's preserved config fields", () => {
+    expect(readMaxOutputSize({ max_output_size: 32768 })).toBe(32768);
+  });
+
+  it("treats absent / malformed / non-positive values as unset", () => {
+    expect(readMaxOutputSize(undefined)).toBeUndefined();
+    expect(readMaxOutputSize(null)).toBeUndefined();
+    expect(readMaxOutputSize({})).toBeUndefined();
+    expect(readMaxOutputSize("32768")).toBeUndefined();
+    expect(readMaxOutputSize({ max_output_size: 0 })).toBeUndefined();
+    expect(readMaxOutputSize({ max_output_size: -1 })).toBeUndefined();
+    expect(readMaxOutputSize({ max_output_size: NaN })).toBeUndefined();
+    expect(readMaxOutputSize([1, 2])).toBeUndefined();
+  });
+});
+
+describe("parseMaxOutputInput", () => {
+  it("keeps positive integers", () => {
+    expect(parseMaxOutputInput("32768")).toBe(32768);
+  });
+
+  it("maps blank / zero / garbage to unset", () => {
+    expect(parseMaxOutputInput("")).toBeUndefined();
+    expect(parseMaxOutputInput("0")).toBeUndefined();
+    expect(parseMaxOutputInput("-5")).toBeUndefined();
+    expect(parseMaxOutputInput("abc")).toBeUndefined();
+  });
+});
+
+describe("setMaxOutputSize", () => {
+  it("sets the key while preserving other preserved fields", () => {
+    expect(setMaxOutputSize({ support_efforts: ["low"] }, 16384)).toEqual({
+      support_efforts: ["low"],
+      max_output_size: 16384,
+    });
+  });
+
+  it("drops the key when the value is unset", () => {
+    expect(setMaxOutputSize({ max_output_size: 16384, force: true }, undefined)).toEqual({
+      force: true,
+    });
+  });
+
+  it("starts from an empty object for null / non-object raw_other", () => {
+    expect(setMaxOutputSize(null, 1024)).toEqual({ max_output_size: 1024 });
+    expect(setMaxOutputSize("junk", 1024)).toEqual({ max_output_size: 1024 });
+    expect(setMaxOutputSize({ max_output_size: 1 }, undefined)).toEqual({});
+  });
+});
+
+describe("withMaxOutputSize", () => {
+  it("returns a new model with the override applied", () => {
+    const before = model({ max_output_size: 4096 });
+    const after = withMaxOutputSize(before, 8192);
+    expect(readMaxOutputSize(after.raw_other)).toBe(8192);
+    expect(readMaxOutputSize(before.raw_other)).toBe(4096);
+    expect(after).not.toBe(before);
+  });
+
+  it("clears the override without touching other fields", () => {
+    const after = withMaxOutputSize(model({ max_output_size: 4096 }), undefined);
+    expect(after.raw_other).toEqual({});
+  });
+});

src/lib/model-max-output.ts

@@ -0,0 +1,59 @@
+import type { Model } from "../types";
+
+/**
+ * Helpers for the `max_output_size` model field.
+ *
+ * kimi-code keeps it in `[models."<alias>"] max_output_size`; it is not a
+ * first-class field of Kimi Switch's Rust `Model` struct, so it round-trips
+ * through `raw_other` (the pass-through bucket for unknown keys). Absent key =
+ * unset = the upstream default applies, so clearing the field removes the key
+ * rather than writing 0.
+ */
+
+function asRecord(value: unknown): Record<string, unknown> {
+  if (value && typeof value === "object" && !Array.isArray(value)) {
+    return value as Record<string, unknown>;
+  }
+  return {};
+}
+
+/** Read the override; undefined when absent or malformed (treated as unset). */
+export function readMaxOutputSize(rawOther: unknown): number | undefined {
+  const value = asRecord(rawOther).max_output_size;
+  if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) {
+    return undefined;
+  }
+  return value;
+}
+
+/** Parse the number input: blank / 0 / NaN mean "unset". */
+export function parseMaxOutputInput(text: string): number | undefined {
+  const value = parseInt(text, 10);
+  if (isNaN(value) || value <= 0) return undefined;
+  return value;
+}
+
+/**
+ * Set (positive value) or drop (undefined) `max_output_size` on a raw_other
+ * blob; every other preserved key is kept.
+ */
+export function setMaxOutputSize(
+  rawOther: unknown,
+  value: number | undefined,
+): Record<string, unknown> {
+  const next = { ...asRecord(rawOther) };
+  if (value === undefined) {
+    delete next.max_output_size;
+  } else {
+    next.max_output_size = Math.floor(value);
+  }
+  return next;
+}
+
+/** Immutable update of a model entry's max_output_size override. */
+export function withMaxOutputSize(
+  model: Model,
+  value: number | undefined,
+): Model {
+  return { ...model, raw_other: setMaxOutputSize(model.raw_other, value) };
+}

src/lib/models-dev.ts

@@ -25,6 +25,8 @@ export interface ModelCost {
 export interface ModelRef {
   name?: string;
   context?: number;
+  /** Upper bound on generated tokens (`limit.output` on models.dev). */
+  output?: number;
   reasoning?: boolean;
   tool_call?: boolean;
   structured_output?: boolean;

src/lib/thinking-efforts.ts

@@ -2,7 +2,7 @@
  * Thinking-effort capability resolution.
  *
  * kimi-code's `[thinking] effort` is a free-form string (low/medium/high/
- * max/xhigh in practice) that the CLI forwards verbatim to OpenAI-compatible
+ * xhigh/max in practice) that the CLI forwards verbatim to OpenAI-compatible
  * upstreams as `reasoning_effort`. Which tiers a given model accepts is
  * declared per model in config.toml (`[models."<alias>"] support_efforts`);
  * models.dev only carries a boolean `reasoning` flag with no tier list. The
@@ -16,8 +16,8 @@ export const THINKING_EFFORTS = [
   "low",
   "medium",
   "high",
-  "max",
   "xhigh",
+  "max",
 ] as const;
 
 export type ThinkingEffort = (typeof THINKING_EFFORTS)[number];