diff --git a/apps/desktop/src/main/maker-host/__tests__/model-catalog-override-store.test.ts b/apps/desktop/src/main/maker-host/__tests__/model-catalog-override-store.test.ts
index e2cd120ef40..94979d7b994 100644
--- a/apps/desktop/src/main/maker-host/__tests__/model-catalog-override-store.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/model-catalog-override-store.test.ts
@@ -10,6 +10,8 @@ import path from 'node:path';
 
 import { describe, expect, it, vi } from 'vitest';
 
+import { MAX_OVERRIDE_ENTRIES_PER_SECTION } from '../model-plane/localCatalogOverrides.js';
+
 const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'model-catalog-overrides-test-'));
 const owner = { current: 'owner-a' };
 
@@ -21,8 +23,14 @@ vi.mock('../../appSessionState.js', () => ({
   ownerScopedUserDataPath: (name: string) => path.join(tmpDir, owner.current, name),
 }));
 
-const { MAX_MODEL_CATALOG_OVERRIDE_FILE_BYTES, readModelCatalogOverrides } =
-  await import('../model-catalog-override-store.js');
+const {
+  MAX_MODEL_CATALOG_OVERRIDE_FILE_BYTES,
+  ModelCatalogOverrideLossError,
+  readModelCatalogOverrides,
+  readModelCatalogImageInput,
+  setModelCatalogImageInput,
+  readModelCatalogThinkingDivergence,
+} = await import('../model-catalog-override-store.js');
 
 const patchEntry = { base: { name: 'Renamed' } };
 
@@ -33,6 +41,167 @@ function writeOwnerFile(value: unknown): string {
   return file;
 }
 
+describe('model-catalog-override-store / 图片输入能力声明', () => {
+  const target = { providerId: 'opencode-go', modelId: 'mimo-v2.6-flash' };
+  const key = 'opencode-go:mimo-v2.6-flash';
+
+  it('写入 true/false 落到 base patch，读回 isCustomized 区分「跟随目录」与「显式声明」', async () => {
+    owner.current = 'owner-image-write';
+    expect(readModelCatalogImageInput(target)).toEqual({ value: null, isCustomized: false });
+
+    await setModelCatalogImageInput(target, true);
+    expect(readModelCatalogOverrides().patches[key]?.base?.supportsImageInput).toBe(true);
+    expect(readModelCatalogImageInput(target)).toEqual({ value: true, isCustomized: true });
+
+    await setModelCatalogImageInput(target, false);
+    expect(readModelCatalogImageInput(target)).toEqual({ value: false, isCustomized: true });
+  });
+
+  it('条目限定了 agents 时把目标引擎并进名单，让声明真正生效', async () => {
+    // 目录合并按 agents 过滤，读回却不检查 —— 不并进 pi 就会「显示已声明、运行期不生效」。
+    owner.current = 'owner-image-agents';
+    writeOwnerFile({
+      version: 1,
+      patches: { [key]: { agents: ['codex'], base: { name: 'Mimo' } } },
+    });
+
+    await setModelCatalogImageInput({ ...target, agent: 'pi' }, true);
+    expect(readModelCatalogOverrides().patches[key]?.agents).toEqual(['codex', 'pi']);
+  });
+
+  it('已有条目无 agents 限制时不加名单', async () => {
+    owner.current = 'owner-image-no-agents';
+    writeOwnerFile({ version: 1, patches: { [key]: { base: { name: 'Mimo' } } } });
+
+    await setModelCatalogImageInput({ ...target, agent: 'pi' }, true);
+    expect(readModelCatalogOverrides().patches[key]?.agents).toBeUndefined();
+  });
+
+  it('写入后条目数超上限时拒绝保存，不留下会被下次读取丢弃的条目', async () => {
+    owner.current = 'owner-image-limit';
+    const full: Record<string, unknown> = {};
+    for (let i = 0; i < MAX_OVERRIDE_ENTRIES_PER_SECTION; i++) full[`p:model-${i}`] = { base: {} };
+    writeOwnerFile({ version: 1, patches: full });
+
+    await expect(
+      setModelCatalogImageInput({ providerId: 'p', modelId: 'one-more' }, true),
+    ).rejects.toThrow(/上限/);
+    // 文件没被改动：原有 1024 条仍完整
+    expect(Object.keys(readModelCatalogOverrides().patches)).toHaveLength(
+      MAX_OVERRIDE_ENTRIES_PER_SECTION,
+    );
+  });
+
+  it('value=null 删除该键回到跟随目录，条目空掉时整条删除', async () => {
+    owner.current = 'owner-image-reset';
+    await setModelCatalogImageInput(target, true);
+    await setModelCatalogImageInput(target, null);
+    expect(readModelCatalogOverrides().patches[key]).toBeUndefined();
+    expect(readModelCatalogImageInput(target)).toEqual({ value: null, isCustomized: false });
+  });
+
+  it('只动 supportsImageInput：同条目其它字段与其它条目原样保留', async () => {
+    owner.current = 'owner-image-preserve';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        [key]: {
+          agents: ['pi'],
+          base: { name: 'Mimo', contextWindow: 128_000 },
+          perAgent: { pi: { maxOutput: 8_192 } },
+        },
+        'openai:gpt-6': { base: { name: 'Renamed' } },
+      },
+    });
+
+    await setModelCatalogImageInput(target, true);
+    const entry = readModelCatalogOverrides().patches[key]!;
+    expect(entry.base).toEqual({
+      name: 'Mimo',
+      contextWindow: 128_000,
+      supportsImageInput: true,
+    });
+    expect(entry.perAgent).toEqual({ pi: { maxOutput: 8_192 } });
+    expect(entry.agents).toEqual(['pi']);
+    expect(readModelCatalogOverrides().patches['openai:gpt-6']).toEqual({ base: { name: 'Renamed' } });
+
+    // 复位后条目里仍留着用户自己写的字段，不能被一并删掉。
+    await setModelCatalogImageInput(target, null);
+    expect(readModelCatalogOverrides().patches[key]).toEqual({
+      agents: ['pi'],
+      base: { name: 'Mimo', contextWindow: 128_000 },
+      perAgent: { pi: { maxOutput: 8_192 } },
+    });
+  });
+
+  it('providerId 里的冒号经 encodeURIComponent 转义，不与 modelId 的分隔符混淆', async () => {
+    owner.current = 'owner-image-escape';
+    await setModelCatalogImageInput({ providerId: 'a:b', modelId: 'm' }, true);
+    expect(readModelCatalogOverrides().patches['a%3Ab:m']?.base?.supportsImageInput).toBe(true);
+  });
+
+  it('一行多个引擎 id 一起写（桥接两端 id 不同，只写一边运行期读不到）', async () => {
+    owner.current = 'owner-image-row';
+    await setModelCatalogImageInput(
+      [
+        { providerId: 'openai', modelId: 'gpt-5.6-sol' },
+        { providerId: 'openai', modelId: 'chatgpt/gpt-5.6-sol' },
+      ],
+      true,
+    );
+    const patches = readModelCatalogOverrides().patches;
+    expect(patches['openai:gpt-5.6-sol']?.base?.supportsImageInput).toBe(true);
+    expect(patches['openai:chatgpt/gpt-5.6-sol']?.base?.supportsImageInput).toBe(true);
+  });
+
+  it('清掉 perAgent 里残留的 supportsImageInput，并让读取按 agent 算有效值', async () => {
+    // perAgent 会遮蔽 base：不清的话 UI（读 base）显示已声明、运行期（按合并语义）仍拒收。
+    owner.current = 'owner-image-peragent';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        [key]: { agents: ['pi'], perAgent: { pi: { supportsImageInput: false } } },
+      },
+    });
+    // 未写入前：读 pi 的有效值仍是显式声明（不假装「跟随目录」）。
+    expect(readModelCatalogImageInput({ ...target, agent: 'pi' })).toEqual({
+      value: false,
+      isCustomized: true,
+    });
+
+    await setModelCatalogImageInput({ ...target, agent: 'pi' }, true);
+    expect(readModelCatalogOverrides().patches[key]?.perAgent).toBeUndefined();
+    expect(readModelCatalogImageInput({ ...target, agent: 'pi' })).toEqual({
+      value: true,
+      isCustomized: true,
+    });
+  });
+
+  it('手改文件里会被 sanitize 丢弃的 patches 条目：拒绝写入而不是静默删掉', async () => {
+    owner.current = 'owner-image-lossy';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        // 未知字段 → 整条被 sanitize 隔离；写盘会把 patches 整段换成清洗后的快照。
+        'openai:typo-model': { base: { name: 'Typo', unknownField: 1 } },
+      },
+    });
+
+    await expect(setModelCatalogImageInput(target, true)).rejects.toThrow(/无法保留/);
+    // 独立错误类型让 IPC 层能把它映射成可执行的 PRECONDITION_FAILED（而不是笼统的 INTERNAL）。
+    await expect(setModelCatalogImageInput(target, true)).rejects.toBeInstanceOf(
+      ModelCatalogOverrideLossError,
+    );
+    // 文件没被动过：那条手改还在（修好后下一次写入自动恢复）。
+    const raw = JSON.parse(
+      fs.readFileSync(path.join(tmpDir, owner.current, 'model-catalog-overrides.json'), 'utf8'),
+    ) as { patches: Record<string, unknown> };
+    expect(raw.patches['openai:typo-model']).toEqual({
+      base: { name: 'Typo', unknownField: 1 },
+    });
+  });
+});
+
 describe('model-catalog-override-store', () => {
   it('读取显式 v1 文件', () => {
     owner.current = 'owner-valid';
@@ -105,4 +274,49 @@ describe('model-catalog-override-store', () => {
     expect(fs.statSync(oversized).size).toBeGreaterThan(MAX_MODEL_CATALOG_OVERRIDE_FILE_BYTES);
     expect(fs.existsSync(oversized)).toBe(true);
   });
+
+  it('readModelCatalogThinkingDivergence 把「同集合不同默认档」报成分叉', () => {
+    // 手改文件留下的分裂：两侧档位集合相同、默认档不同。diverged 不置位的话，
+    // 抽屉的「重选当前项」会被同值 no-op 挡掉，默认档的分叉永远修不掉。
+    owner.current = 'owner-thinking-diverged-default';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        'prov:model': {
+          agents: ['pi', 'codex'],
+          perAgent: {
+            pi: { efforts: ['low', 'max'], defaultEffort: 'low' },
+            codex: { efforts: ['low', 'max'], defaultEffort: 'max' },
+          },
+        },
+      },
+    });
+    expect(
+      readModelCatalogThinkingDivergence([
+        { providerId: 'prov', modelId: 'model', agent: 'pi' },
+        { providerId: 'prov', modelId: 'model', agent: 'codex' },
+      ]),
+    ).toEqual({ diverged: true });
+
+    // 默认档一致时不算分叉（集合相同 + 默认档相同）。
+    owner.current = 'owner-thinking-consistent';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        'prov:model': {
+          agents: ['pi', 'codex'],
+          perAgent: {
+            pi: { efforts: ['low', 'max'], defaultEffort: 'low' },
+            codex: { efforts: ['low', 'max'], defaultEffort: 'low' },
+          },
+        },
+      },
+    });
+    expect(
+      readModelCatalogThinkingDivergence([
+        { providerId: 'prov', modelId: 'model', agent: 'pi' },
+        { providerId: 'prov', modelId: 'model', agent: 'codex' },
+      ]),
+    ).toEqual({ diverged: false });
+  });
 });
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelCatalogThinkingOverride.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelCatalogThinkingOverride.test.ts
new file mode 100644
index 00000000000..0c8998e2019
--- /dev/null
+++ b/apps/desktop/src/main/maker-host/__tests__/modelCatalogThinkingOverride.test.ts
@@ -0,0 +1,217 @@
+/**
+ * model-catalog-override-store 的思考档位声明面 —— 与图片输入声明同构：
+ * 三态（跟随 / 自定义档位 / 复位）、只动自己那两个字段、perAgent 残留遮蔽要清掉、
+ * 非法组合在写入面就拒。
+ *
+ * 为什么这条面必须存在：目录把「会推理」的模型声明成不支持思考时，Pi 客户端就没有
+ * thinking 通道（`piThinkingLevels.mjs`：`reasoning !== true` → 零档位），模型推理只能随
+ * content 返回、被 Cindy 当普通 assistant 正文渲染。用户此前没有任何入口能声明这件事。
+ */
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import { describe, expect, it, vi } from 'vitest';
+
+import type { Effort } from '@cindy/model-providers';
+
+const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'model-catalog-thinking-test-'));
+const owner = { current: 'owner-a' };
+
+vi.mock('electron', () => ({ app: { getPath: () => '/tmp/never-used-here' } }));
+vi.mock('../logger-adapter.js', () => ({
+  desktopMakerLogger: { child: () => ({ info: () => {}, warn: () => {}, error: () => {} }) },
+}));
+vi.mock('../../appSessionState.js', () => ({
+  ownerScopedUserDataPath: (name: string) => path.join(tmpDir, owner.current, name),
+}));
+
+const {
+  readModelCatalogOverrides,
+  readModelCatalogThinking,
+  readModelCatalogThinkingDivergence,
+  setModelCatalogThinking,
+} = await import('../model-catalog-override-store.js');
+
+const target = { providerId: 'opencode-go', modelId: 'space-bunny-free' };
+const key = 'opencode-go:space-bunny-free';
+/** 厂商模型卡：5 档、默认 max、强制推理。 */
+const TIERS = ['low', 'medium', 'high', 'xhigh', 'max'] as const;
+
+function writeOwnerFile(value: unknown): void {
+  const file = path.join(tmpDir, owner.current, 'model-catalog-overrides.json');
+  fs.mkdirSync(path.dirname(file), { recursive: true });
+  fs.writeFileSync(file, typeof value === 'string' ? value : JSON.stringify(value), 'utf8');
+}
+
+describe('model-catalog-override-store / 思考档位声明', () => {
+  it('写入落到 base patch，读回 isCustomized 区分「跟随目录」与「显式声明」', async () => {
+    owner.current = 'owner-thinking-write';
+    // 未声明时读回不带 defaultEffort 键（没有声明就没有默认档可言）。
+    expect(readModelCatalogThinking(target)).toEqual({ value: null, isCustomized: false });
+
+    await setModelCatalogThinking(target, [...TIERS], 'max');
+    expect(readModelCatalogOverrides().patches[key]?.base).toEqual({
+      efforts: [...TIERS],
+      defaultEffort: 'max',
+    });
+    // 读回带上默认档：写入侧靠它沿用既有默认，而不是猜「排序第一档」。
+    expect(readModelCatalogThinking(target)).toEqual({
+      value: [...TIERS],
+      defaultEffort: 'max',
+      isCustomized: true,
+    });
+
+    // 换一组档位 = 数组整体替换，不与旧值合并。
+    await setModelCatalogThinking(target, ['low', 'high'], 'high');
+    expect(readModelCatalogOverrides().patches[key]?.base?.efforts).toEqual(['low', 'high']);
+  });
+
+  it('tiers=null 删除该键回到跟随目录；条目空掉时整条删除', async () => {
+    owner.current = 'owner-thinking-reset';
+    await setModelCatalogThinking(target, [...TIERS], 'max');
+    await setModelCatalogThinking(target, null);
+    expect(readModelCatalogOverrides().patches[key]).toBeUndefined();
+    // 未声明时读回不带 defaultEffort 键（没有声明就没有默认档可言）。
+    expect(readModelCatalogThinking(target)).toEqual({ value: null, isCustomized: false });
+  });
+
+  it('只动 efforts/defaultEffort：同条目其它字段与其它条目原样保留', async () => {
+    owner.current = 'owner-thinking-preserve';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        [key]: {
+          agents: ['pi'],
+          base: { name: 'Space Bunny', contextWindow: 1_000_000 },
+          perAgent: { pi: { maxOutput: 8_192 } },
+        },
+        'openai:gpt-6': { base: { name: 'Renamed' } },
+      },
+    });
+
+    await setModelCatalogThinking(target, [...TIERS], 'max');
+    const entry = readModelCatalogOverrides().patches[key]!;
+    expect(entry.base).toEqual({
+      name: 'Space Bunny',
+      contextWindow: 1_000_000,
+      efforts: [...TIERS],
+      defaultEffort: 'max',
+    });
+    expect(entry.perAgent).toEqual({ pi: { maxOutput: 8_192 } });
+    expect(entry.agents).toEqual(['pi']);
+    expect(readModelCatalogOverrides().patches['openai:gpt-6']).toEqual({ base: { name: 'Renamed' } });
+
+    // 复位后条目里仍留着用户自己写的字段，不能被一并删掉。
+    await setModelCatalogThinking(target, null);
+    expect(readModelCatalogOverrides().patches[key]).toEqual({
+      agents: ['pi'],
+      base: { name: 'Space Bunny', contextWindow: 1_000_000 },
+      perAgent: { pi: { maxOutput: 8_192 } },
+    });
+  });
+
+  it('perAgent 里残留的 efforts 会遮蔽 base（UI 显示与运行期不一致），写入时一并清掉', async () => {
+    owner.current = 'owner-thinking-peragent';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        [key]: {
+          perAgent: { pi: { efforts: ['low'], defaultEffort: 'low' } },
+        },
+      },
+    });
+
+    await setModelCatalogThinking(target, [...TIERS], 'max');
+    const entry = readModelCatalogOverrides().patches[key]!;
+    // base 有值、perAgent 被清空 → 读回不再被遮蔽。
+    expect(entry.base?.efforts).toEqual([...TIERS]);
+    expect(entry.perAgent).toBeUndefined();
+    expect(readModelCatalogThinking({ ...target, agent: 'pi' }).value).toEqual([...TIERS]);
+  });
+
+  it('按 agent 读回时合成 perAgent 引擎例外（base + perAgent[agent]）', async () => {
+    owner.current = 'owner-thinking-agent-merge';
+    writeOwnerFile({
+      version: 1,
+      patches: {
+        [key]: {
+          base: { efforts: [...TIERS], defaultEffort: 'max' },
+          perAgent: { codex: { efforts: ['low', 'high'], defaultEffort: 'high' } },
+        },
+      },
+    });
+
+    expect(readModelCatalogThinking({ ...target, agent: 'pi' }).value).toEqual([...TIERS]);
+    expect(readModelCatalogThinking({ ...target, agent: 'codex' }).value).toEqual(['low', 'high']);
+  });
+
+  it('多引擎 id 分叉可被检出，交给 UI 允许「重选当前项」写回一致', async () => {
+    owner.current = 'owner-thinking-diverge';
+    const piTarget = { providerId: 'openai', agent: 'pi' as const, modelId: 'chatgpt/gpt-5.6-sol' };
+    const codexTarget = { providerId: 'openai', agent: 'codex' as const, modelId: 'gpt-5.6-sol' };
+
+    await setModelCatalogThinking(piTarget, [...TIERS], 'max');
+    expect(readModelCatalogThinkingDivergence([piTarget, codexTarget]).diverged).toBe(true);
+
+    await setModelCatalogThinking(codexTarget, [...TIERS], 'max');
+    expect(readModelCatalogThinkingDivergence([piTarget, codexTarget]).diverged).toBe(false);
+  });
+
+  it('空数组 = 显式「不支持思考」必须能存下来（三态的第三态）', async () => {
+    owner.current = 'owner-thinking-explicit-off';
+    // sanitize 明确支持 efforts: [] 配 defaultEffort: null；写入面不能把它当非法输入。
+    await setModelCatalogThinking(target, [], null);
+    // 空集合必须显式带 defaultEffort: null：sanitize 的自洽校验要求「空集合 ⇒ 无默认档」。
+    expect(readModelCatalogOverrides().patches[key]?.base).toEqual({ efforts: [], defaultEffort: null });
+    expect(readModelCatalogThinking(target)).toMatchObject({ value: [], isCustomized: true });
+
+    // 复位回「跟随目录」后条目整条删除。
+    await setModelCatalogThinking(target, null);
+    expect(readModelCatalogOverrides().patches[key]).toBeUndefined();
+  });
+
+  it('读回带上默认档，写入侧才能不猜（否则会把厂商默认静默改成排序第一档）', async () => {
+    owner.current = 'owner-thinking-default-echo';
+    await setModelCatalogThinking(target, [...TIERS], 'max');
+    // 厂商默认是 max，读回必须原样带回，UI 才能沿用而不是猜 low。
+    expect(readModelCatalogThinking(target)).toMatchObject({
+      value: [...TIERS],
+      defaultEffort: 'max',
+    });
+  });
+
+  it('含 ultra 的声明不会被后续写入静默删掉', async () => {
+    owner.current = 'owner-thinking-ultra';
+    const withUltra: Effort[] = [...TIERS, 'ultra'];
+    await setModelCatalogThinking(target, withUltra, 'max');
+    // 换成不含 ultra 的集合是用户的显式选择，应当照做。
+    await setModelCatalogThinking(target, ['low', 'high'], 'low');
+    expect(readModelCatalogThinking(target).value).toEqual(['low', 'high']);
+    // 但仅声明默认档、集合不变时，ultra 必须原样留着。
+    await setModelCatalogThinking(target, withUltra, 'max');
+    expect(readModelCatalogThinking(target).value).toContain('ultra');
+  });
+
+  it('分叉判定按内容比较：两侧声明完全一致不算分叉', async () => {
+    owner.current = 'owner-thinking-div-content';
+    const piTarget = { providerId: 'openai', agent: 'pi' as const, modelId: 'chatgpt/gpt-5.6-sol' };
+    const codexTarget = { providerId: 'openai', agent: 'codex' as const, modelId: 'gpt-5.6-sol' };
+    // 两个不同 key 各写一次、内容相同：每次读盘都是新数组，引用比较会永远误判分叉。
+    await setModelCatalogThinking(piTarget, [...TIERS], 'max');
+    await setModelCatalogThinking(codexTarget, [...TIERS], 'max');
+    expect(readModelCatalogThinkingDivergence([piTarget, codexTarget]).diverged).toBe(false);
+  });
+
+  it('写入面就拒掉不自洽的档位组合，不让 sanitize 事后把整条判无效', async () => {
+    owner.current = 'owner-thinking-invalid';
+    // 默认档不在档位集合里。
+    await expect(
+      setModelCatalogThinking(target, ['low', 'high'], 'max'),
+    ).rejects.toThrow(/not in the declared tiers/);
+    // 空档位集合却带默认档。
+    await expect(setModelCatalogThinking(target, [], 'max')).rejects.toThrow(/cannot be empty/);
+    // 失败的写入不得留下任何痕迹。
+    expect(readModelCatalogOverrides().patches[key]).toBeUndefined();
+  });
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelMetadataLayers.integration.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelMetadataLayers.integration.test.ts
index 6f1962f3756..5e961b57546 100644
--- a/apps/desktop/src/main/maker-host/__tests__/modelMetadataLayers.integration.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/modelMetadataLayers.integration.test.ts
@@ -1,5 +1,9 @@
 import { afterEach, describe, expect, it } from 'vitest';
 
+import { deriveAvailableModels } from '../catalog-to-descriptors.js';
+import { resolveSendToSessionExecutionConfig } from '../../maker-ipc/sendToSessionExecutionConfig.js';
+import type { OrcaWorkerProviderRoutingContext } from '../../maker-ipc/orcaWorkerCreationService.js';
+
 it('isolates excess additions and duplicate tags without discarding valid local overrides', () => {
   const model = (n: number) => ({
     id: 'local-' + n,
@@ -451,6 +455,76 @@ describe('metadata layers through the active catalog', () => {
     });
     expect(getActiveCatalog().providers.find((p) => p.id === 'xd')?.models.codex).toHaveLength(1);
   });
+  it('clears the main-side valid: none gate once the user declares the effort tiers', () => {
+    // issue #5762 的实报路径：自定义供应商手填模型 ID、不配任何推理选项。此时 efforts 只是
+    // 占位空表，任何显式档位都被 main 准入按 `effort "…" not supported … valid: none` 拒掉。
+    // 用户在本 PR 的「思考」档位多选里声明真实档位后，同一份目录投影出去的 efforts 必须
+    // 非空 —— store 侧（modelCatalogThinkingOverride）与准入侧（sendToSessionExecutionConfig）
+    // 各有一半测试，中间的 active-catalog → descriptor → effortMetaByModel 投影此前无人看守，
+ // 任一侧回归都会让用户在真机上重新撞 valid: none 而单测全绿。
+    const runtime = { baseUrl: 'https://relay.example/v1', models: [{ id: 'hand-filled', name: 'Hand filled' }] };
+    setCustomProviders([
+      buildUserProvider({ id: 'relay', name: 'Relay', runtimes: { pi: runtime } }),
+    ]);
+    const available = () => deriveAvailableModels(getActiveCatalog(), 'pi');
+    // 与 orcaProviderRoutingContext 同一构造：准入真正读的是 effortMetaByModel，
+    // 不是 availableModels 上的 efforts（见 sendToSessionExecutionConfig 的 routeEffortMeta）。
+    const routing = (): OrcaWorkerProviderRoutingContext => ({
+      availability: {
+        'claude-code': [],
+        codex: [],
+        pi: [{
+          id: 'relay',
+          name: 'Relay',
+          models: available().map((m) => m.id),
+          fastModels: [],
+          effortMetaByModel: Object.fromEntries(
+            available().map((m) => [
+              m.id,
+              {
+                efforts: m.efforts,
+                defaultEffort: m.defaultEffort,
+                ...(m.effortsUnknown === true ? { effortsUnknown: true } : {}),
+              },
+            ]),
+          ),
+        }],
+      },
+      resolveDefaultProviderIdForModel: () => 'relay',
+    });
+    const submit = () =>
+      resolveSendToSessionExecutionConfig({
+        source: {
+          agentKind: 'pi',
+          model: 'hand-filled',
+          effort: 'high',
+          fastMode: false,
+          providerId: 'relay',
+        },
+        overrides: { effort: 'high' },
+        availableModels: available(),
+        providerRouting: routing(),
+        hasCindyAiApiKey: true,
+      });
+
+    // 基线：占位空表 → 显式 high 被拒（本 issue 的报错原文）。
+    const before = submit();
+    expect(before).toMatchObject({ ok: false, errorCode: 'INVALID_ARGS' });
+    expect(before.ok === false && before.message).toContain(
+      'effort "high" not supported by model "hand-filled". valid: none',
+    );
+
+    // 声明真实档位后，同一份投影出的 efforts 非空，显式档位原样放行。
+    setLocalCatalogOverrides(
+      sanitizeModelCatalogOverrides({
+        version: 1,
+        patches: { 'relay:hand-filled': { base: { efforts: ['high'] } } },
+      }).overrides,
+    );
+    expect(available().find((m) => m.id === 'hand-filled')?.efforts).toEqual(['high']);
+    expect(submit()).toMatchObject({ ok: true, config: { effort: 'high', model: 'hand-filled' } });
+  });
+
   it('keeps explicit custom Pi values above force and inherits public metadata for an unknown supplier', () => {
     const r = registry();
     setActiveCatalog({ ...BUNDLED_CATALOG, modelRegistry: r });
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
index 07d76065c79..776d5031dc2 100644
--- a/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/modelPlane.test.ts
@@ -29,6 +29,7 @@ import {
   setXdGatewayModels,
 } from '../active-catalog.js';
 import {
+  applyExistingModelLocalPatch,
   EMPTY_MODEL_CATALOG_OVERRIDES,
   sanitizeModelCatalogOverrides,
   type ModelCatalogOverrides,
@@ -798,6 +799,57 @@ describe('retired tombstone 与 discovery 回补', () => {
   });
 });
 
+describe('图片输入能力的本地声明(未声明 → 显式声明)', () => {
+  // 自定义连接（例如 opencode-go）的模型常常没有声明 supportsImageInput，
+  // 运行期 pi/index.ts 的 assertImageInputSupported 只认 `=== true`，于是带图消息在
+  // 客户端就被拒收。这里证明「写本机 patch」确实能把它变成显式声明 ——
+  // 用户无需改动连接配置（那样会让 preset 连接脱离官方目录）。
+  const customModel = (id: string): CatalogModel =>
+    ({ id, name: id, contextWindow: 200_000, efforts: [], defaultEffort: null }) as CatalogModel;
+
+  const patchFor = (modelId: string, value: boolean): ModelCatalogOverrides =>
+    sanitizeModelCatalogOverrides({
+      version: 1,
+      additions: {},
+      patches: {
+        [`opencode-go:${modelId}`]: { base: { supportsImageInput: value } },
+      },
+    }).overrides;
+
+  it('把未声明的模型改成显式支持，false 也如实落成明确关闭', () => {
+    const overrides = patchFor('mimo-v2.6-flash', true);
+    const patched = applyExistingModelLocalPatch(
+      'opencode-go',
+      'pi',
+      customModel('mimo-v2.6-flash'),
+      overrides,
+    );
+    expect(patched.supportsImageInput).toBe(true);
+    // 运行期门读的就是这个字段（`=== true` 才放行图片）。
+    expect(patched.supportsImageInput === true).toBe(true);
+
+    const denied = applyExistingModelLocalPatch(
+      'opencode-go',
+      'pi',
+      customModel('mimo-v2.6-flash'),
+      patchFor('mimo-v2.6-flash', false),
+    );
+    expect(denied.supportsImageInput).toBe(false);
+  });
+
+  it('不影响同 provider 的其它模型，也不影响别的 provider', () => {
+    const overrides = patchFor('mimo-v2.6-flash', true);
+    expect(
+      applyExistingModelLocalPatch('opencode-go', 'pi', customModel('other'), overrides)
+        .supportsImageInput,
+    ).toBeUndefined();
+    expect(
+      applyExistingModelLocalPatch('openai', 'pi', customModel('mimo-v2.6-flash'), overrides)
+        .supportsImageInput,
+    ).toBeUndefined();
+  });
+});
+
 describe('本地 override(local 永远最高)', () => {
   it('账号顺序压过 Registry sortOrder,本地 sortOrder patch 仍最高;xAI 双 root 保留声明顺序', () => {
     setActiveCatalog(
diff --git a/apps/desktop/src/main/maker-host/model-catalog-override-store.ts b/apps/desktop/src/main/maker-host/model-catalog-override-store.ts
index 42863b7a465..cc2758a9385 100644
--- a/apps/desktop/src/main/maker-host/model-catalog-override-store.ts
+++ b/apps/desktop/src/main/maker-host/model-catalog-override-store.ts
@@ -15,25 +15,42 @@
  * active-catalog 合并,是路由/能力派生的输入,MCP create_worker / scheduler 等
  * 无窗口路径也要一致生效,真源必须 main 可靠可读。
  *
- * 本轮(PR-C)刻意**不加 IPC/写 API**:设置 UI 是后续 PR,现在没有 renderer
- * 消费者,提前开 channel 或留死写入口只会扩大 attack surface。用户仍可直接编辑
- * 文件；坏 JSON/超限文件原样保留，修正后下一次同步自动恢复。
+ * 写入面:只开「模型级图片输入能力」(setModelCatalogImageInput)与
+ * 「模型级思考档位」(setModelCatalogThinking)两组字段——「未声明」与「声明错误」
+ * 都是真实状态,用户必须能自己显式声明(见 model-metadata-precedence.md 的字段
+ * 继承与用户显式覆盖)。仍然**不开通用 patch 写入口**:那会让任意目录字段可从
+ * renderer 改写,超出当前需求;价格、defaultEnabled、routing 依旧不在此列。用户仍
+ * 可直接手改文件;坏 JSON/超限文件原样保留,修正后下一次同步自动恢复。
  */
 
+import fs from 'node:fs';
+
 import { desktopMakerLogger } from './logger-adapter.js';
 import { createOverrideSettingsFile } from './override-settings-file.js';
 import {
   EMPTY_MODEL_CATALOG_OVERRIDES,
+  MAX_OVERRIDE_ENTRIES_PER_SECTION,
   sanitizeModelCatalogOverrides,
+  type ModelCatalogOverrideEntry,
   type ModelCatalogOverrides,
 } from './model-plane/localCatalogOverrides.js';
 import { ownerScopedUserDataPath } from '../appSessionState.js';
+import type { AgentKind, Effort } from '@cindy/model-providers';
+import type { ModelCatalogThinkingView } from '../../shared/modelCatalogThinking.js';
 
 const log = desktopMakerLogger.child('model-catalog-overrides');
 
 /** main 同步读取的硬上限；目录 override 正常只有数 KB。 */
 export const MAX_MODEL_CATALOG_OVERRIDE_FILE_BYTES = 1_048_576;
 
+/**
+ * 写入会丢掉手改条目时抛出（写在锁内、不落盘）。用独立类型而不是普通 Error，
+ * 让 IPC 层能把它映射成可执行的 PRECONDITION_FAILED（带指引），而不是笼统的 INTERNAL。
+ */
+export class ModelCatalogOverrideLossError extends Error {
+  override readonly name = 'ModelCatalogOverrideLossError';
+}
+
 function normalize(raw: unknown): ModelCatalogOverrides {
   const { overrides, invalid } = sanitizeModelCatalogOverrides(raw);
   if (invalid.length > 0) {
@@ -46,13 +63,35 @@ function normalize(raw: unknown): ModelCatalogOverrides {
 }
 
 const store = createOverrideSettingsFile<ModelCatalogOverrides>({
-  filePath: () => ownerScopedUserDataPath('model-catalog-overrides.json'),
+  filePath: overrideFilePath,
   defaults: EMPTY_MODEL_CATALOG_OVERRIDES,
   normalize,
   log,
   label: 'model-catalog-overrides',
   maxBytes: MAX_MODEL_CATALOG_OVERRIDE_FILE_BYTES,
   preserveUnreadableFile: true,
+  // patch 只替换被写入键的快照，其余原样落盘：不加这条时 writePatch 会用 sanitize 后的
+  // 整段 patches 覆盖原文件，手改引入的未知字段/超限条目会被静默删除。
+  mergeOverrides: ({ patch, overrides }) => {
+    const next: Record<string, unknown> = { ...overrides };
+    for (const key of Object.keys(patch)) {
+      const value = patch[key as keyof ModelCatalogOverrides];
+      if (key === 'patches' && isPlainObject(value) && Object.keys(value).length === 0) {
+        delete next.patches;
+      } else {
+        next[key] = value;
+      }
+    }
+    // 空段不落盘（与 reset 语义一致：没有条目时不留 `{}` 噪音）。
+    for (const section of ['patches', 'additions'] as const) {
+      if (isPlainObject(next[section]) && Object.keys(next[section]).length === 0) {
+        delete next[section];
+      }
+    }
+    // 除 version 外什么都不剩 → 整文件收回「无覆盖」。
+    if (Object.keys(next).every((key) => key === 'version')) return {};
+    return next;
+  },
 });
 
 /** 当前 override 快照(注入 active-catalog 合并;mtime 守卫让手改文件下次读取生效)。 */
@@ -60,3 +99,327 @@ export function readModelCatalogOverrides(): ModelCatalogOverrides {
   store.invalidateIfChanged();
   return store.read();
 }
+
+/** 模型级图片输入能力的 override 目标。agent 只影响读取时的有效值(base + perAgent 合并)。 */
+export interface ModelCatalogImageInputTarget {
+  providerId: string;
+  modelId: string;
+  agent?: AgentKind;
+}
+
+/**
+ * 模型级思考档位的 override 目标。形状与图片输入那份**逐字段相同**，语义也相同:
+ * 写入只用 (providerId, modelId) 拼 key；`agent` 只影响读取时的有效值合成
+ * (base + perAgent 合并)与目录成员校验。
+ */
+export interface ModelCatalogThinkingStoreTarget {
+  providerId: string;
+  modelId: string;
+  agent?: AgentKind;
+}
+
+function overrideFilePath(): string {
+  return ownerScopedUserDataPath('model-catalog-overrides.json');
+}
+
+function isPlainObject(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+
+/** patch 条目的键格式，与 sanitize/合并侧保持一致。 */
+function patchKey(providerId: string, modelId: string): string {
+  return `${encodeURIComponent(providerId)}:${modelId}`;
+}
+
+/**
+ * 写盘会把 `patches` 整段换成 sanitize 后的快照 —— 手改文件里被隔离的条目（未知字段、
+ * 超上限的尾部条目）会被静默丢掉。在锁内現读原文件，发现会丢就显式拒绝，而不是把用户的
+ * 手改当垃圾清掉（与 sanitize 的「显式契约」一致）。
+ */
+function assertNoPatchEntriesWouldBeLost(): void {
+  let rawText: string;
+  try {
+    rawText = fs.readFileSync(overrideFilePath(), 'utf-8');
+  } catch {
+    // 文件不存在/不可读：没东西可丢；真不可读时 store 自己会拒绝覆写。
+    return;
+  }
+  let parsed: unknown;
+  try {
+    parsed = JSON.parse(rawText);
+  } catch {
+    return; // 坏 JSON 由 store 的 preserveUnreadableFile 语义处理。
+  }
+  if (!isPlainObject(parsed) || !isPlainObject(parsed.patches)) return;
+  const rawPatchKeys = Object.keys(parsed.patches);
+  if (rawPatchKeys.length === 0) return;
+  const { overrides, invalid } = sanitizeModelCatalogOverrides(parsed);
+  const keptPatchKeys = new Set(Object.keys(overrides.patches));
+  const lostKeys = rawPatchKeys.filter((key) => !keptPatchKeys.has(key));
+  const patchSectionInvalid = invalid.filter(
+    (entry) => entry === 'patches' || entry.startsWith('patches:'),
+  );
+  if (lostKeys.length > 0 || patchSectionInvalid.length > 0) {
+    throw new ModelCatalogOverrideLossError(
+      'model-catalog-overrides.json 里有本版本无法保留的 patches 条目' +
+        `（${[...new Set([...lostKeys, ...patchSectionInvalid])].slice(0, 5).join(', ')}）；` +
+        '请先修正该文件再在设置里改动图片输入能力，以免手改内容被静默丢弃。',
+    );
+  }
+}
+
+/**
+ * 条目被限定了作用范围时（如手改的 `agents:['codex']`），把公共字段写进 base
+ * 对不在名单里的引擎不会生效：目录合并按 agents 过滤，而读回不检查 agents ——
+ * 抽屉显示「已声明」，运行期仍用旧能力（评审 P1）。把本次目标引擎并进名单，
+ * 让声明真正生效，而不是静默丢弃。
+ *
+ * 名单为空（无限制）或目标没带 agent 时不做任何事：base 本来就对该条目全部引擎生效。
+ */
+function widenAgentsForDeclaredTargets(
+  entry: ModelCatalogOverrideEntry,
+  agents: readonly AgentKind[],
+): void {
+  const declared = entry.agents;
+  if (!declared || agents.length === 0) return;
+  const missing = agents.filter((agent) => !declared.includes(agent));
+  if (missing.length === 0) return;
+  entry.agents = [...declared, ...missing];
+}
+
+/** 从条目里拿指定 agent 的有效图片能力声明（perAgent 覆盖 base，与合并侧一致）。 */
+function effectiveImageInput(
+  entry: ModelCatalogOverrideEntry | undefined,
+  agent: AgentKind | undefined,
+): boolean | undefined {
+  const perAgentValue = agent ? entry?.perAgent?.[agent]?.supportsImageInput : undefined;
+  const baseValue = entry?.base?.supportsImageInput;
+  const value = perAgentValue !== undefined ? perAgentValue : baseValue;
+  return typeof value === 'boolean' ? value : undefined;
+}
+
+/**
+ * 声明图片输入能力：`true`/`false` 写 base patch，`null` 删除该覆盖回到「跟随供应商」。
+ * 缺字段继承、false 明确关闭，所以三态必须都能表达。
+ *
+ * 写 base 而不是 perAgent：图片能力是公共字段(model-metadata-precedence.md 的公共
+ * 字段范围)，运行期门按解析后的模型读它；perAgent 留给真正逐引擎不同的场景。
+ * 调用方一次传该行的**全部引擎 id**（桥接两端 id 不同，只写主展示引擎会让运行期读不到）。
+ * perAgent 里残留的 supportsImageInput 会遮蔽 base（导致 UI 显示与运行期不一致），连同
+ * 本次写入一起清掉；条目里其它字段与其它条目原样保留。
+ */
+export async function setModelCatalogImageInput(
+  target: ModelCatalogImageInputTarget | readonly ModelCatalogImageInputTarget[],
+  value: boolean | null,
+): Promise<ModelCatalogOverrides> {
+  const targets = Array.isArray(target) ? target : [target as ModelCatalogImageInputTarget];
+  const keys = [...new Set(targets.map((t) => patchKey(t.providerId, t.modelId)))];
+  // 一个 key 可能对应多个引擎（调用方传该行的全部引擎），按键归组供下面拓宽作用范围。
+  const agentsByKey = new Map<string, AgentKind[]>();
+  for (const t of targets) {
+    if (!t.agent) continue;
+    const key = patchKey(t.providerId, t.modelId);
+    agentsByKey.set(key, [...(agentsByKey.get(key) ?? []), t.agent]);
+  }
+  const agentsForKey = (key: string): AgentKind[] => agentsByKey.get(key) ?? [];
+  return store.updateAtomic((current) => {
+    assertNoPatchEntriesWouldBeLost();
+    const patches: Record<string, ModelCatalogOverrideEntry> = { ...current.value.patches };
+    for (const key of keys) {
+      const entry: ModelCatalogOverrideEntry = { ...patches[key] };
+      const base = { ...entry.base };
+      if (value === null) delete base.supportsImageInput;
+      else base.supportsImageInput = value;
+      if (Object.keys(base).length > 0) entry.base = base;
+      else delete entry.base;
+      widenAgentsForDeclaredTargets(entry, agentsForKey(key));
+      if (entry.perAgent) {
+        const perAgent = { ...entry.perAgent };
+        for (const agent of Object.keys(perAgent) as AgentKind[]) {
+          const fields = { ...perAgent[agent] };
+          if (fields.supportsImageInput === undefined) continue;
+          delete fields.supportsImageInput;
+          if (Object.keys(fields).length > 0) perAgent[agent] = fields;
+          else delete perAgent[agent];
+        }
+        if (Object.keys(perAgent).length > 0) entry.perAgent = perAgent;
+        else delete entry.perAgent;
+      }
+      // 条目空掉就整条删除：不留 "{} 条目" 噪音，也让文件能回到"无覆盖"状态。
+      if (entry.base || entry.perAgent || entry.agents) patches[key] = entry;
+      else delete patches[key];
+    }
+    assertPatchCountWithinLimit(patches);
+    return { patches };
+  });
+}
+
+/**
+ * 写入面挡住条目数超限：sanitize 只保留按键排序后的前 MAX_OVERRIDE_ENTRIES_PER_SECTION 条，
+ * 超出的尾部（含本次新写的）下次读就被丢掉，且丢失检查随后会把文件判为不可写，
+ * 用户必须手改文件才能继续在设置里保存（评审 P2）。这里提前拒绝本次保存。
+ */
+function assertPatchCountWithinLimit(patches: Record<string, ModelCatalogOverrideEntry>): void {
+  const count = Object.keys(patches).length;
+  if (count <= MAX_OVERRIDE_ENTRIES_PER_SECTION) return;
+  throw new ModelCatalogOverrideLossError(
+    `model-catalog-overrides.json 的 patches 条目已达上限 ${MAX_OVERRIDE_ENTRIES_PER_SECTION}，` +
+      `本次保存后将有 ${count} 条；超出上限的条目下次启动就会被丢弃。` +
+      '请先在文件里删掉不再需要的条目再保存。',
+  );
+}
+
+/**
+ * 读该模型的图片输入 override。`isCustomized` 让 UI 能区分「跟随供应商」与
+ * 「显式声明了一个刚好等于目录的值」—— 只报 value 会让「恢复跟随供应商」无从表达。
+ * 给了 agent 就按合并语义算有效值(base + perAgent[agent])，否则只看 base。
+ */
+export function readModelCatalogImageInput(target: ModelCatalogImageInputTarget): {
+  value: boolean | null;
+  isCustomized: boolean;
+} {
+  const overrides = readModelCatalogOverrides();
+  const entry = overrides.patches[patchKey(target.providerId, target.modelId)];
+  const value = effectiveImageInput(entry, target.agent);
+  return value === undefined ? { value: null, isCustomized: false } : { value, isCustomized: true };
+}
+
+
+/** 从条目里拿指定 agent 的有效思考档位声明（perAgent 覆盖 base，与合并侧一致）。 */
+function effectiveThinking(
+  entry: ModelCatalogOverrideEntry | undefined,
+  agent: AgentKind | undefined,
+): { efforts: Effort[]; defaultEffort: Effort | null } | undefined {
+  const perAgent = agent ? entry?.perAgent?.[agent] : undefined;
+  const fields = { ...entry?.base, ...perAgent };
+  if (fields.efforts === undefined) return undefined;
+  return { efforts: fields.efforts, defaultEffort: fields.defaultEffort ?? null };
+}
+
+/**
+ * 声明思考档位：`efforts` 写 base patch，`null` 删除该覆盖回到「跟随供应商」。
+ * 缺字段继承、数组整体替换，所以「跟随 / 自定义档位 / 明确不支持」三态都能表达。
+ *
+ * 为什么要这条:目录里的档位决定 Pi 能不能开真正的 thinking 通道
+ * (`piThinkingLevels.mjs`:`reasoning !== true` → 零档位)。上游把会推理的模型声明成
+ * 「不支持思考」时，模型推理只能随 content 返回、被 Cindy 当**普通 assistant 正文**
+ * 渲染（实报：协同 worker 切到该模型后正文变成推理叙述）。这条让用户自己把事实补回来。
+ *
+ * `defaultEffort` 缺省时由 sanitize 侧的「继承宿主默认」语义处理，落到目录当前默认档。
+ * 调用方一次传该行的**全部引擎 id**（桥接两端 id 不同，只写主展示引擎会让运行期读不到）。
+ * perAgent 里残留的 efforts/defaultEffort 会遮蔽 base（UI 显示与运行期不一致），连同本次
+ * 写入一起清掉；条目里其它字段与其它条目原样保留。
+ */
+export async function setModelCatalogThinking(
+  target: ModelCatalogThinkingStoreTarget | readonly ModelCatalogThinkingStoreTarget[],
+  value: readonly Effort[] | null,
+  defaultEffort?: Effort | null,
+): Promise<ModelCatalogOverrides> {
+  const targets = Array.isArray(target) ? target : [target as ModelCatalogThinkingStoreTarget];
+  const keys = [...new Set(targets.map((t) => patchKey(t.providerId, t.modelId)))];
+  const agentsByKey = new Map<string, AgentKind[]>();
+  for (const t of targets) {
+    if (!t.agent) continue;
+    const key = patchKey(t.providerId, t.modelId);
+    agentsByKey.set(key, [...(agentsByKey.get(key) ?? []), t.agent]);
+  }
+  const agentsForKey = (key: string): AgentKind[] => agentsByKey.get(key) ?? [];
+  // 档位与默认档的自洽性在写入面就挡住：非法组合会让 sanitize 整条判无效，用户看着
+  // "保存了但没生效"，而真正原因（档位不自洽）不会出现在任何界面上。
+  if (value !== null) {
+    if (value.length === 0 && defaultEffort) {
+      throw new Error('thinking tiers cannot be empty while a default tier is set');
+    }
+    if (value.length > 0 && defaultEffort && !value.includes(defaultEffort)) {
+      throw new Error(`default tier '${defaultEffort}' is not in the declared tiers`);
+    }
+  }
+  return store.updateAtomic((current) => {
+    assertNoPatchEntriesWouldBeLost();
+    const patches: Record<string, ModelCatalogOverrideEntry> = { ...current.value.patches };
+    for (const key of keys) {
+      const entry: ModelCatalogOverrideEntry = { ...patches[key] };
+      const base = { ...entry.base };
+      if (value === null) {
+        delete base.efforts;
+        delete base.defaultEffort;
+      } else {
+        base.efforts = [...value];
+        if (defaultEffort !== undefined) base.defaultEffort = defaultEffort;
+        else delete base.defaultEffort;
+      }
+      if (Object.keys(base).length > 0) entry.base = base;
+      else delete entry.base;
+      widenAgentsForDeclaredTargets(entry, agentsForKey(key));
+      if (entry.perAgent) {
+        const perAgent = { ...entry.perAgent };
+        for (const agent of Object.keys(perAgent) as AgentKind[]) {
+          const fields = { ...perAgent[agent] };
+          if (fields.efforts === undefined && fields.defaultEffort === undefined) continue;
+          delete fields.efforts;
+          delete fields.defaultEffort;
+          if (Object.keys(fields).length > 0) perAgent[agent] = fields;
+          else delete perAgent[agent];
+        }
+        if (Object.keys(perAgent).length > 0) entry.perAgent = perAgent;
+        else delete entry.perAgent;
+      }
+      // 条目空掉就整条删除：不留 "{} 条目" 噪音，也让文件能回到"无覆盖"状态。
+      if (entry.base || entry.perAgent || entry.agents) patches[key] = entry;
+      else delete patches[key];
+    }
+    assertPatchCountWithinLimit(patches);
+    return { patches };
+  });
+}
+
+/**
+ * 读该模型的思考档位 override。`isCustomized` 让 UI 能区分「跟随供应商」与
+ * 「显式声明了一个刚好等于目录的值」—— 只报 value 会让「恢复跟随供应商」无从表达。
+ * 给了 agent 就按合并语义算有效值(base + perAgent[agent])，否则只看 base。
+ */
+export function readModelCatalogThinking(target: ModelCatalogThinkingStoreTarget): ModelCatalogThinkingView {
+  const overrides = readModelCatalogOverrides();
+  const entry = overrides.patches[patchKey(target.providerId, target.modelId)];
+  const value = effectiveThinking(entry, target.agent);
+  return value === undefined
+    ? { value: null, isCustomized: false }
+    : { value: value.efforts, defaultEffort: value.defaultEffort, isCustomized: true };
+}
+
+/**
+ * 该行各引擎 id 的声明是否**不一致**（手工改文件 / 旧版单键写入留下的存量数据）。
+ * 展示值取运行期真正消费档位的引擎(Pi)那一侧，但 UI 必须允许用户「重选当前项」把分裂
+ * 写回一致 —— 否则同值 no-op 会让分叉永远修不掉（与图片输入 override 同口径）。
+ */
+export function readModelCatalogThinkingDivergence(
+  targets: readonly ModelCatalogThinkingStoreTarget[],
+): { diverged: boolean } {
+  const overrides = readModelCatalogOverrides();
+  const views = targets.map((target) => {
+    const declared = effectiveThinking(
+      overrides.patches[patchKey(target.providerId, target.modelId)],
+      target.agent,
+    );
+    // 「一边声明、一边跟随目录」也是分叉：运行期真正消费档位的那一侧可能正好是跟随的那侧，
+    // 用户却以为已经声明。与图片输入 override 的判定同口径。
+    return {
+      // 按内容比而不是按引用：每次读盘都会造新数组，引用比较会把「两侧声明一致」误判成分叉。
+      tiers: declared ? JSON.stringify(declared.efforts) : null,
+      // 默认档也参与分叉判定：两侧集合相同但默认档不同（手改文件留下的分裂）时，
+      // 「重选当前项」同样要能放行，否则默认档的分叉永远修不掉。
+      defaultEffort: declared ? (declared.defaultEffort ?? null) : null,
+      customized: declared !== undefined,
+    };
+  });
+  if (views.length <= 1) return { diverged: false };
+  const head = views[0]!;
+  return {
+    diverged: views.some(
+      (view) =>
+        view.tiers !== head.tiers ||
+        view.defaultEffort !== head.defaultEffort ||
+        view.customized !== head.customized,
+    ),
+  };
+}
diff --git a/apps/desktop/src/main/maker-host/pi-host.ts b/apps/desktop/src/main/maker-host/pi-host.ts
index 2dbc006a4eb..652ba25549a 100644
--- a/apps/desktop/src/main/maker-host/pi-host.ts
+++ b/apps/desktop/src/main/maker-host/pi-host.ts
@@ -2104,6 +2104,45 @@ export function buildPiAgent(opts: BuildPiAgentOpts): PiAgent | null {
     },
     registerPiProxySession,
     resolvePiNativeProviders: (ctx) => resolvePiNativeProviders(ctx),
+    // 旧会话切模时的能力对账读这里：活动目录已含本机 override 合并(applyExistingModelLocalPatch
+    // 在末位，local 永远最高)，所以用户刚声明的「图片输入」能被读到。查不到(provider/model
+    // 不在目录)返回 undefined → 调用方按「未声明」不动会话快照。
+    readModelImageInput: (providerId, modelId) =>
+      getActiveCatalog()
+        .providers.find((provider) => provider.id === providerId)
+        ?.models.pi?.find((model) => model.id === modelId)?.supportsImageInput,
+    // 目录未声明时的**最终结论**（与 resolvePiNativeProviders 的 input 映射同一口径：
+    // supportsImageInput 未声明 → 默认支持，显式 false → 不支持）。能力对账靠它区分
+    // 「用户改回跟随供应商，该恢复支持」与「目录自己标了不支持，应保持纯文本」—— 这两种
+    // 状态在 supportsImageInput 上都是 undefined，只有最终 input 能分辨。
+    // provider/model 不在目录里 → undefined，调用方保持现状不翻转。
+    readModelCatalogImageCapability: (providerId, modelId) => {
+      const model = getActiveCatalog()
+        .providers.find((provider) => provider.id === providerId)
+        ?.models.pi?.find((entry) => entry.id === modelId);
+      if (!model) return undefined;
+      if (model.supportsImageInput !== undefined) return model.supportsImageInput;
+      // 未声明 → 落到默认结论。口径与 resolvePiNativeProviders 的 input 映射一致
+      // (`supportsImageInput === undefined ? template?.input ?? ['text','image']`)：
+      // 目录若用 modalities.input 明确列了能力就以它为准，否则默认支持。
+      const declaredInput = model.modalities?.input;
+      return declaredInput === undefined ? true : declaredInput.includes('image');
+    },
+    // 该模型在活动目录里最终的思考档位集合（目录与本机 override 合并后的结论）。
+    // 活会话的 activeEffortSnapshot 是启动时解析的，Pi 的 set_model 不重读 models.json，
+    // 用户声明档位后必须靠这个钩子在切模点对齐，否则 thinking 通道开不出来。
+    // provider/model 不在目录里 → undefined，调用方保持现状不猜。
+    readModelCatalogThinkingTiers: (providerId, modelId) => {
+      const model = getActiveCatalog()
+        .providers.find((provider) => provider.id === providerId)
+        ?.models.pi?.find((entry) => entry.id === modelId);
+      // 模型不在目录里 → 拿不到结论，调用方保持快照不变。
+      if (!model) return undefined;
+      // `efforts: []` 是**有效结论**：用户在设置里显式声明了「该模型不支持思考」。
+      // 把它折成 undefined 会让对账退回「保持旧快照」，于是刚保存的关闭声明在当前
+      // 会话不生效（仍按启动时的非空档位继续发 set_thinking_level）。
+      return [...model.efforts];
+    },
     resolvePiRuntimeModelDescriptor: opts.resolvePiRuntimeModelDescriptor,
     resolvePiRuntimeModels: opts.resolvePiRuntimeModels,
     resolvePiGatewayModelDescriptor: opts.resolvePiGatewayModelDescriptor,
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
index 5b21fb658ba..b753c5bd306 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/providerHandlers.test.ts
@@ -3579,6 +3579,390 @@ describe('provider:custom:* CRUD handlers', () => {
   });
 });
 
+describe('model catalog thinking handlers', () => {
+  const target = { providerId: 'opencode-go', agent: 'pi' as const, modelId: 'space-bunny-free' };
+
+  function thinkingDeps(over: Partial<ProviderHandlerDeps> = {}): ProviderHandlerDeps {
+    return makeDeps({
+      listProviders: async () => [catalogView('opencode-go', { pi: ['space-bunny-free'] })],
+      readModelCatalogThinking: vi.fn(() => ({ value: null, isCustomized: false })),
+      writeModelCatalogThinking: vi.fn(async () => {}),
+      syncLocalCatalogOverrides: vi.fn(() => {}),
+      ...over,
+    });
+  }
+
+  it('writes the declared tiers, re-syncs the catalog and broadcasts', async () => {
+    const harness = new IpcHarness();
+    const deps = thinkingDeps();
+    registerProviderHandlers(harness, deps);
+
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+      ...target,
+      value: ['low', 'medium', 'high', 'xhigh', 'max'],
+      defaultEffort: 'max',
+    });
+    expect(deps.writeModelCatalogThinking).toHaveBeenCalledWith(
+      [target],
+      ['low', 'medium', 'high', 'xhigh', 'max'],
+      'max',
+    );
+    // 写盘不会自动生效：必须重读 override 注入活动目录并刷新/广播。
+    expect(deps.syncLocalCatalogOverrides).toHaveBeenCalledTimes(1);
+    expect(deps.broadcastChanged).toHaveBeenCalledTimes(1);
+  });
+
+  it('passes null through as the "follow the catalog" reset without a membership check', async () => {
+    const harness = new IpcHarness();
+    const deps = thinkingDeps();
+    registerProviderHandlers(harness, deps);
+
+    // 目录漂移后指向已下架 id 的陈旧 override 必须能清掉，因此复位不做成员校验。
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+      ...target,
+      modelId: 'retired-model',
+      value: null,
+    });
+    expect(deps.writeModelCatalogThinking).toHaveBeenCalledWith(
+      [{ providerId: 'opencode-go', agent: 'pi', modelId: 'retired-model' }],
+      null,
+      null,
+    );
+  });
+
+  it('rejects enterprise-managed providers so the organization stays authoritative', async () => {
+    // 组织托管供应商的能力由管理员下发；override 是本机最高优先级，不加这道门就能让运行期
+    // 能力偏离企业配置。价格 override 已有同款拒绝，这里必须对齐。
+    const harness = new IpcHarness();
+    const deps = thinkingDeps({
+      listProviders: async () => [catalogView('corp-gateway', { pi: ['corp-model'] })],
+      isOrganizationManagedProviderId: (providerId: string) => providerId === 'corp-gateway',
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+        ...target,
+        providerId: 'corp-gateway',
+        modelId: 'corp-model',
+        value: ['high'],
+        defaultEffort: 'high',
+      }),
+    ).rejects.toThrow(/PERMISSION_DENIED.*managed by your organization/);
+    expect(deps.writeModelCatalogThinking).not.toHaveBeenCalled();
+  });
+
+  it('rejects gateway rows and self-inconsistent tier sets at the IPC boundary', async () => {
+    const harness = new IpcHarness();
+    const deps = thinkingDeps();
+    registerProviderHandlers(harness, deps);
+
+    // 网关模型的能力由服务端目录控制。
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+        ...target,
+        providerId: 'xd',
+        value: ['high'],
+        defaultEffort: 'high',
+      }),
+    ).rejects.toThrow(/INVALID_PARAMS.*server-controlled/);
+
+    // 默认档不在声明的档位集合里 → 写下去只会被 sanitize 判整条无效，必须在边界就拒。
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+        ...target,
+        value: ['low', 'high'],
+        defaultEffort: 'max',
+      }),
+    ).rejects.toThrow(/INVALID_PARAMS.*must be one of the declared tiers/);
+
+    // 非法档位名 / 重复档位 / 空集合。
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, { ...target, value: ['turbo'] }),
+    ).rejects.toThrow(/INVALID_PARAMS.*invalid thinking tier/);
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, { ...target, value: ['low', 'low'] }),
+    ).rejects.toThrow(/INVALID_PARAMS.*must not repeat/);
+    expect(deps.writeModelCatalogThinking).not.toHaveBeenCalled();
+  });
+
+  it('accepts an empty tier array as the explicit "no thinking" state', async () => {
+    // 三态的第三态必须能存下来：底层 sanitize 支持 efforts: [] 配 defaultEffort: null，
+    // 把它当非法输入会让用户永远无法声明「该模型不支持思考」。
+    const harness = new IpcHarness();
+    const deps = thinkingDeps();
+    registerProviderHandlers(harness, deps);
+
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+      ...target,
+      value: [],
+      defaultEffort: null,
+    });
+    expect(deps.writeModelCatalogThinking).toHaveBeenCalledWith([target], [], null);
+  });
+
+  it('rejects a default tier that is absent from an explicitly empty declaration', async () => {
+    const harness = new IpcHarness();
+    const deps = thinkingDeps();
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_SET, {
+        ...target,
+        value: [],
+        defaultEffort: 'high',
+      }),
+    ).rejects.toThrow(/INVALID_PARAMS.*must be one of the declared tiers/);
+    expect(deps.writeModelCatalogThinking).not.toHaveBeenCalled();
+  });
+
+  it('reads every engine id of the row and prefers the Pi side for display', async () => {
+    const harness = new IpcHarness();
+    const read = vi.fn((t: { agent?: string }) =>
+      t.agent === 'pi'
+        ? { value: ['low', 'high'], isCustomized: true }
+        : { value: null, isCustomized: false },
+    );
+    const deps = thinkingDeps({ readModelCatalogThinking: read as never });
+    registerProviderHandlers(harness, deps);
+
+    const view = await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_GET, {
+      ...target,
+      relatedTargets: [{ providerId: 'opencode-go', agent: 'codex', modelId: 'gpt-5.6-sol' }],
+    });
+    // 展示值取运行期真正消费 thinking 档位的那一侧（Pi），并把「一边声明、一边跟随」报成分叉。
+    expect(view).toMatchObject({ value: ['low', 'high'], isCustomized: true, diverged: true });
+  });
+
+  it('reports divergence when only the default tier differs across engines', async () => {
+    // 两侧档位集合相同、默认档不同（手改文件留下的分裂）：diverged 不置位会让「重选当前项」
+    // 被同值 no-op 挡掉，默认档的分叉永远修不掉。
+    const harness = new IpcHarness();
+    const read = vi.fn((t: { agent?: string }) =>
+      t.agent === 'pi'
+        ? { value: ['low', 'max'], defaultEffort: 'low', isCustomized: true }
+        : { value: ['low', 'max'], defaultEffort: 'max', isCustomized: true },
+    );
+    const deps = thinkingDeps({ readModelCatalogThinking: read as never });
+    registerProviderHandlers(harness, deps);
+
+    const view = await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_THINKING_GET, {
+      ...target,
+      relatedTargets: [{ providerId: 'opencode-go', agent: 'codex', modelId: 'gpt-5.6-sol' }],
+    });
+    expect(view).toMatchObject({
+      value: ['low', 'max'],
+      defaultEffort: 'low',
+      isCustomized: true,
+      diverged: true,
+    });
+  });
+});
+
+describe('model catalog image input handlers', () => {
+  const target = { providerId: 'opencode-go', agent: 'pi' as const, modelId: 'mimo-v2.6-flash' };
+
+  function imageDeps(over: Partial<ProviderHandlerDeps> = {}): ProviderHandlerDeps {
+    return makeDeps({
+      listProviders: async () => [catalogView('opencode-go', { pi: ['mimo-v2.6-flash'] })],
+      readModelCatalogImageInput: vi.fn(() => ({ value: null, isCustomized: false })),
+      writeModelCatalogImageInput: vi.fn(async () => {}),
+      syncLocalCatalogOverrides: vi.fn(() => {}),
+      ...over,
+    });
+  }
+
+  it('writes the declaration, re-syncs the catalog and broadcasts', async () => {
+    const harness = new IpcHarness();
+    const deps = imageDeps();
+    registerProviderHandlers(harness, deps);
+
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, { ...target, value: true });
+    expect(deps.writeModelCatalogImageInput).toHaveBeenCalledWith([target], true);
+    // 写盘不会自动生效：必须重读 override 注入活动目录并刷新/广播。
+    expect(deps.syncLocalCatalogOverrides).toHaveBeenCalledTimes(1);
+    expect(deps.refreshCatalog).toHaveBeenCalledTimes(1);
+    // 能力变更必须走 PROVIDER_CHANGED，renderer 靠它重拉 provider 视图。
+    expect(deps.broadcastChanged).toHaveBeenCalledTimes(1);
+    expect(deps.broadcastPricingChanged).not.toHaveBeenCalled();
+  });
+
+  it('passes null through as the "follow the catalog" reset', async () => {
+    const harness = new IpcHarness();
+    const deps = imageDeps();
+    registerProviderHandlers(harness, deps);
+
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, { ...target, value: null });
+    expect(deps.writeModelCatalogImageInput).toHaveBeenCalledWith([target], null);
+  });
+
+  it('writes every harness id of the row so the Pi-side key is covered', async () => {
+    // 桥接投影两端 id 不同（openai 行：codex 用 gpt-5.6-sol、pi 用 chatgpt/gpt-5.6-sol）。
+    // override 按 providerId:modelId 精确匹配，只写主展示引擎的 id 时运行期（读 pi 侧 id）
+    // 完全看不到声明，UI 却会显示「已声明」。
+    const harness = new IpcHarness();
+    const deps = imageDeps({
+      listProviders: async () => [
+        catalogView('openai', { codex: ['gpt-5.6-sol'], pi: ['chatgpt/gpt-5.6-sol'] }),
+      ],
+    });
+    registerProviderHandlers(harness, deps);
+
+    const codexTarget = { providerId: 'openai', agent: 'codex' as const, modelId: 'gpt-5.6-sol' };
+    const piTarget = { providerId: 'openai', agent: 'pi' as const, modelId: 'chatgpt/gpt-5.6-sol' };
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, {
+      ...codexTarget,
+      relatedTargets: [piTarget],
+      value: true,
+    });
+    expect(deps.writeModelCatalogImageInput).toHaveBeenCalledWith([codexTarget, piTarget], true);
+  });
+
+  it('clears a stale declaration without requiring catalog membership', async () => {
+    // value=null 是「恢复跟随目录」：目录漂移（模型下架）后 UI 仍要能清掉陈旧 override。
+    const harness = new IpcHarness();
+    const deps = imageDeps({ listProviders: async () => [] });
+    registerProviderHandlers(harness, deps);
+
+    await harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, { ...target, value: null });
+    expect(deps.writeModelCatalogImageInput).toHaveBeenCalledWith([target], null);
+  });
+
+  it('rejects the gateway provider even though the button is disabled in the UI', async () => {
+    // 按钮 disable 挡不住受信 renderer 直接调 preload：网关能力由服务端目录控制。
+    const harness = new IpcHarness();
+    const deps = imageDeps({
+      listProviders: async () => [catalogView('xd', { pi: ['mimo-v2.6-flash'] })],
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, {
+        providerId: 'xd',
+        agent: 'pi',
+        modelId: 'mimo-v2.6-flash',
+        value: true,
+      }),
+    ).rejects.toThrow(/INVALID_PARAMS/);
+    expect(deps.writeModelCatalogImageInput).not.toHaveBeenCalled();
+  });
+
+  it('surfaces the "refuse rather than drop hand-edited entries" guard as an actionable precondition', async () => {
+    // 写入会静默丢掉手改条目时 store 拒绝落盘；不能退化成笼统的「保存失败」——用户需要知道
+    // 该去修哪个文件。PRECONDITION_FAILED 携带 store 给出的指引。
+    const { ModelCatalogOverrideLossError } = await import(
+      '../../maker-host/model-catalog-override-store.js'
+    );
+    const harness = new IpcHarness();
+    const deps = imageDeps({
+      writeModelCatalogImageInput: vi.fn(async () => {
+        throw new ModelCatalogOverrideLossError('请先修正 model-catalog-overrides.json');
+      }),
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, { ...target, value: true }),
+    ).rejects.toThrow(/PRECONDITION_FAILED.*请先修正/);
+    expect(deps.broadcastChanged).not.toHaveBeenCalled();
+  });
+
+  it('rejects enterprise-managed providers so the organization stays authoritative', async () => {
+    // 组织托管供应商的能力由管理员下发；override 是本机最高优先级，不加这道门就能让运行期
+    // 能力偏离企业配置。价格 override 已有同款拒绝，这里必须对齐。
+    const harness = new IpcHarness();
+    const deps = imageDeps({
+      listProviders: async () => [
+        catalogView('corp-gateway', { pi: ['corp-model'] }),
+      ],
+      isOrganizationManagedProviderId: (providerId: string) => providerId === 'corp-gateway',
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, {
+        ...target,
+        providerId: 'corp-gateway',
+        modelId: 'corp-model',
+        value: true,
+      }),
+    ).rejects.toThrow(/PERMISSION_DENIED.*managed by your organization/);
+    expect(deps.writeModelCatalogImageInput).not.toHaveBeenCalled();
+  });
+
+  it('reports a split declaration and shows the runtime-effective (Pi) value', async () => {
+    // 存量数据：手工改文件或旧版单键写入可能让同一行两个引擎键各存一个值。展示值必须取运行期
+    // 真正消费该能力的 Pi 那一侧，否则 UI 显示“支持”而 Pi 实际拒收；同时用 diverged 标出分叉，
+    // 让抽屉允许“重选当前项”把两个键写回一致。
+    const harness = new IpcHarness();
+    const piTarget = { providerId: 'openai', agent: 'pi' as const, modelId: 'chatgpt/gpt-5.6-sol' };
+    const codexTarget = { providerId: 'openai', agent: 'codex' as const, modelId: 'gpt-5.6-sol' };
+    const deps = imageDeps({
+      readModelCatalogImageInput: vi.fn((candidate) =>
+        candidate.modelId === 'chatgpt/gpt-5.6-sol'
+          ? { value: true, isCustomized: true }
+          : { value: false, isCustomized: true },
+      ),
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_GET, {
+        ...codexTarget,
+        relatedTargets: [piTarget],
+      }),
+    ).resolves.toEqual({ value: true, isCustomized: true, diverged: true });
+  });
+
+  it('does not flag divergence when every engine key agrees', async () => {
+    const harness = new IpcHarness();
+    const deps = imageDeps({
+      readModelCatalogImageInput: vi.fn(() => ({ value: true, isCustomized: true })),
+    });
+    registerProviderHandlers(harness, deps);
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_GET, {
+        ...target,
+        relatedTargets: [{ ...target, agent: 'codex' as const }],
+      }),
+    ).resolves.toEqual({ value: true, isCustomized: true });
+  });
+
+  it('reads without a catalog membership check so stale overrides stay visible and clearable', async () => {
+    const harness = new IpcHarness();
+    const deps = imageDeps({ listProviders: async () => [] });
+    registerProviderHandlers(harness, deps);
+
+    await expect(harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_GET, target)).resolves.toEqual({
+      value: null,
+      isCustomized: false,
+    });
+  });
+
+  it('rejects a non-catalog target, a malformed value, and an unwired host', async () => {
+    const harness = new IpcHarness();
+    registerProviderHandlers(harness, imageDeps());
+
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, {
+        ...target,
+        modelId: 'missing',
+        value: true,
+      }),
+    ).rejects.toThrow(/INVALID_PARAMS/);
+    await expect(
+      harness.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET, { ...target, value: 'yes' }),
+    ).rejects.toThrow(/INVALID_PARAMS/);
+
+    const bare = new IpcHarness();
+    registerProviderHandlers(bare, makeDeps());
+    await expect(bare.invoke(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_GET, target)).rejects.toThrow(
+      /INTERNAL/,
+    );
+  });
+});
+
 describe('model price override handlers', () => {
   const target = {
     providerId: 'openrouter',
diff --git a/apps/desktop/src/main/maker-ipc/channels.ts b/apps/desktop/src/main/maker-ipc/channels.ts
index d02d95f8747..70f1f71ea59 100644
--- a/apps/desktop/src/main/maker-ipc/channels.ts
+++ b/apps/desktop/src/main/maker-ipc/channels.ts
@@ -312,6 +312,17 @@ export const MAKER_INVOKE = {
   MODEL_COMPACTION_SET: 'maker:model-compaction:set',
   MODEL_CONTEXT_LIMIT_SET: 'maker:model-context-limit:set',
   MODEL_CONTEXT_LIMIT_RESET: 'maker:model-context-limit:reset',
+  /**
+   * 单模型图片输入能力的本地 override(设置 → 模型 → 高级设置)。
+   * 上游未声明该能力时「未知」是真实状态，用户必须能显式声明支持/不支持；
+   * 写入本机目录 override(不写回连接配置，preset 连接因此仍跟随官方目录)。
+   * value=null 表示删除 override、回到跟随供应商。设置类写操作：仅本机主页面可调，
+   * **不进 device-link allowlist**(远程改被控端全局设置越权)。
+   */
+  MODEL_CATALOG_THINKING_GET: 'maker:model-catalog-thinking:get',
+  MODEL_CATALOG_THINKING_SET: 'maker:model-catalog-thinking:set',
+  MODEL_CATALOG_IMAGE_INPUT_GET: 'maker:model-catalog-image-input:get',
+  MODEL_CATALOG_IMAGE_INPUT_SET: 'maker:model-catalog-image-input:set',
   // 附加只读引用目录 — 走 closure 推送; DB 持久化由 renderer 同步调
   // local-db:sessions:update (跟 SET_MODEL / sessionService.update 双 IPC 协调先例一致)
   SET_EXTRA_DIRS: 'maker:set-extra-dirs',
diff --git a/apps/desktop/src/main/maker-ipc/providerHandlers.ts b/apps/desktop/src/main/maker-ipc/providerHandlers.ts
index 1b582b6425b..93448473cf6 100644
--- a/apps/desktop/src/main/maker-ipc/providerHandlers.ts
+++ b/apps/desktop/src/main/maker-ipc/providerHandlers.ts
@@ -3,6 +3,7 @@ import { setProviderPresentation, retainProviderPresentationAfterAuthChange } fr
 import type { DeploymentDetectionResult } from '../../shared/providerDetection.js';
 import { pruneProviderGroupBindings } from '../provider-group/bindings.js';
 import { writeProviderGroup } from '../provider-group/store.js';
+import { ModelCatalogOverrideLossError } from '../maker-host/model-catalog-override-store.js';
 /**
  * provider:* IPC handlers。
  *
@@ -31,6 +32,7 @@ import {
   type AgentKind,
   type Catalog,
   type CustomProviderConfig,
+  type Effort,
   type ProviderModelDiscoveryFailure,
   type ProviderPreset,
   type ProviderView,
@@ -52,6 +54,7 @@ import type {
   ModelPriceOverrideTarget,
   ModelPriceOverrideView,
 } from '../../shared/modelPriceOverride.js';
+import type { ModelCatalogThinkingView } from '../../shared/modelCatalogThinking.js';
 import type { MoneyCurrency } from '../../shared/regionalMoney.js';
 import {
   MAX_PROVIDER_ORDER_ID_LENGTH,
@@ -483,9 +486,47 @@ export interface ProviderHandlerDeps {
   readModelContextLimit?(target: ModelPriceOverrideTarget): ModelContextLimitView;
   validateModelContextLimit?(targets: readonly ModelPriceOverrideTarget[], limit: number): Promise<void>;
   writeModelContextLimit?(targets: readonly ModelPriceOverrideTarget[], limit: number | null): void | Promise<void>;
+  /**
+   * 单模型图片输入能力的本地目录 override。value=null 删除 override(回到跟随目录)；
+   * isCustomized 让 UI 区分「跟随目录」与「显式声明了一个等于目录的值」。
+   * 可选 —— 未注入(单测最小桩)时对应 handler 报 INTERNAL 而不是静默成功。
+   */
+  /**
+   * 单模型思考档位的本地目录 override。tiers=null 删除 override(回到跟随目录)；
+   * isCustomized 让 UI 区分「跟随目录」与「显式声明了一个等于目录的值」。
+   * 可选 —— 未注入(单测最小桩)时对应 handler 报 INTERNAL 而不是静默成功。
+   */
+  readModelCatalogThinking?(target: ModelPriceOverrideTarget): ModelCatalogThinkingView;
+  /**
+   * 写入该行**全部引擎**的 id(桥接两端 id 不同：运行期消费 thinking 档位的 pi 侧与主展示
+   * 引擎不同，只写一边会让声明对运行期完全无效)。单次调用内原子写所有 key。
+   */
+  writeModelCatalogThinking?(
+    targets: readonly ModelPriceOverrideTarget[],
+    tiers: readonly Effort[] | null,
+    defaultEffort?: Effort | null,
+  ): void | Promise<void>;
+  readModelCatalogImageInput?(target: ModelPriceOverrideTarget): ModelCatalogImageInputView;
+  /**
+   * 写入该行**全部引擎**的 id(桥接两端 id 不同：运行期消费能力的 pi 侧与主展示引擎不同，
+   * 只写一边会让声明对运行期完全无效)。单次调用内原子写所有 key。
+   */
+  writeModelCatalogImageInput?(
+    targets: readonly ModelPriceOverrideTarget[],
+    value: boolean | null,
+  ): void | Promise<void>;
+  /** 写入后把 override 文件重新注入 active-catalog(写盘不会自动生效)。 */
   readModelCompaction?(target: ModelPriceOverrideTarget): { pct: number; isCustomized: boolean };
   writeModelCompaction?(targets: readonly ModelPriceOverrideTarget[], pct: number | null): Promise<void>;
   stageClearProviderModelCompaction?(providerId: string): () => boolean;
+  syncLocalCatalogOverrides?(): void;
+}
+
+/** 图片输入能力的读回视图。 */
+export interface ModelCatalogImageInputView {
+  /** null = 没有本地声明(跟随目录)。 */
+  value: boolean | null;
+  isCustomized: boolean;
 }
 
 /** 上下文上限的读回视图(与写入返回同形，UI 一次拿齐当前值与是否自定义)。 */
@@ -2026,6 +2067,138 @@ export function registerProviderHandlers(
     },
   );
 
+  const requireCatalogImageInputDeps = (): {
+    read: NonNullable<ProviderHandlerDeps['readModelCatalogImageInput']>;
+    write: NonNullable<ProviderHandlerDeps['writeModelCatalogImageInput']>;
+  } => {
+    if (!deps.readModelCatalogImageInput || !deps.writeModelCatalogImageInput) {
+      throwIpcError('INTERNAL', 'model catalog image input override is not wired');
+    }
+    return {
+      read: deps.readModelCatalogImageInput,
+      write: deps.writeModelCatalogImageInput,
+    };
+  };
+  const parseCatalogImageInputTarget = (input: unknown): ModelPriceOverrideTarget =>
+    parsePriceTarget(input);
+  /**
+   * SET/GET 的行目标：主目标 + 同行的其它引擎 id（relatedTargets）。与上下文上限同一套校验：
+   * 必须同 provider、agent 不重复。写/读都覆盖全部 id —— 运行期读的是 pi 侧 id，主展示引擎
+   * 的 id 与它不一定相同。
+   */
+  const parseCatalogImageInputTargets = (input: unknown): ModelPriceOverrideTarget[] => {
+    const target = parseCatalogImageInputTarget(input);
+    const related = (input as { relatedTargets?: unknown }).relatedTargets;
+    if (related === undefined) return [target];
+    if (!Array.isArray(related) || related.length > 2) {
+      throwIpcError('INVALID_PARAMS', 'invalid related image input targets');
+    }
+    const targets = [target, ...related.map(parsePriceTarget)];
+    if (
+      targets.some((candidate) => candidate.providerId !== target.providerId) ||
+      new Set(targets.map((candidate) => candidate.agent)).size !== targets.length
+    ) {
+      throwIpcError(
+        'INVALID_PARAMS',
+        'image input targets must name distinct harnesses of one provider',
+      );
+    }
+    return targets;
+  };
+  /** 多目标读回：任一声明即视为已声明，取第一个非空值（写入总是一起写，正常情况下同步）。 */
+  const readCatalogImageInputTargets = (targets: readonly ModelPriceOverrideTarget[]) => {
+    const { read } = requireCatalogImageInputDeps();
+    const views = targets.map((target) => read(target));
+    // 展示值优先取运行期真正消费该能力的引擎（Pi）那一侧：它才是“会不会真的收图”的答案。
+    // 仅当各引擎声明不一致（手改文件/旧版单键写入的存量数据）时两者才会不同。
+    const piIndex = targets.findIndex((target) => target.agent === 'pi');
+    const ordered = piIndex >= 0 ? [views[piIndex]!, ...views.filter((_, i) => i !== piIndex)] : views;
+    const firstNonNull = ordered.find((view) => view.value !== null)?.value ?? null;
+    const head = views[0]!;
+    return {
+      value: firstNonNull,
+      isCustomized: views.some((view) => view.isCustomized),
+      // 任一引擎与首个引擎的 (value, isCustomized) 不同 = 分叉（含“一边声明、一边跟随目录”）。
+      // UI 据此允许“重选当前项”把各引擎键写回一致。
+      ...(views.some((view) => view.value !== head.value || view.isCustomized !== head.isCustomized)
+        ? { diverged: true }
+        : {}),
+    };
+  };
+  // value 只在 SET 上校验：GET 请求本来就没有这个字段。
+  const parseCatalogImageInputValue = (input: unknown): boolean | null => {
+    const value = (input as { value?: unknown }).value;
+    if (value !== null && typeof value !== 'boolean') {
+      throwIpcError('INVALID_PARAMS', 'image input override value must be boolean or null');
+    }
+    return value;
+  };
+
+  registry.handle(MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_GET, async (event, input: unknown) => {
+    assertTrustedProviderMutationSender(event);
+    const targets = parseCatalogImageInputTargets(input);
+    // 读不做目录成员校验：目录漂移后 UI 仍要能显示并清掉指向已下架 id 的陈旧 override。
+    return readCatalogImageInputTargets(targets);
+  });
+
+  registry.handle(
+    MAKER_INVOKE.MODEL_CATALOG_IMAGE_INPUT_SET,
+    async (event, input: unknown) => {
+      assertTrustedProviderMutationSender(event);
+      const targets = parseCatalogImageInputTargets(input);
+      const target = targets[0]!;
+      const value = parseCatalogImageInputValue(input);
+      const { write } = requireCatalogImageInputDeps();
+      // 网关模型的能力声明由服务端目录控制，不开放本机 override：按钮已 disable，但受信
+      // renderer 能绕过按钮直接调 preload，所以在这里再拒一次（与价格 override 同一道门）。
+      if (target.providerId === 'xd') {
+        throwIpcError('INVALID_PARAMS', 'Cindy AI Gateway model capabilities are server-controlled');
+      }
+      // 组织托管供应商的能力由管理员下发：override 是本机最高优先级，不加这道门就能让运行期
+      // 能力偏离企业配置。与价格 override 同一位置、同一错误码（UI 侧也一并禁用控件）。
+      if (deps.isOrganizationManagedProviderId(target.providerId)) {
+        throwIpcError('PERMISSION_DENIED', 'Enterprise model capabilities are managed by your organization');
+      }
+      const ownerAtIngress = captureProviderOwnerSession();
+      return withProviderConfigMutation(target.providerId, () =>
+        enqueuePriceMutation(async () => {
+          // 目录成员校验必须在队列内：它带着 await，放在队列外会拿旧账号目录校验、再把结果
+          // 落到新账号的文件里（与价格/上下文上限两个姊妹 handler 同一顺序）。
+          // value=null 是「恢复跟随目录」，不校验成员 —— 目录漂移后的陈旧 override 必须能清掉。
+          if (value !== null) {
+            for (const candidate of targets) await requirePriceTargetModel(candidate);
+          }
+          assertProviderMutationOwner(
+            ownerAtIngress,
+            'active account changed before persisting image input override',
+          );
+          try {
+            await write(targets, value);
+          } catch (err) {
+            log.warn('model catalog image input override persist failed', {
+              providerId: target.providerId,
+              agent: target.agent,
+              modelId: target.modelId,
+              error: err instanceof Error ? err.message : String(err),
+            });
+            // 手改文件里有本版本无法保留的条目：把指引带给用户（可执行），而不是「保存失败」。
+            if (err instanceof ModelCatalogOverrideLossError) {
+              throwIpcError('PRECONDITION_FAILED', err.message);
+            }
+            throwIpcError('INTERNAL', 'failed to persist model catalog image input override');
+          }
+          // 写盘不会自动生效：先重读 override 注入活动目录，再刷新目录并广播
+          // PROVIDER_CHANGED（不是 pricing 通道 —— 能力变了要让 renderer 重拉 provider
+          // 视图，抽屉据此刷新；价格通道只推价格）。
+          deps.syncLocalCatalogOverrides?.();
+          await refreshCatalogAfterCommit();
+          deps.broadcastChanged();
+          return readCatalogImageInputTargets(targets);
+        }),
+      );
+    },
+  );
+
   // One mutation implementation for manual forms and confirmed secret-bearing imports.
   async function createProviderFromInput(
     event: unknown,
@@ -2100,6 +2273,155 @@ export function registerProviderHandlers(
       }
     });
   }
+  /**
+   * 多目标读回。展示值**只认运行期真正消费档位的那一侧（Pi）**：Pi 侧「跟随供应商」
+   * 就显示跟随，不能回落到别的引擎的声明上 —— 那会把 Codex 的值说成 Pi 的实际状态。
+   * 没有任何 Pi 侧目标时（该行没有 Pi 投影）才退回首个非空值。
+   *
+   * 档位数组按**内容**比较而非引用：每次读盘都会造新数组，引用比较会把「两侧声明
+   * 完全一致」永远误判成 diverged，进而让「重选当前项」触发无效落盘与整目录刷新。
+   */
+  const readCatalogThinkingTargets = (targets: readonly ModelPriceOverrideTarget[]) => {
+    const { read } = requireCatalogThinkingDeps();
+    const views = targets.map((target) => read(target));
+    const piIndex = targets.findIndex((target) => target.agent === 'pi');
+    const sameTiers = (a: string[] | null, b: string[] | null): boolean =>
+      a === b || (a !== null && b !== null && a.length === b.length && a.every((t, i) => t === b[i]));
+    // 默认档也参与分叉判定（undefined 与 null 同归「未设」）：两侧集合相同但默认档不同时，
+    // 「重选当前项」同样要能放行，否则默认档的分叉永远修不掉。
+    const sameDefault = (a: string | null | undefined, b: string | null | undefined): boolean =>
+      (a ?? null) === (b ?? null);
+    const primary = piIndex >= 0 ? views[piIndex]! : views[0]!;
+    const value = primary.value ?? (piIndex >= 0 ? null : views.find((view) => view.value !== null)?.value ?? null);
+    return {
+      value,
+      // 默认档跟着展示值那一侧走：写入侧靠它避免猜「排序第一档」把厂商默认改掉。
+      ...(value !== null ? { defaultEffort: primary.defaultEffort ?? null } : {}),
+      isCustomized: views.some((view) => view.isCustomized),
+      ...(views.some(
+          (view) =>
+            !sameTiers(view.value, primary.value) ||
+            !sameDefault(view.defaultEffort, primary.defaultEffort) ||
+            view.isCustomized !== primary.isCustomized,
+        )
+        ? { diverged: true }
+        : {}),
+    };
+  };
+  /** 目录认可的思考档位全集；IPC 边界用它挡住非法档位，避免写出必然被 sanitize 丢弃的条目。 */
+  const VALID_THINKING_TIERS = new Set<Effort>([
+    'minimal', 'low', 'medium', 'high', 'xhigh', 'max', 'ultra',
+  ]);
+  const requireCatalogThinkingDeps = (): {
+    read: NonNullable<ProviderHandlerDeps['readModelCatalogThinking']>;
+    write: NonNullable<ProviderHandlerDeps['writeModelCatalogThinking']>;
+  } => {
+    if (!deps.readModelCatalogThinking || !deps.writeModelCatalogThinking) {
+      throwIpcError('INTERNAL', 'model catalog thinking override is not wired');
+    }
+    return {
+      read: deps.readModelCatalogThinking,
+      write: deps.writeModelCatalogThinking,
+    };
+  };
+  /**
+   * 思考档位的合法档位集合与目录同源(Effort)。非法档位在 IPC 边界就拒 —— 写下去只会被
+   * sanitize 判成整条无效，用户界面表现为「保存了但没生效」，真正原因不落在任何界面上。
+   */
+  const parseCatalogThinkingTiers = (input: unknown): Effort[] | null => {
+    const value = (input as { value?: unknown }).value;
+    if (value === null) return null;
+    // 空数组是三态里的「显式不支持思考」，不是非法输入：底层 sanitize 明确支持
+    // `efforts: []` 配 `defaultEffort: null`（localCatalogOverrides 的 additionModelFor
+    // 同款分支）。把它一并拒掉会让抽屉的第三态永远存不下来，用户只看到保存失败。
+    if (!Array.isArray(value)) {
+      throwIpcError('INVALID_PARAMS', 'thinking tiers must be an array or null');
+    }
+    const tiers = value.map((tier) => {
+      if (typeof tier !== 'string' || !VALID_THINKING_TIERS.has(tier as Effort)) {
+        throwIpcError('INVALID_PARAMS', `invalid thinking tier '${String(tier)}'`);
+      }
+      return tier as Effort;
+    });
+    if (new Set(tiers).size !== tiers.length) {
+      throwIpcError('INVALID_PARAMS', 'thinking tiers must not repeat');
+    }
+    return tiers;
+  };
+  const parseCatalogThinkingDefault = (input: unknown, tiers: Effort[] | null): Effort | null => {
+    const raw = (input as { defaultEffort?: unknown }).defaultEffort;
+    if (raw === undefined || raw === null) return null;
+    if (typeof raw !== 'string' || !VALID_THINKING_TIERS.has(raw as Effort)) {
+      throwIpcError('INVALID_PARAMS', `invalid default thinking tier '${String(raw)}'`);
+    }
+    if (tiers === null || !tiers.includes(raw as Effort)) {
+      throwIpcError('INVALID_PARAMS', 'default thinking tier must be one of the declared tiers');
+    }
+    // unreachable：上面的 includes 已保证 tiers 非空时 defaultEffort 在集合内。
+    return raw as Effort;
+  };
+
+  registry.handle(MAKER_INVOKE.MODEL_CATALOG_THINKING_GET, async (event, input: unknown) => {
+    assertTrustedProviderMutationSender(event);
+    const targets = parseCatalogImageInputTargets(input);
+    // 读不做目录成员校验：目录漂移后 UI 仍要能显示并清掉指向已下架 id 的陈旧 override。
+    return readCatalogThinkingTargets(targets);
+  });
+
+  registry.handle(
+    MAKER_INVOKE.MODEL_CATALOG_THINKING_SET,
+    async (event, input: unknown) => {
+      assertTrustedProviderMutationSender(event);
+      const targets = parseCatalogImageInputTargets(input);
+      const target = targets[0]!;
+      const tiers = parseCatalogThinkingTiers(input);
+      const defaultEffort = parseCatalogThinkingDefault(input, tiers);
+      const { write } = requireCatalogThinkingDeps();
+      // 网关模型的能力声明由服务端目录控制，不开放本机 override：按钮已 disable，但受信
+      // renderer 能绕过按钮直接调 preload，所以在这里再拒一次（与图片输入 override 同一道门）。
+      if (target.providerId === 'xd') {
+        throwIpcError('INVALID_PARAMS', 'Cindy AI Gateway model capabilities are server-controlled');
+      }
+      // 组织托管供应商的能力由管理员下发：override 是本机最高优先级，不加这道门就能让运行期
+      // 能力偏离企业配置。与价格 override 同一位置、同一错误码（UI 侧也一并禁用控件）。
+      if (deps.isOrganizationManagedProviderId(target.providerId)) {
+        throwIpcError('PERMISSION_DENIED', 'Enterprise model capabilities are managed by your organization');
+      }
+      const ownerAtIngress = captureProviderOwnerSession();
+      return withProviderConfigMutation(target.providerId, () =>
+        enqueuePriceMutation(async () => {
+          // 目录成员校验必须在队列内：它带着 await，放在队列外会拿旧账号目录校验、再把结果
+          // 落到新账号的文件里（与价格/上下文上限/图片输入这几个姊妹 handler 同一顺序）。
+          // value=null 是「恢复跟随目录」，不校验成员 —— 目录漂移后的陈旧 override 必须能清掉。
+          if (tiers !== null) {
+            for (const candidate of targets) await requirePriceTargetModel(candidate);
+          }
+          assertProviderMutationOwner(
+            ownerAtIngress,
+            'active account changed before persisting thinking override',
+          );
+          try {
+            await write(targets, tiers, defaultEffort);
+          } catch (err) {
+            log.warn('model catalog thinking override persist failed', {
+              providerId: target.providerId,
+              modelId: target.modelId,
+              tiers,
+              error: err instanceof Error ? err.message : String(err),
+            });
+            throwIpcError('INTERNAL', 'failed to persist model catalog thinking override');
+          }
+          // 写盘不会自动生效：先重读 override 注入活动目录，再刷新目录并广播
+          // PROVIDER_CHANGED（不是 pricing 通道 —— 档位变了要让 renderer 重拉 provider
+          // 视图，抽屉据此把「跟随供应商」的括注刷新成新值）。
+          deps.syncLocalCatalogOverrides?.();
+          await refreshCatalogAfterCommit();
+          deps.broadcastChanged();
+        }),
+      );
+    },
+  );
+
   registry.handle(MAKER_INVOKE.PROVIDER_CUSTOM_CREATE, createProviderFromInput);
 
   async function updateProviderFromInput(
diff --git a/apps/desktop/src/main/maker-ipc/register.ts b/apps/desktop/src/main/maker-ipc/register.ts
index 4999bd08089..2eeb239e98d 100644
--- a/apps/desktop/src/main/maker-ipc/register.ts
+++ b/apps/desktop/src/main/maker-ipc/register.ts
@@ -1193,6 +1193,7 @@ import {
   getDesktopSelectableCatalog,
   refreshActiveCatalogFromSource,
   refreshCustomProvidersIntoCatalog,
+  syncLocalCatalogOverridesIntoActiveCatalog,
 } from '../maker-host/createDesktopProviderService.js';
 import {
   deviceAvailableModels,
@@ -1223,6 +1224,12 @@ import {
   writeModelContextLimitsWithRefresh,
 } from '../maker-host/model-context-limit-store.js';
 import { refreshAnthropicModelsFromProbe } from '../maker-host/model-discovery/anthropic.js';
+import {
+  readModelCatalogImageInput,
+  readModelCatalogThinking,
+  setModelCatalogThinking,
+  setModelCatalogImageInput,
+} from '../maker-host/model-catalog-override-store.js';
 import { refreshOpenAiMediaModels } from '../maker-host/model-discovery/openai-media.js';
 import { refreshXaiMediaModels } from '../maker-host/model-discovery/xai-media.js';
 import { testProviderConnection } from '../maker-host/provider-diagnostics.js';
@@ -7681,6 +7688,38 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)
         target.modelId,
       ),
     }),
+    readModelCatalogThinking: (target) =>
+      readModelCatalogThinking({
+        providerId: target.providerId,
+        modelId: target.modelId,
+        agent: target.agent,
+      }),
+    writeModelCatalogThinking: async (targets, tiers, defaultEffort) => {
+      await setModelCatalogThinking(
+        targets.map((target) => ({
+          providerId: target.providerId,
+          modelId: target.modelId,
+        })),
+        tiers,
+        defaultEffort,
+      );
+    },
+    readModelCatalogImageInput: (target) =>
+      readModelCatalogImageInput({
+        providerId: target.providerId,
+        modelId: target.modelId,
+        agent: target.agent,
+      }),
+    writeModelCatalogImageInput: async (targets, value) => {
+      await setModelCatalogImageInput(
+        targets.map((target) => ({
+          providerId: target.providerId,
+          modelId: target.modelId,
+        })),
+        value,
+      );
+    },
+    syncLocalCatalogOverrides: () => syncLocalCatalogOverridesIntoActiveCatalog(),
     readModelCompaction: (target) => {
       const provider = getActiveCatalog().providers.find((entry) => entry.id === target.providerId);
       const identity = provider ? modelCompactionTarget(provider, target.agent, target.modelId) : target;
diff --git a/apps/desktop/src/preload/preload.ts b/apps/desktop/src/preload/preload.ts
index 096ebd7ed14..98aba3105ca 100644
--- a/apps/desktop/src/preload/preload.ts
+++ b/apps/desktop/src/preload/preload.ts
@@ -6858,6 +6858,34 @@ contextBridge.exposeInMainWorld('electronAPI', {
       owner: import('../shared/modelContextLimit').ModelContextLimitOwner,
     ): Promise<import('../shared/modelContextLimit').ModelContextLimitView> =>
       ipcRenderer.invoke('maker:model-context-limit:set', target, limit, owner),
+    /**
+     * 单模型思考档位的本地目录 override(设置 → 模型 → 高级设置)。
+     * tiers=null 表示「跟随供应商」;写入后 main 会重注 active-catalog 并广播。
+     */
+    getModelCatalogThinking: (
+      target: import('../shared/modelCatalogThinking').ModelCatalogThinkingTarget,
+    ): Promise<import('../shared/modelCatalogThinking').ModelCatalogThinkingView> =>
+      ipcRenderer.invoke('maker:model-catalog-thinking:get', target),
+    setModelCatalogThinking: (
+      target: import('../shared/modelCatalogThinking').ModelCatalogThinkingTarget,
+      tiers: string[] | null,
+      defaultTier?: string | null,
+    ): Promise<void> =>
+      ipcRenderer.invoke('maker:model-catalog-thinking:set', {
+        ...target,
+        value: tiers,
+        defaultEffort: defaultTier ?? null,
+      }),
+
+    getModelCatalogImageInput: (
+      target: import('../shared/modelCatalogImageInput').ModelCatalogImageInputTarget,
+    ): Promise<import('../shared/modelCatalogImageInput').ModelCatalogImageInputView> =>
+      ipcRenderer.invoke('maker:model-catalog-image-input:get', target),
+    setModelCatalogImageInput: (
+      target: import('../shared/modelCatalogImageInput').ModelCatalogImageInputTarget,
+      value: boolean | null,
+    ): Promise<import('../shared/modelCatalogImageInput').ModelCatalogImageInputView> =>
+      ipcRenderer.invoke('maker:model-catalog-image-input:set', { ...target, value }),
     resetModelContextLimit: (
       target: import('../shared/modelContextLimit').ModelContextLimitTarget,
       owner: import('../shared/modelContextLimit').ModelContextLimitOwner,
diff --git a/apps/desktop/src/renderer/components/chat/ErrorBanner.tsx b/apps/desktop/src/renderer/components/chat/ErrorBanner.tsx
index 2aff87c8614..8db6b4554cd 100644
--- a/apps/desktop/src/renderer/components/chat/ErrorBanner.tsx
+++ b/apps/desktop/src/renderer/components/chat/ErrorBanner.tsx
@@ -62,7 +62,10 @@ import {
   CLAUDE_GATEWAY_OPUS_PLAN_MISMATCH_REASON,
   CLAUDE_SUBSCRIPTION_OPUS_PLAN_MISMATCH_REASON,
 } from '../../../shared/claudeGatewayError';
-import { isPiImageInputUnsupportedError } from '../../../shared/inputError';
+import {
+  isPiImageCapabilityRefreshFailedError,
+  isPiImageInputUnsupportedError,
+} from '../../../shared/inputError';
 import type { ToolLoopErrorDetails } from '@cindy/maker-core';
 import type { ErrorDiagnosisSource } from '@/lib/errorDiagnosisPrompt';
 
@@ -467,6 +470,9 @@ export function ErrorBanner({
     displayError = t('chat.errorBanner.codexResumeNotReady');
   } else if (isPiImageInputUnsupportedError(error)) {
     displayError = t('ipcError.PI_IMAGE_INPUT_UNSUPPORTED');
+  } else if (isPiImageCapabilityRefreshFailedError(error)) {
+    // 能力声明已写入目录、但子进程还没加载（回合在跑或重载被拒）：可重试，不是模型不支持。
+    displayError = t('ipcError.PI_IMAGE_CAPABILITY_REFRESH_FAILED');
   } else if (isCredentialSwitchBusy) {
     displayError = t('chat.errorBanner.credentialSwitchBusy');
   } else if (isAccountBoundaryPending) {
diff --git a/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx b/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
index 68ac7664a2b..fc0b507448d 100644
--- a/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
+++ b/apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx
@@ -28,7 +28,16 @@ import { useTranslation } from 'react-i18next';
 import { SegmentedControl } from '@/components/ui/segmented-control';
 import { cn } from '@/lib/utils';
 import { providerViewToCustomProviderConfig, updateCustomProvider } from '@/lib/customProviders';
-import { DropdownMenu, DropdownMenuTrigger, DropdownMenuContent, DropdownMenuRadioGroup, DropdownMenuRadioItem } from '@/components/ui/dropdown-menu';
+import {
+  DropdownMenu,
+  DropdownMenuTrigger,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuCheckboxItem,
+  DropdownMenuRadioGroup,
+  DropdownMenuRadioItem,
+} from '@/components/ui/dropdown-menu';
 import { MODEL_HARNESS_COLOR } from '@/lib/modelHarnessPresentation';
 import { ModelCompatibilityNotice } from '@/components/new-chat/ModelCompatibilityNotice';
 import { MODEL_PROTOCOL_LABEL, MODEL_PROTOCOL_OPTIONS } from '@/lib/modelProtocolLabel';
@@ -39,6 +48,8 @@ import { ClaudeMark } from '@/components/icons/ClaudeMark';
 import { CodexMark } from '@/components/icons/CodexMark';
 import { PiMark } from '@/components/icons/PiMark';
 import { useModelContextLimit } from '@/hooks/useModelContextLimit';
+import { useModelCatalogImageInput } from '@/hooks/useModelCatalogImageInput';
+import { useModelCatalogThinking } from '@/hooks/useModelCatalogThinking';
 import { ModelCompactionSettings } from './ModelCompactionSettings';
 import { modelPriceDetailRows, type ModelPricePresentation } from '@/lib/modelPriceFormat';
 import {
@@ -284,6 +295,102 @@ export function ModelAdvancedDrawer({
     [contextAgent, contextModel, provider.id, row, chatAgents],
   );
   const ctx = useModelContextLimit(open ? contextTarget : null);
+  // 图片输入能力的本地声明。与上下文上限同一个目标形状；agent 只用于目录成员校验。
+  // 必须覆盖该行全部引擎的 id：运行期真正消费这个能力的只有 Pi，而桥接投影两端 id 不同
+  // （openai 行：codex 用 gpt-5.6-sol、pi 用 chatgpt/gpt-5.6-sol）—— 只写主展示引擎的 id，
+  // Pi 侧读不到（override 按 providerId:modelId 精确匹配），UI 却会显示「已声明」。
+  const imageInputTarget = useMemo(() => {
+    if (!open || !primaryAgent || !primaryModel) return null;
+    const relatedTargets = (row?.avail ?? [])
+      .filter((agent) => agent !== primaryAgent)
+      .flatMap((agent) => {
+        const model = row?.byAgent[agent];
+        return model ? [{ providerId: provider.id, agent, modelId: model.id }] : [];
+      });
+    return {
+      providerId: provider.id,
+      agent: primaryAgent,
+      modelId: primaryModel.id,
+      ...(relatedTargets.length > 0 ? { relatedTargets } : {}),
+    };
+  }, [open, primaryAgent, primaryModel, provider.id, row]);
+  const imageInput = useModelCatalogImageInput(imageInputTarget);
+  // 思考档位的本地声明。目标形状与图片输入完全一致：必须覆盖该行全部引擎的 id ——
+  // 运行期真正消费 thinking 档位的是 Pi，桥接投影两端 id 不同，只写主展示引擎会让声明
+  // 对运行期无效、UI 却显示「已声明」。
+  const thinkingTarget = useMemo(() => {
+    if (!open || !primaryAgent || !primaryModel) return null;
+    const relatedTargets = (row?.avail ?? [])
+      .filter((agent) => agent !== primaryAgent)
+      .flatMap((agent) => {
+        const model = row?.byAgent[agent];
+        return model ? [{ providerId: provider.id, agent, modelId: model.id }] : [];
+      });
+    return {
+      providerId: provider.id,
+      agent: primaryAgent,
+      modelId: primaryModel.id,
+      ...(relatedTargets.length > 0 ? { relatedTargets } : {}),
+    };
+  }, [open, primaryAgent, primaryModel, provider.id, row]);
+  const thinking = useModelCatalogThinking(thinkingTarget);
+  /**
+   * 档位组合用固定候选（与目录 Effort 全集同序）而不是自由输入：组合爆炸（2^7）无法
+   * 列举，而真实诉求只有「沿用厂商 / 声明常见的几档」两类。勾选语义 = 该模型支持这一档。
+   */
+  // 与目录 Effort 全集同序（含 ultra：它是合法 Effort，IPC 与 Pi 映射都接受）。
+  // 漏掉它会让「已有含 ultra 的声明」在用户勾选其它档位时被按这份列表重建 → 静默删除。
+  const THINKING_TIER_CHOICES = ['minimal', 'low', 'medium', 'high', 'xhigh', 'max', 'ultra'] as const;
+  const setThinkingTiers = async (tiers: string[] | null) => {
+    // 已经是该状态就别再写一次：免掉一次空写入引起的全量目录刷新与广播。各引擎键分叉时
+    // 必须放行 —— 否则「重选当前项」被 no-op 挡掉，分叉永远修不掉。
+    const already =
+      tiers === null
+        ? !thinking.isCustomized
+        : thinking.isCustomized &&
+          tiers.length === (thinking.value?.length ?? -1) &&
+          tiers.every((tier) => thinking.value?.includes(tier));
+    if (already && !thinking.diverged) return;
+    // 默认档**不猜**：沿用读回回来的既有默认档（仍在新集合内才继续沿用），用户没有
+    // 单独改默认档的入口，猜「排序第一档」会把厂商声明的默认（如 max）静默改成 low。
+    const keptDefault =
+      tiers !== null && thinking.defaultEffort && tiers.includes(thinking.defaultEffort)
+        ? thinking.defaultEffort
+        : null;
+    const persisted = await thinking.setTiers(tiers, keptDefault);
+    if (!persisted) {
+      const reason = thinking.errorReason;
+      toast.error(
+        reason
+          ? t('settings.providers.models.advanced.thinkingOverride.saveFailedWithReason', { reason })
+          : t('settings.providers.models.advanced.thinkingOverride.saveFailed'),
+      );
+    }
+  };
+  // 'inherit' = 删除本机 override，回到跟随供应商。失败由 hook 回读真值，这里只负责提示。
+  const setImageInput = async (next: string) => {
+    const value = next === 'inherit' ? null : next === 'true';
+    // 已经是该状态就别再写一次：免掉一次空写入引起的全量目录刷新与广播。但各引擎键分叉时
+    // 必须放行 —— 否则“重选当前项”被 no-op 挡掉，分叉永远修不掉（UI 显示一侧、运行期读另一侧）。
+    const already =
+      value === null ? !imageInput.isCustomized : imageInput.isCustomized && value === imageInput.value;
+    if (already && !imageInput.diverged) {
+      return;
+    }
+    const persisted = await imageInput.setValue(value);
+    if (!persisted) {
+      // main 在这条路径上会给出可执行指引（手改文件里有无法保留的条目时要先修文件）：
+      // 把它带进提示，否则用户反复保存失败却看不到唯一的修复方式。
+      const reason = imageInput.errorReason;
+      toast.error(
+        reason
+          ? t('settings.providers.models.advanced.imageInputOverride.saveFailedWithReason', {
+              message: reason,
+            })
+          : t('settings.providers.models.advanced.imageInputOverride.saveFailed'),
+      );
+    }
+  };
   const canEditProtocol = provider.source === 'user' && !provider.auth?.native && !isManagedSidecarProviderId(provider.id);
   const imageAgents = canEditProtocol
     ? provider.agents.filter(agent => provider.routing?.[agent]?.imageApi === 'openai-images')
@@ -354,6 +461,14 @@ export function ModelAdvancedDrawer({
 
 
   const [ctxDraft, setCtxDraft] = useState('');
+  // 提交后写入是异步的：hook 先置 loading、等 IPC 回声才落新值。若提交瞬间就丢掉草稿，
+  // 说明行会先退回**旧的** effectiveLimit、回声到了再跳新值 —— 用户实测到的「闪一下旧值」。
+  // 这里把「提交在途」也当成还在编辑：显示值继续取草稿，直到写入收口。
+  const [ctxCommitting, setCtxCommitting] = useState(false);
+  // 提交代次：目标（模型/agent/provider）一变或又一次提交开始，旧提交的 finally 就不许再
+  // 清掉新的 committing 状态 —— 否则 A 写未完成时切到 B 并提交，A 完成会把 B 的「提交在途」
+  // 提前收掉，B 在写入期间重新显示旧的 effectiveLimit（本次修改要消除的闪回原地复活）。
+  const ctxCommitGen = useRef(0);
   const ctxDirtyRef = useRef(false);
   const defaultWindow = contextAgent === 'codex' && ctx.codexContext
     ? ctx.codexContext.contextWindow : contextModel?.contextWindow ?? 0;
@@ -365,9 +480,28 @@ export function ModelAdvancedDrawer({
   const minimumContextK = isLocalRuntimeBetaProviderId(provider.id)
     ? 1 : Math.max(1, Math.floor(Math.min(100_000, ...modelWindows) / 1000));
   const routeWindow = supportsMillionContext ? 1_000_000 : primaryModel?.contextWindowMax ?? primaryModel?.contextWindow ?? 0;
+  // 「上游最大上下文」这一行只印上游**确实下发过**的窗口，来源按优先级：
+  //   1) contextWindowMax —— 路由/网关声明的容量；
+  //   2) contextWindowVerified === true 的 contextWindow —— 用户或预设显式配置的窗口；
+  //   3) 都没有 → 未声明。
+  // 第 3 条必须排除「自定义模型缺元数据时的 200K 展示兜底」：那个常量在
+  // user-provider.ts 里明确写着「仅用于展示」、不带 contextWindowVerified —— 把它印成
+  // 上游下发的窗口，会让用户以为容量只有 200K，而运行期窗口取模型级上下文上限
+  // （实测报障：圆环 1.0M、这里 200K）。
+  // 与 routeWindow 的区别：routeWindow 是路由展示用的乐观值（供百万上下文档位等
+  // 比较），declaredWindow 是这一行的**如实声明**口径。
+  const declaredCapacity = primaryModel?.contextWindowMax ?? 0;
+  const declaredWindow =
+    declaredCapacity > 0
+      ? declaredCapacity
+      : primaryModel?.contextWindowVerified === true
+        ? (primaryModel.contextWindow ?? 0)
+        : 0;
   const effectiveLimit = ctx.limit ?? (defaultWindow > 0 ? defaultWindow : null);
   useEffect(() => {
     ctxDirtyRef.current = false;
+    ctxCommitGen.current += 1;
+    setCtxCommitting(false);
     setCtxDraft('');
     setPriceDialogOpen(false);
   }, [open, primaryAgent, primaryModel?.id, provider.id]);
@@ -383,11 +517,20 @@ export function ModelAdvancedDrawer({
   const commitCtxDraft = useCallback(() => {
     if (!ctxDirtyRef.current || ctxInvalid || ctx.loading || localConfigurationBlocked) return;
     ctxDirtyRef.current = false;
-    void ctx.setLimit(ctxDraft.trim() === '' ? null : parsedTokens);
+    // 提交在途期间锁住 displayedLimit：hook 的 promise 在它把新值（或失败回滚后的值）
+    // 写进 state 之后才 resolve，不锁会闪回一帧旧值。仅当没有更新的提交/重置发生时
+    // 才清状态（ctxCommitGen 递增即代表有更新的那次）。
+    const generation = (ctxCommitGen.current += 1);
+    setCtxCommitting(true);
+    void Promise.resolve(ctx.setLimit(ctxDraft.trim() === '' ? null : parsedTokens)).finally(() => {
+      if (ctxCommitGen.current === generation) setCtxCommitting(false);
+    });
   }, [ctx, ctxDraft, ctxInvalid, parsedTokens, localConfigurationBlocked]);
   const resetCtx = useCallback(() => {
     if (localConfigurationBlocked) return;
     ctxDirtyRef.current = false;
+    ctxCommitGen.current += 1;
+    setCtxCommitting(false);
     setCtxDraft(defaultWindow > 0 ? editableContextK(defaultWindow) : '');
     void ctx.reset();
   }, [ctx, defaultWindow, localConfigurationBlocked]);
@@ -401,9 +544,11 @@ export function ModelAdvancedDrawer({
   }
 
   const capabilityModel = primaryModel.mode === 'video_generation' ? videoModel ?? primaryModel : imageModel ?? videoModel ?? primaryModel;
-  const vision = imageModel?.modalities
-    ? imageModel.modalities.input.includes('image') ? 'vision' : 'no-vision'
-    : primaryModel.supportsImageInput !== undefined
+  // 本行的能力事实必须属于**被编辑的那个模型**：图片输入控件把声明写到该行选中模型的
+  // override，图标/文案却显示图片/视频生成模型（上游 capabilityModel）会变成「编辑 A 却
+  // 显示 B」。生成模型的能力展示由上游新增的图片/视频能力行与 modalities 行承担。
+  const vision =
+    primaryModel.supportsImageInput !== undefined
       ? primaryModel.supportsImageInput
         ? 'vision'
         : 'no-vision'
@@ -450,12 +595,22 @@ export function ModelAdvancedDrawer({
   const protocols = modelProtocolComparison(provider, row.byAgent);
   const protocolLabel = (api: PiModelApi | null) =>
     api ? MODEL_PROTOCOL_LABEL[api] ?? null : t('settings.providers.models.advanced.undeclared');
-  const displayedLimit = ctxDirtyRef.current
+  const displayedLimit = ctxDirtyRef.current || ctxCommitting
     ? ctxDraft.trim() === ''
       ? defaultWindow
       : parsedTokens
     : effectiveLimit;
-  const editRouteWindow = contextModel?.contextWindowMax ?? defaultWindow;
+  // 只有「上游声明的容量」比它小时才该告警；未声明容量时拿工作默认值当窗口会误报
+  // （自定义连接上 200K 只是兜底，不是任何人的声明）。来源与上方「上游最大上下文」同一套：
+  // contextWindowMax → 已验证的 contextWindow → 未声明，否则「已验证窗口但无 max」的模型
+  // 把限制填到窗口以上也不会告警。
+  const editCapacity = contextModel?.contextWindowMax ?? 0;
+  const editRouteWindow =
+    editCapacity > 0
+      ? editCapacity
+      : contextModel?.contextWindowVerified === true
+        ? (contextModel.contextWindow ?? 0)
+        : 0;
   const overRouteWindow =
     editRouteWindow > 0 &&
     displayedLimit !== null &&
@@ -877,18 +1032,18 @@ export function ModelAdvancedDrawer({
 
                   <Section title={t('settings.providers.models.advanced.spec')}>
                     <Row label={t('settings.providers.models.advanced.contextWindow')}>
-                      {routeWindow > 0 ? (
+                      {declaredWindow > 0 ? (
                         <>
-                          {formatExactTokens(routeWindow, locale)}
+                          {formatExactTokens(declaredWindow, locale)}
                           <span className="ml-1 text-11 text-[var(--text-tertiary)]">
                             {t('settings.providers.models.advanced.tokensApprox', {
-                              approx: approxTokens(routeWindow),
+                              approx: approxTokens(declaredWindow),
                             })}
                           </span>
                         </>
                       ) : (
                         <span className="text-[var(--text-tertiary)]">
-                          {t('settings.providers.models.advanced.catalogMissing')}
+                          {t('settings.providers.models.advanced.undeclared')}
                         </span>
                       )}
                     </Row>
@@ -955,20 +1110,150 @@ export function ModelAdvancedDrawer({
                       </>
                     )}
                     <Row label={t('settings.providers.models.advanced.imageInput')}>
-                      <span
-                        title={
-                          capabilityModel.modalities || capabilityModel.supportsImageInput !== undefined
-                            ? t('settings.providers.models.advanced.catalogCapabilities')
-                            : t(`settings.providers.models.advanced.visionSource.${vision}`)
-                        }
-                      >
-                        <CapabilityValue
-                          state={
-                            vision === 'vision' ? true : vision === 'no-vision' ? false : undefined
-                          }
-                          label={t(`settings.providers.models.advanced.vision.${vision}`)}
-                        />
-                      </span>
+                      {/*
+                        三态声明：目录未声明图片能力时，Pi 会在客户端就拒收图片（请求根本不出网），
+                        用户此前没有任何可点的入口去声明它 —— 报错文案让人去开的开关并不存在。
+                        这里把状态与声明合成一个控件：图标+文案保留「目录怎么说」的事实，
+                        下拉写本机目录 override（不改连接配置，preset 连接因此仍跟随官方目录）。
+                      */}
+                      <DropdownMenu>
+                        <DropdownMenuTrigger asChild>
+                          <button
+                            type="button"
+                            // 只在真写入期间禁用。刷新期间(hook 收到 PROVIDER_CHANGED 会重读)
+                            // 值原样保留，禁用它只会造成一次多余的明暗跳变；透明度也不动 ——
+                            // 与同抽屉协议控件一致(disabled 但不改 opacity)。
+                            // xd 的能力由服务端目录决定、组织托管供应商的能力由管理员下发：两者
+                            // 本机都不可覆盖，控件保留但禁用（与同抽屉其他不可改项一致）；
+                            // Main IPC 侧另有 INVALID_PARAMS / PERMISSION_DENIED 拒写兜底。
+                            disabled={
+                              imageInput.saving ||
+                              provider.id === 'xd' ||
+                              isOrganizationManagedProvider(provider)
+                            }
+                            aria-label={t('settings.providers.models.advanced.imageInputOverride.label')}
+                            title={
+                              primaryModel.modalities || primaryModel.supportsImageInput !== undefined
+                                ? t('settings.providers.models.advanced.catalogCapabilities')
+                                : t(`settings.providers.models.advanced.visionSource.${vision}`)
+                            }
+                            className="inline-flex items-center gap-1 rounded-full px-2 py-1 hover:bg-[var(--surface-hover)]"
+                          >
+                            <CapabilityValue
+                              state={
+                                vision === 'vision' ? true : vision === 'no-vision' ? false : undefined
+                              }
+                              label={t(`settings.providers.models.advanced.vision.${vision}`)}
+                            />
+                            <span className="text-11 text-[var(--text-tertiary)]">
+                              {imageInput.isCustomized
+                                ? imageInput.value === true
+                                  ? t('settings.providers.models.advanced.imageInputOverride.declaredTrue')
+                                  : t('settings.providers.models.advanced.imageInputOverride.declaredFalse')
+                                : t('settings.providers.models.advanced.imageInputOverride.inherit')}
+                            </span>
+                            <ChevronDown size={12} className="inline" aria-hidden />
+                          </button>
+                        </DropdownMenuTrigger>
+                        <DropdownMenuContent className="z-[10003]" align="end">
+                          <DropdownMenuRadioGroup
+                            value={
+                              imageInput.isCustomized ? String(imageInput.value) : 'inherit'
+                            }
+                            onValueChange={(next) => void setImageInput(next)}
+                          >
+                            <DropdownMenuRadioItem value="inherit">
+                              {t('settings.providers.models.advanced.imageInputOverride.inherit')}
+                            </DropdownMenuRadioItem>
+                            <DropdownMenuRadioItem value="true">
+                              {t('settings.providers.models.advanced.imageInputOverride.declaredTrue')}
+                            </DropdownMenuRadioItem>
+                            <DropdownMenuRadioItem value="false">
+                              {t('settings.providers.models.advanced.imageInputOverride.declaredFalse')}
+                            </DropdownMenuRadioItem>
+                          </DropdownMenuRadioGroup>
+                        </DropdownMenuContent>
+                      </DropdownMenu>
+                    </Row>
+                    <Row label={t('settings.providers.models.advanced.thinking')}>
+                      {/*
+                        思考档位三态声明：目录说「不支持思考」时，Pi 客户端就没有 thinking
+                        通道，模型推理只能随 content 返回、被当成普通正文渲染（用户看到的是
+                        「思考被写进正文」）。此前没有任何入口能声明这件事。勾选语义 = 该模型
+                        支持这一档；「跟随供应商」删除本机 override。
+                      */}
+                      <DropdownMenu>
+                        <DropdownMenuTrigger asChild>
+                          <button
+                            type="button"
+                            // xd 的能力由服务端目录决定、组织托管供应商的能力由管理员下发：两者
+                            // 本机都不可覆盖，控件保留但禁用（与同抽屉其他不可改项一致）；
+                            // Main IPC 侧另有 INVALID_PARAMS / PERMISSION_DENIED 拒写兜底。
+                            disabled={
+                              // 首次读回前空值只是占位：此时勾一档会把「只有这一档」当成完整集合
+                              // 写盘，靜默覆盖本机已声明的档位与默认档。后续后台刷新不受影响。
+                              !thinking.loaded ||
+                              thinking.saving ||
+                              provider.id === 'xd' ||
+                              isOrganizationManagedProvider(provider)
+                            }
+                            aria-label={t('settings.providers.models.advanced.thinkingOverride.label')}
+                            title={
+                              (primaryModel.efforts?.length ?? 0) > 0
+                                ? t('settings.providers.models.advanced.catalogCapabilities')
+                                : t('settings.providers.models.advanced.thinkingOverride.sourceHint')
+                            }
+                            className="inline-flex items-center gap-1 rounded-full px-2 py-1 hover:bg-[var(--surface-hover)]"
+                          >
+                            <CapabilityValue
+                              state={(primaryModel.efforts?.length ?? 0) > 0}
+                              label={
+                                (primaryModel.efforts?.length ?? 0) > 0
+                                  ? t('settings.providers.models.advanced.thinkingCapability.on')
+                                  : t('settings.providers.models.advanced.thinkingCapability.off')
+                              }
+                            />
+                            <span className="text-11 text-[var(--text-tertiary)]">
+                              {thinking.isCustomized
+                                ? t('settings.providers.models.advanced.thinkingOverride.declared')
+                                : t('settings.providers.models.advanced.thinkingOverride.inherit')}
+                            </span>
+                            <ChevronDown size={12} className="inline" aria-hidden />
+                          </button>
+                        </DropdownMenuTrigger>
+                        <DropdownMenuContent className="z-[10003]" align="end">
+                          <DropdownMenuItem onSelect={() => void setThinkingTiers(null)}>
+                            {t('settings.providers.models.advanced.thinkingOverride.inherit')}
+                          </DropdownMenuItem>
+                          <DropdownMenuSeparator />
+                          {THINKING_TIER_CHOICES.map((tier) => {
+                            const active = thinking.value?.includes(tier) ?? false;
+                            return (
+                              <DropdownMenuCheckboxItem
+                                key={tier}
+                                checked={active}
+                                onSelect={(event) => {
+                                  // 多选：preventDefault 关掉 Radix 的自动关闭，用户可以连续勾多档。
+                                  event.preventDefault();
+                                  const next = active
+                                    ? (thinking.value ?? []).filter((item) => item !== tier)
+                                    : THINKING_TIER_CHOICES.filter((item) =>
+                                        item === tier || (thinking.value ?? []).includes(item),
+                                      );
+                                  // 空组合 = 「明确不支持思考」，必须走 null 之外的显式空数组：
+                                  // null 是「跟随供应商」，两者语义不同。
+                                  void setThinkingTiers(next.length > 0 ? next : []);
+                                }}
+                              >
+                                {/* 档位名一律用既有正本 effortLevels：与「默认推理强度」
+                                    选择器逐字一致，避免同一档位在两处叫不同名字（也避免
+                                    ultra 被译成与 xhigh 同名的「超高」）。 */}
+                                {t(`effortLevels.${tier}`)}
+                              </DropdownMenuCheckboxItem>
+                            );
+                          })}
+                        </DropdownMenuContent>
+                      </DropdownMenu>
                     </Row>
                     {capabilityModel.modalities &&
                       (['input', 'output'] as const).map((direction) => (
diff --git a/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx b/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
index a42b01f7411..3a15eb16f34 100644
--- a/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
+++ b/apps/desktop/src/renderer/components/settings/__tests__/ModelAdvancedDrawer.test.tsx
@@ -30,6 +30,34 @@ vi.mock('@/hooks/useModelContextLimit', () => ({
     };
   },
 }));
+const imageInputMocks = vi.hoisted(() => ({
+  setValue: vi.fn(async () => true),
+  target: vi.fn(),
+  value: null as boolean | null,
+  isCustomized: false,
+  diverged: false,
+  errorReason: undefined as string | undefined,
+  saving: false,
+  loading: false,
+}));
+vi.mock('@/hooks/useModelCatalogImageInput', () => ({
+  useModelCatalogImageInput: (target: unknown) => {
+    imageInputMocks.target(target);
+    return {
+      value: imageInputMocks.value,
+      isCustomized: imageInputMocks.isCustomized,
+      diverged: imageInputMocks.diverged,
+      errorReason: imageInputMocks.errorReason,
+      loading: imageInputMocks.loading,
+      saving: imageInputMocks.saving,
+      error: false,
+      setValue: imageInputMocks.setValue,
+    };
+  },
+}));
+vi.mock('@/lib/toast', () => ({
+  toast: { error: vi.fn(), success: vi.fn() },
+}));
 vi.mock('@/state/modelVisibilityPrefs', () => ({
   useModelVisibilityVersion: () => 0,
   isModelEnabled: (_agent: unknown, _provider: unknown, model: { defaultEnabled?: boolean }) => model.defaultEnabled !== false,
@@ -97,6 +125,12 @@ afterEach(cleanup);
 beforeEach(() => {
   vi.clearAllMocks();
   mocks.limit = null;
+  imageInputMocks.value = null;
+  imageInputMocks.isCustomized = false;
+  imageInputMocks.diverged = false;
+  imageInputMocks.errorReason = undefined;
+  imageInputMocks.saving = false;
+  imageInputMocks.loading = false;
   vi.mocked(getProviderModelEffort).mockReset();
 });
 
@@ -274,6 +308,94 @@ describe('model advanced editor', () => {
     expect(screen.getByText(/Chat Completions/)).toBeTruthy();
   });
 
+  it('keeps the thinking declaration disabled until the first read lands', async () => {
+    // 首次读回前 value 是空占位：此时能点开菜单并勾一档，就会把「只有这一档」当成完整
+    // 集合写盘，覆盖本机已声明的档位与默认档。
+    let release!: (view: unknown) => void;
+    const pending = new Promise((resolve) => {
+      release = resolve;
+    });
+    const setTiers = vi.fn(async () => undefined);
+    const previous = window.electronAPI;
+    Object.defineProperty(window, 'electronAPI', { configurable: true, value: { maker: {
+      updateCustomProvider: vi.fn(async () => ({ ok: true })),
+      getModelCatalogThinking: vi.fn(() => pending),
+      setModelCatalogThinking: setTiers,
+    } } });
+    const source = { ...buildUserProvider({ id: 'fixture', name: 'Fixture', runtimes: {
+      codex: { baseUrl: 'https://supplier.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat',
+        models: [{ id: 'gpt-6', name: 'GPT-6', efforts: [] as never, route: { baseUrl: 'https://supplier.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat' } }] },
+    } }), connected: true } as ProviderView;
+    try {
+      render(drawer(source.models.codex![0], 'high', ['high'], source));
+      const trigger = screen.getByRole('button', { name: 'settings.providers.models.advanced.thinkingOverride.label' });
+      expect((trigger as HTMLButtonElement).disabled).toBe(true);
+      release({ value: ['low', 'high', 'max'], defaultEffort: 'max', isCustomized: true });
+      await waitFor(() => expect((trigger as HTMLButtonElement).disabled).toBe(false));
+      expect(setTiers).not.toHaveBeenCalled();
+    } finally { Object.defineProperty(window, 'electronAPI', { configurable: true, value: previous }); }
+  });
+
+  it('disables the thinking declaration for organization-managed providers', () => {
+    // 企业托管：能力由管理员下发，本机 override 是最高优先级，必须在控件上就拦住。
+    // （按钮禁用挡不住受信 renderer 直调，IPC 侧另有 PERMISSION_DENIED 兜底。）
+    const previous = window.electronAPI;
+    Object.defineProperty(window, 'electronAPI', { configurable: true, value: { maker: {
+      updateCustomProvider: vi.fn(async () => ({ ok: true })),
+      getModelCatalogThinking: vi.fn(async () => ({ value: null, isCustomized: false })),
+      setModelCatalogThinking: vi.fn(async () => undefined),
+    } } });
+    // 企业托管的判定口径是 source === 'organization'（isOrganizationManagedProvider）。
+    const managed = { ...buildUserProvider({ id: 'corp', name: 'Corp', runtimes: {
+      codex: { baseUrl: 'https://corp.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat',
+        models: [{ id: 'corp-model', name: 'Corp Model', route: { baseUrl: 'https://corp.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat' } }] },
+    } }), connected: true, source: 'organization' } as ProviderView;
+    try {
+      render(drawer(managed.models.codex![0], 'high', ['high'], managed));
+      const trigger = screen.getByRole('button', { name: 'settings.providers.models.advanced.thinkingOverride.label' });
+      expect((trigger as HTMLButtonElement).disabled).toBe(true);
+    } finally { Object.defineProperty(window, 'electronAPI', { configurable: true, value: previous }); }
+  });
+
+  it('declares thinking tiers as a local catalog override instead of editing the connection', async () => {
+    // 目录把该模型声明成零档位（不支持思考），用户从高级设置里补上真实档位。
+    const setTiers = vi.fn(async () => undefined);
+    const getView = vi.fn(async () => ({ value: null, isCustomized: false }));
+    const previous = window.electronAPI;
+    Object.defineProperty(window, 'electronAPI', { configurable: true, value: { maker: {
+      updateCustomProvider: vi.fn(async () => ({ ok: true })),
+      getModelCatalogThinking: getView,
+      setModelCatalogThinking: setTiers,
+    } } });
+    const source = { ...buildUserProvider({ id: 'fixture', name: 'Fixture', runtimes: {
+      codex: { baseUrl: 'https://supplier.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat',
+        models: [{ id: 'gpt-6', name: 'GPT-6', efforts: [] as never, route: { baseUrl: 'https://supplier.example/v1', wireProtocol: 'openai-responses', requestPath: '/custom/chat' } }] },
+    } }), connected: true } as ProviderView;
+    try {
+      render(drawer(source.models.codex![0], 'high', ['high'], source));
+      fireEvent.keyDown(
+        screen.getByRole('button', { name: 'settings.providers.models.advanced.thinkingOverride.label' }),
+        { key: 'ArrowDown' },
+      );
+      // 档位名走既有正本 effortLevels.*：与「默认推理强度」选择器逐字一致。
+      const high = await screen.findByRole('menuitemcheckbox', { name: 'effortLevels.high' });
+      fireEvent.click(high);
+      await waitFor(() => expect(setTiers).toHaveBeenCalledTimes(1));
+      // 写入的是目录 override（不是连接配置），且带上目标行全部引擎的 id。
+      const [target, tiers, defaultTier] = setTiers.mock.calls[0]! as unknown as [
+        { providerId: string; modelId: string; relatedTargets?: unknown[] }, string[] | null, string | null,
+      ];
+      expect(target.providerId).toBe('fixture');
+      expect(target.modelId).toBe('gpt-6');
+      expect(tiers).toEqual(['high']);
+      // 首次声明没有既有默认档可沿用 → 留 null 让运行期按目录默认回退，
+      // 而不是猜「排序第一档」（那会把厂商默认如 max 静默改成 high）。
+      expect(defaultTier).toBeNull();
+      // 连接配置不得被这条路径改写。
+      expect(window.electronAPI.maker.updateCustomProvider).not.toHaveBeenCalled();
+    } finally { Object.defineProperty(window, 'electronAPI', { configurable: true, value: previous }); }
+  });
+
   it('saves protocol selection to the selected engine and model without writing derived model limits', async () => {
     const update = vi.fn(async (...args: unknown[]) => {
       void args;
@@ -370,6 +492,33 @@ describe('model advanced editor', () => {
     expect(mocks.setLimit).toHaveBeenCalledTimes(1);
   });
 
+  it('keeps the trailing token readout on the typed value until the write lands', async () => {
+    // 提交后写入是异步的（hook 先 loading、仍持有旧 limit）。修复前抽屉在 blur 瞬间丢掉草稿，
+    // 说明行会先退回旧的 400,000、等回声到了再跳 500,000 —— 用户实测到的「闪一下旧值」。
+    mocks.limit = 400_000;
+    let release: () => void = () => {};
+    mocks.setLimit.mockImplementationOnce(
+      () =>
+        new Promise<void>((resolve) => {
+          release = resolve;
+        }),
+    );
+    draw();
+    const input = screen.getByRole('textbox') as HTMLInputElement;
+    expect(input.value).toBe('400');
+    fireEvent.change(input, { target: { value: '500' } });
+    fireEvent.blur(input);
+    expect(mocks.setLimit).toHaveBeenLastCalledWith(500_000);
+    // 写入在途：说明行必须继续显示刚输入的值。
+    expect(screen.getByText(/advanced\.contextLimitRoute 500,000/)).toBeTruthy();
+    expect(screen.queryByText(/advanced\.contextLimitRoute 400,000/)).toBeNull();
+    // 回声落地（hook 此时已持有新值）后再收口：仍然是新值，不回退。
+    mocks.limit = 500_000;
+    release();
+    await waitFor(() => expect(screen.getByText(/advanced\.contextLimitRoute 500,000/)).toBeTruthy());
+    expect(screen.queryByText(/advanced\.contextLimitRoute 400,000/)).toBeNull();
+  });
+
   it.each(['cindy-local-ollama', 'ollama', 'cindy-local-lmstudio'])('keeps %s editable at 1K even when the catalog advertises a large window', (id) => {
     render(drawer(model, model.defaultEffort, model.efforts, { ...provider, id, source: 'user' }));
     const input = screen.getByRole('textbox');
@@ -394,6 +543,243 @@ describe('model advanced editor', () => {
   });
 
 
+  // 上游未声明容量时不能把「工作默认值」印成「上游最大上下文」：自定义连接的 200K 是兜底，
+  // 而运行期窗口其实取模型级上下文上限（用户实测报障：圆环 1.0M / 这里 200K）。
+  function withoutCapacity(over: Partial<CatalogModel> = {}): CatalogModel {
+    const { contextWindowMax: _omit, ...rest } = model;
+    return { ...rest, contextWindow: 200_000, ...over };
+  }
+
+  it('shows "not declared" instead of the working default when the provider declares no capacity', () => {
+    draw(withoutCapacity());
+    // 这一行只报「容量未声明」：不再把工作默认值（200K 兜底，或目录给的 1M）当成容量塞进这一格。
+    expect(
+      screen.getAllByText('settings.providers.models.advanced.undeclared').length,
+    ).toBeGreaterThan(0);
+    expect(screen.queryByText('200,000')).toBeNull();
+    expect(screen.queryByText('settings.providers.models.advanced.tokensApprox')).toBeNull();
+    // 工作窗口仍然可见 —— 在「上下文上限」那行的说明里（未设上限时取目录默认值）。
+    // 那行带插值（t 的 mock 会拼上 tokens），所以用正则而非精确匹配。
+    expect(screen.getByText(/advanced\.contextLimitRoute/)).toBeTruthy();
+  });
+
+  it('shows the window upstream handed down even when no capacity is declared', () => {
+    // opencode-go 的 deepseek-v4.1-flash 就是这种：没有 contextWindowMax，但预设/服务端目录
+    // 下发了窗口（buildUserProvider 会带上 contextWindowVerified）→ 这行要显示该窗口，而不是「未声明」。
+    draw({ ...withoutCapacity(), contextWindow: 1_048_576, contextWindowVerified: true });
+    expect(screen.getByText('1,048,576')).toBeTruthy();
+    // 「约 X」只由规格那一行渲染（「未声明」这个 key 在抽屉里另有多处，不能拿它判负）。
+    expect(screen.getByText('settings.providers.models.advanced.tokensApprox')).toBeTruthy();
+  });
+
+  it('lets the user declare an undeclared image capability from the drawer', async () => {
+    // 上游未声明图片能力时 Pi 会在客户端拒收图片，而此前没有任何可点的声明入口。
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    // 未声明时显示「跟随目录」。
+    expect(trigger.textContent).toContain('settings.providers.models.advanced.imageInputOverride.inherit');
+
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    await waitFor(() => expect(imageInputMocks.setValue).toHaveBeenCalledWith(true));
+  });
+
+  it('shows a declared capability and can return to following the catalog', async () => {
+    imageInputMocks.value = false;
+    imageInputMocks.isCustomized = true;
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    expect(trigger.textContent).toContain('settings.providers.models.advanced.imageInputOverride.declaredFalse');
+
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.inherit',
+      }),
+    );
+    await waitFor(() => expect(imageInputMocks.setValue).toHaveBeenCalledWith(null));
+  });
+
+  it('allows re-selecting the current value to repair a split across engine keys', async () => {
+    // 各引擎键分叉（手工改文件/旧版单键写入的存量数据）时，同值 no-op 会把“重选当前项”
+    // 挡掉 —— 分叉永远修不掉，UI 显示一侧而运行期读另一侧。diverged 时必须放行写入。
+    imageInputMocks.value = true;
+    imageInputMocks.isCustomized = true;
+    imageInputMocks.diverged = true;
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    await waitFor(() => expect(imageInputMocks.setValue).toHaveBeenCalledWith(true));
+  });
+
+  it('skips the write when the user picks the state it is already in', async () => {
+    imageInputMocks.value = true;
+    imageInputMocks.isCustomized = true;
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    expect(imageInputMocks.setValue).not.toHaveBeenCalled();
+  });
+
+  it('keeps the control usable and visually stable while a provider refresh refetches', () => {
+    // 写入后 main 广播 PROVIDER_CHANGED，hook 会重读；重读期间值原样保留，
+    // 既不禁用也不改透明度 —— 否则一次切换会出现两次明暗跳变（用户报障：闪一下）。
+    imageInputMocks.loading = true;
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    expect(trigger).toHaveProperty('disabled', false);
+    expect(trigger.className).not.toContain('disabled:opacity');
+    expect(trigger.className).not.toContain('opacity-60');
+  });
+
+  it('reports a failed declaration instead of silently keeping the optimistic value', async () => {
+    const { toast } = await import('@/lib/toast');
+    imageInputMocks.setValue.mockResolvedValueOnce(false);
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    await waitFor(() =>
+      expect(toast.error).toHaveBeenCalledWith(
+        'settings.providers.models.advanced.imageInputOverride.saveFailed',
+      ),
+    );
+  });
+
+  it('disables the declaration control for organization-managed providers', () => {
+    // 组织托管供应商的能力由管理员下发：本机不可覆盖。控件保留但禁用（与 xd 网关同一约定），
+    // Main IPC 侧另有 PERMISSION_DENIED 拒写兜底 —— 按钮禁用挡不住受信 renderer 直调 preload。
+    const orgProvider = { ...provider, id: 'corp-gateway', source: 'organization' } as ProviderView;
+    render(drawer(model, model.defaultEffort, model.efforts, orgProvider));
+    expect(
+      screen.getByRole('button', {
+        name: 'settings.providers.models.advanced.imageInputOverride.label',
+      }),
+    ).toHaveProperty('disabled', true);
+  });
+
+  it('shows the actionable reason from main when a hand-edited file blocks the write', async () => {
+    // store 拒绝写入时给出「先修正 model-catalog-overrides.json」的指引；只显示通用
+    // 「保存失败」会让用户反复失败却看不到唯一的修复方式。
+    const { toast } = await import('@/lib/toast');
+    imageInputMocks.setValue.mockResolvedValueOnce(false);
+    imageInputMocks.errorReason = '请先修正 model-catalog-overrides.json';
+    draw();
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    await waitFor(() =>
+      expect(toast.error).toHaveBeenCalledWith(
+        'settings.providers.models.advanced.imageInputOverride.saveFailedWithReason',
+      ),
+    );
+  });
+
+  it('covers every harness id of the row when declaring image capability', async () => {
+    // openai 行的主展示引擎是 codex（gpt-6），而运行期消费该能力的 Pi 侧 id 是 chatgpt/gpt-6。
+    // 只写主引擎的 id 时 Pi 永远读不到声明，UI 却会显示「已声明」。
+    const piModel = { ...model, id: 'chatgpt/gpt-6' };
+    render(
+      <ModelAdvancedDrawer
+        provider={provider}
+        row={{
+          id: model.id,
+          name: model.name,
+          avail: ['codex', 'pi'],
+          byAgent: { codex: model, pi: piModel },
+        }}
+        open
+        onOpenChange={vi.fn()}
+        pricePresentationOf={() => null}
+        onDisable={vi.fn()}
+        disabled={false}
+        paymentRequired={false}
+      />,
+    );
+
+    const trigger = screen.getByRole('button', {
+      name: 'settings.providers.models.advanced.imageInputOverride.label',
+    });
+    fireEvent.keyDown(trigger, { key: 'ArrowDown' });
+    fireEvent.click(
+      await screen.findByRole('menuitemradio', {
+        name: 'settings.providers.models.advanced.imageInputOverride.declaredTrue',
+      }),
+    );
+    await waitFor(() => expect(imageInputMocks.setValue).toHaveBeenCalledWith(true));
+    // 声明目标必须带上 Pi 侧 id（写盘/读取都按该 id 精确匹配）。
+    const targets = imageInputMocks.target.mock.calls.map((call) => call[0]);
+    expect(targets).toContainEqual(
+      expect.objectContaining({
+        providerId: 'openai',
+        agent: 'codex',
+        modelId: 'gpt-6',
+        relatedTargets: [expect.objectContaining({ agent: 'pi', modelId: 'chatgpt/gpt-6' })],
+      }),
+    );
+  });
+
+  it('warns when the limit exceeds a verified upstream window without a declared max', () => {
+    // 「已验证窗口但无 contextWindowMax」的模型此前不会告警：填到窗口以上也静默接受。
+    mocks.limit = 1_200_000;
+    draw({ ...withoutCapacity(), contextWindow: 1_048_576, contextWindowVerified: true });
+    expect(
+      screen.getByText('settings.providers.models.advanced.contextLimitOverWindow'),
+    ).toBeTruthy();
+  });
+
+  it('does not offer the local declaration for the server-controlled gateway provider', () => {
+    render(drawer(model, model.defaultEffort, model.efforts, { ...provider, id: 'xd' }));
+    expect(
+      screen.getByRole('button', {
+        name: 'settings.providers.models.advanced.imageInputOverride.label',
+      }),
+    ).toHaveProperty('disabled', true);
+  });
+
+  it('does not warn about the limit exceeding an undeclared capacity', () => {
+    // 未声明容量时拿 200K 工作默认值当告警基线会误报「已超出当前显示的窗口」。
+    mocks.limit = 1_200_000;
+    draw(withoutCapacity());
+    expect(screen.queryByText('settings.providers.models.advanced.contextLimitOverWindow')).toBeNull();
+  });
+
   it('keeps useful identity fields without exposing internal defaults, normal lifecycle or raw descriptions', () => {
     draw({ ...model, status: 'active', defaultEnabled: false, description: 'GPT for coding tasks' });
     expect(screen.queryByText('active')).toBeNull();
diff --git a/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogImageInput.test.ts b/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogImageInput.test.ts
new file mode 100644
index 00000000000..f8e172ca737
--- /dev/null
+++ b/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogImageInput.test.ts
@@ -0,0 +1,250 @@
+// @vitest-environment jsdom
+import { act, renderHook } from '@testing-library/react';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+import { useModelCatalogImageInput } from '../useModelCatalogImageInput';
+
+function deferred<T>() {
+  let resolve!: (value: T) => void;
+  const promise = new Promise<T>((r) => {
+    resolve = r;
+  });
+  return { promise, resolve };
+}
+
+const target = { agent: 'pi' as const, providerId: 'opencode-go', modelId: 'mimo-v2.6-flash' };
+const view = (value: boolean | null) => ({ value, isCustomized: value !== null });
+const get = vi.fn();
+const set = vi.fn();
+const ownerState = { current: true };
+let providersChanged: (() => void) | undefined;
+
+vi.mock('@/contexts/dataOwnerGeneration', () => ({
+  getDataOwnerGeneration: () => ({ dataOwnerId: 'owner', ownerGeneration: 1 }),
+  isDataOwnerGenerationCurrent: () => ownerState.current,
+}));
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  ownerState.current = true;
+  providersChanged = undefined;
+  Object.assign(window, {
+    electronAPI: {
+      maker: {
+        getModelCatalogImageInput: get,
+        setModelCatalogImageInput: set,
+        onProvidersChanged: (callback: () => void) => {
+          providersChanged = callback;
+          return () => {};
+        },
+      },
+    },
+  });
+});
+
+describe('model catalog image input override', () => {
+  it('loads the current declaration', async () => {
+    get.mockResolvedValue(view(null));
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    expect(get).toHaveBeenCalledWith(target);
+    expect(hook.result.current).toMatchObject({ value: null, isCustomized: false, loading: false });
+  });
+
+  it('applies the written declaration', async () => {
+    get.mockResolvedValue(view(null));
+    set.mockResolvedValue(view(true));
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    let persisted = false;
+    await act(async () => {
+      persisted = await hook.result.current.setValue(true);
+    });
+    expect(set).toHaveBeenCalledWith(target, true);
+    expect(persisted).toBe(true);
+    expect(hook.result.current).toMatchObject({ value: true, isCustomized: true, saving: false });
+  });
+
+  it('reflects the new value before the write echoes back', async () => {
+    get.mockResolvedValue(view(null));
+    const pending = deferred<ReturnType<typeof view>>();
+    set.mockReturnValue(pending.promise);
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    // 不等待回声：调用后立即就是新值（消除「点完卡一下」的观感）。
+    act(() => {
+      void hook.result.current.setValue(true);
+    });
+    expect(hook.result.current.value).toBe(true);
+    expect(hook.result.current.isCustomized).toBe(true);
+    expect(hook.result.current.saving).toBe(true);
+    await act(async () => {
+      pending.resolve(view(true));
+    });
+    expect(hook.result.current).toMatchObject({ value: true, saving: false });
+  });
+
+  it('reverts to the committed value when the write fails instead of keeping the optimistic one', async () => {
+    get.mockResolvedValue(view(false));
+    set.mockRejectedValue(new Error('persist failed'));
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    let persisted = true;
+    await act(async () => {
+      persisted = await hook.result.current.setValue(true);
+    });
+    // 回读拿到的真值（false）必须盖掉乐观值，并标 error；返回值告诉调用方别提示成功。
+    expect(persisted).toBe(false);
+    expect(hook.result.current).toMatchObject({ value: false, isCustomized: true, error: true });
+  });
+
+  it('falls back to "follow the catalog" when both the write and the recovery read fail', async () => {
+    get.mockResolvedValueOnce(view(true)).mockRejectedValue(new Error('read failed'));
+    set.mockRejectedValue(new Error('persist failed'));
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    await act(async () => {
+      await hook.result.current.setValue(null);
+    });
+    expect(hook.result.current).toMatchObject({ value: null, isCustomized: false, error: true });
+  });
+
+  it('ignores broadcasts that arrive while a write is in flight', async () => {
+    // 写入成功后 main 会广播 PROVIDER_CHANGED；而其它窗口期的广播可能在写盘前就到达。
+    // 若让它触发的 GET 顶掉 generation，写入自己的回声会被丢弃，UI 停在旧值而写其实成功。
+    get.mockResolvedValue(view(null));
+    const pending = deferred<ReturnType<typeof view>>();
+    set.mockReturnValue(pending.promise);
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    const callsAfterMount = get.mock.calls.length;
+
+    act(() => {
+      void hook.result.current.setValue(true);
+    });
+    expect(hook.result.current.value).toBe(true);
+
+    // 写在途时的广播：不得再发 GET（发了就会抢走写入的回声）。
+    await act(async () => {
+      providersChanged?.();
+    });
+    expect(get.mock.calls.length).toBe(callsAfterMount);
+
+    await act(async () => {
+      pending.resolve(view(true));
+    });
+    expect(hook.result.current).toMatchObject({ value: true, isCustomized: true, saving: false });
+  });
+
+  it('still reads a different model while a write for another row is in flight', async () => {
+    // 写在途时跳过回读只对同一个 target 生效：否则写 A 行后立刻打开 B 行，B 的回读会被跳过、
+    // 界面永远停在 loading（显示成「跟随供应商」且点击无效）。
+    const pendingWrite = deferred<ReturnType<typeof view>>();
+    get.mockResolvedValue(view(null));
+    set.mockReturnValue(pendingWrite.promise);
+    const hook = renderHook(({ modelId }) => useModelCatalogImageInput({ ...target, modelId }), {
+      initialProps: { modelId: 'mimo-v2.6-flash' },
+    });
+    await act(async () => {});
+    act(() => {
+      void hook.result.current.setValue(true);
+    });
+
+    get.mockResolvedValue(view(false));
+    await act(async () => {
+      hook.rerender({ modelId: 'gpt-6' });
+    });
+    // 新行的读取没有被在途写吞掉，拿到的是自己的值。
+    expect(get).toHaveBeenLastCalledWith({ ...target, modelId: 'gpt-6' });
+    expect(hook.result.current).toMatchObject({
+      value: false,
+      isCustomized: true,
+      loading: false,
+    });
+
+    await act(async () => {
+      pendingWrite.resolve(view(true));
+    });
+    hook.unmount();
+  });
+
+  it('re-reads the row when the user comes back while its write is still in flight', async () => {
+    // 写 A 未完成 → 切 B → 切回 A：A 的首次读取当时被跳过，旧写入回声又被 generation 守卫
+    // 丢弃。不补读的话 A 会永远停在 loading 占位（显示成「跟随供应商」且点击无效）。
+    const pendingWrite = deferred<ReturnType<typeof view>>();
+    get.mockResolvedValue(view(null));
+    set.mockReturnValue(pendingWrite.promise);
+    const hook = renderHook(({ modelId }) => useModelCatalogImageInput({ ...target, modelId }), {
+      initialProps: { modelId: 'mimo-v2.6-flash' },
+    });
+    await act(async () => {});
+    act(() => {
+      void hook.result.current.setValue(true);
+    });
+
+    await act(async () => {
+      hook.rerender({ modelId: 'gpt-6' });
+    });
+    get.mockResolvedValue(view(false));
+    await act(async () => {
+      hook.rerender({ modelId: 'mimo-v2.6-flash' });
+    });
+
+    const readsBefore = get.mock.calls.length;
+    await act(async () => {
+      pendingWrite.resolve(view(true));
+    });
+    // 写入落定时回声已被丢弃：必须补一次读取，界面不能停在 loading。
+    expect(get.mock.calls.length).toBeGreaterThan(readsBefore);
+    expect(hook.result.current).toMatchObject({ loading: false });
+    hook.unmount();
+  });
+
+  it('surfaces the main-side failure reason instead of only reporting false', async () => {
+    // main 在这条路径上带可执行指引（手改文件里有无法保留的条目时要先修文件）；
+    // 只回 false 会让用户反复保存失败却看不到唯一的修复方式。
+    get.mockResolvedValue(view(null));
+    set.mockRejectedValue(
+      Object.assign(
+        new Error('[PRECONDITION_FAILED] 请先修正 model-catalog-overrides.json'),
+        { code: 'PRECONDITION_FAILED' },
+      ),
+    );
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+    await act(async () => {});
+    await act(async () => {
+      await hook.result.current.setValue(true);
+    });
+    // 展示用文案去掉机器码前缀。
+    expect(hook.result.current.errorReason).toBe('请先修正 model-catalog-overrides.json');
+  });
+
+  it('drops a response that lands after the data owner changed', async () => {
+    const pending = deferred<ReturnType<typeof view>>();
+    get.mockReturnValue(pending.promise);
+    const hook = renderHook(() => useModelCatalogImageInput(target));
+
+    // 切号（账号/session owner 代次变化）后旧响应不得写进状态。
+    ownerState.current = false;
+    await act(async () => {
+      pending.resolve(view(true));
+    });
+    expect(hook.result.current.value).toBeNull();
+  });
+
+  it('does not let a stale read for a previous model win', async () => {
+    const stale = deferred<ReturnType<typeof view>>();
+    get.mockReturnValueOnce(stale.promise).mockResolvedValueOnce(view(true));
+    const hook = renderHook(({ modelId }) => useModelCatalogImageInput({ ...target, modelId }), {
+      initialProps: { modelId: 'old' },
+    });
+    await act(async () => {
+      hook.rerender({ modelId: 'new' });
+    });
+    expect(hook.result.current.value).toBe(true);
+    await act(async () => {
+      stale.resolve(view(false));
+    });
+    expect(hook.result.current.value).toBe(true);
+    hook.unmount();
+  });
+});
diff --git a/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogThinking.test.ts b/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogThinking.test.ts
new file mode 100644
index 00000000000..938dad71092
--- /dev/null
+++ b/apps/desktop/src/renderer/hooks/__tests__/useModelCatalogThinking.test.ts
@@ -0,0 +1,138 @@
+// @vitest-environment jsdom
+import { act, renderHook } from '@testing-library/react';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+import { useModelCatalogThinking } from '../useModelCatalogThinking';
+
+function deferred<T = void>() {
+  let resolve!: (value: T) => void;
+  const promise = new Promise<T>((r) => {
+    resolve = r;
+  });
+  return { promise, resolve };
+}
+
+const get = vi.fn();
+const set = vi.fn();
+const ownerState = { current: true };
+let subscriptions = 0;
+let unsubscribes = 0;
+let broadcast: (() => void) | undefined;
+
+vi.mock('@/contexts/dataOwnerGeneration', () => ({
+  getDataOwnerGeneration: () => ({ dataOwnerId: 'owner', ownerGeneration: 1 }),
+  isDataOwnerGenerationCurrent: () => ownerState.current,
+}));
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  ownerState.current = true;
+  subscriptions = 0;
+  unsubscribes = 0;
+  broadcast = undefined;
+  Object.assign(window, {
+    electronAPI: {
+      maker: {
+        getModelCatalogThinking: get,
+        setModelCatalogThinking: set,
+        onProvidersChanged: (callback: () => void) => {
+          subscriptions += 1;
+          broadcast = callback;
+          return () => {
+            unsubscribes += 1;
+          };
+        },
+      },
+    },
+  });
+});
+
+const target = { agent: 'pi' as const, providerId: 'p', modelId: 'm' };
+type Props = { target: typeof target | null };
+
+describe('model catalog thinking override hook', () => {
+  it('returns the cleanup on the empty-target cycle too, so open/close cycles do not leak subscriptions', async () => {
+    // 回归（greptile P2）：抽屉关闭（target=null）的 effect 周期先前只订阅、直接 return，
+    // 不交还清理函数 —— 每开关一次抽屉泄漏一个 onProvidersChanged 订阅。泄漏回调在之后的
+    // 保存广播里调 run()：setState(EMPTY) 清掉勾选、递增 generation 顶掉写入回声，补读前
+    // 界面错误显示「跟随供应商」。
+    get.mockResolvedValue({ tiers: ['low'], defaultEffort: 'low', isCustomized: false });
+    const hook = renderHook((props: Props) => useModelCatalogThinking(props.target), {
+      initialProps: { target } as Props,
+    });
+    await act(async () => {});
+    expect(subscriptions).toBe(1);
+
+    // 开 → 关 → 再开：每个 effect 周期留下的订阅都必须在其清理函数里注销。
+    hook.rerender({ target: null });
+    await act(async () => {});
+    hook.rerender({ target });
+    await act(async () => {});
+    expect(subscriptions).toBe(3);
+    expect(unsubscribes).toBe(2); // 前两个周期已注销；泄漏时这里是 1
+
+    hook.unmount();
+    expect(unsubscribes).toBe(subscriptions); // 卸载后一个不剩
+  });
+
+  it('a broadcast after reopen reads through exactly once instead of being clobbered by a stale callback', async () => {
+    // 泄漏的可观察后果：重开后的广播会同时触发泄漏回调（空目标 → setState(EMPTY)）与当前
+    // 回调（真读回），EMPTY 后到还会把读回值盖掉。修复后每个周期只有一份订阅，广播恰好
+    // 触发一次 GET，界面停在真值上。
+    get.mockResolvedValue({ tiers: ['low'], defaultEffort: 'low', isCustomized: false });
+    const hook = renderHook((props: Props) => useModelCatalogThinking(props.target), {
+      initialProps: { target } as Props,
+    });
+    await act(async () => {});
+    hook.rerender({ target: null });
+    await act(async () => {});
+    hook.rerender({ target });
+    await act(async () => {});
+    get.mockClear();
+
+    await act(async () => {
+      broadcast?.();
+    });
+    expect(get).toHaveBeenCalledTimes(1);
+    expect(hook.result.current).toMatchObject({ tiers: ['low'], loading: false, error: false });
+  });
+
+  it('keeps the committed defaultEffort in optimistic state so a consecutive checkbox write does not null it', async () => {
+    // 回归（greptile P1）：乐观状态先前只保留 value/isCustomized，丢掉 defaultEffort。
+    // 下拉菜单在保存期间保持可用（多选不自动关），用户在第一笔落盘前再勾一档时，抽屉
+    // 拿 state.defaultEffort 当 keptDefault —— 丢了它，第二笔就会把 null 写回文件，
+    // 静默清掉用户原有的默认档。
+    get.mockResolvedValue({ tiers: ['low'], defaultEffort: 'low', isCustomized: false });
+    const firstWrite = deferred<void>();
+    let firstSettled = false;
+    set.mockImplementationOnce(() => {
+      firstSettled = true;
+      void firstWrite;
+      return firstWrite.promise as Promise<unknown>;
+    });
+    const hook = renderHook((props: Props) => useModelCatalogThinking(props.target), {
+      initialProps: { target } as Props,
+    });
+    await act(async () => {});
+
+    // 第一笔写入保持在途；等 React 冲出乐观渲染后，模拟用户再勾一档（抽屉按当前 state
+    // 计算 keptDefault）。第二笔立即落盘。
+    let second: Promise<boolean> | undefined;
+    await act(async () => {
+      void hook.result.current.setTiers(['low', 'high'], 'low');
+    });
+    const keptDefault = hook.result.current.defaultEffort;
+    set.mockResolvedValue({ tiers: ['low', 'high', 'max'], defaultEffort: 'low', isCustomized: true });
+    await act(async () => {
+      second = hook.result.current.setTiers(['low', 'high', 'max'], keptDefault);
+      await second;
+    });
+    await act(async () => {
+      firstWrite.resolve();
+    });
+    expect(firstSettled).toBe(true);
+
+    // 第二笔的默认档必须还是 low（被乐观状态保留），而不是 null（被丢掉后抽屉传回 null）。
+    expect(keptDefault).toBe('low');
+    expect(set.mock.calls[1]).toEqual([target, ['low', 'high', 'max'], 'low']);
+  });
+});
diff --git a/apps/desktop/src/renderer/hooks/useModelCatalogImageInput.ts b/apps/desktop/src/renderer/hooks/useModelCatalogImageInput.ts
new file mode 100644
index 00000000000..5a5964dd1ad
--- /dev/null
+++ b/apps/desktop/src/renderer/hooks/useModelCatalogImageInput.ts
@@ -0,0 +1,159 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+import {
+  getDataOwnerGeneration,
+  isDataOwnerGenerationCurrent,
+} from '@/contexts/dataOwnerGeneration';
+import { createLogger } from '@/lib/logger';
+import type {
+  ModelCatalogImageInputTarget,
+  ModelCatalogImageInputView,
+} from '../../shared/modelCatalogImageInput';
+
+const log = createLogger('UseModelCatalogImageInput');
+const EMPTY: ModelCatalogImageInputView = { value: null, isCustomized: false };
+
+/**
+ * 取给用户看的失败原因：main 的 `[CODE] message` 去掉机器码前缀。main 只在这条路径上带
+ * 可执行指引（如「先修正 model-catalog-overrides.json」），把它吞掉会让用户反复保存失败
+ * 却看不到唯一的修复方式。
+ */
+function toDisplayReason(error: unknown): string | undefined {
+  const raw = error instanceof Error ? error.message : typeof error === 'string' ? error : '';
+  const message = raw.replace(/^\[[A-Z0-9_]+\]\s*/, '').trim();
+  return message.length > 0 ? message : undefined;
+}
+
+/**
+ * 单模型图片输入能力的本地目录 override。
+ *
+ * 与 `useModelContextLimit` 同一套纪律：请求不跨 target 存活(generation 守卫)、不跨账号
+ * 存活(owner 守卫)、失败时回读真值而不是把乐观值留在界面上。写入成功后 main 会广播
+ * PROVIDER_CHANGED，这里同时订阅该事件，让「跟随供应商」的括注随目录变化刷新。
+ */
+export function useModelCatalogImageInput(target: ModelCatalogImageInputTarget | null) {
+  const key = JSON.stringify(target);
+  const stableTarget = useMemo<ModelCatalogImageInputTarget | null>(() => JSON.parse(key), [key]);
+  const [state, setState] = useState({
+    ...EMPTY,
+    key,
+    loading: true,
+    saving: false,
+    error: false,
+    errorReason: undefined as string | undefined,
+  });
+  const generation = useRef(0);
+  /** 在途写入数：广播触发的回读可能在写盘前拿到旧值，不得顶掉写入自己的回声。 */
+  const writesInFlight = useRef(0);
+  /** 在途写入的目标 key：只有在途写与本次读取是同一个 target 时才跳过回读（见下）。 */
+  const writesInFlightKey = useRef<string | null>(null);
+  /** 当前渲染周期的 target key 与 run：写入落定时用它判断「用户是否已经切回/切走」。 */
+  const latestKey = useRef(key);
+  latestKey.current = key;
+  const latestRun = useRef<(write?: { value: boolean | null }) => Promise<boolean>>(async () => true);
+
+  const run = useCallback(
+    async (write?: { value: boolean | null }): Promise<boolean> => {
+      // 写在途时不发 GET：广播（含本次写入自己触发的那次）可能在写盘前到达，其结果比写入的
+      // 回声旧，却会顶掉 generation 使写入结果被丢弃 —— UI 会停在旧值而写其实已经成功。
+      // 只对**同一个 target**的在途写生效：否则刚写 A 行就打开 B 行时，B 的回读会被跳过、
+      // 界面永远停在 loading（显示成「跟随供应商」且点击无效）。
+      if (!write && writesInFlight.current > 0 && writesInFlightKey.current === key) return true;
+      const request = ++generation.current;
+      if (!stableTarget) {
+        setState({ ...EMPTY, key, loading: false, saving: false, error: false, errorReason: undefined });
+        return false;
+      }
+      const owner = getDataOwnerGeneration();
+      const current = () => request === generation.current && isDataOwnerGenerationCurrent(owner);
+      // 乐观更新：写入时立即反映新值，标签点击即变；失败时由下方的回读把真值盖回来，
+      // 不会把乐观值留在界面上冒充已保存。刷新(GET)保留原值，不产生明暗/文案跳变。
+      setState((prev) => ({
+        ...(write ? { value: write.value, isCustomized: write.value !== null } : prev),
+        key,
+        loading: !write,
+        saving: write !== undefined,
+        error: false,
+        errorReason: undefined,
+      }));
+      if (write) {
+        writesInFlight.current += 1;
+        writesInFlightKey.current = key;
+      }
+      // 写入回声是否真的写进了状态（generation 被切走/切回就会丢）。
+      let echoLanded = true;
+      try {
+        const view = write
+          ? await window.electronAPI.maker.setModelCatalogImageInput(stableTarget, write.value)
+          : await window.electronAPI.maker.getModelCatalogImageInput(stableTarget);
+        echoLanded = current();
+        if (echoLanded) {
+          setState({ ...view, key, loading: false, saving: false, error: false, errorReason: undefined });
+        }
+        return true;
+      } catch (error) {
+        log.warn('model catalog image input request failed', error);
+        // 带给人看的失败原因（main 在这条路径上会给出可执行指引），不再只回一个 false。
+        const reason = toDisplayReason(error);
+        // 失败时回读已提交的值；回读也失败就退回「跟随供应商」并标记 error，
+        // 绝不让乐观值留在界面上冒充已保存。
+        if (write && current()) {
+          try {
+            const view = await window.electronAPI.maker.getModelCatalogImageInput(stableTarget);
+            echoLanded = current();
+            if (echoLanded) {
+              setState({ ...view, key, loading: false, saving: false, error: true, errorReason: reason });
+            }
+            return false;
+          } catch (readError) {
+            log.warn('model catalog image input recovery read failed', readError);
+          }
+        }
+        echoLanded = current();
+        if (echoLanded) {
+          setState({ ...EMPTY, key, loading: false, saving: false, error: true, errorReason: reason });
+        }
+        return false;
+      } finally {
+        if (write) {
+          writesInFlight.current -= 1;
+          if (writesInFlight.current <= 0) writesInFlightKey.current = null;
+          // 写入回声丢失且用户又回到同一行（写 A → 切 B → 切回 A）：这行的首次读取当时被
+          // 跳过、回声又被 generation 守卫丢弃，界面会永远停在 loading 占位。补一次读取。
+          // 用户停在其他行时不需要补：那一行的 key 不同，本就没有被跳过。
+          if (!echoLanded && latestKey.current === key) void latestRun.current();
+        }
+      }
+    },
+    [stableTarget, key],
+  );
+  latestRun.current = run;
+
+  useEffect(() => {
+    setState({
+      ...EMPTY,
+      key,
+      loading: stableTarget !== null,
+      saving: false,
+      error: false,
+      errorReason: undefined,
+    });
+    const unsubscribe = window.electronAPI?.maker?.onProvidersChanged?.(() => {
+      void run();
+    });
+    void run();
+    return () => {
+      unsubscribe?.();
+      generation.current += 1;
+    };
+  }, [run, stableTarget, key]);
+
+  /** 返回值 = 是否真正落盘。调用方据此提示失败，不要读渲染期的 error（stale closure）。 */
+  const setValue = useCallback((value: boolean | null) => run({ value }), [run]);
+  return {
+    ...(state.key === key
+      ? state
+      : { ...EMPTY, loading: true, saving: false, error: false, errorReason: undefined }),
+    setValue,
+  };
+}
diff --git a/apps/desktop/src/renderer/hooks/useModelCatalogThinking.ts b/apps/desktop/src/renderer/hooks/useModelCatalogThinking.ts
new file mode 100644
index 00000000000..525d7ced2ca
--- /dev/null
+++ b/apps/desktop/src/renderer/hooks/useModelCatalogThinking.ts
@@ -0,0 +1,224 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+import {
+  getDataOwnerGeneration,
+  isDataOwnerGenerationCurrent,
+} from '@/contexts/dataOwnerGeneration';
+import { createLogger } from '@/lib/logger';
+import type {
+  ModelCatalogThinkingTarget,
+  ModelCatalogThinkingView,
+} from '../../shared/modelCatalogThinking';
+
+const log = createLogger('UseModelCatalogThinking');
+const EMPTY: ModelCatalogThinkingView = { value: null, defaultEffort: null, isCustomized: false };
+
+/**
+ * 取给用户看的失败原因：main 的 `[CODE] message` 去掉机器码前缀。main 只在这条路径上带
+ * 可执行指引（如「先修正 model-catalog-overrides.json」），把它吞掉会让用户反复保存失败
+ * 却看不到唯一的修复方式。
+ */
+function toDisplayReason(error: unknown): string | undefined {
+  const raw = error instanceof Error ? error.message : typeof error === 'string' ? error : '';
+  const message = raw.replace(/^\[[A-Z0-9_]+\]\s*/, '').trim();
+  return message.length > 0 ? message : undefined;
+}
+
+/**
+ * 单模型思考档位的本地目录 override。
+ *
+ * 与 `useModelCatalogImageInput` 同一套纪律：请求不跨 target 存活(generation 守卫)、
+ * 不跨账号存活(owner 守卫)、失败时回读真值而不是把乐观值留在界面上。写入成功后 main 会
+ * 广播 PROVIDER_CHANGED，这里同时订阅该事件，让「跟随供应商」的括注随目录变化刷新。
+ *
+ * 语义差异：档位是**数组**而非布尔，所以乐观值是整组档位 + 可选默认档；「跟随供应商」
+ * 用 `null` 表达（与图片输入的 `value: null` 同义）。
+ */
+export function useModelCatalogThinking(target: ModelCatalogThinkingTarget | null) {
+  const key = JSON.stringify(target);
+  const stableTarget = useMemo<ModelCatalogThinkingTarget | null>(() => JSON.parse(key), [key]);
+  const [state, setState] = useState({
+    ...EMPTY,
+    key,
+    loading: true,
+    /** 首次读回是否完成：未完成时界面上的空值只是占位，不能据此写入。 */
+    loaded: false,
+    saving: false,
+    error: false,
+    errorReason: undefined as string | undefined,
+  });
+  const generation = useRef(0);
+  /** 在途写入数：广播触发的回读可能在写盘前拿到旧值，不得顶掉写入自己的回声。 */
+  const writesInFlight = useRef(0);
+  /** 在途写入的目标 key：只有在途写与本次读取是同一个 target 时才跳过回读（见下）。 */
+  const writesInFlightKey = useRef<string | null>(null);
+  /** 当前渲染周期的 target key 与 run：写入落定时用它判断「用户是否已经切回/切走」。 */
+  const latestKey = useRef(key);
+  latestKey.current = key;
+  const latestRun = useRef<(write?: WriteShape) => Promise<boolean>>(async () => true);
+
+  type WriteShape = { tiers: string[] | null; defaultTier: string | null };
+
+  const run = useCallback(
+    async (write?: WriteShape): Promise<boolean> => {
+      // 写在途时不发 GET：广播（含本次写入自己触发的那次）可能在写盘前到达，其结果比写入的
+      // 回声旧，却会顶掉 generation 使写入结果被丢弃 —— UI 会停在旧值而写其实已经成功。
+      if (!write && writesInFlight.current > 0 && writesInFlightKey.current === key) return true;
+      const request = ++generation.current;
+      if (!stableTarget) {
+        setState({
+          ...EMPTY,
+          key,
+          loading: false,
+          loaded: true,
+          saving: false,
+          error: false,
+          errorReason: undefined,
+        });
+        return false;
+      }
+      const owner = getDataOwnerGeneration();
+      const current = () => request === generation.current && isDataOwnerGenerationCurrent(owner);
+      // 乐观更新：写入时立即反映新值；失败时由下方的回读把真值盖回来，不会把乐观值留在
+      // 界面上冒充已保存。刷新(GET)保留原值，不产生文案跳变。
+      setState((prev) => ({
+        ...(write
+          ? {
+              value: write.tiers,
+              isCustomized: write.tiers !== null,
+              // 乐观期必须保留本次提交的默认档：下拉菜单在保存期间保持可用（多选不自动关），
+              // 用户在第一笔落盘前再勾一档时，抽屉会拿 state.defaultEffort 当 keptDefault ——
+              // 这里丢了它，第二笔就会把 null 写回文件，静默清掉用户原有的默认档。
+              defaultEffort: write.defaultTier,
+            }
+          : prev),
+        key,
+        // 首次读回前就写入也不会回退到「未加载」：loaded 是「已拿到过真值」的意思，
+        // 乐观写入本身就是一次提交，不该把控件重新锁上。
+        loaded: write ? true : prev.loaded,
+        loading: !write,
+        saving: write !== undefined,
+        error: false,
+        errorReason: undefined,
+      }));
+      if (write) {
+        writesInFlight.current += 1;
+        writesInFlightKey.current = key;
+      }
+      // 写入回声是否真的写进了状态（generation 被切走/切回就会丢）。
+      let echoLanded = true;
+      try {
+        const view = write
+          ? await (async () => {
+              await window.electronAPI.maker.setModelCatalogThinking(
+                stableTarget,
+                write.tiers,
+                write.defaultTier,
+              );
+              return window.electronAPI.maker.getModelCatalogThinking(stableTarget);
+            })()
+          : await window.electronAPI.maker.getModelCatalogThinking(stableTarget);
+        echoLanded = current();
+        if (echoLanded) {
+          setState({
+            ...view,
+            key,
+            loading: false,
+            loaded: true,
+            saving: false,
+            error: false,
+            errorReason: undefined,
+          });
+        }
+        return true;
+      } catch (error) {
+        log.warn('model catalog thinking request failed', error);
+        const reason = toDisplayReason(error);
+        // 失败时回读已提交的值；回读也失败就退回「跟随供应商」并标记 error，
+        // 绝不让乐观值留在界面上冒充已保存。
+        if (write && current()) {
+          try {
+            const view = await window.electronAPI.maker.getModelCatalogThinking(stableTarget);
+            echoLanded = current();
+            if (echoLanded) {
+              setState({
+                ...view,
+                key,
+                loading: false,
+                loaded: true,
+                saving: false,
+                error: true,
+                errorReason: reason,
+              });
+            }
+            return false;
+          } catch (readError) {
+            log.warn('model catalog thinking recovery read failed', readError);
+          }
+        }
+        echoLanded = current();
+        if (echoLanded) {
+          setState({
+            ...EMPTY,
+            key,
+            loading: false,
+            loaded: true,
+            saving: false,
+            error: true,
+            errorReason: reason,
+          });
+        }
+        return false;
+      } finally {
+        if (write) {
+          writesInFlight.current -= 1;
+          if (writesInFlight.current <= 0) writesInFlightKey.current = null;
+          // 写入回声丢失且用户又回到同一行（写 A → 切 B → 切回 A）：这行的首次读取当时被
+          // 跳过、回声又被 generation 守卫丢弃，界面会永远停在 loading 占位。补一次读取。
+          if (!echoLanded && latestKey.current === key) void latestRun.current();
+        }
+      }
+    },
+    [stableTarget, key],
+  );
+  latestRun.current = run;
+
+  useEffect(() => {
+    setState({
+      ...EMPTY,
+      key,
+      loading: stableTarget !== null,
+      loaded: false,
+      saving: false,
+      error: false,
+      errorReason: undefined,
+    });
+    // 目录刷新(远端 sync / 账号切换)后重新读回：跟随供应商的显示值可能已经变了。
+    // 可选链是必需的：设置抽屉的既有测试只桩了部分 electronAPI，硬调用会让整个抽屉崩掉
+    // （与 useModelCatalogImageInput 同一处理）。
+    const unsubscribe = window.electronAPI?.maker?.onProvidersChanged?.(() => {
+      void run();
+    });
+    // 订阅在任何 target 状态下都必须交还清理函数：抽屉关闭（target=null）时若直接 return，
+    // 订阅会逐次开关累积泄漏 —— 之后每次保存广播都会让泄漏回调调 run()，它 setState(EMPTY)
+    // 并递增 generation，把正在保存的回声顶掉，补读前界面会错误显示「跟随供应商」。
+    // 空目标只跳过初始 GET（run() 内部对空目标本就是 no-op）。
+    if (!stableTarget) {
+      return () => {
+        unsubscribe?.();
+      };
+    }
+    void run();
+    return () => {
+      unsubscribe?.();
+      generation.current += 1;
+    };
+  }, [run, stableTarget, key]);
+
+  return {
+    ...state,
+    /** 声明一组档位；`tiers=null` 恢复「跟随供应商」。 */
+    setTiers: (tiers: string[] | null, defaultTier: string | null = null) =>
+      run({ tiers, defaultTier }),
+  };
+}
diff --git a/apps/desktop/src/renderer/i18n/locales/en/common.json b/apps/desktop/src/renderer/i18n/locales/en/common.json
index 402707264c5..45a7321742c 100644
--- a/apps/desktop/src/renderer/i18n/locales/en/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/en/common.json
@@ -1241,9 +1241,10 @@
     "MODEL_SWITCH_TASK_RUNNING": "Model not switched: this task is still running. Wait for the current response to finish, then try again.",
     "MODEL_WINDOW_PREPARATION_IN_PROGRESS": "Model not switched: Cindy is already preparing this task’s context. Wait for it to finish, then try again.",
     "SESSION_RUNNING": "This session is currently running a task. Please try again shortly.",
-    "PI_IMAGE_INPUT_UNSUPPORTED": "Image input is not enabled for the current Pi model. Switch to a model that supports images. For a custom model, enable “Supports image input” in provider settings, then start a new Pi task.",
+    "PI_IMAGE_INPUT_UNSUPPORTED": "Image input is not enabled for the current Pi model. Switch to a model that supports images. For a custom model, enable “Supports image input” in provider settings, then switch the model once or start a new Pi task.",
     "CREDENTIAL_SWITCH_BUSY": "Another Codex task is still running, so this provider switch was not applied (the session keeps its previous provider). Try again after those tasks finish.",
-    "PI_CATALOG_RELOAD_UNCONFIRMED": "Model catalog reload was not confirmed, so this task was stopped. Reopen the task and switch models again.",
+    "PI_CATALOG_RELOAD_UNCONFIRMED": "Model catalog reload was not confirmed, so this task was stopped. Reopen the task and retry.",
+    "PI_IMAGE_CAPABILITY_REFRESH_FAILED": "Failed to apply the image input capability change. The change was rolled back or will be retried; retry the action, and reopen the task if it keeps failing.",
     "REMOTE_WORKDIR_INVALID": "The remote project path is invalid. Choose the project directory again.",
     "REMOTE_WORKDIR_NOT_FOUND": "The remote project directory no longer exists. Choose it again.",
     "REMOTE_WORKDIR_NOT_DIRECTORY": "The remote project path is not a folder. Choose a project directory.",
@@ -3760,6 +3761,14 @@
           },
           "capabilityHint": "Uses capabilities declared by this provider. Image support falls back to model naming when undeclared.",
           "imageInput": "Image input",
+          "imageInputOverride": {
+            "label": "Image input capability",
+            "inherit": "Follow provider",
+            "declaredTrue": "Supports image input",
+            "declaredFalse": "No image input",
+            "saveFailed": "Failed to save image input capability",
+            "saveFailedWithReason": "Failed to save image input capability: {{message}}"
+          },
           "supported": "Supported",
           "unsupported": "Not supported",
           "undeclared": "Not declared",
@@ -3835,7 +3844,20 @@
           },
           "codexContextUnknown": "CLI configuration unavailable",
           "codexContextValues": "Total {{total}} · Usable {{usable}} · Auto-compact {{compact}} tokens",
-          "contextLimitGptHint": "GPT defaults to at most 272K; you can set it to 1M here. Codex applies the window and automatic compaction threshold together, retaining 5% headroom. Actual capacity depends on the model service."
+          "contextLimitGptHint": "GPT defaults to at most 272K; you can set it to 1M here. Codex applies the window and automatic compaction threshold together, retaining 5% headroom. Actual capacity depends on the model service.",
+          "thinking": "Thinking",
+          "thinkingOverride": {
+            "label": "Declare thinking tiers",
+            "inherit": "Follow provider",
+            "declared": "Declared",
+            "sourceHint": "The catalog reports no thinking tiers; declare them here if the model actually reasons",
+            "saveFailed": "Failed to save the thinking declaration",
+            "saveFailedWithReason": "Failed to save the thinking declaration: {{reason}}"
+          },
+          "thinkingCapability": {
+            "on": "Supports thinking",
+            "off": "No thinking"
+          }
         },
         "kindFilter": {
           "all": "All",
diff --git a/apps/desktop/src/renderer/i18n/locales/ja/common.json b/apps/desktop/src/renderer/i18n/locales/ja/common.json
index 15c4b50d8da..7c28f59da25 100644
--- a/apps/desktop/src/renderer/i18n/locales/ja/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ja/common.json
@@ -1229,9 +1229,10 @@
     "MODEL_SWITCH_TASK_RUNNING": "モデルは切り替わっていません。現在のタスクはまだ実行中です。今回の応答が完了してから再試行してください。",
     "MODEL_WINDOW_PREPARATION_IN_PROGRESS": "モデルは切り替わっていません。このタスクのコンテキストを整理しています。完了してから再試行してください。",
     "SESSION_RUNNING": "このセッションは現在タスクを実行中です。しばらくしてからもう一度お試しください。",
-    "PI_IMAGE_INPUT_UNSUPPORTED": "現在の Pi モデルでは画像入力が有効になっていません。画像に対応するモデルへ切り替えてください。カスタムモデルの場合は、プロバイダー設定で「画像入力に対応」を有効にしてから、新しい Pi タスクを開始してください。",
+    "PI_IMAGE_INPUT_UNSUPPORTED": "現在の Pi モデルでは画像入力が有効になっていません。画像に対応するモデルへ切り替えてください。カスタムモデルの場合は、プロバイダー設定で「画像入力に対応」を有効にしてから、モデルを一度切り替えるか、新しい Pi タスクを開始してください。",
     "CREDENTIAL_SWITCH_BUSY": "他の Codex タスクが実行中のため、今回のプロバイダー切り替えは適用されませんでした（セッションは元のプロバイダーのままです）。他のタスクが終わってからもう一度お試しください。",
-    "PI_CATALOG_RELOAD_UNCONFIRMED": "モデルカタログの再読み込みを確認できなかったため、このタスクを終了しました。タスクを開き直してからモデルを切り替えてください。",
+    "PI_CATALOG_RELOAD_UNCONFIRMED": "モデルカタログの再読み込みを確認できなかったため、このタスクを終了しました。タスクを開き直してから再試行してください。",
+    "PI_IMAGE_CAPABILITY_REFRESH_FAILED": "画像入力能力の変更を適用できませんでした。変更はロールバックされたか、あとで再試行されます。もう一度操作するか、改善しない場合はタスクを開き直してください。",
     "REMOTE_WORKDIR_INVALID": "リモートプロジェクトのパスが無効です。プロジェクトディレクトリを選び直してください。",
     "REMOTE_WORKDIR_NOT_FOUND": "リモートプロジェクトのディレクトリが見つかりません。選び直してください。",
     "REMOTE_WORKDIR_NOT_DIRECTORY": "リモートプロジェクトのパスはディレクトリではありません。プロジェクトディレクトリを選び直してください。",
@@ -3730,6 +3731,14 @@
           },
           "capabilityHint": "このプロバイダーが宣言した機能を優先します。画像対応が未宣言の場合はモデル名から推定します。",
           "imageInput": "画像入力",
+          "imageInputOverride": {
+            "label": "画像入力能力の宣言",
+            "inherit": "プロバイダーに従う",
+            "declaredTrue": "画像入力に対応",
+            "declaredFalse": "画像入力に非対応",
+            "saveFailed": "画像入力能力の保存に失敗しました",
+            "saveFailedWithReason": "画像入力能力の保存に失敗しました：{{message}}"
+          },
           "supported": "対応",
           "unsupported": "非対応",
           "undeclared": "宣言なし",
@@ -3805,7 +3814,20 @@
           },
           "codexContextUnknown": "CLI 設定を取得できません",
           "codexContextValues": "総容量 {{total}} · 使用可能 {{usable}} · 自動圧縮 {{compact}} tokens",
-          "contextLimitGptHint": "GPT は既定で最大 272K を使用します。ここで 1M に変更できます。Codex はウィンドウと自動圧縮のしきい値を同時に適用し、5% の余裕を確保します。実際の容量はモデルサービスに依存します。"
+          "contextLimitGptHint": "GPT は既定で最大 272K を使用します。ここで 1M に変更できます。Codex はウィンドウと自動圧縮のしきい値を同時に適用し、5% の余裕を確保します。実際の容量はモデルサービスに依存します。",
+          "thinking": "思考",
+          "thinkingOverride": {
+            "label": "思考ティアを宣言",
+            "inherit": "プロバイダーに従う",
+            "declared": "宣言済み",
+            "sourceHint": "カタログに思考ティアの宣言がありません。モデルが実際に推論する場合はここで宣言できます",
+            "saveFailed": "思考の宣言を保存できませんでした",
+            "saveFailedWithReason": "思考の宣言を保存できませんでした: {{reason}}"
+          },
+          "thinkingCapability": {
+            "on": "思考に対応",
+            "off": "思考に非対応"
+          }
         },
         "kindFilter": {
           "all": "すべて",
diff --git a/apps/desktop/src/renderer/i18n/locales/ko/common.json b/apps/desktop/src/renderer/i18n/locales/ko/common.json
index 4a73770ea06..cfbbd759150 100644
--- a/apps/desktop/src/renderer/i18n/locales/ko/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/ko/common.json
@@ -1229,9 +1229,10 @@
     "MODEL_SWITCH_TASK_RUNNING": "모델이 전환되지 않았습니다. 현재 작업이 아직 실행 중입니다. 이번 응답이 끝난 후 다시 시도하세요.",
     "MODEL_WINDOW_PREPARATION_IN_PROGRESS": "모델이 전환되지 않았습니다. 현재 작업의 컨텍스트를 정리하고 있습니다. 완료된 후 다시 시도하세요.",
     "SESSION_RUNNING": "이 세션은 현재 작업을 실행 중입니다. 잠시 후 다시 시도해 주세요.",
-    "PI_IMAGE_INPUT_UNSUPPORTED": "현재 Pi 모델에는 이미지 입력이 활성화되어 있지 않습니다. 이미지를 지원하는 모델로 전환하세요. 사용자 지정 모델이라면 제공자 설정에서 ‘이미지 입력 지원’을 켠 다음 새 Pi 작업을 시작하세요.",
+    "PI_IMAGE_INPUT_UNSUPPORTED": "현재 Pi 모델에는 이미지 입력이 활성화되어 있지 않습니다. 이미지를 지원하는 모델로 전환하세요. 사용자 지정 모델이라면 제공자 설정에서 ‘이미지 입력 지원’을 켠 다음 모델을 한 번 전환하거나 새 Pi 작업을 시작하세요.",
     "CREDENTIAL_SWITCH_BUSY": "다른 Codex 작업이 실행 중이어서 이번 제공자 전환은 적용되지 않았습니다(세션은 기존 제공자를 유지합니다). 다른 작업이 끝난 후 다시 시도해 주세요.",
-    "PI_CATALOG_RELOAD_UNCONFIRMED": "모델 카탈로그 다시 로드가 확인되지 않아 이 작업을 종료했습니다. 작업을 다시 연 뒤 모델을 전환하세요.",
+    "PI_CATALOG_RELOAD_UNCONFIRMED": "모델 카탈로그 다시 로드가 확인되지 않아 이 작업을 종료했습니다. 작업을 다시 연 뒤 다시 시도하세요.",
+    "PI_IMAGE_CAPABILITY_REFRESH_FAILED": "이미지 입력 능력 변경을 적용하지 못했습니다. 변경이 롤백되었거나 나중에 다시 시도됩니다. 다시 시도하고, 계속 실패하면 작업을 다시 열어 주세요.",
     "REMOTE_WORKDIR_INVALID": "원격 프로젝트 경로가 올바르지 않습니다. 프로젝트 폴더를 다시 선택하세요.",
     "REMOTE_WORKDIR_NOT_FOUND": "원격 프로젝트 폴더가 없습니다. 프로젝트 폴더를 다시 선택하세요.",
     "REMOTE_WORKDIR_NOT_DIRECTORY": "원격 프로젝트 경로가 폴더가 아닙니다. 프로젝트 폴더를 다시 선택하세요.",
@@ -3730,6 +3731,14 @@
           },
           "capabilityHint": "이 공급업체가 명시한 기능을 우선합니다. 이미지 지원 정보가 없으면 모델 이름으로 추정합니다.",
           "imageInput": "이미지 입력",
+          "imageInputOverride": {
+            "label": "이미지 입력 능력 선언",
+            "inherit": "제공자 따르기",
+            "declaredTrue": "이미지 입력 지원",
+            "declaredFalse": "이미지 입력 미지원",
+            "saveFailed": "이미지 입력 능력 저장 실패",
+            "saveFailedWithReason": "이미지 입력 능력 저장 실패: {{message}}"
+          },
           "supported": "지원",
           "unsupported": "지원 안 함",
           "undeclared": "선언되지 않음",
@@ -3805,7 +3814,20 @@
           },
           "codexContextUnknown": "CLI 설정을 읽을 수 없음",
           "codexContextValues": "전체 {{total}} · 사용 가능 {{usable}} · 자동 압축 {{compact}} tokens",
-          "contextLimitGptHint": "GPT는 기본적으로 최대 272K를 사용하며 여기서 1M으로 변경할 수 있습니다. Codex는 창과 자동 압축 임계값을 함께 적용하고 5%의 여유를 유지합니다. 실제 용량은 모델 서비스에 따라 다릅니다."
+          "contextLimitGptHint": "GPT는 기본적으로 최대 272K를 사용하며 여기서 1M으로 변경할 수 있습니다. Codex는 창과 자동 압축 임계값을 함께 적용하고 5%의 여유를 유지합니다. 실제 용량은 모델 서비스에 따라 다릅니다.",
+          "thinking": "사고",
+          "thinkingOverride": {
+            "label": "사고 등급 선언",
+            "inherit": "제공자 따르기",
+            "declared": "선언됨",
+            "sourceHint": "카탈로그에 사고 등급이 없습니다. 모델이 실제로 추론한다면 여기에서 선언하세요",
+            "saveFailed": "사고 선언을 저장하지 못했습니다",
+            "saveFailedWithReason": "사고 선언을 저장하지 못했습니다: {{reason}}"
+          },
+          "thinkingCapability": {
+            "on": "사고 지원",
+            "off": "사고 미지원"
+          }
         },
         "kindFilter": {
           "all": "전체",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
index aaf35373ed2..5ec911b940b 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-CN/common.json
@@ -1229,9 +1229,10 @@
     "MODEL_SWITCH_TASK_RUNNING": "模型未切换：当前任务仍在运行。请等待本次回复结束后再试",
     "MODEL_WINDOW_PREPARATION_IN_PROGRESS": "模型未切换：正在整理当前任务的上下文。请等待完成后再试",
     "SESSION_RUNNING": "当前任务正在运行，请稍后再试。",
-    "PI_IMAGE_INPUT_UNSUPPORTED": "当前 Pi 模型未启用图片输入。请切换到支持图片的模型；如果这是自定义模型，请在供应商设置中开启“支持图片输入”，然后新建 Pi 任务。",
+    "PI_IMAGE_INPUT_UNSUPPORTED": "当前 Pi 模型未启用图片输入。请切换到支持图片的模型；如果这是自定义模型，请在供应商设置中开启“支持图片输入”，然后切换一次模型或新建 Pi 任务。",
     "CREDENTIAL_SWITCH_BUSY": "Codex 正在其它任务中运行，本次来源切换未生效（本任务仍使用原来源）。等它结束后再试。",
-    "PI_CATALOG_RELOAD_UNCONFIRMED": "模型目录重载未确认，已终止本任务。请重新打开任务后再切换模型。",
+    "PI_CATALOG_RELOAD_UNCONFIRMED": "模型目录重载未确认，已终止本任务。请重新打开任务后重试。",
+    "PI_IMAGE_CAPABILITY_REFRESH_FAILED": "图片输入能力更新未生效（已回滚或稍后自动重试）。请重试；若持续失败请重新打开任务。",
     "REMOTE_MODEL_SWITCH_ROUTE_CHANGE": "目标模型或来源需要不同的远端路由，请关闭并重新创建远程任务后生效。",
     "REMOTE_PROVIDER_UPDATING": "供应商凭证正在更新，请稍后重试",
     "REMOTE_PROVIDER_UNSUPPORTED": "该供应商不支持 SSH 远程会话，请更换来源",
@@ -3730,6 +3731,14 @@
           },
           "capabilityHint": "优先使用当前供应商声明的能力；未声明时，图片能力参考模型名称。",
           "imageInput": "图片输入",
+          "imageInputOverride": {
+            "label": "图片输入能力声明",
+            "inherit": "跟随供应商",
+            "declaredTrue": "支持图片输入",
+            "declaredFalse": "不支持图片输入",
+            "saveFailed": "保存图片输入能力失败",
+            "saveFailedWithReason": "保存图片输入能力失败：{{message}}"
+          },
           "supported": "支持",
           "unsupported": "不支持",
           "undeclared": "未声明",
@@ -3805,7 +3814,20 @@
           },
           "codexContextUnknown": "未读取到 CLI 配置",
           "codexContextValues": "总窗口 {{total}} · 可用 {{usable}} · 自动压缩 {{compact}} tokens",
-          "contextLimitGptHint": "GPT 默认最多使用 272K；可在此自定义到 1M。Codex 会同步应用窗口与自动压缩阈值，保留 5% 余量；实际容量受模型服务限制。"
+          "contextLimitGptHint": "GPT 默认最多使用 272K；可在此自定义到 1M。Codex 会同步应用窗口与自动压缩阈值，保留 5% 余量；实际容量受模型服务限制。",
+          "thinking": "思考",
+          "thinkingOverride": {
+            "label": "声明思考档位",
+            "inherit": "跟随供应商",
+            "declared": "已声明",
+            "sourceHint": "目录没有声明思考档位；若该模型实际会推理，可在此声明",
+            "saveFailed": "保存思考声明失败",
+            "saveFailedWithReason": "保存思考声明失败：{{reason}}"
+          },
+          "thinkingCapability": {
+            "on": "支持思考",
+            "off": "不支持思考"
+          }
         },
         "kindFilter": {
           "all": "全部",
diff --git a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
index 565e46a6aad..cd6350f3c3a 100644
--- a/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
+++ b/apps/desktop/src/renderer/i18n/locales/zh-TW/common.json
@@ -1229,9 +1229,10 @@
     "MODEL_SWITCH_TASK_RUNNING": "模型未切換：目前任務仍在執行。請等待本次回覆結束後再試",
     "MODEL_WINDOW_PREPARATION_IN_PROGRESS": "模型未切換：正在整理目前任務的上下文。請等待完成後再試",
     "SESSION_RUNNING": "當前任務正在執行，請稍後再試。",
-    "PI_IMAGE_INPUT_UNSUPPORTED": "當前 Pi 模型未啟用圖片輸入。請切換到支援圖片的模型；如果這是自定義模型，請在供應商設定中開啟“支援圖片輸入”，然後新建 Pi 任務。",
+    "PI_IMAGE_INPUT_UNSUPPORTED": "當前 Pi 模型未啟用圖片輸入。請切換到支援圖片的模型；如果這是自定義模型，請在供應商設定中開啟“支援圖片輸入”，然後切換一次模型或新建 Pi 任務。",
     "CREDENTIAL_SWITCH_BUSY": "Codex 正在其它任務中執行，本次來源切換未生效（本任務仍使用原來源）。等它結束後再試。",
-    "PI_CATALOG_RELOAD_UNCONFIRMED": "模型目錄重載未確認，已終止本任務。請重新開啟任務後再切換模型。",
+    "PI_CATALOG_RELOAD_UNCONFIRMED": "模型目錄重載未確認，已終止本任務。請重新開啟任務後重試。",
+    "PI_IMAGE_CAPABILITY_REFRESH_FAILED": "圖片輸入能力更新未生效（已回滾或稍後自動重試）。請重試；若持續失敗請重新開啟任務。",
     "REMOTE_MODEL_SWITCH_ROUTE_CHANGE": "目標模型或來源需要不同的遠端路由，請關閉並重新建立遠端任務後生效。",
     "REMOTE_PROVIDER_UPDATING": "供應商憑證正在更新，請稍後重試",
     "REMOTE_PROVIDER_UNSUPPORTED": "該供應商不支援 SSH 遠端會話，請更換來源",
@@ -3730,6 +3731,14 @@
           },
           "capabilityHint": "優先使用目前供應商宣告的能力；未宣告時，圖片能力參考模型名稱。",
           "imageInput": "圖片輸入",
+          "imageInputOverride": {
+            "label": "圖片輸入能力宣告",
+            "inherit": "跟隨供應商",
+            "declaredTrue": "支援圖片輸入",
+            "declaredFalse": "不支援圖片輸入",
+            "saveFailed": "儲存圖片輸入能力失敗",
+            "saveFailedWithReason": "儲存圖片輸入能力失敗：{{message}}"
+          },
           "supported": "支援",
           "unsupported": "不支援",
           "undeclared": "未宣告",
@@ -3805,7 +3814,20 @@
           },
           "codexContextUnknown": "未讀取到 CLI 設定",
           "codexContextValues": "總視窗 {{total}} · 可用 {{usable}} · 自動壓縮 {{compact}} tokens",
-          "contextLimitGptHint": "GPT 預設最多使用 272K；可在此自訂為 1M。Codex 會同步套用視窗與自動壓縮門檻，保留 5% 餘量；實際容量受模型服務限制。"
+          "contextLimitGptHint": "GPT 預設最多使用 272K；可在此自訂為 1M。Codex 會同步套用視窗與自動壓縮門檻，保留 5% 餘量；實際容量受模型服務限制。",
+          "thinking": "思考",
+          "thinkingOverride": {
+            "label": "宣告思考檔位",
+            "inherit": "跟隨供應商",
+            "declared": "已宣告",
+            "sourceHint": "目錄沒有宣告思考檔位；若該模型實際會推理，可在此宣告",
+            "saveFailed": "儲存思考宣告失敗",
+            "saveFailedWithReason": "儲存思考宣告失敗：{{reason}}"
+          },
+          "thinkingCapability": {
+            "on": "支援思考",
+            "off": "不支援思考"
+          }
         },
         "kindFilter": {
           "all": "全部",
diff --git a/apps/desktop/src/renderer/vite-env.d.ts b/apps/desktop/src/renderer/vite-env.d.ts
index 51ff58a86b1..0c112254aaf 100644
--- a/apps/desktop/src/renderer/vite-env.d.ts
+++ b/apps/desktop/src/renderer/vite-env.d.ts
@@ -6003,6 +6003,21 @@ interface ElectronAPI {
       limit: number | null,
       owner: import('../shared/modelContextLimit').ModelContextLimitOwner,
     ) => Promise<import('../shared/modelContextLimit').ModelContextLimitView>;
+    getModelCatalogThinking: (
+      target: import('../shared/modelCatalogThinking').ModelCatalogThinkingTarget,
+    ) => Promise<import('../shared/modelCatalogThinking').ModelCatalogThinkingView>;
+    setModelCatalogThinking: (
+      target: import('../shared/modelCatalogThinking').ModelCatalogThinkingTarget,
+      tiers: string[] | null,
+      defaultTier?: string | null,
+    ) => Promise<void>;
+    getModelCatalogImageInput: (
+      target: import('../shared/modelCatalogImageInput').ModelCatalogImageInputTarget,
+    ) => Promise<import('../shared/modelCatalogImageInput').ModelCatalogImageInputView>;
+    setModelCatalogImageInput: (
+      target: import('../shared/modelCatalogImageInput').ModelCatalogImageInputTarget,
+      value: boolean | null,
+    ) => Promise<import('../shared/modelCatalogImageInput').ModelCatalogImageInputView>;
     resetModelContextLimit: (
       target: import('../shared/modelContextLimit').ModelContextLimitTarget,
       owner: import('../shared/modelContextLimit').ModelContextLimitOwner,
diff --git a/apps/desktop/src/shared/__tests__/inputError.test.ts b/apps/desktop/src/shared/__tests__/inputError.test.ts
index 50dbc94836e..9c877fcbb81 100644
--- a/apps/desktop/src/shared/__tests__/inputError.test.ts
+++ b/apps/desktop/src/shared/__tests__/inputError.test.ts
@@ -1,6 +1,11 @@
 import { describe, expect, it } from 'vitest';
 
-import { PI_IMAGE_INPUT_UNSUPPORTED_MARKER, isPiImageInputUnsupportedError } from '../inputError';
+import {
+  PI_IMAGE_CAPABILITY_REFRESH_FAILED_MARKER,
+  PI_IMAGE_INPUT_UNSUPPORTED_MARKER,
+  isPiImageCapabilityRefreshFailedError,
+  isPiImageInputUnsupportedError,
+} from '../inputError';
 
 describe('isPiImageInputUnsupportedError', () => {
   it('recognizes both the structured code and the wire-safe marker', () => {
@@ -20,3 +25,26 @@ describe('isPiImageInputUnsupportedError', () => {
     );
   });
 });
+
+describe('isPiImageCapabilityRefreshFailedError', () => {
+  it('recognizes both the structured code and the wire-safe marker', () => {
+    expect(
+      isPiImageCapabilityRefreshFailedError(
+        Object.assign(new Error('pending'), { code: 'PI_IMAGE_CAPABILITY_REFRESH_FAILED' }),
+      ),
+    ).toBe(true);
+    expect(
+      isPiImageCapabilityRefreshFailedError(`${PI_IMAGE_CAPABILITY_REFRESH_FAILED_MARKER} pending`),
+    ).toBe(true);
+  });
+
+  it('does not confuse the retryable refresh failure with the unsupported verdict', () => {
+    // 两个码在 UI 上是两种处置：前者可重试，后者要用户换模型/开能力。混判会给出错误指引。
+    expect(
+      isPiImageCapabilityRefreshFailedError(`${PI_IMAGE_INPUT_UNSUPPORTED_MARKER} disabled`),
+    ).toBe(false);
+    expect(
+      isPiImageInputUnsupportedError(`${PI_IMAGE_CAPABILITY_REFRESH_FAILED_MARKER} pending`),
+    ).toBe(false);
+  });
+});
diff --git a/apps/desktop/src/shared/inputError.ts b/apps/desktop/src/shared/inputError.ts
index cc9671adfdc..48a76e0d011 100644
--- a/apps/desktop/src/shared/inputError.ts
+++ b/apps/desktop/src/shared/inputError.ts
@@ -1,5 +1,7 @@
 export const PI_IMAGE_INPUT_UNSUPPORTED_CODE = 'PI_IMAGE_INPUT_UNSUPPORTED';
 export const PI_IMAGE_INPUT_UNSUPPORTED_MARKER = `[${PI_IMAGE_INPUT_UNSUPPORTED_CODE}]`;
+export const PI_IMAGE_CAPABILITY_REFRESH_FAILED_CODE = 'PI_IMAGE_CAPABILITY_REFRESH_FAILED';
+export const PI_IMAGE_CAPABILITY_REFRESH_FAILED_MARKER = `[${PI_IMAGE_CAPABILITY_REFRESH_FAILED_CODE}]`;
 
 /** Stable Desktop main -> renderer/device-link marker; presentation localizes at display time. */
 export function isPiImageInputUnsupportedError(value: unknown): boolean {
@@ -13,6 +15,19 @@ export function isPiImageInputUnsupportedError(value: unknown): boolean {
   );
 }
 
+/** 图片能力已改但子进程还没加载（回合在跑 / 重载被拒），可重试。 */
+export function isPiImageCapabilityRefreshFailedError(value: unknown): boolean {
+  const message = value instanceof Error ? value.message : String(value ?? '');
+  const code =
+    value && typeof value === 'object' && !Array.isArray(value)
+      ? (value as { code?: unknown }).code
+      : null;
+  return (
+    code === PI_IMAGE_CAPABILITY_REFRESH_FAILED_CODE ||
+    message.includes(PI_IMAGE_CAPABILITY_REFRESH_FAILED_MARKER)
+  );
+}
+
 /** The existing Pi prompt RPC expiry signal, shared by native recovery and group settlement. */
 export function isPiPromptRpcTimeoutError(data: unknown): boolean {
   if (!data || typeof data !== 'object') return false;
diff --git a/apps/desktop/src/shared/ipc-errors.ts b/apps/desktop/src/shared/ipc-errors.ts
index 172be8209a2..71b031feaaf 100644
--- a/apps/desktop/src/shared/ipc-errors.ts
+++ b/apps/desktop/src/shared/ipc-errors.ts
@@ -133,6 +133,8 @@ export type IpcErrorCode =
   | 'CREDENTIAL_SWITCH_BUSY'
   // Pi 热切 SuperGrok 时 switch_session 未确认，任务已终止。
   | 'PI_CATALOG_RELOAD_UNCONFIRMED'
+  // Pi 图片输入能力对账失败:写 models.json / 子进程重载被拒。会话仍可用(已回滚或保留 pending)。
+  | 'PI_IMAGE_CAPABILITY_REFRESH_FAILED'
   // 远端 Claude 路由 materialization 失败(remote-claude-route.ts):
   // 供应商凭证 mutation 窗口(稍后重试)/ 远端不可表达(换来源)/ 订阅未连接(连接 Claude.ai)。
   | 'REMOTE_PROVIDER_UPDATING'
@@ -395,6 +397,7 @@ const IPC_ERROR_CODES: ReadonlySet<IpcErrorCode> = new Set<IpcErrorCode>([
   'SESSION_RUNNING',
   'CREDENTIAL_SWITCH_BUSY',
   'PI_CATALOG_RELOAD_UNCONFIRMED',
+  'PI_IMAGE_CAPABILITY_REFRESH_FAILED',
   'REMOTE_PROVIDER_UPDATING',
   'REMOTE_PROVIDER_UNSUPPORTED',
   'REMOTE_NATIVE_OAUTH_UNAVAILABLE',
diff --git a/apps/desktop/src/shared/modelCatalogImageInput.ts b/apps/desktop/src/shared/modelCatalogImageInput.ts
new file mode 100644
index 00000000000..7cb53e613f7
--- /dev/null
+++ b/apps/desktop/src/shared/modelCatalogImageInput.ts
@@ -0,0 +1,29 @@
+import type { ModelPriceOverrideTarget } from './modelPriceOverride';
+
+/**
+ * 单模型图片输入能力的本地目录 override(设置 → 模型 → 高级设置)。
+ *
+ * 与价格/上下文上限 override 共用 (providerId, agent, modelId) 目标形状；agent 只用于
+ * 目录成员校验，实际写入的是**模型级** base patch(图片能力是公共字段，见
+ * docs/product-rules/model-metadata-precedence.md)。
+ */
+export interface ModelCatalogImageInputTarget extends ModelPriceOverrideTarget {
+  /**
+   * 同一行在**其它引擎**下的 wire id。桥接投影两端 id 不同（例如 openai 行：codex 用
+   * `gpt-5.6-sol`、pi 用 `chatgpt/gpt-5.6-sol`），而本机 override 按 (providerId, modelId)
+   * 精确匹配 —— 只写主展示引擎的 id，运行期真正消费该能力的 Pi 读不到，UI 却会显示已声明。
+   */
+  relatedTargets?: ModelPriceOverrideTarget[];
+}
+
+/** 读回视图。value=null 表示没有本地声明(跟随供应商)。 */
+export interface ModelCatalogImageInputView {
+  value: boolean | null;
+  isCustomized: boolean;
+  /**
+   * 该行各引擎的声明**不一致**(手工改文件/旧版单键写入留下的存量数据)。展示值取运行期
+   * 真正消费该能力的引擎(Pi)那一侧，但 UI 必须允许用户“重选当前项”把分裂写回一致 ——
+   * 否则同值 no-op 会让分叉永远修不掉。
+   */
+  diverged?: boolean;
+}
diff --git a/apps/desktop/src/shared/modelCatalogThinking.ts b/apps/desktop/src/shared/modelCatalogThinking.ts
new file mode 100644
index 00000000000..2be536a3a67
--- /dev/null
+++ b/apps/desktop/src/shared/modelCatalogThinking.ts
@@ -0,0 +1,42 @@
+import type { ModelPriceOverrideTarget } from './modelPriceOverride';
+
+/**
+ * 单模型思考档位的本地目录 override(设置 → 模型 → 高级设置)。
+ *
+ * 与价格 / 上下文上限 / 图片输入 override 共用 `(providerId, agent, modelId)` 目标形状。
+ * 写入的是模型级 base patch 的 `efforts` + `defaultEffort`(思考档位是公共字段，见
+ * `docs/product-rules/model-metadata-precedence.md`)。
+ *
+ * 为什么需要它:目录里的思考档位决定 Pi 能不能开真正的 thinking 通道
+ * (`piThinkingLevels.mjs`:`reasoning !== true` → 零档位)。上游把一个**会推理**的模型
+ * 声明成「不支持思考」时,Pi 没有通道可用,模型推理只能随 content 返回、被 Cindy 当
+ * **普通 assistant 正文**渲染(实报:协同 worker 切到该模型后正文变成推理叙述)。
+ * 这是「未声明」与「声明错误」两种真实状态,用户必须能自己显式声明。
+ */
+export interface ModelCatalogThinkingTarget extends ModelPriceOverrideTarget {
+  /**
+   * 同一行在**其它引擎**下的 wire id。桥接投影两端 id 不同(例如 openai 行:codex 用
+   * `gpt-5.6-sol`、pi 用 `chatgpt/gpt-5.6-sol`),而本机 override 按 (providerId, modelId)
+   * 精确匹配 —— 只写主展示引擎的 id,运行期真正消费档位的 Pi 读不到,UI 却显示已声明。
+   */
+  relatedTargets?: ModelPriceOverrideTarget[];
+}
+
+/** 读回视图。value=null 表示没有本地声明(跟随供应商)。 */
+export interface ModelCatalogThinkingView {
+  /** null = 没有本地声明(跟随供应商实报)。 */
+  value: string[] | null;
+  /**
+   * 本机声明里的默认档。`value` 非 null 时才有意义。
+   * 必须读回：否则用户在抽屉里改档位时无从知道当前默认是哪一档，写入侧只能猜一个
+   * （例如「排序第一档」），那会把厂商声明的默认静默改掉且用户无从修正。
+   */
+  defaultEffort?: string | null;
+  isCustomized: boolean;
+  /**
+   * 该行各引擎的声明**不一致**(手工改文件 / 旧版单键写入留下的存量数据)。展示值取运行期
+   * 真正消费档位的引擎(Pi)那一侧,但 UI 必须允许用户「重选当前项」把分裂写回一致 ——
+   * 否则同值 no-op 会让分叉永远修不掉。
+   */
+  diverged?: boolean;
+}
diff --git a/docs/design-rules/design-inventory.md b/docs/design-rules/design-inventory.md
index 857a86c945e..7c8ae8db879 100644
--- a/docs/design-rules/design-inventory.md
+++ b/docs/design-rules/design-inventory.md
@@ -38,7 +38,7 @@
 | `desktop.overlay.toast` | desktop | Toast | App 常驻 ToastContainer（用户可见出口，不展开业务逻辑） | Toast, ToastContainer | apps/desktop/src/renderer/components/ui/toast/DownloadToast.tsx, apps/desktop/src/renderer/components/ui/toast/Toast.tsx, apps/desktop/src/renderer/components/ui/toast/ToastContainer.tsx, apps/desktop/src/renderer/components/ui/toast/index.ts | 7 | 4 | 3 |
 | `desktop.plugins.app-main` | desktop | 插件主视图 | hash `/apps/:ghostId`（GhostMainViewFeatureLayout） | GhostMainViewFeatureLayout, GhostMainViewHost | apps/desktop/src/renderer/features/plugin/GhostMainViewFeatureLayout.tsx, apps/desktop/src/renderer/features/plugin/GhostMainViewHost.tsx | 0 | 0 | 0 |
 | `desktop.plugins.installed` | desktop | 已装插件 | hash `/plugins`（GhostPluginPage） | AddMarketplaceDialog, GhostPluginActions, GhostPluginDetailView, GhostPluginPage, MarketPluginDetailView, MyPublishesSection, PluginDevicePicker, PluginManagementLayout, PluginProjectMenu, RemotePluginConfiguration, RemotePluginPage, RemotePluginReadOnlyDetail, UpdateAllDialog | apps/desktop/src/renderer/features/plugin/AddMarketplaceDialog.tsx, apps/desktop/src/renderer/features/plugin/GhostPluginActions.tsx, apps/desktop/src/renderer/features/plugin/GhostPluginDetailView.tsx, apps/desktop/src/renderer/features/plugin/GhostPluginPage.tsx, apps/desktop/src/renderer/features/plugin/MarketPluginDetailView.tsx, apps/desktop/src/renderer/features/plugin/MyPublishesSection.tsx, apps/desktop/src/renderer/features/plugin/PluginDevicePicker.tsx, apps/desktop/src/renderer/features/plugin/PluginManagementLayout.tsx, apps/desktop/src/renderer/features/plugin/PluginScopePicker.tsx, apps/desktop/src/renderer/features/plugin/RemotePluginConfiguration.tsx, apps/desktop/src/renderer/features/plugin/RemotePluginPage.tsx, apps/desktop/src/renderer/features/plugin/RemotePluginReadOnlyDetail.tsx, apps/desktop/src/renderer/features/plugin/UpdateAllDialog.tsx, apps/desktop/src/renderer/features/plugin/plugin-motion.css | 28 | 0 | 48 |
-| `desktop.settings` | desktop | 设置 | hash `/settings`（SettingsView；tab 含 general / personalization / providers / billing / usage / voice-input / im-bot / shortcuts / agent-island / import / remote-control / ghosts / builtin-tools / computer-use / help / about） | SettingsView | apps/desktop/src/renderer/components/settings/AboutSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AboutSection.tsx, apps/desktop/src/renderer/components/settings/AccountDeletionSection.tsx, apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx, apps/desktop/src/renderer/components/settings/AddSpeechModelsDialog.tsx, apps/desktop/src/renderer/components/settings/AgentIslandSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AgentIslandSection.tsx, apps/desktop/src/renderer/components/settings/AgentResourceSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AgentResourceSection.tsx, apps/desktop/src/renderer/components/settings/AppearanceSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AppearanceSection.tsx, apps/desktop/src/renderer/components/settings/AuxiliaryModelSection.tsx, apps/desktop/src/renderer/components/settings/AvatarCropEditor.tsx, apps/desktop/src/renderer/components/settings/BetaChannelCell.tsx, apps/desktop/src/renderer/components/settings/BillingSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/BrowserBackendSubsection.tsx, apps/desktop/src/renderer/components/settings/BrowserRealProfileSubsection.tsx, apps/desktop/src/renderer/components/settings/BuiltinToolsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/BuiltinToolsSection.tsx, apps/desktop/src/renderer/components/settings/ChatEmbeddingCell.tsx, apps/desktop/src/renderer/components/settings/ChipMetricsSection.tsx, apps/desktop/src/renderer/components/settings/CindyMakeSection.settings-search.ts, apps/desktop/src/renderer/components/settings/CindyMakeSection.tsx, apps/desktop/src/renderer/components/settings/CollaborationSection.settings-search.ts, apps/desktop/src/renderer/components/settings/CollaborationSection.tsx, apps/desktop/src/renderer/components/settings/CompactionSection.tsx, apps/desktop/src/renderer/components/settings/ComposerSendShortcutSection.tsx, apps/desktop/src/renderer/components/settings/ComputerPermissionGuideWindow.tsx, apps/desktop/src/renderer/components/settings/ComputerPermissionRow.tsx, apps/desktop/src/renderer/components/settings/ComputerUseSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ComputerUseSection.tsx, apps/desktop/src/renderer/components/settings/CustomProviderRuntimeFillOverlay.tsx, apps/desktop/src/renderer/components/settings/DefaultOverrideControls.tsx, apps/desktop/src/renderer/components/settings/DialogueDirectorySection.tsx, apps/desktop/src/renderer/components/settings/DingTalkBotSection.tsx, apps/desktop/src/renderer/components/settings/DiscordBotSection.tsx, apps/desktop/src/renderer/components/settings/DownloadMeter.tsx, apps/desktop/src/renderer/components/settings/ExperimentalSection.tsx, apps/desktop/src/renderer/components/settings/FeishuBotNotificationSection.tsx, apps/desktop/src/renderer/components/settings/FeishuBotSection.tsx, apps/desktop/src/renderer/components/settings/FontFamilyPicker.tsx, apps/desktop/src/renderer/components/settings/GeneralSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/GenericGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/GitSafetySection.settings-search.ts, apps/desktop/src/renderer/components/settings/GitSafetySection.tsx, apps/desktop/src/renderer/components/settings/HarnessVersionMenuRow.tsx, apps/desktop/src/renderer/components/settings/HelpAssistantPanel.tsx, apps/desktop/src/renderer/components/settings/HelpSection.settings-search.ts, apps/desktop/src/renderer/components/settings/HelpSection.tsx, apps/desktop/src/renderer/components/settings/HelpThreadView.tsx, apps/desktop/src/renderer/components/settings/HookConnectionsSection.tsx, apps/desktop/src/renderer/components/settings/HookWorkspacePrefsEditor.tsx, apps/desktop/src/renderer/components/settings/ImBotSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ImBotSection.tsx, apps/desktop/src/renderer/components/settings/ImChannelSettingsCard.tsx, apps/desktop/src/renderer/components/settings/ImConnectionRecovery.tsx, apps/desktop/src/renderer/components/settings/ImDefaultSettingsSection.tsx, apps/desktop/src/renderer/components/settings/ImLifecycleAnnouncementSection.tsx, apps/desktop/src/renderer/components/settings/InputDeviceConnectionStatus.tsx, apps/desktop/src/renderer/components/settings/JoyConGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/KeyboardShortcutsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/KeyboardShortcutsSection.tsx, apps/desktop/src/renderer/components/settings/LanguageSection.tsx, apps/desktop/src/renderer/components/settings/LayoutResetControl.tsx, apps/desktop/src/renderer/components/settings/LinkOpenSection.tsx, apps/desktop/src/renderer/components/settings/LlamaCppCandidates.tsx, apps/desktop/src/renderer/components/settings/LlamaCppProviderDetail.tsx, apps/desktop/src/renderer/components/settings/LocalAgentSection.settings-search.ts, apps/desktop/src/renderer/components/settings/LocalAgentSection.tsx, apps/desktop/src/renderer/components/settings/LocalModelDownloadUI.tsx, apps/desktop/src/renderer/components/settings/LocalOllamaInstall.tsx, apps/desktop/src/renderer/components/settings/LocalPackagingTag.tsx, apps/desktop/src/renderer/components/settings/LogoutSection.tsx, apps/desktop/src/renderer/components/settings/LspBetaCell.tsx, apps/desktop/src/renderer/components/settings/McpServerDialog.tsx, apps/desktop/src/renderer/components/settings/McpServersSection.tsx, apps/desktop/src/renderer/components/settings/MemorySection.tsx, apps/desktop/src/renderer/components/settings/MessageNavRailCell.tsx, apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx, apps/desktop/src/renderer/components/settings/ModelCompactionSettings.tsx, apps/desktop/src/renderer/components/settings/ModelPriceOverrideDialog.tsx, apps/desktop/src/renderer/components/settings/MyDevicesPanel.tsx, apps/desktop/src/renderer/components/settings/NotificationSection.settings-search.ts, apps/desktop/src/renderer/components/settings/NotificationSection.tsx, apps/desktop/src/renderer/components/settings/OAuthDeviceCodeCard.tsx, apps/desktop/src/renderer/components/settings/OllamaProviderDetail.tsx, apps/desktop/src/renderer/components/settings/OwnRemoteProviders.tsx, apps/desktop/src/renderer/components/settings/PersonalizationDetails.settings-search.ts, apps/desktop/src/renderer/components/settings/PersonalizationSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/PiKernelVersionRow.tsx, apps/desktop/src/renderer/components/settings/PiPackagesSection.tsx, apps/desktop/src/renderer/components/settings/PlayStationGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/ProfileEditDialog.tsx, apps/desktop/src/renderer/components/settings/ProviderAccountUsageModule.tsx, apps/desktop/src/renderer/components/settings/ProviderConnectionDialog.tsx, apps/desktop/src/renderer/components/settings/ProviderImportDialog.tsx, apps/desktop/src/renderer/components/settings/ProvidersSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ProvidersSection.tsx, apps/desktop/src/renderer/components/settings/ReadOnlyProviderModels.tsx, apps/desktop/src/renderer/components/settings/RemoteControlSection.settings-search.ts, apps/desktop/src/renderer/components/settings/RemoteControlSection.tsx, apps/desktop/src/renderer/components/settings/RemoteDesktopPermissions.tsx, apps/desktop/src/renderer/components/settings/RemoteDesktopSetting.tsx, apps/desktop/src/renderer/components/settings/RemoteHostDetail.tsx, apps/desktop/src/renderer/components/settings/RemoteSection.tsx, apps/desktop/src/renderer/components/settings/SessionImportSection.settings-search.ts, apps/desktop/src/renderer/components/settings/SessionImportSection.tsx, apps/desktop/src/renderer/components/settings/SessionRuntimeFallbackCell.tsx, apps/desktop/src/renderer/components/settings/SessionShareImportWizard.tsx, apps/desktop/src/renderer/components/settings/SettingsCatalogPanel.settings-search.ts, apps/desktop/src/renderer/components/settings/SettingsCatalogPanel.tsx, apps/desktop/src/renderer/components/settings/SettingsSearchBox.tsx, apps/desktop/src/renderer/components/settings/SettingsSearchNavigation.ts, apps/desktop/src/renderer/components/settings/SettingsSidebarNav.tsx, apps/desktop/src/renderer/components/settings/SettingsTextInput.tsx, apps/desktop/src/renderer/components/settings/SettingsView.tsx, apps/desktop/src/renderer/components/settings/SharedTasks.settings-search.ts, apps/desktop/src/renderer/components/settings/SilentEncryptedRetryCell.tsx, apps/desktop/src/renderer/components/settings/SpeechModelChoices.tsx, apps/desktop/src/renderer/components/settings/SshKeySetupDialog.tsx, apps/desktop/src/renderer/components/settings/StorageManagementCard.settings-search.ts, apps/desktop/src/renderer/components/settings/StorageManagementCard.tsx, apps/desktop/src/renderer/components/settings/StreamFadeSection.tsx, apps/desktop/src/renderer/components/settings/SubagentModelSection.settings-search.ts, apps/desktop/src/renderer/components/settings/SubagentModelSection.tsx, apps/desktop/src/renderer/components/settings/SwitchProGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/TelegramBehaviorSettings.tsx, apps/desktop/src/renderer/components/settings/TelegramBotSection.tsx, apps/desktop/src/renderer/components/settings/TelegramRemoteDevices.tsx, apps/desktop/src/renderer/components/settings/TerminalShellSection.tsx, apps/desktop/src/renderer/components/settings/TipsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/TipsSection.tsx, apps/desktop/src/renderer/components/settings/UnifiedModelList.tsx, apps/desktop/src/renderer/components/settings/UserProfileCard.tsx, apps/desktop/src/renderer/components/settings/UserPromptSection.tsx, apps/desktop/src/renderer/components/settings/VisionBridgeSection.settings-search.ts, apps/desktop/src/renderer/components/settings/VisionBridgeSection.tsx, apps/desktop/src/renderer/components/settings/VoiceInputSection.settings-search.ts, apps/desktop/src/renderer/components/settings/VoiceInputSection.tsx, apps/desktop/src/renderer/components/settings/VoiceSpeakerEnrollmentDialog.tsx, apps/desktop/src/renderer/components/settings/VoiceSpeakerFilterCard.tsx, apps/desktop/src/renderer/components/settings/WallpaperSection.tsx, apps/desktop/src/renderer/components/settings/WechatBotSection.tsx, apps/desktop/src/renderer/components/settings/WecomBotSection.tsx, apps/desktop/src/renderer/components/settings/WindowBehaviorSection.settings-search.ts, apps/desktop/src/renderer/components/settings/WindowBehaviorSection.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexKeyboardLayout.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexKeycapGlyphs.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexSettings.tsx, apps/desktop/src/renderer/components/settings/WorktreeRecycleCard.settings-search.ts, apps/desktop/src/renderer/components/settings/WorktreeRecycleCard.tsx, apps/desktop/src/renderer/components/settings/XUsageGuide.tsx, apps/desktop/src/renderer/components/settings/XboxGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/XboxGamepadSettings.tsx, apps/desktop/src/renderer/components/settings/androidStatusPresentation.ts, apps/desktop/src/renderer/components/settings/avatarCropGeometry.ts, apps/desktop/src/renderer/components/settings/avatarCropImage.ts, apps/desktop/src/renderer/components/settings/billingVisibility.ts, apps/desktop/src/renderer/components/settings/browserOpenForLoginError.ts, apps/desktop/src/renderer/components/settings/cindyMakeVisibility.ts, apps/desktop/src/renderer/components/settings/computerPermissionFlow.ts, apps/desktop/src/renderer/components/settings/contacts/ContactDetailPane.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsImportDialog.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsListPane.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsManagerDialog.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsSection.tsx, apps/desktop/src/renderer/components/settings/contacts/startContactsAiSession.ts, apps/desktop/src/renderer/components/settings/dialogueDirectoryDisplay.ts, apps/desktop/src/renderer/components/settings/dualSenseSilhouette.ts, apps/desktop/src/renderer/components/settings/feishuBotPresentation.ts, apps/desktop/src/renderer/components/settings/fontFamilyValue.ts, apps/desktop/src/renderer/components/settings/gamepadLayoutPrimitives.tsx, apps/desktop/src/renderer/components/settings/gamepadSilhouetteGeom.ts, apps/desktop/src/renderer/components/settings/hookWorkspacePrefsLogic.ts, apps/desktop/src/renderer/components/settings/imBotVisibility.ts, apps/desktop/src/renderer/components/settings/imDefaultSettingsLogic.ts, apps/desktop/src/renderer/components/settings/joyConSilhouette.ts, apps/desktop/src/renderer/components/settings/modelManagementPresentation.ts, apps/desktop/src/renderer/components/settings/modelManagementState.ts, apps/desktop/src/renderer/components/settings/myDevicesModel.ts, apps/desktop/src/renderer/components/settings/providerAssetModule.ts, apps/desktop/src/renderer/components/settings/realProfilePermissionGuide.ts, apps/desktop/src/renderer/components/settings/settingsRowClasses.ts, apps/desktop/src/renderer/components/settings/settingsSearchCatalog.ts, apps/desktop/src/renderer/components/settings/settingsSearchTypes.ts, apps/desktop/src/renderer/components/settings/switchProSilhouette.ts, apps/desktop/src/renderer/components/settings/ultimateC1Silhouette.ts, apps/desktop/src/renderer/components/settings/usage/UsageBarsTooltip.tsx, apps/desktop/src/renderer/components/settings/usage/UsageBreakdownTables.tsx, apps/desktop/src/renderer/components/settings/usage/UsageDeviceSelect.tsx, apps/desktop/src/renderer/components/settings/usage/UsageFilterSelect.tsx, apps/desktop/src/renderer/components/settings/usage/UsageHistorySection.settings-search.ts, apps/desktop/src/renderer/components/settings/usage/UsageHistorySection.tsx, apps/desktop/src/renderer/components/settings/usage/UsageStatRow.tsx, apps/desktop/src/renderer/components/settings/usage/UsageTaskTable.tsx, apps/desktop/src/renderer/components/settings/usage/UsageTokenBars.tsx, apps/desktop/src/renderer/components/settings/usage/formatUsagePercent.ts, apps/desktop/src/renderer/components/settings/usage/usageCharts.css, apps/desktop/src/renderer/components/settings/usage/usageHistoryColors.ts, apps/desktop/src/renderer/components/settings/usage/usageHistoryStats.ts, apps/desktop/src/renderer/components/settings/usageVisibility.ts, apps/desktop/src/renderer/components/settings/useAutomationSettingsRead.ts, apps/desktop/src/renderer/components/settings/useProviderSubscriptionCard.ts, apps/desktop/src/renderer/components/settings/useSettingsEscape.ts, apps/desktop/src/renderer/components/settings/useSettingsSectionTarget.ts, apps/desktop/src/renderer/components/settings/wizardRecommend.ts, apps/desktop/src/renderer/components/settings/workLouderCodexCommandCopy.ts, apps/desktop/src/renderer/components/settings/xboxSeriesSilhouette.ts, apps/desktop/src/renderer/styles/sortable.css | 137 | 95 | 552 |
+| `desktop.settings` | desktop | 设置 | hash `/settings`（SettingsView；tab 含 general / personalization / providers / billing / usage / voice-input / im-bot / shortcuts / agent-island / import / remote-control / ghosts / builtin-tools / computer-use / help / about） | SettingsView | apps/desktop/src/renderer/components/settings/AboutSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AboutSection.tsx, apps/desktop/src/renderer/components/settings/AccountDeletionSection.tsx, apps/desktop/src/renderer/components/settings/AddProviderWizard.tsx, apps/desktop/src/renderer/components/settings/AddSpeechModelsDialog.tsx, apps/desktop/src/renderer/components/settings/AgentIslandSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AgentIslandSection.tsx, apps/desktop/src/renderer/components/settings/AgentResourceSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AgentResourceSection.tsx, apps/desktop/src/renderer/components/settings/AppearanceSection.settings-search.ts, apps/desktop/src/renderer/components/settings/AppearanceSection.tsx, apps/desktop/src/renderer/components/settings/AuxiliaryModelSection.tsx, apps/desktop/src/renderer/components/settings/AvatarCropEditor.tsx, apps/desktop/src/renderer/components/settings/BetaChannelCell.tsx, apps/desktop/src/renderer/components/settings/BillingSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/BrowserBackendSubsection.tsx, apps/desktop/src/renderer/components/settings/BrowserRealProfileSubsection.tsx, apps/desktop/src/renderer/components/settings/BuiltinToolsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/BuiltinToolsSection.tsx, apps/desktop/src/renderer/components/settings/ChatEmbeddingCell.tsx, apps/desktop/src/renderer/components/settings/ChipMetricsSection.tsx, apps/desktop/src/renderer/components/settings/CindyMakeSection.settings-search.ts, apps/desktop/src/renderer/components/settings/CindyMakeSection.tsx, apps/desktop/src/renderer/components/settings/CollaborationSection.settings-search.ts, apps/desktop/src/renderer/components/settings/CollaborationSection.tsx, apps/desktop/src/renderer/components/settings/CompactionSection.tsx, apps/desktop/src/renderer/components/settings/ComposerSendShortcutSection.tsx, apps/desktop/src/renderer/components/settings/ComputerPermissionGuideWindow.tsx, apps/desktop/src/renderer/components/settings/ComputerPermissionRow.tsx, apps/desktop/src/renderer/components/settings/ComputerUseSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ComputerUseSection.tsx, apps/desktop/src/renderer/components/settings/CustomProviderRuntimeFillOverlay.tsx, apps/desktop/src/renderer/components/settings/DefaultOverrideControls.tsx, apps/desktop/src/renderer/components/settings/DialogueDirectorySection.tsx, apps/desktop/src/renderer/components/settings/DingTalkBotSection.tsx, apps/desktop/src/renderer/components/settings/DiscordBotSection.tsx, apps/desktop/src/renderer/components/settings/DownloadMeter.tsx, apps/desktop/src/renderer/components/settings/ExperimentalSection.tsx, apps/desktop/src/renderer/components/settings/FeishuBotNotificationSection.tsx, apps/desktop/src/renderer/components/settings/FeishuBotSection.tsx, apps/desktop/src/renderer/components/settings/FontFamilyPicker.tsx, apps/desktop/src/renderer/components/settings/GeneralSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/GenericGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/GitSafetySection.settings-search.ts, apps/desktop/src/renderer/components/settings/GitSafetySection.tsx, apps/desktop/src/renderer/components/settings/HarnessVersionMenuRow.tsx, apps/desktop/src/renderer/components/settings/HelpAssistantPanel.tsx, apps/desktop/src/renderer/components/settings/HelpSection.settings-search.ts, apps/desktop/src/renderer/components/settings/HelpSection.tsx, apps/desktop/src/renderer/components/settings/HelpThreadView.tsx, apps/desktop/src/renderer/components/settings/HookConnectionsSection.tsx, apps/desktop/src/renderer/components/settings/HookWorkspacePrefsEditor.tsx, apps/desktop/src/renderer/components/settings/ImBotSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ImBotSection.tsx, apps/desktop/src/renderer/components/settings/ImChannelSettingsCard.tsx, apps/desktop/src/renderer/components/settings/ImConnectionRecovery.tsx, apps/desktop/src/renderer/components/settings/ImDefaultSettingsSection.tsx, apps/desktop/src/renderer/components/settings/ImLifecycleAnnouncementSection.tsx, apps/desktop/src/renderer/components/settings/InputDeviceConnectionStatus.tsx, apps/desktop/src/renderer/components/settings/JoyConGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/KeyboardShortcutsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/KeyboardShortcutsSection.tsx, apps/desktop/src/renderer/components/settings/LanguageSection.tsx, apps/desktop/src/renderer/components/settings/LayoutResetControl.tsx, apps/desktop/src/renderer/components/settings/LinkOpenSection.tsx, apps/desktop/src/renderer/components/settings/LlamaCppCandidates.tsx, apps/desktop/src/renderer/components/settings/LlamaCppProviderDetail.tsx, apps/desktop/src/renderer/components/settings/LocalAgentSection.settings-search.ts, apps/desktop/src/renderer/components/settings/LocalAgentSection.tsx, apps/desktop/src/renderer/components/settings/LocalModelDownloadUI.tsx, apps/desktop/src/renderer/components/settings/LocalOllamaInstall.tsx, apps/desktop/src/renderer/components/settings/LocalPackagingTag.tsx, apps/desktop/src/renderer/components/settings/LogoutSection.tsx, apps/desktop/src/renderer/components/settings/LspBetaCell.tsx, apps/desktop/src/renderer/components/settings/McpServerDialog.tsx, apps/desktop/src/renderer/components/settings/McpServersSection.tsx, apps/desktop/src/renderer/components/settings/MemorySection.tsx, apps/desktop/src/renderer/components/settings/MessageNavRailCell.tsx, apps/desktop/src/renderer/components/settings/ModelAdvancedDrawer.tsx, apps/desktop/src/renderer/components/settings/ModelCompactionSettings.tsx, apps/desktop/src/renderer/components/settings/ModelPriceOverrideDialog.tsx, apps/desktop/src/renderer/components/settings/MyDevicesPanel.tsx, apps/desktop/src/renderer/components/settings/NotificationSection.settings-search.ts, apps/desktop/src/renderer/components/settings/NotificationSection.tsx, apps/desktop/src/renderer/components/settings/OAuthDeviceCodeCard.tsx, apps/desktop/src/renderer/components/settings/OllamaProviderDetail.tsx, apps/desktop/src/renderer/components/settings/OwnRemoteProviders.tsx, apps/desktop/src/renderer/components/settings/PersonalizationDetails.settings-search.ts, apps/desktop/src/renderer/components/settings/PersonalizationSettings.settings-search.ts, apps/desktop/src/renderer/components/settings/PiKernelVersionRow.tsx, apps/desktop/src/renderer/components/settings/PiPackagesSection.tsx, apps/desktop/src/renderer/components/settings/PlayStationGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/ProfileEditDialog.tsx, apps/desktop/src/renderer/components/settings/ProviderAccountUsageModule.tsx, apps/desktop/src/renderer/components/settings/ProviderConnectionDialog.tsx, apps/desktop/src/renderer/components/settings/ProviderImportDialog.tsx, apps/desktop/src/renderer/components/settings/ProvidersSection.settings-search.ts, apps/desktop/src/renderer/components/settings/ProvidersSection.tsx, apps/desktop/src/renderer/components/settings/ReadOnlyProviderModels.tsx, apps/desktop/src/renderer/components/settings/RemoteControlSection.settings-search.ts, apps/desktop/src/renderer/components/settings/RemoteControlSection.tsx, apps/desktop/src/renderer/components/settings/RemoteDesktopPermissions.tsx, apps/desktop/src/renderer/components/settings/RemoteDesktopSetting.tsx, apps/desktop/src/renderer/components/settings/RemoteHostDetail.tsx, apps/desktop/src/renderer/components/settings/RemoteSection.tsx, apps/desktop/src/renderer/components/settings/SessionImportSection.settings-search.ts, apps/desktop/src/renderer/components/settings/SessionImportSection.tsx, apps/desktop/src/renderer/components/settings/SessionRuntimeFallbackCell.tsx, apps/desktop/src/renderer/components/settings/SessionShareImportWizard.tsx, apps/desktop/src/renderer/components/settings/SettingsCatalogPanel.settings-search.ts, apps/desktop/src/renderer/components/settings/SettingsCatalogPanel.tsx, apps/desktop/src/renderer/components/settings/SettingsSearchBox.tsx, apps/desktop/src/renderer/components/settings/SettingsSearchNavigation.ts, apps/desktop/src/renderer/components/settings/SettingsSidebarNav.tsx, apps/desktop/src/renderer/components/settings/SettingsTextInput.tsx, apps/desktop/src/renderer/components/settings/SettingsView.tsx, apps/desktop/src/renderer/components/settings/SharedTasks.settings-search.ts, apps/desktop/src/renderer/components/settings/SilentEncryptedRetryCell.tsx, apps/desktop/src/renderer/components/settings/SpeechModelChoices.tsx, apps/desktop/src/renderer/components/settings/SshKeySetupDialog.tsx, apps/desktop/src/renderer/components/settings/StorageManagementCard.settings-search.ts, apps/desktop/src/renderer/components/settings/StorageManagementCard.tsx, apps/desktop/src/renderer/components/settings/StreamFadeSection.tsx, apps/desktop/src/renderer/components/settings/SubagentModelSection.settings-search.ts, apps/desktop/src/renderer/components/settings/SubagentModelSection.tsx, apps/desktop/src/renderer/components/settings/SwitchProGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/TelegramBehaviorSettings.tsx, apps/desktop/src/renderer/components/settings/TelegramBotSection.tsx, apps/desktop/src/renderer/components/settings/TelegramRemoteDevices.tsx, apps/desktop/src/renderer/components/settings/TerminalShellSection.tsx, apps/desktop/src/renderer/components/settings/TipsSection.settings-search.ts, apps/desktop/src/renderer/components/settings/TipsSection.tsx, apps/desktop/src/renderer/components/settings/UnifiedModelList.tsx, apps/desktop/src/renderer/components/settings/UserProfileCard.tsx, apps/desktop/src/renderer/components/settings/UserPromptSection.tsx, apps/desktop/src/renderer/components/settings/VisionBridgeSection.settings-search.ts, apps/desktop/src/renderer/components/settings/VisionBridgeSection.tsx, apps/desktop/src/renderer/components/settings/VoiceInputSection.settings-search.ts, apps/desktop/src/renderer/components/settings/VoiceInputSection.tsx, apps/desktop/src/renderer/components/settings/VoiceSpeakerEnrollmentDialog.tsx, apps/desktop/src/renderer/components/settings/VoiceSpeakerFilterCard.tsx, apps/desktop/src/renderer/components/settings/WallpaperSection.tsx, apps/desktop/src/renderer/components/settings/WechatBotSection.tsx, apps/desktop/src/renderer/components/settings/WecomBotSection.tsx, apps/desktop/src/renderer/components/settings/WindowBehaviorSection.settings-search.ts, apps/desktop/src/renderer/components/settings/WindowBehaviorSection.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexKeyboardLayout.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexKeycapGlyphs.tsx, apps/desktop/src/renderer/components/settings/WorkLouderCodexSettings.tsx, apps/desktop/src/renderer/components/settings/WorktreeRecycleCard.settings-search.ts, apps/desktop/src/renderer/components/settings/WorktreeRecycleCard.tsx, apps/desktop/src/renderer/components/settings/XUsageGuide.tsx, apps/desktop/src/renderer/components/settings/XboxGamepadLayout.tsx, apps/desktop/src/renderer/components/settings/XboxGamepadSettings.tsx, apps/desktop/src/renderer/components/settings/androidStatusPresentation.ts, apps/desktop/src/renderer/components/settings/avatarCropGeometry.ts, apps/desktop/src/renderer/components/settings/avatarCropImage.ts, apps/desktop/src/renderer/components/settings/billingVisibility.ts, apps/desktop/src/renderer/components/settings/browserOpenForLoginError.ts, apps/desktop/src/renderer/components/settings/cindyMakeVisibility.ts, apps/desktop/src/renderer/components/settings/computerPermissionFlow.ts, apps/desktop/src/renderer/components/settings/contacts/ContactDetailPane.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsImportDialog.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsListPane.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsManagerDialog.tsx, apps/desktop/src/renderer/components/settings/contacts/ContactsSection.tsx, apps/desktop/src/renderer/components/settings/contacts/startContactsAiSession.ts, apps/desktop/src/renderer/components/settings/dialogueDirectoryDisplay.ts, apps/desktop/src/renderer/components/settings/dualSenseSilhouette.ts, apps/desktop/src/renderer/components/settings/feishuBotPresentation.ts, apps/desktop/src/renderer/components/settings/fontFamilyValue.ts, apps/desktop/src/renderer/components/settings/gamepadLayoutPrimitives.tsx, apps/desktop/src/renderer/components/settings/gamepadSilhouetteGeom.ts, apps/desktop/src/renderer/components/settings/hookWorkspacePrefsLogic.ts, apps/desktop/src/renderer/components/settings/imBotVisibility.ts, apps/desktop/src/renderer/components/settings/imDefaultSettingsLogic.ts, apps/desktop/src/renderer/components/settings/joyConSilhouette.ts, apps/desktop/src/renderer/components/settings/modelManagementPresentation.ts, apps/desktop/src/renderer/components/settings/modelManagementState.ts, apps/desktop/src/renderer/components/settings/myDevicesModel.ts, apps/desktop/src/renderer/components/settings/providerAssetModule.ts, apps/desktop/src/renderer/components/settings/realProfilePermissionGuide.ts, apps/desktop/src/renderer/components/settings/settingsRowClasses.ts, apps/desktop/src/renderer/components/settings/settingsSearchCatalog.ts, apps/desktop/src/renderer/components/settings/settingsSearchTypes.ts, apps/desktop/src/renderer/components/settings/switchProSilhouette.ts, apps/desktop/src/renderer/components/settings/ultimateC1Silhouette.ts, apps/desktop/src/renderer/components/settings/usage/UsageBarsTooltip.tsx, apps/desktop/src/renderer/components/settings/usage/UsageBreakdownTables.tsx, apps/desktop/src/renderer/components/settings/usage/UsageDeviceSelect.tsx, apps/desktop/src/renderer/components/settings/usage/UsageFilterSelect.tsx, apps/desktop/src/renderer/components/settings/usage/UsageHistorySection.settings-search.ts, apps/desktop/src/renderer/components/settings/usage/UsageHistorySection.tsx, apps/desktop/src/renderer/components/settings/usage/UsageStatRow.tsx, apps/desktop/src/renderer/components/settings/usage/UsageTaskTable.tsx, apps/desktop/src/renderer/components/settings/usage/UsageTokenBars.tsx, apps/desktop/src/renderer/components/settings/usage/formatUsagePercent.ts, apps/desktop/src/renderer/components/settings/usage/usageCharts.css, apps/desktop/src/renderer/components/settings/usage/usageHistoryColors.ts, apps/desktop/src/renderer/components/settings/usage/usageHistoryStats.ts, apps/desktop/src/renderer/components/settings/usageVisibility.ts, apps/desktop/src/renderer/components/settings/useAutomationSettingsRead.ts, apps/desktop/src/renderer/components/settings/useProviderSubscriptionCard.ts, apps/desktop/src/renderer/components/settings/useSettingsEscape.ts, apps/desktop/src/renderer/components/settings/useSettingsSectionTarget.ts, apps/desktop/src/renderer/components/settings/wizardRecommend.ts, apps/desktop/src/renderer/components/settings/workLouderCodexCommandCopy.ts, apps/desktop/src/renderer/components/settings/xboxSeriesSilhouette.ts, apps/desktop/src/renderer/styles/sortable.css | 137 | 95 | 554 |
 | `desktop.shell.main-layout` | desktop | 主窗口壳（标题栏 / 左右栏 / 内容区） | main BrowserWindow → renderer/index.tsx → main-entry.tsx → App → MainLayout (`#/` 受保护壳) | ChromeActions, DeviceSectionHeader, FeishuConflictDialogHost, GhostMediaLightboxHost, MainLayout, ProjectsSection, RightSidebar, RightSidebarShell, SessionShareImportWizard, Sidebar, UpdateNoticeDialog, WindowControls | apps/desktop/src/renderer/cindy-brain/GhostMediaLightboxHost.tsx, apps/desktop/src/renderer/components/UpdateNoticeDialog.tsx, apps/desktop/src/renderer/components/feishuBot/FeishuConflictDialog.tsx, apps/desktop/src/renderer/components/feishuBot/FeishuConflictDialogHost.tsx, apps/desktop/src/renderer/components/layout/AppBadgeAttentionSync.tsx, apps/desktop/src/renderer/components/layout/BrowserWebviewPool.tsx, apps/desktop/src/renderer/components/layout/ChromeActions.tsx, apps/desktop/src/renderer/components/layout/ContentHeader.tsx, apps/desktop/src/renderer/components/layout/CredentialStoreBanner.tsx, apps/desktop/src/renderer/components/layout/FadeSwitcher.tsx, apps/desktop/src/renderer/components/layout/GlobalDropImportListener.tsx, apps/desktop/src/renderer/components/layout/MainEntryRedirect.tsx, apps/desktop/src/renderer/components/layout/MainLayout.tsx, apps/desktop/src/renderer/components/layout/RightSidebar.tsx, apps/desktop/src/renderer/components/layout/RightSidebarDetach.tsx, apps/desktop/src/renderer/components/layout/RightSidebarMaximize.tsx, apps/desktop/src/renderer/components/layout/RightSidebarToggle.tsx, apps/desktop/src/renderer/components/layout/SidebarTabWindowLayout.tsx, apps/desktop/src/renderer/components/layout/SidebarWindowLayout.tsx, apps/desktop/src/renderer/components/layout/WallpaperVideo.tsx, apps/desktop/src/renderer/components/layout/chromeActionsGeometry.ts, apps/desktop/src/renderer/components/layout/chromeActionsLayout.ts, apps/desktop/src/renderer/components/layout/railChromeActions.ts, apps/desktop/src/renderer/components/layout/windowDrag.tsx, apps/desktop/src/renderer/components/settings/SessionShareImportWizard.tsx, apps/desktop/src/renderer/components/sidebar/AccountSwitcherDialog.tsx, apps/desktop/src/renderer/components/sidebar/AttentionDot.tsx, apps/desktop/src/renderer/components/sidebar/DatabaseSizeWarningBanner.tsx, apps/desktop/src/renderer/components/sidebar/GhostMainViewNavEntries.tsx, apps/desktop/src/renderer/components/sidebar/MobileDownloadDialog.tsx, apps/desktop/src/renderer/components/sidebar/NavigationCountBadge.tsx, apps/desktop/src/renderer/components/sidebar/NewTag.tsx, apps/desktop/src/renderer/components/sidebar/Sidebar.tsx, apps/desktop/src/renderer/components/sidebar/SidebarIconButton.tsx, apps/desktop/src/renderer/components/sidebar/SidebarNavigationCustomize.tsx, apps/desktop/src/renderer/components/sidebar/SidebarTopNav.tsx, apps/desktop/src/renderer/components/sidebar/SortableList.tsx, apps/desktop/src/renderer/components/sidebar/UpdateBanner.tsx, apps/desktop/src/renderer/components/sidebar/UserInfoSection.tsx, apps/desktop/src/renderer/components/sidebar/VendorIcon.tsx, apps/desktop/src/renderer/components/sidebar/VendorReadinessBadge.tsx, apps/desktop/src/renderer/components/sidebar/WorktreeBadge.tsx, apps/desktop/src/renderer/components/sidebar/sidebarNavigationPrefs.ts, apps/desktop/src/renderer/components/sidebar/useHoverOpenMenu.ts, apps/desktop/src/renderer/components/title-bar/ChromeIconButton.tsx, apps/desktop/src/renderer/components/title-bar/MenuButton.tsx, apps/desktop/src/renderer/components/title-bar/WindowControls.tsx, apps/desktop/src/renderer/features/cc-agent/sidebar/DeviceSectionHeader.tsx, apps/desktop/src/renderer/features/cc-agent/sidebar/sections/ProjectsSection.tsx, apps/desktop/src/renderer/features/right-sidebar/AddTabDropdown.tsx, apps/desktop/src/renderer/features/right-sidebar/EmptyState.tsx, apps/desktop/src/renderer/features/right-sidebar/RightSidebarShell.tsx, apps/desktop/src/renderer/features/right-sidebar/TabBar.tsx, apps/desktop/src/renderer/features/right-sidebar/TabBodyErrorBoundary.tsx, apps/desktop/src/renderer/features/right-sidebar/TabDragController.tsx, apps/desktop/src/renderer/features/right-sidebar/TabGroupPanel.tsx, apps/desktop/src/renderer/features/right-sidebar/botMenuOwnerContext.ts, apps/desktop/src/renderer/features/right-sidebar/globalTabs.ts, apps/desktop/src/renderer/features/right-sidebar/hooks/useBrowserWebview.ts, apps/desktop/src/renderer/features/right-sidebar/hooks/useNativePopupSurface.ts, apps/desktop/src/renderer/features/right-sidebar/hooks/useRetainedBrowserTabHosts.ts, apps/desktop/src/renderer/features/right-sidebar/lib/browserPartition.ts, apps/desktop/src/renderer/features/right-sidebar/lib/browserWebviewPool.ts, apps/desktop/src/renderer/features/right-sidebar/lib/detachedSidebarRouting.ts, apps/desktop/src/renderer/features/right-sidebar/lib/executeSidebarCommand.ts, apps/desktop/src/renderer/features/right-sidebar/lib/lastTurnChangedFiles.ts, apps/desktop/src/renderer/features/right-sidebar/lib/mailPreview.test.ts, apps/desktop/src/renderer/features/right-sidebar/lib/mailPreview.ts, apps/desktop/src/renderer/features/right-sidebar/lib/mergeSidebarWindowBack.ts, apps/desktop/src/renderer/features/right-sidebar/lib/nativePopupTabs.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openBackgroundTasksTab.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openInSidebarBrowser.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openInSidebarFileBrowser.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openPluginTaskTab.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openRoutinesTab.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openSideChatTab.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openSubagentsTab.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openTerminalShortcut.ts, apps/desktop/src/renderer/features/right-sidebar/lib/openTurnReview.ts, apps/desktop/src/renderer/features/right-sidebar/lib/popupRouter.ts, apps/desktop/src/renderer/features/right-sidebar/lib/popupTabs.ts, apps/desktop/src/renderer/features/right-sidebar/lib/rsbBrowserBridge.ts, apps/desktop/src/renderer/features/right-sidebar/lib/sidebarCommands.ts, apps/desktop/src/renderer/features/right-sidebar/lib/sidebarHostSession.tsx, apps/desktop/src/renderer/features/right-sidebar/lib/tabHosting.ts, apps/desktop/src/renderer/features/right-sidebar/lib/useTabCreationMenu.ts, apps/desktop/src/renderer/features/right-sidebar/moveTab.ts, apps/desktop/src/renderer/features/right-sidebar/pluginPlacement.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/BackgroundTasksBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/WorkflowAgentStrip.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/WorkflowProgressTree.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/chatTaskFocusIntent.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/listSessionTasks.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/background-tasks/workflowProgressModel.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/cindy-make/CindyMakeTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/cindy-make/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/FileBrowserBody.css, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/FileBrowserBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/dropExternalFile.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/fileTreeImagePreview.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/file-browser/useSessionScopedTreeWidth.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/index.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/mail-preview/MailCardBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/mail-preview/MailConversationView.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/mail-preview/MailParticipant.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/mail-preview/MailPreview.test.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/mail-preview/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/orca-workers/OrcaWorkersAttentionIcon.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/orca-workers/OrcaWorkersTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/orca-workers/actions.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/orca-workers/closeDecision.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/orca-workers/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/plugin-task/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/plugin-view/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/resource-usage/ResourceUsageBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/resource-usage/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/resource-usage/resource-usage.css, apps/desktop/src/renderer/features/right-sidebar/plugins/resource-usage/subscription.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/retired-feature/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/ImageDiffPreview.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/MarkdownDiffPreview.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/PlainUnifiedDiff.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/diffRows.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/highlight.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/inlineDiff.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/DiffViewer/useDiffHighlights.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/ReviewTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/diffExpansionPreference.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/fileTree.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/gitApplyCommand.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/review/markdownPreview.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/useLastTurnFilter.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/useReviewGitState.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/review/useReviewSource.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/routines/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/side-chat/SideChatTabBody.test.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/side-chat/SideChatTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/side-chat/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/side-chat/state.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/ConversationStream.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/DetailView.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/RunList.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/SubagentChrome.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/SubagentToolCard.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/SubagentsBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/subagentChangeFence.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/subagentConversation.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/subagents/subagentFormat.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/terminal/TerminalTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/terminal/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/terminal/lib/terminalFont.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/terminal/lib/useTerminalFontSettings.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/terminal/lib/xtermPool.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/BrowserChrome.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/BrowserCommentPopover.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/BrowserTabBody.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/browserCommentEditorDraft.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/index.tsx, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/lib/parseOmnibox.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/useBrowserComment.ts, apps/desktop/src/renderer/features/right-sidebar/plugins/web-browser/useLocalHtmlAutoReload.ts, apps/desktop/src/renderer/features/right-sidebar/registry.ts, apps/desktop/src/renderer/features/right-sidebar/store.ts, apps/desktop/src/renderer/features/right-sidebar/tabAvailability.ts, apps/desktop/src/renderer/features/right-sidebar/tabGroupHost.tsx, apps/desktop/src/renderer/features/right-sidebar/tabGroupPanes.ts, apps/desktop/src/renderer/features/right-sidebar/tabGroups.ts, apps/desktop/src/renderer/features/right-sidebar/types.ts, apps/desktop/src/renderer/layout/LayoutRoot.tsx, apps/desktop/src/renderer/layout/PanelDragController.tsx, apps/desktop/src/renderer/layout/PanelDragPrototype.tsx, apps/desktop/src/renderer/layout/collapsePrefs.ts, apps/desktop/src/renderer/layout/layoutDevTools.ts, apps/desktop/src/renderer/layout/panePlacement.tsx, apps/desktop/src/renderer/layout/paneWidths.tsx, apps/desktop/src/renderer/layout/panelMaximize.tsx, apps/desktop/src/renderer/styles/generated/tokens.css, apps/desktop/src/renderer/styles/globals.css | 147 | 46 | 199 |
 | `desktop.skillhub.local` | desktop | SkillHub 本地技能 | hash `/skillhub/local` 及统一详情 `/skillhub/detail`（旧详情链接重定向） | InstallTargetPicker, PluginDevicePicker, PluginManagementLayout, RemoteSkillsPage, SkillCategoryFilterBar, SkillDetailLayout, SkillhubDetailRoute, SkillhubDetailView, SkillhubFeatureLayout, SkillhubHomeView, SkillhubLocalLayout, SkillhubMarketDetailView | apps/desktop/src/renderer/features/plugin/PluginDevicePicker.tsx, apps/desktop/src/renderer/features/plugin/PluginManagementLayout.tsx, apps/desktop/src/renderer/features/skillhub/RemoteSkillsPage.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubDetailRoute.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubDetailView.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubFeatureLayout.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubHomeView.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubLocalLayout.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubMarketDetailView.tsx, apps/desktop/src/renderer/features/skillhub/components/InstallTargetPicker.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillCategoryFilterBar.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillDetailLayout.tsx, apps/desktop/src/renderer/features/skillhub/skillhub-home.css | 39 | 0 | 45 |
 | `desktop.skillhub.market` | desktop | SkillHub 市场 | hash `/skillhub/market`（SkillhubMarketListView） | InstallTargetPicker, MarketCard, MarketInfoEditDialog, SkillhubMarketDetailView, SkillhubMarketListView, VisibilityEditorDialog | apps/desktop/src/renderer/features/skillhub/SkillhubMarketDetailView.tsx, apps/desktop/src/renderer/features/skillhub/SkillhubMarketListView.tsx, apps/desktop/src/renderer/features/skillhub/components/HomeMarketCard.tsx, apps/desktop/src/renderer/features/skillhub/components/InstallTargetPicker.tsx, apps/desktop/src/renderer/features/skillhub/components/LocalSkillControls.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketCard.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketInfoEditDialog.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketInstallStatus.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketListError.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketLocalSkills.tsx, apps/desktop/src/renderer/features/skillhub/components/MarketPreviewTree.tsx, apps/desktop/src/renderer/features/skillhub/components/OfficialSkillBadge.tsx, apps/desktop/src/renderer/features/skillhub/components/PlatformTagSelector.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillCategoryFilterBar.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillDetailLayout.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillIcon.tsx, apps/desktop/src/renderer/features/skillhub/components/SkillTagList.tsx, apps/desktop/src/renderer/features/skillhub/components/TeamScopePicker.tsx, apps/desktop/src/renderer/features/skillhub/components/VisibilityEditorDialog.tsx | 42 | 0 | 35 |
diff --git a/packages/maker-core/src/agents/base-agent.ts b/packages/maker-core/src/agents/base-agent.ts
index 99811e03238..82b75f20298 100644
--- a/packages/maker-core/src/agents/base-agent.ts
+++ b/packages/maker-core/src/agents/base-agent.ts
@@ -876,6 +876,42 @@ export interface AgentDeps {
     },
   ) => Promise<PiNativeProvidersResult | null>;
 
+  /**
+   * Pi-only:读**活动目录**(含本机 override 合并)里某模型的图片输入能力声明。
+   * `undefined` = 目录未声明。
+   * 只用于旧会话切模时的能力对账(见 pi/index.ts refreshImageCapabilityOnSwitch),
+   * 不参与准入 —— 准入仍由会话快照的 input 与 assertImageInputSupported 负责。
+   */
+  readModelImageInput?: (providerId: string, modelId: string) => boolean | undefined;
+
+  /**
+   * Pi-only:该模型在**活动目录**里最终的图片输入能力(目录与本机 override 合并后的结论)。
+   * 与 readModelImageInput 的区别:后者只回报“显式声明了什么”(未声明即 undefined)，而本项
+   * 回报“最终会怎样”——未声明时按上游 #4854 的语义落到默认支持(见 resolvePiNativeProviders
+   * 的 `supportsImageInput === undefined ? ['text','image'] : ...`)。
+   *
+   * 为什么需要它:能力对账要在“用户把显式声明改回跟随供应商(override 被删、声明读回
+   * undefined)”时，把旧会话快照从纯文本恢复成支持。但这种状态与“目录自己就把该行标成
+   * input:['text']”在 readModelImageInput 看来完全一样(都是 undefined)，快照自身又已被
+   * 上一轮对账改成了纯文本，两者无法区分。只有活动目录的最终结论能判定。
+   *
+   * 拿不到(undefined)时调用方必须保持现状、不翻转快照。
+   */
+  readModelCatalogImageCapability?: (providerId: string, modelId: string) => boolean | undefined;
+
+  /**
+   * Pi-only:该模型在**活动目录**里最终的思考档位集合(目录与本机 override 合并后的结论)。
+   *
+   * 为什么需要它:`activeEffortSnapshot` 是 startSession 一次性解析的,而 Pi 的 set_model
+   * 不重读 models.json —— 用户在设置里给当前模型声明档位后,活着的子进程仍是启动时的
+   * 零档位快照,于是 thinking 通道开不出来(思考继续漏进正文),必须重开会话才生效。
+   * 与 readModelCatalogImageCapability 同一形状:切模是用户可见的自然同步点,在这里
+   * 重新读一次活动目录即可对齐,无需重建子进程。
+   *
+   * 返回 undefined = 拿不到结论,调用方保持现状不猜。
+   */
+  readModelCatalogThinkingTiers?: (providerId: string, modelId: string) => readonly string[] | undefined;
+
   /**
    * Pi-only:按实际 provider/model 路由解析运行时描述符。用于启动前校验已持久化 effort，
    * 以及恢复已 retired 模型时补齐当前 session 的私有 models.json；结果不得进入公开
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
index e67b608d3fa..c4d32785475 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
@@ -6399,6 +6399,1129 @@ describe("Pi provider-aware model routing", () => {
     await handle.close();
   });
 
+  it("picks up an image-capability declaration made after session start when the user switches models", async () => {
+    // 会话启动时目录未声明图片能力(快照 input=['text'])，带图消息被客户端门拒收；用户随后在
+    // 设置里声明(本机目录 override，只落在活动目录)。旧会话的快照与 models.json 都还是启动时
+    // 那份，所以要借切模这个同步点：同路由 setModel(心跳走的也是这条)先做能力对账，不一致就
+    // 热写 models.json 并 switch_session 让子进程 ModelConfig.load 重载。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-declared",
+              name: "Native Declared",
+              baseUrl: "http://declared.test",
+              api: "openai-completions",
+              models: [{ id: "local-model", input: ["text"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      // 活动目录(含本机 override)的声明：未声明 undefined / 声明后 true。
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-declared" && modelId === "local-model" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-declared-after-start",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-declared",
+    });
+    const imagePath = path.join(cwd, "declared.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    const imageMessage = {
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/declared.png",
+        },
+      ],
+    };
+    const readNativeInput = (): string[] => {
+      const modelsJson = JSON.parse(
+        readFileSync(
+          path.join(captured.env.PI_CODING_AGENT_DIR as string, "models.json"),
+          "utf8",
+        ),
+      ) as {
+        providers: Record<string, { models: Array<{ id: string; input: string[] }> }>;
+      };
+      const block =
+        modelsJson.providers["native-declared"] ??
+        modelsJson.providers["cindy-byom-native-declared"];
+      return block!.models.find((entry) => entry.id === "local-model")!.input;
+    };
+
+    // 基线：声明前客户端门拒收，请求根本不出去，models.json 也没有 image。
+    captured.requests.length = 0;
+    await expect(handle.send(imageMessage)).rejects.toMatchObject({
+      code: "PI_IMAGE_INPUT_UNSUPPORTED",
+    });
+    expect(captured.requests).toHaveLength(0);
+    expect(readNativeInput()).toEqual(["text"]);
+
+    // 用户在设置里声明「支持图片输入」(目录变化，快照不变)。
+    declared.image = true;
+
+    // 切模(目标 = 当前同路由，与心跳同路径)触发对账：热写 + 子进程重载。
+    captured.requests.length = 0;
+    await handle.setModel!("local-model", { providerId: "native-declared" });
+    expect(readNativeInput()).toEqual(["text", "image"]);
+    expect(captured.requests.map((request) => request.type)).toContain("switch_session");
+
+    // 同一条消息现在放行。
+    captured.requests.length = 0;
+    await handle.send(imageMessage);
+    expect(captured.requests).toContainEqual(
+      expect.objectContaining({
+        type: "prompt",
+        images: [expect.objectContaining({ type: "image" })],
+      }),
+    );
+    await handle.close();
+  });
+
+  it("keeps the session context budget in settings.json across an image capability reload", async () => {
+    // 图片声明重写 settings.json 时必须带上 sessionContextWindowBudget：漏传会把逐模型
+    // compaction.modelOverrides 按模型自身窗口重算（fallback 16_384），任务级预算静默失效。
+    const declared = { image: false };
+    const deps = byomDeps(
+      async () => ({
+        providers: [
+          {
+            id: "native-budget",
+            name: "Native Budget",
+            baseUrl: "http://budget.test",
+            api: "openai-completions",
+            models: [{ id: "model-a", input: ["text"], contextWindow: 200_000 }],
+          },
+        ],
+        env: {},
+      }),
+    );
+    deps.runtimeConfig = { ...deps.runtimeConfig, piAutoCompactThresholdPct: 75 };
+    const agent = new PiAgent({
+      ...deps,
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-budget" && modelId === "model-a" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-reload-budget",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-budget",
+      contextWindowBudget: 50_000,
+    });
+    const settingsPath = path.join(captured.env.PI_CODING_AGENT_DIR!, "settings.json");
+    const readReserve = () =>
+      (JSON.parse(readFileSync(settingsPath, "utf8")) as {
+        compaction?: { modelOverrides?: Record<string, { reserveTokens?: number }> };
+      }).compaction?.modelOverrides?.["native-budget/model-a"]?.reserveTokens;
+    // 预算口径：200_000 - 50_000 × 75% = 162_500。
+    expect(readReserve()).toBe(162_500);
+
+    // 图片声明变更触发的对账重写 settings.json，预算不能被冲掉。
+    declared.image = true;
+    await handle.setModel!("model-a", { providerId: "native-budget" });
+    expect(captured.runtimeModel).toBe("model-a");
+    expect(readReserve()).toBe(162_500);
+    await handle.close();
+  });
+
+  it("reapplies the session thinking level after an image capability reload reissues set_model", async () => {
+    // reload 的 switch_session/set_model 会把思考档位重置成模型默认（pi CLI 行为），而重载
+    // 路径不经过 switchModel 末尾的档位恢复 —— 发图触发的对账、同路由切模触发的对账早退后，
+    // 子进程实际跑在默认档而 UI 仍显示原档位。重载确认路由后必须把会话当前档位重发一遍。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-reload-tier",
+              name: "Native Reload Tier",
+              baseUrl: "http://reload-tier.test",
+              api: "openai-completions",
+              models: [
+                {
+                  id: "model-a",
+                  input: ["text"],
+                  reasoning: true,
+                  thinkingLevelMap: { low: "low", max: "max" },
+                },
+                {
+                  id: "model-b",
+                  input: ["text"],
+                  reasoning: true,
+                  thinkingLevelMap: { low: "low", max: "max" },
+                },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-reload-tier" && modelId === "model-b" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-reload-thinking-restore",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-reload-tier",
+      effort: "max",
+    });
+    await handle.setModel!("model-b", { providerId: "native-reload-tier", effort: "max" });
+    expect(captured.runtimeModel).toBe("model-b");
+
+    // 用户此刻声明图片能力，同路由 setModel 触发对账：switch_session 重建会话 + 重发
+    // set_model（档位被重置）→ 修复后必须跟着一条 set_thinking_level 把会话档位找回来。
+    declared.image = true;
+    captured.requests.length = 0;
+    await handle.setModel!("model-b", { providerId: "native-reload-tier" });
+
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("switch_session");
+    expect(requestTypes.indexOf("set_model")).toBeGreaterThan(
+      requestTypes.indexOf("switch_session"),
+    );
+    expect(requestTypes.indexOf("set_thinking_level")).toBeGreaterThan(
+      requestTypes.indexOf("set_model"),
+    );
+    expect(
+      captured.requests.find((request) => request.type === "set_thinking_level"),
+    ).toEqual({ type: "set_thinking_level", level: "max" });
+    await handle.close();
+  });
+
+  it("restores the confirmed thinking-off state after an image capability reload", async () => {
+    // 关闭开关只发 off、不更新 mutableEffort；reload 恢复若只看 mutableEffort，就会把用户
+    // 关掉的思考悄悄重开（等待时间与费用都涨）。恢复时优先还原已确认的关闭状态。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-off-restore",
+              name: "Native Off Restore",
+              baseUrl: "http://off-restore.test",
+              api: "openai-completions",
+              models: [
+                {
+                  id: "local-model",
+                  input: ["text"],
+                  reasoning: true,
+                  thinkingLevelMap: { low: "low", max: "max" },
+                },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-off-restore" && modelId === "local-model" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-reload-thinking-off",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-off-restore",
+      effort: "max",
+    });
+    // 用户关掉思考；随后声明图片能力，直接发图触发对账重载。
+    await handle.setThinkingEnabled!(false);
+    declared.image = true;
+    const imagePath = path.join(cwd, "off-restore.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    captured.requests.length = 0;
+    await handle.send({
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/off-restore.png",
+        },
+      ],
+    });
+
+    // 重载发生了（switch_session），之后的恢复必须是 off（而不是 mutableEffort 的 max）。
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("switch_session");
+    const restores = captured.requests.filter(
+      (request) => request.type === "set_thinking_level",
+    );
+    expect(restores).toEqual([{ type: "set_thinking_level", level: "off" }]);
+    await handle.close();
+  });
+
+  it("restores the applied effort level, not the stale off state, when thinking was re-enabled via setEffort", async () => {
+    // 关闭思考后用 setEffort 应用新档位：应用档位本身就意味着思考开着，关闭记录必须被
+    // 覆盖。否则 reload 恢复会把刚应用的档位用 off 关掉，后续回答不再按用户选的档位跑。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-re-enable",
+              name: "Native Re-enable",
+              baseUrl: "http://re-enable.test",
+              api: "openai-completions",
+              models: [
+                {
+                  id: "local-model",
+                  input: ["text"],
+                  reasoning: true,
+                  thinkingLevelMap: { low: "low", max: "max" },
+                },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-re-enable" && modelId === "local-model" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-reload-re-enable",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-re-enable",
+      effort: "max",
+    });
+    await handle.setThinkingEnabled!(false);
+    await handle.setEffort!("max");
+    declared.image = true;
+    const imagePath = path.join(cwd, "re-enable.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    captured.requests.length = 0;
+    await handle.send({
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/re-enable.png",
+        },
+      ],
+    });
+
+    // 重载后的恢复必须是刚应用的 max，而不是过期的关闭状态。
+    expect(captured.requests.map((request) => request.type)).toContain("switch_session");
+    expect(
+      captured.requests.filter((request) => request.type === "set_thinking_level"),
+    ).toEqual([{ type: "set_thinking_level", level: "max" }]);
+    await handle.close();
+  });
+
+  it("does not rewrite models.json or reload the runtime when the catalog already matches the snapshot", async () => {
+    // 心跳会反复以同路由下发 setModel；声明与会话快照一致时必须零 I/O —— 不热写、不 switch_session。
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-vision",
+              name: "Native Vision",
+              baseUrl: "http://vision.test",
+              api: "openai-completions",
+              models: [{ id: "local-model", input: ["text", "image"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: () => true,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-already-matches",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-vision",
+    });
+    const modelsPath = path.join(
+      captured.env.PI_CODING_AGENT_DIR as string,
+      "models.json",
+    );
+    const before = readFileSync(modelsPath, "utf8");
+    captured.requests.length = 0;
+    await handle.setModel!("local-model", { providerId: "native-vision" });
+    expect(readFileSync(modelsPath, "utf8")).toBe(before);
+    expect(captured.requests.map((request) => request.type)).not.toContain("switch_session");
+    await handle.close();
+  });
+
+  it("refreshes on the first image send when a capability declaration appears after session start without a model switch", async () => {
+    // 声明可能发生在会话启动之后、且用户不切模直接发图：客户端门在拒收前就地补一次能力
+    // 对账(与切模路径同一函数、同一失败语义)，补上即放行；盘上 models.json 同步热写、
+    // 子进程 switch_session 重载。未声明时仍拒收(见另一用例的基线)，且对账零副作用。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-send-refresh",
+              name: "Native Send Refresh",
+              baseUrl: "http://send-refresh.test",
+              api: "openai-completions",
+              models: [{ id: "local-model", input: ["text"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: () => (declared.image ? true : undefined),
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-declared-send-only",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-send-refresh",
+    });
+    const imagePath = path.join(cwd, "send-refresh.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    const imageMessage = {
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/send-refresh.png",
+        },
+      ],
+    };
+    const modelsPath = path.join(
+      captured.env.PI_CODING_AGENT_DIR as string,
+      "models.json",
+    );
+    const readSendRefreshInput = (): string[] => {
+      const modelsJson = JSON.parse(readFileSync(modelsPath, "utf8")) as {
+        providers: Record<string, { models: Array<{ id: string; input: string[] }> }>;
+      };
+      const block =
+        modelsJson.providers["native-send-refresh"] ??
+        modelsJson.providers["cindy-byom-native-send-refresh"];
+      return block!.models.find((entry) => entry.id === "local-model")!.input;
+    };
+    expect(readSendRefreshInput()).toEqual(["text"]);
+
+    // 会话启动后才声明；不切模，直接发图。
+    declared.image = true;
+    captured.requests.length = 0;
+    await handle.send(imageMessage);
+
+    // 对账生效：盘上热写 + 子进程重载 + prompt 带图放行。
+    expect(readSendRefreshInput()).toEqual(["text", "image"]);
+    expect(captured.requests.map((request) => request.type)).toContain("switch_session");
+    expect(captured.requests).toContainEqual(
+      expect.objectContaining({
+        type: "prompt",
+        images: [expect.objectContaining({ type: "image" })],
+      }),
+    );
+    await handle.close();
+  });
+
+  it("keeps the session route when a late capability declaration is reconciled on a same-route model switch", async () => {
+    // switch_session 按进程启动时的 --provider/--model 重建 AgentSession：对账里如果只
+    // switch_session 而不重放当前路由，会话内切过模的会话会被静默拽回启动路由 —— 同路由
+    // no-op 与发图都不会再打 set_model，提示词发给错误的模型且没有任何报错。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-route-keep",
+              name: "Native Route Keep",
+              baseUrl: "http://route-keep.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"] },
+                { id: "model-b", input: ["text"] },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-route-keep" && modelId === "model-b" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-route-keep",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-route-keep",
+    });
+    await handle.setModel!("model-b", { providerId: "native-route-keep" });
+    expect(captured.runtimeModel).toBe("model-b");
+
+    // 会话里已切到 model-b；用户此刻声明图片能力，下一次心跳以同路由 setModel(model-b)
+    // 触发对账（与 no-op 分支同路径）。
+    declared.image = true;
+    captured.requests.length = 0;
+    await handle.setModel!("model-b", { providerId: "native-route-keep" });
+
+    // 对账生效，且子进程仍在 model-b（而不是被 switch_session 拽回 model-a）。
+    expect(captured.runtimeModel).toBe("model-b");
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("switch_session");
+    expect(requestTypes.indexOf("set_model")).toBeGreaterThan(
+      requestTypes.indexOf("switch_session"),
+    );
+    await handle.close();
+  });
+
+  it("realigns the thinking tier snapshot when the user declares tiers on the current model", async () => {
+    // 用户在设置里给**当前**模型声明思考档位后，活着的子进程仍是启动时的零档位快照：
+    // activeEffortSnapshot 是 startSession 一次性解析的，Pi 的 set_model 不重读 models.json。
+    // 切模（含同路由 no-op）是用户可见的自然同步点，必须在这里对齐，否则 thinking 通道
+    // 永远开不出来、推理继续漏进正文，只有重开会话才生效。
+    const declared = { tiers: [] as string[] | undefined };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "tier-realign",
+              name: "Tier Realign",
+              baseUrl: "http://tier-realign.test",
+              api: "openai-completions",
+              // 目录声明支持思考但只给 high 一档：启动快照非空，便于观察对齐行为。
+              models: [{ id: "model-a", input: ["text"], thinkingLevelMap: { high: "high" } }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelCatalogThinkingTiers: (providerId, modelId) =>
+        providerId === "tier-realign" && modelId === "model-a" ? declared.tiers : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "thinking-tier-realign",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "tier-realign",
+    });
+
+    // 用户声明了更多档位 → 同路由 setModel(no-op) 必须把快照对齐过去。
+    declared.tiers = ["low", "high", "max"];
+    captured.requests.length = 0;
+    await handle.setModel!("model-a", { providerId: "tier-realign" });
+
+    // 快照对齐后 setEffort 不再被「零档位快照」挡住：会真的发出 set_thinking_level。
+    await handle.setEffort!("max");
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("set_thinking_level");
+    await handle.close();
+  });
+
+  it("preflights a freshly declared effort tier against the catalog, not the frozen startup snapshot", async () => {
+    // 用户在设置里给 model-b 声明了新档位 max，随即带 effort=max 切模。启动时冻结的
+    // nextEffortSnapshot 里没有 max，若预检用它就会在 RPC 发出前误判「不可用」→ 切模失败。
+    // 预检必须用对账读到的目录结论；活动快照仍只在切模成功后落定。
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "tier-preflight",
+              name: "Tier Preflight",
+              baseUrl: "http://tier-preflight.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"], thinkingLevelMap: { high: "high" } },
+                { id: "model-b", input: ["text"], thinkingLevelMap: { low: "low" } },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      // model-b 的目录最终档位含 max（用户刚声明的），而随包目录只给了 low。
+      readModelCatalogThinkingTiers: (providerId, modelId) =>
+        providerId === "tier-preflight" && modelId === "model-b"
+          ? ["low", "medium", "high", "max"]
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "thinking-tier-preflight",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "tier-preflight",
+    });
+
+    // 带新声明的 effort 切模：预检放行、set_model 发出、会话落到 model-b。
+    const beforeSwitch = captured.requests.length;
+    await handle.setModel!("model-b", { providerId: "tier-preflight", effort: "max" });
+    expect(captured.runtimeModel).toBe("model-b");
+
+    // 切模请求本身就必须应用用户显式指定的档位：pi CLI 在 set_model 后会把思考档位重置
+    // 成新模型自己的默认，恢复门若仍按启动时冻结的快照（只有 low）判断，就会静默不发
+    // set_thinking_level，实际跑在默认档上。恢复门必须用落定后的活动快照（含新声明的 max）。
+    expect(
+      captured.requests.slice(beforeSwitch).filter(
+        (request) => request.type === "set_thinking_level",
+      ),
+    ).toEqual([{ type: "set_thinking_level", level: "max" }]);
+
+    // 切模成功后活动快照已含新档位，随后 setEffort(max) 能真的下发。
+    captured.requests.length = 0;
+    await handle.setEffort!("max");
+    expect(
+      captured.requests.filter((request) => request.type === "set_thinking_level"),
+    ).not.toHaveLength(0);
+    await handle.close();
+  });
+
+  it("still realigns when the target catalog equals the current snapshot, not the frozen one", async () => {
+    // 当前模型的快照与目标模型目录结论碰巧相同时，对账也不能跳过：目标模型在本进程
+    // 启动时冻结的快照还停在旧声明，set_model 后 pi 又会重置到目标默认档。若用「相同」
+    // 判空，preflight 退回冻结快照 → 带新档切入直接被误拒；不带 effort 切则落定过期值、
+    // 静默掉默认档。
+    const declaredB = { tiers: ["high"] as string[] | undefined };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "tier-equal",
+              name: "Tier Equal",
+              baseUrl: "http://tier-equal.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"], reasoning: true, thinkingLevelMap: { high: "high", max: "max" } },
+                { id: "model-b", input: ["text"], reasoning: true, thinkingLevelMap: { high: "high" } },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelCatalogThinkingTiers: (providerId, modelId) =>
+        providerId === "tier-equal" && modelId === "model-b" ? declaredB.tiers : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "thinking-tier-equal",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "tier-equal",
+    });
+
+    // 会话期间 model-b 的声明改为与当前模型 model-a 的快照完全相同的集合（启动时
+    // model-b 冻结的快照还没有 max）。宣告集合碰巧相同，对账也不能跳过。
+    declaredB.tiers = ["minimal", "low", "medium", "high", "max"];
+    const beforeSwitch = captured.requests.length;
+    await handle.setModel!("model-b", { providerId: "tier-equal", effort: "max" });
+    expect(captured.runtimeModel).toBe("model-b");
+
+    // 恢复门按落定快照（含 max）发出档位，而不是静默跑在 pi 默认档上。
+    expect(
+      captured.requests.slice(beforeSwitch).filter(
+        (request) => request.type === "set_thinking_level",
+      ),
+    ).toEqual([{ type: "set_thinking_level", level: "max" }]);
+    await handle.close();
+  });
+
+  it("clears the explicitly-off marker when a thinkingEnabled carrier re-opens thinking", async () => {
+    // 用户先关思考，再带 thinkingEnabled:true + 原档位切到另一个推理模型：恢复门与
+    // 归一化算出同一档，去重分支只省重复 RPC，仍必须清掉 thinkingExplicitlyOff——
+    // 否则紧随的无载体切模仍按「用户明确关过」跳过恢复，刚重新开启的思考被误判回关闭。
+    const denseMap = {
+      minimal: "minimal",
+      low: "low",
+      medium: "medium",
+      high: "high",
+      xhigh: "xhigh",
+      max: "max",
+    } as const;
+    const agent = new PiAgent(
+      byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-a",
+              name: "Native A",
+              baseUrl: "http://a.test",
+              api: "openai-responses",
+              models: [
+                { id: "model-a", reasoning: true, thinkingLevelMap: { ...denseMap } },
+                { id: "model-b", reasoning: true, thinkingLevelMap: { ...denseMap } },
+                { id: "model-c", reasoning: true, thinkingLevelMap: { ...denseMap } },
+              ],
+            },
+          ],
+          env: {},
+        }),
+        [
+          { id: "model-a", displayName: "A", contextWindow: 200_000, efforts: ["low", "high"], defaultEffort: "low" },
+          { id: "model-b", displayName: "B", contextWindow: 200_000, efforts: ["low", "high"], defaultEffort: "low" },
+          { id: "model-c", displayName: "C", contextWindow: 200_000, efforts: ["low", "high"], defaultEffort: "low" },
+        ],
+      ),
+    );
+    const handle = await agent.startSession({
+      sessionId: "thinking-reopen-clears-off-marker",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-a",
+      effort: "high",
+    });
+
+    await handle.setThinkingEnabled!(false);
+    const beforeReopen = captured.requests.length;
+    await handle.setModel!("model-b", {
+      providerId: "native-a",
+      effort: "high",
+      thinkingEnabled: true,
+    });
+    // 载体切模本身必须应用 high（恢复门）。
+    expect(
+      captured.requests.slice(beforeReopen).filter(
+        (request) => request.type === "set_thinking_level",
+      ),
+    ).toEqual([{ type: "set_thinking_level", level: "high" }]);
+
+    // 紧随的无载体切模：关闭标记已被清除，档位恢复必须继续生效。
+    const beforeCarrierless = captured.requests.length;
+    await handle.setModel!("model-c", { providerId: "native-a" });
+    expect(
+      captured.requests.slice(beforeCarrierless).filter(
+        (request) => request.type === "set_thinking_level",
+      ),
+    ).toEqual([{ type: "set_thinking_level", level: "high" }]);
+    await handle.close();
+  });
+
+  it("does not realign the thinking tier snapshot when the model switch fails", async () => {
+    // 对账发生在切模的最早入口（成员校验 / effort 校验 / set_model RPC 之前）。若在那里
+    // 就改写 activeEffortSnapshot，set_model 被拒后子进程仍停在旧模型，快照却已是失败
+    // 目标的档位 —— 随后的思考开关与 effort 校验会按错误模型执行。落定只允许在成功路径。
+    //
+    // 用 set_model RPC 失败（success:false）而不是「模型不在目录」：后者对账返回
+    // undefined、压根不进污染分支，测不到东西。
+    const declared = { tiers: ["low", "medium", "high", "max"] as string[] | undefined };
+    // 只拦 set_model；get_state 等其余命令必须走默认桩，否则会话起不来。
+    const defaultHandler = captured.requestHandler;
+    captured.requestHandler = async (command) => {
+      if (command.type === "set_model") return { success: false, error: "injected failure" };
+      return defaultHandler
+        ? await defaultHandler(command)
+        : command.type === "get_state"
+          ? { success: true, data: { sessionFile: "/mock/s.jsonl", model: { contextWindow: 200_000 } } }
+          : { success: true, data: {} };
+    };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "tier-rollback",
+              name: "Tier Rollback",
+              baseUrl: "http://tier-rollback.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"], thinkingLevelMap: { high: "high" } },
+                { id: "model-b", input: ["text"], thinkingLevelMap: { low: "low" } },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelCatalogThinkingTiers: (providerId, modelId) =>
+        providerId === "tier-rollback" && modelId === "model-b" ? declared.tiers : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "thinking-tier-rollback",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "tier-rollback",
+    });
+
+    // 切到 model-b（目录里有，声明了四档），但 set_model 失败 → 仍停在 model-a。
+    await expect(handle.setModel!("model-b", { providerId: "tier-rollback" })).rejects.toThrow();
+    expect(captured.runtimeModel).toBe("model-a");
+
+    // 快照必须仍是 model-a 的（只有 high）。若被 model-b 的四档污染，max 会被当合法档位
+    // 直接下发；快照仍是 model-a 时 assertStartupEffortAllowed 必须拒 —— 抛错即证据。
+    // （报错文案随上游演进，这里只断言「拒绝」这个事实，不锁死具体措辞。）
+    captured.requests.length = 0;
+    await expect(handle.setEffort!("max")).rejects.toThrow(/set_thinking_level refused/);
+    expect(
+      captured.requests.filter((request) => request.type === "set_thinking_level"),
+    ).toHaveLength(0);
+    await handle.close();
+  });
+
+  it("rejects images when the declaration flips to unsupported without a model switch", async () => {
+    // 反向对账：会话启动时快照是「支持图片」，用户后来把声明改成不支持且不切模 ——
+    // 先查快照就放行会把图片发给一个已声明不支持的模型。对账必须在信任快照之前跑。
+    const declared = { image: true };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-revoke",
+              name: "Native Revoke",
+              baseUrl: "http://revoke.test",
+              api: "openai-completions",
+              models: [{ id: "local-model", input: ["text", "image"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: () => declared.image,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-revoked-after-start",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-revoke",
+    });
+    const imagePath = path.join(cwd, "revoke.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    const imageMessage = {
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/revoke.png",
+        },
+      ],
+    };
+    const readRevokeInput = (): string[] => {
+      const modelsJson = JSON.parse(
+        readFileSync(path.join(captured.env.PI_CODING_AGENT_DIR as string, "models.json"), "utf8"),
+      ) as { providers: Record<string, { models: Array<{ id: string; input: string[] }> }> };
+      const block =
+        modelsJson.providers["native-revoke"] ?? modelsJson.providers["cindy-byom-native-revoke"];
+      return block!.models.find((entry) => entry.id === "local-model")!.input;
+    };
+
+    // 基线：声明支持时放行。
+    captured.requests.length = 0;
+    await handle.send(imageMessage);
+    expect(captured.requests.map((request) => request.type)).toContain("prompt");
+
+    // 改成「不支持图片」，不切模直接发图：必须拒收，且盘上快照同步回纯文本。
+    declared.image = false;
+    captured.requests.length = 0;
+    await expect(handle.send(imageMessage)).rejects.toMatchObject({
+      code: "PI_IMAGE_INPUT_UNSUPPORTED",
+    });
+    expect(captured.requests.map((request) => request.type)).not.toContain("prompt");
+    expect(readRevokeInput()).toEqual(["text"]);
+    await handle.close();
+  });
+
+  it("restores image support when the user returns to following the catalog", async () => {
+    // 用户先显式声明「不支持」（会话快照被写成纯文本），随后改回「跟随供应商」：override 被删、
+    // 声明读回 undefined。此时目录未声明 = 默认支持（上游 #4854），旧会话必须恢复支持图片 ——
+    // 否则只能重开任务。关键在于区分「改回继承」与「目录自己标了 input:['text']」：两者在
+    // readModelImageInput 看来都是 undefined，只有活动目录的最终结论(readModelCatalogImageCapability)
+    // 能分辨，所以对账在目录未声明时必须查这个钩子。
+    const declared = { image: true as boolean | null };  // null = 改回跟随供应商（未声明）
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-restore",
+              name: "Native Restore",
+              baseUrl: "http://restore.test",
+              api: "openai-completions",
+              // 目录本身不标 input（未声明 = 支持），所以“恢复继承”应当回到支持。
+              models: [{ id: "local-model", input: ["text", "image"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      // 目录 spec 说支持（input 含 image），用户先显式声明不支持、随后改回继承。
+      // readModelImageInput 只回报“显式声明了什么”，两种状态在它看来都是 undefined。
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-restore" && modelId === "local-model" && declared.image !== null
+          ? declared.image
+          : undefined,
+      // 活动目录的最终结论：未声明 → 默认支持。
+      readModelCatalogImageCapability: (providerId, modelId) =>
+        providerId === "native-restore" && modelId === "local-model" ? true : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-restored-after-inherit",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-restore",
+    });
+    const imagePath = path.join(cwd, "restore.png");
+    writeFileSync(
+      imagePath,
+      Buffer.from(
+        "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY4YAAAAASUVORK5CYII=",
+        "base64",
+      ),
+    );
+    const imageMessage = {
+      type: "user" as const,
+      content: [
+        {
+          type: "image" as const,
+          path: imagePath,
+          managedUrl: "xdt-image://pi-managed/restore.png",
+        },
+      ],
+    };
+    const readRestoreInput = (): string[] => {
+      const modelsJson = JSON.parse(
+        readFileSync(path.join(captured.env.PI_CODING_AGENT_DIR as string, "models.json"), "utf8"),
+      ) as { providers: Record<string, { models: Array<{ id: string; input: string[] }> }> };
+      const block =
+        modelsJson.providers["native-restore"] ?? modelsJson.providers["cindy-byom-native-restore"];
+      return block!.models.find((entry) => entry.id === "local-model")!.input;
+    };
+
+    // 先显式声明「不支持」：目录 spec 说支持，但用户声明压过它 → 快照被写成纯文本，带图被拒收。
+    declared.image = false;
+    captured.requests.length = 0;
+    await handle.setModel!("local-model", { providerId: "native-restore" });
+    await expect(handle.send(imageMessage)).rejects.toMatchObject({
+      code: "PI_IMAGE_INPUT_UNSUPPORTED",
+    });
+    expect(readRestoreInput()).toEqual(["text"]);
+
+    // 用户改回「跟随供应商」：声明读回 undefined，目录未声明 = 默认支持。
+    declared.image = null;
+    captured.requests.length = 0;
+    await handle.setModel!("local-model", { providerId: "native-restore" });
+    expect(readRestoreInput()).toEqual(["text", "image"]);
+
+    // 旧会话恢复支持图片输入，无需重开任务。
+    captured.requests.length = 0;
+    await handle.send(imageMessage);
+    expect(captured.requests).toContainEqual(
+      expect.objectContaining({
+        type: "prompt",
+        images: [expect.objectContaining({ type: "image" })],
+      }),
+    );
+    await handle.close();
+  });
+
+  it("skips the route replay when read-back already matches (spawn-only routes)", async () => {
+    // spawn 能靠 custom model id 跑在 Pi 自带目录没有的路由上，对这种路由重发 set_model 会
+    // 被 Pi 拒（见 switchModel 开头同路由 no-op 的说明）；而重放失败会 terminate 整个任务 ——
+    // 只改一个能力声明就杀掉正在跑的任务。switch_session 后先 get_state 读回，读回已经等于
+    // 当前路由时不得再发 set_model。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-spawn-only",
+              name: "Native Spawn Only",
+              baseUrl: "http://spawn-only.test",
+              api: "openai-completions",
+              models: [{ id: "local-model", input: ["text"] }],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-spawn-only" && modelId === "local-model" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-spawn-only-route",
+      workingDir: cwd,
+      model: "local-model",
+      providerId: "native-spawn-only",
+    });
+
+    declared.image = true;
+    captured.requests.length = 0;
+    // 真机上的 spawn-only 路由：任何 set_model 都会被拒。
+    captured.requestHandler = async (command) =>
+      command.type === "set_model"
+        ? { success: false, error: 'Model "local-model" not found for provider "xai"' }
+        : { success: true, data: {} };
+    await handle.setModel!("local-model", { providerId: "native-spawn-only" });
+
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("switch_session");
+    expect(requestTypes).not.toContain("set_model");
+    expect(captured.closes).toBe(0);
+    // 会话仍然可用：对账后路由没被拽走。
+    expect(captured.runtimeModel).toBe("local-model");
+    await handle.send({ type: "user", content: "still alive" });
+    expect(captured.requests.map((request) => request.type)).toContain("prompt");
+    await handle.close();
+  });
+
+  it("terminates the session when the route cannot be replayed onto a different model", async () => {
+    // 会话内切过模：switch_session 会把子进程拽回启动路由，读回与当前路由不一致就必须重放。
+    // 重放被拒 = 子进程确实跑在错误的路由上，继续用会把提示词静默发给错误的模型 ——
+    // 保留 pending 重试也修不好，只能 fail-closed 终止（且不留下“已重载”的错误结论）。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-route-lost",
+              name: "Native Route Lost",
+              baseUrl: "http://route-lost.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"] },
+                { id: "model-b", input: ["text"] },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-route-lost" && modelId === "model-b" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-route-replay-rejected",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-route-lost",
+    });
+    await handle.setModel!("model-b", { providerId: "native-route-lost" });
+    expect(captured.runtimeModel).toBe("model-b");
+
+    declared.image = true;
+    const baselineSetModels = captured.requests.filter(
+      (request) => request.type === "set_model",
+    ).length;
+    captured.requestHandler = async (command) =>
+      command.type === "get_state" && captured.runtimeModel === "model-b"
+        ? { success: true, data: { model: { provider: "native-route-lost", id: "model-b" } } }
+        : command.type === "set_model"
+          ? { success: false, error: 'Model "model-b" not found for provider "native-route-lost"' }
+          : { success: true, data: {} };
+    await expect(
+      handle.setModel!("model-b", { providerId: "native-route-lost" }),
+    ).rejects.toThrow(/PI_CATALOG_RELOAD_UNCONFIRMED/);
+    expect(captured.closes).toBe(1);
+    expect(
+      captured.requests.filter((request) => request.type === "set_model").length,
+    ).toBeGreaterThan(baselineSetModels);
+    await handle.close();
+  });
+
+  it("replays the previous route but lands on the new model when switching models", async () => {
+    // 非 no-op 切模：对账在 set_model 之前跑，重放恢复的是**旧**路由，随后切模再指向目标 ——
+    // 最终必须落在目标模型上，不能因为对账的重放把切模吃掉。
+    const declared = { image: false };
+    const agent = new PiAgent({
+      ...byomDeps(
+        async () => ({
+          providers: [
+            {
+              id: "native-switch-route",
+              name: "Native Switch Route",
+              baseUrl: "http://switch-route.test",
+              api: "openai-completions",
+              models: [
+                { id: "model-a", input: ["text"] },
+                { id: "model-b", input: ["text"] },
+              ],
+            },
+          ],
+          env: {},
+        }),
+      ),
+      readModelImageInput: (providerId, modelId) =>
+        providerId === "native-switch-route" && modelId === "model-b" && declared.image
+          ? true
+          : undefined,
+    });
+    const handle = await agent.startSession({
+      sessionId: "image-switch-lands-on-target",
+      workingDir: cwd,
+      model: "model-a",
+      providerId: "native-switch-route",
+    });
+
+    // 目标模型 model-b 的声明与会话快照不一致 → 切模路径里先对账后切模。
+    declared.image = true;
+    captured.requests.length = 0;
+    await handle.setModel!("model-b", { providerId: "native-switch-route" });
+    expect(captured.runtimeModel).toBe("model-b");
+    const requestTypes = captured.requests.map((request) => request.type);
+    expect(requestTypes).toContain("switch_session");
+    expect(requestTypes.lastIndexOf("set_model")).toBeGreaterThan(
+      requestTypes.indexOf("switch_session"),
+    );
+    await handle.close();
+  });
+
   it("waits through Pi preflight compaction when accepting a prompt", async () => {
     const agent = new PiAgent(
       byomDeps(async () => ({ providers: [], env: {} })),
diff --git a/packages/maker-core/src/agents/pi/index.ts b/packages/maker-core/src/agents/pi/index.ts
index 96e7e445778..10d2cb5877c 100644
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@ -4211,6 +4211,10 @@ export class PiAgent extends BaseAgent {
     };
     let activeEffortSnapshot = initialEffortSnapshot;
     let mutableEffort: Effort | null = startupEffort ?? null;
+    // 已确认的思考开关状态：setThinkingEnabled(false) 只发 off，不动 mutableEffort（off 不是
+    // 档位）。路由重建类操作（switch_session/set_model）会把档位重置成模型默认，恢复时需要
+    // 知道「用户是否关了思考」，否则会把关掉的思考悄悄重开（等待时间与费用都涨）。
+    let mutableThinkingEnabled = true;
     // 用户明确关掉思考（setThinkingEnabled(false) / 切模显式 thinkingEnabled:false /
     // 启动 opts.thinkingEnabled === false）的意图标记：无载体切模不得据此悄悄重新开启。
     // 只跟随显式意图变化；目标不支持思考而下发 off 不算用户意图，不置位。
@@ -6707,16 +6711,39 @@ export class PiAgent extends BaseAgent {
       }
     };
 
-    const assertImageInputSupported = (images: readonly PiPromptImage[]): void => {
+    const assertImageInputSupported = async (images: readonly PiPromptImage[]): Promise<void> => {
       if (images.length === 0) return;
-      const nativeModel = nativeProviderById
-        .get(mutablePiProviderId)
-        ?.models.find((candidate) => candidate.id === resolveNativeModelId(mutablePiProviderId, mutableModel));
-      const supportsImageInput =
-        mutablePiProviderId === PI_PROVIDER_ID
-          ? gatewayImageInputByModel.get(mutableModel) === true
-          : nativeModel !== undefined && (nativeModel.input ?? ['text', 'image']).includes('image');
-      if (supportsImageInput) return;
+      // 目录未声明 input 时按**支持**图片处理（上游 #4854：未知能力默认放行，不再一律拒收），
+      // 与本文件其余判定口径一致；本机声明面（override）只用来把「确实不支持」的模型挡下来。
+      const supportsNow = (): boolean => {
+        if (mutablePiProviderId === PI_PROVIDER_ID) {
+          return gatewayImageInputByModel.get(mutableModel) === true;
+        }
+        const nativeModel = nativeProviderById
+          .get(mutablePiProviderId)
+          ?.models.find(
+            (candidate) => candidate.id === resolveNativeModelId(mutablePiProviderId, mutableModel),
+          );
+        return nativeModel !== undefined && (nativeModel.input ?? ['text', 'image']).includes('image');
+      };
+      // 双向对账必须先于信任快照：声明可能从「支持」被改成「不支持」，此时旧快照仍是 true，
+      // 先查 supportsNow() 会直接把图片放给一个已声明不支持的模型。对账在判定之前跑
+      // (目录未声明 / 与快照一致时只花一次内存查找)。
+      // send / steer 不在会话串行链上，这里补链：switch_session / set_model 不得与在途控制
+      // RPC(心跳 setModel、compact)交错；switchModel 路径已在链内，故不在那边重复加锁。
+      // 无图消息在第一行 return，纯文本发送仍是零成本。
+      await runExclusivePiRpc(() =>
+        refreshImageCapabilityOnSwitch(mutableModel, mutablePiProviderId),
+      );
+      // 宿主快照已对齐但子进程还没吃到新清单（回合在跑，重载被延后）：本次不得放行 —— 图片
+      // 会发给一个仍按旧清单判定的子进程。缓存已对齐，下一轮回合结束/下一次同步点自动生效。
+      if (imageCapabilityReloadPending && supportsNow()) {
+        throw new Error(
+          '[PI_IMAGE_CAPABILITY_REFRESH_FAILED] 图片输入能力已更新，但子进程要等当前回合结束后才能加载新清单；' +
+            '请稍后重发这张图（或切一次模型）。',
+        );
+      }
+      if (supportsNow()) return;
       throw new PiImageInputUnsupportedError();
     };
 
@@ -6740,9 +6767,55 @@ export class PiAgent extends BaseAgent {
         );
       }
       throw new Error(
-        '[PI_CATALOG_RELOAD_UNCONFIRMED] 模型目录重载未确认，已终止本任务。请重新打开任务后再切换模型。',
+        '[PI_CATALOG_RELOAD_UNCONFIRMED] 模型目录重载未确认，已终止本任务。请重新打开任务后重试。',
       );
     };
+    const restoreNativeCatalog = async (
+      previousProviders: PiNativeProviderSpec[],
+    ): Promise<void> => {
+      // 先写盘再改内存：写失败时磁盘仍是新目录，内存也保持新目录，避免分叉后再抛。
+      const written = await this.writeModelsJson(
+        configHome,
+        previousProviders,
+        retainedRuntimeModel,
+        authProviderId,
+        {
+          remote,
+          fileOps,
+          contextWindow: ctx.contextWindow || startupContextWindow,
+          workingContextWindow: ctx.workingContextWindow,
+          // 任务级预算必须随 settings.json 重写下发：漏传会把逐模型
+          // compaction.modelOverrides 按模型自身窗口重算，预算静默失效。
+          sessionContextWindowBudget,
+          piCompactionPct: sessionPiAutoCompactPct,
+          packages: nativePackagePaths,
+          disabledSkills: disabledSkillLaunch,
+        },
+      );
+      nativeProviders = previousProviders;
+      nativeProviderById.clear();
+      nativeProviderBySourceId.clear();
+      for (const spec of previousProviders) {
+        if (spec.id === PI_PROVIDER_ID) continue;
+        nativeProviderById.set(spec.id, spec);
+        nativeProviderBySourceId.set(spec.sourceProviderId ?? spec.id, spec);
+      }
+      gatewayApiByModel.clear();
+      for (const [key, value] of written.gatewayApiByModel) gatewayApiByModel.set(key, value);
+      gatewayImageInputByModel.clear();
+      for (const [key, value] of written.gatewayImageInputByModel) {
+        gatewayImageInputByModel.set(key, value);
+      }
+    };
+    const restoreNativeCatalogOrTerminate = async (
+      previousProviders: PiNativeProviderSpec[],
+    ): Promise<void> => {
+      try {
+        await restoreNativeCatalog(previousProviders);
+      } catch (err) {
+        await terminateUnconfirmedCatalogReload(err);
+      }
+    };
     let liveRuntimeSettings: PiNativeRuntimeSettings | undefined;
     const queryNativeRuntime = async (
       operation: 'inspect' | 'refresh',
@@ -7195,6 +7268,347 @@ export class PiAgent extends BaseAgent {
       pendingCurrentDescriptorRefresh = true;
       if (nextGateway) retainedLiveGatewayModel = nextDescriptor ?? previousGatewayModel;
     };
+    /**
+     * 图片能力对账的「宿主快照已改、子进程清单未重载」标记。回合在跑时不重建子会话
+     * (switch_session 会按 CLI 路由重建 AgentSession，打断在跑的回合)，此时只对齐宿主
+     * 快照(准入门立即生效)，由下一次空闲对账(心跳 / 发图 / 切模)补重载。
+     */
+    let imageCapabilityReloadPending = false;
+    /**
+     * 把 models.json 的新清单交给子进程。与 xAI 热刷新同口径：set_model 不重读 models.json，
+     * 只有 switch_session → createRuntime → ModelConfig.load 才让子进程加载新清单。
+     * 无 sdkSessionId 时无法 switch_session，只保证 host 侧(准入门读的快照)已对齐。
+     */
+    const reloadImageCapabilityIntoChild = async (
+      model: string,
+      rollbackBaseline?: PiNativeProviderSpec[],
+    ): Promise<void> => {
+      const sessionPath = sdkSessionId;
+      if (!sessionPath) {
+        imageCapabilityReloadPending = false;
+        return;
+      }
+      let reloaded;
+      try {
+        reloaded = await proc.request({ type: 'switch_session', sessionPath });
+      } catch (err) {
+        // reject/超时 = 不知道子进程有没有吃到新 models.json，回滚和放行都可能分叉。
+        return await terminateUnconfirmedCatalogReload(err);
+      }
+      if (!reloaded.success) {
+        // 本次对账刚写过目录就回滚回旧快照；若是早先推迟的补重载（没有本次基线），保留
+        // pending —— 宿主快照已生效，下一次空闲对账继续重试。
+        if (rollbackBaseline) await restoreNativeCatalogOrTerminate(rollbackBaseline);
+        imageCapabilityReloadPending = rollbackBaseline === undefined;
+        throw new Error(
+          `[PI_IMAGE_CAPABILITY_REFRESH_FAILED] 图片输入能力更新未生效：子进程重载被拒（${reloaded.error ?? 'unknown'}）。` +
+            (rollbackBaseline
+              ? '本次变更已回滚，可重试或切换模型后再试。'
+              : '稍后会自动重试；若持续失败请重新打开任务。'),
+        );
+      }
+      // switch_session 按进程启动时的 --provider/--model 重建 AgentSession（与本文件其余 4 处
+      // switch_session 同一坑，见 switchModel 内注释）：重建后子进程回到 spawn 路由，而宿主与
+      // 子代理快照仍以为在当前路由上 —— 同路由 no-op、发图、steer 之后都不会再打 set_model，
+      // 提示词会发给错误的模型且没有任何报错。
+      // 但重放**不能无条件发**：spawn 能靠 custom model id 跑在 Pi 自带目录没有的路由上
+      // （如 SuperGrok），对这种路由重发 set_model 反而会被 Pi 拒（见 switchModel 开头同路由
+      // no-op 的说明与对应测试）—— 先 get_state 读回，读回就等于当前路由时什么都不做。
+      let verifiedModel: { provider?: unknown; id?: unknown } | null | undefined;
+      try {
+        const verified = await proc.request({ type: 'get_state' });
+        if (!verified.success) throw new Error(`get_state failed: ${verified.error ?? 'unknown'}`);
+        verifiedModel = (
+          verified.data as { model?: { provider?: unknown; id?: unknown } | null } | undefined
+        )?.model;
+      } catch (err) {
+        // 读不到状态 = 不知道子进程在哪条路由上，放行与回滚都可能分叉。
+        return await terminateUnconfirmedCatalogReload(err);
+      }
+      const routeMatches =
+        verifiedModel?.provider === mutablePiProviderId && verifiedModel.id === mutableWireModel;
+      if (!routeMatches) {
+        let reapplied;
+        try {
+          reapplied = await proc.request({
+            type: 'set_model',
+            provider: mutablePiProviderId,
+            modelId: mutableWireModel,
+          });
+        } catch (err) {
+          // 请求本身失败 = 变更结果未知。
+          return await terminateUnconfirmedCatalogReload(err);
+        }
+        if (!reapplied.success) {
+          // 子进程确实没被切回当前路由：继续用它会静默发到错误的模型上，而保留 pending
+          // 重试也修不好“已经跑在错路由上”这个事实 —— 终止（fail-closed）。
+          return await terminateUnconfirmedCatalogReload(
+            reapplied.error ?? 'set_model rejected during image capability reload',
+          );
+        }
+        let reverifiedModel: { provider?: unknown; id?: unknown } | null | undefined;
+        try {
+          const reverified = await proc.request({ type: 'get_state' });
+          if (!reverified.success) {
+            throw new Error(`get_state failed: ${reverified.error ?? 'unknown'}`);
+          }
+          reverifiedModel = (
+            reverified.data as { model?: { provider?: unknown; id?: unknown } | null } | undefined
+          )?.model;
+        } catch (err) {
+          return await terminateUnconfirmedCatalogReload(err);
+        }
+        if (
+          reverifiedModel?.provider !== mutablePiProviderId ||
+          reverifiedModel.id !== mutableWireModel
+        ) {
+          return await terminateUnconfirmedCatalogReload(
+            `route read-back mismatch (expected ${String(mutablePiProviderId)}/${mutableWireModel}, ` +
+              `got ${String(reverifiedModel?.provider)}/${String(reverifiedModel?.id)})`,
+          );
+        }
+      }
+      // switch_session / set_model 都会把思考档位重置成模型默认（pi CLI 行为，见 switchModel
+      // 内注释与转录实证）。重载路径不会经过 switchModel 末尾的档位恢复 —— 发图触发的重载
+      // 或同路由切模触发的重载早退后，子进程实际跑在默认档而 UI 仍显示原档位。这里把会话
+      // 当前档位重新下发；失败只记 warn（与 switchModel 同一纪律：档位救不回也不能让发图/对账失败）。
+      if (mutableEffort && activeEffortSnapshot?.includes(mutableEffort)) {
+        // 关闭开关只发 off、不更新 mutableEffort（见 setThinkingEnabled）：恢复时优先还原
+        // 已确认的关闭状态，否则重载会把用户关掉的思考悄悄重开。
+        const restoreLevel = mutableThinkingEnabled
+          ? effortToPiThinkingLevel(mutableEffort)
+          : ('off' as const);
+        try {
+          const restore = await proc.request({
+            type: 'set_thinking_level',
+            level: restoreLevel,
+          });
+          if (!restore.success) {
+            deps.logger.warn('pi: restore thinking level after image capability reload failed', {
+              level: restoreLevel,
+              error: restore.error,
+            });
+          }
+        } catch (err) {
+          deps.logger.warn('pi: restore thinking level after image capability reload errored', {
+            level: restoreLevel,
+            message: err instanceof Error ? err.message : String(err),
+          });
+        }
+      }
+      imageCapabilityReloadPending = false;
+      this.deps.logger.debug('pi image capability refresh reloaded', {
+        model,
+        provider: mutablePiProviderId,
+      });
+    };
+    /**
+     * 会话启动后用户可能才在设置里声明图片输入(本机目录 override)。快照是 startSession 一次
+     * 性解析的、Pi 的 set_model 又不重读 models.json —— 旧会话里「声明完仍被
+     * assertImageInputSupported 拒收」。切模是用户可见的自然同步点:目标模型的**目录声明**与
+     * 会话快照不一致时,只对齐该模型的 input(不整体换 providers:活会话的 env/代理 token 是
+     * 启动时注入的,整体换会踩 xAI 热刷新已知的坑),原子热写 models.json,再让子进程加载
+     * (见 reloadImageCapabilityIntoChild;回合在跑时只对齐宿主快照).
+     * 心跳会反复下发同路由(见 switchModel 开头的同路由 no-op),所以对账只做一次目录内存查找:
+     * 无变化零 I/O。目录未声明(undefined)按「默认支持图片」对齐(上游 #4854 的准入语义)——
+     * 包括「用户先声明不支持、再改回跟随供应商」时把快照从纯文本恢复回支持。
+     * 网关(xd)模型的图片能力不在本机声明面内(设置里也没有该入口),不走此路径。
+     */
+    const refreshImageCapabilityOnSwitch = async (
+      nextModel: string,
+      nextRequestedProviderId: string | null | undefined,
+    ): Promise<void> => {
+      const specProviderId = resolveProviderForModel(nextModel, nextRequestedProviderId);
+      if (specProviderId === PI_PROVIDER_ID) {
+        this.deps.logger.debug('pi image capability refresh skipped', {
+          reason: 'gateway-route',
+          model: nextModel,
+        });
+        return;
+      }
+      const fresh = this.deps.readModelImageInput?.(
+        resolveSourceProvider(specProviderId),
+        nextModel,
+      );
+      const specModelId = resolveNativeModelId(specProviderId, nextModel);
+      const currentSpec = nativeProviderById.get(specProviderId)
+        ?.models.find((candidate) => candidate.id === specModelId);
+      // 快照基线口径与准入门一致（上游 #4854）：`input` 未声明时按「支持图片」算。
+      const hadImage = (currentSpec?.input ?? ['text', 'image']).includes('image');
+      // 目录已声明（本机 override，或目录自己标了不支持）→ 声明就是权威，直接对齐。
+      // 目录未声明（undefined）时需要活动目录的**最终结论**才能判定：此时
+      // 「用户曾声明不支持、刚改回跟随供应商」(应恢复支持) 与「目录自己标 input:['text']」
+      // (应保持纯文本) 在声明面上完全一样。拿不到结论就不翻转快照（保守，不猜）。
+      const catalogCapability =
+        fresh === undefined
+          ? this.deps.readModelCatalogImageCapability?.(
+              resolveSourceProvider(specProviderId),
+              nextModel,
+            )
+          : undefined;
+      const wantsImage = fresh === undefined ? catalogCapability : fresh;
+      const shouldAlign = wantsImage !== undefined && hadImage !== wantsImage;
+      if (!shouldAlign && !imageCapabilityReloadPending) {
+        this.deps.logger.debug('pi image capability refresh skipped', {
+          reason: fresh === undefined ? 'catalog-undeclared' : 'matches-snapshot',
+          model: nextModel,
+          provider: specProviderId,
+        });
+        return;
+      }
+      if (shouldAlign) {
+        // 会话快照里根本没有这条模型行（启动后才新增的 BYOM/目录行）：nextProviders 是在旧
+        // 快照的 models 数组上 map，写不出这一行 —— 写盘是空操作，却会置 pending 并触发一次
+        // 有终止风险的子进程重载，而准入门读的快照始终没变（声明永远不会生效）。显式跳过。
+        if (currentSpec === undefined) {
+          this.deps.logger.warn('pi image capability refresh skipped', {
+            reason: 'row-not-in-snapshot',
+            model: nextModel,
+            provider: specProviderId,
+          });
+          return;
+        }
+        // inheritModels 的订阅行只有带 api/catalogAddition 才写进 models.json(见 writeModelsJson
+        // 的 models 过滤):这类行改 input 落不到子进程,只翻宿主快照会造成 host/child 分叉。
+        // 当前目录数据里订阅行都带 api,这里显式挡住未来的例外。
+        const providerSpec = nativeProviders.find((candidate) => candidate.id === specProviderId);
+        if (
+          providerSpec?.inheritModels === true &&
+          currentSpec.api === undefined &&
+          currentSpec.catalogAddition !== true
+        ) {
+          this.deps.logger.warn('pi image capability refresh skipped', {
+            reason: 'row-not-materialized',
+            model: nextModel,
+            provider: specProviderId,
+          });
+          return;
+        }
+        this.deps.logger.info('pi image capability refresh applying', {
+          model: nextModel,
+          provider: specProviderId,
+          wantsImage,
+          hadImage,
+        });
+      }
+
+      const rollbackBaseline = shouldAlign ? nativeProviders.slice() : undefined;
+      if (rollbackBaseline) {
+        const previousProviders = rollbackBaseline;
+        const nextProviders: PiNativeProviderSpec[] = previousProviders.map((provider) =>
+          provider.id !== specProviderId
+            ? provider
+            : {
+                ...provider,
+                models: provider.models.map((modelSpec) =>
+                  modelSpec.id !== specModelId
+                    ? modelSpec
+                    : {
+                        ...modelSpec,
+                        input: (wantsImage
+                          ? ['text', 'image']
+                          : ['text']) as Array<'text' | 'image'>,
+                      },
+                ),
+              },
+        );
+        try {
+          const written = await this.writeModelsJson(
+            configHome,
+            nextProviders,
+            retainedRuntimeModel,
+            authProviderId,
+            {
+              remote,
+              fileOps,
+              contextWindow: ctx.contextWindow || startupContextWindow,
+              workingContextWindow: ctx.workingContextWindow,
+              // 任务级预算必须随 settings.json 重写下发：漏传会把逐模型
+              // compaction.modelOverrides 按模型自身窗口重算，预算静默失效。
+              sessionContextWindowBudget,
+              piCompactionPct: sessionPiAutoCompactPct,
+              packages: nativePackagePaths,
+              disabledSkills: disabledSkillLaunch,
+            },
+          );
+          nativeProviders = nextProviders;
+          nativeProviderById.clear();
+          nativeProviderBySourceId.clear();
+          for (const spec of nextProviders) {
+            if (spec.id === PI_PROVIDER_ID) continue;
+            nativeProviderById.set(spec.id, spec);
+            nativeProviderBySourceId.set(spec.sourceProviderId ?? spec.id, spec);
+          }
+          gatewayApiByModel.clear();
+          for (const [key, value] of written.gatewayApiByModel) gatewayApiByModel.set(key, value);
+          gatewayImageInputByModel.clear();
+          for (const [key, value] of written.gatewayImageInputByModel) {
+            gatewayImageInputByModel.set(key, value);
+          }
+          imageCapabilityReloadPending = true;
+        } catch (err) {
+          // 原子写保证盘上要么是旧的完整文件要么是新的完整文件；回滚把两种情况都收敛成旧状态，
+          // 与内存一致(写失败时内存还没换)。
+          this.deps.logger.warn('pi image capability refresh persist failed', {
+            model: nextModel,
+            provider: specProviderId,
+            message: err instanceof Error ? err.message : String(err),
+          });
+          await restoreNativeCatalogOrTerminate(previousProviders);
+          throw new Error(
+            `[PI_IMAGE_CAPABILITY_REFRESH_FAILED] 图片输入能力更新写入失败，本次变更已回滚：` +
+              `${err instanceof Error ? err.message : String(err)}。请重试；若持续失败请重新打开任务。`,
+          );
+        }
+      }
+      if (!imageCapabilityReloadPending) return;
+      // 回合运行中不重建子会话：switch_session 会按 CLI 路由重建 AgentSession，把在跑的回合
+      // 打断/错配。宿主快照已对齐(准入门立即生效)，子进程清单留到下一次空闲对账再加载。
+      if (ctx.isStreaming || ctx.pendingHostTurnStartToken) {
+        this.deps.logger.debug('pi image capability refresh deferred', {
+          reason: 'turn-running',
+          model: nextModel,
+        });
+        return;
+      }
+      await reloadImageCapabilityIntoChild(nextModel, rollbackBaseline);
+    };
+
+    /**
+     * 思考档位对账:重新读活动目录里该模型**最终**的档位集合,对齐活会话的
+     * `activeEffortSnapshot`。
+     *
+     * 为什么必须做:`activeEffortSnapshot` 是 startSession 一次性解析的,Pi 的 set_model 又
+     * 不重读 models.json —— 用户在设置里声明档位后,活着的子进程仍是零档位快照,thinking
+     * 通道开不出来(推理继续漏进正文),只有重开会话才生效。切模是用户可见的自然同步点,
+     * 与图片能力对账放在同一位置(no-op 早退之前)。
+     *
+     * 只对齐**档位集合**(决定通道是否存在),不代用户改档位记忆:`set_thinking_level` 的
+     * 具体档位仍由会话自己的 effort 决定。拿不到结论(undefined)时保持快照不变,不猜。
+     */
+    const resolveThinkingTiersForSwitch = (
+      nextModel: string,
+      nextRequestedProviderId: string | null | undefined,
+    ): { tiers: Effort[]; from: readonly Effort[] } | null => {
+      const specProviderId = resolveProviderForModel(nextModel, nextRequestedProviderId);
+      if (specProviderId === PI_PROVIDER_ID) {
+        // 网关路由的档位不在本机声明面内(设置里没有该入口)。
+        return null;
+      }
+      const declared = this.deps.readModelCatalogThinkingTiers?.(
+        resolveSourceProvider(specProviderId),
+        nextModel,
+      );
+      if (!declared) return null;
+      // 空数组 = 显式「不支持思考」，同样是有效结论（上一轮把 efforts:[] 当成 undefined
+      // 会让「刚保存的关闭声明」在对账层又退回保留旧快照，关闭不生效）。
+      // 不能用「与 activeEffortSnapshot 相同」判空：那是**当前模型**的快照，与目标模型的
+      // 子进程状态无关（目标模型在本进程启动时冻结的快照可能还停在旧声明，set_model 后
+      // pi 又会重置到目标默认档）。碰巧相同也必须返回结论，否则 preflight 退回冻结快照
+      // （声明后带新档切入被误拒）、落定回退同样过期的冻结值。返回结论不会多发 RPC：
+      // 恢复门仍按落定快照决定是否 set_thinking_level。
+      return { tiers: [...declared] as Effort[], from: activeEffortSnapshot ?? [] };
+    };
     /**
      * 应用思考开关（调用方必须已在 `runExclusivePiRpc` 队列内）。
      *
@@ -7232,6 +7646,9 @@ export class PiAgent extends BaseAgent {
       // 只在真的打开时更新档位记忆：关掉思考不丢档位——档位是用户的独立选择，
       // 再打开应回到原档位，而不是回落到启动时的档位。
       if (wanted) mutableEffort = targetEffort ?? 'xhigh';
+      // 已确认的思考开关状态（reloadImageCapabilityIntoChild 的档位恢复靠它区分
+      // 「档位在 max 的思考」与「已关闭的思考」）。
+      mutableThinkingEnabled = enabled;
     };
     const switchModel = async (
       model: string,
@@ -7258,6 +7675,27 @@ export class PiAgent extends BaseAgent {
       // docs/research/pi-native-model-refresh.md 与 docs/dev-rules/pi-harness.md
       // 「只适配官方 Pi」）。任务级窗口预算因此不走热改：预算变化由宿主判定有效窗口变化后
       // 重启引擎（createOpts 重新注入），或在预演里判定需要重建后走既有恢复分支。
+// 上游已把压缩预留的重算/比对交给 queryNativeRuntime + liveRuntimeSettings（本次
+      // 306 个提交里重构），本 PR 不再自带 projectedReserve 预检，避免两套口径重复。
+      // 思考档位对账必须在同路由 no-op **之前**(与图片能力同一位置):否则用户
+      // 「声明完不动模型」永远走那条早返回,活着的子进程仍是启动时的零档位快照。
+      //
+      // **只算不落定**:这里是切模的最早入口,后面还有目录成员校验 / effort 校验 /
+      // set_model RPC,任何一步失败都会停在旧模型 —— 提前改写快照会让随后的思考开关与
+      // effort 校验按「失败目标」的档位执行。落定交给成功路径,失败则完全不动
+      // (不猜、不回滚,因为从没改过)。
+      const pendingThinkingTiers = resolveThinkingTiersForSwitch(model, requestedProviderId);
+      if (pendingThinkingTiers) {
+        deps.logger.info('pi: thinking tiers will realign from the active catalog', {
+          model,
+          providerId: resolveProviderForModel(model, requestedProviderId),
+          from: pendingThinkingTiers.from,
+          to: pendingThinkingTiers.tiers,
+        });
+      }
+      // 能力对账必须在同路由 no-op **之前**:心跳与「声明后切回同模型」走的都是那条早返回。
+      // 无目录变化时这里只花一次内存查找(见函数注释),心跳不受影响。
+      await refreshImageCapabilityOnSwitch(model, requestedProviderId);
       if (preview.action === 'refresh') await refreshLiveCatalog(model, effectiveProviderId);
       // Never let a requested native source silently fall through to Cindy's
       // gateway when a provider disappeared or a model was removed.
@@ -7274,9 +7712,17 @@ export class PiAgent extends BaseAgent {
         wireModel === mutableWireModel;
       // effort 能力校验必须排在写路由快照**之前**:它会抛错中止本次切换,而快照一旦落盘就
       // 指向了新 provider —— 那正是父子路由分叉的形状(upstream #1451 与本 PR 的合并点)。
+      //
+      // 预检用**对账读到的目录结论**(pendingThinkingTiers):用户刚为目标模型声明了新档位、
+      // 随即带着新 effort 切模时,启动时冻结的 nextEffortSnapshot 里还没有这一档,
+      // 会在 RPC 发出前把它误判为不可用、让切换失败。校验只是预检,活动快照
+      // 仍只在切模成功后落定(见 routeUnchanged 早退与成功落定两处)。
       const nextEffortSnapshot = resolveAppliedEffortSnapshot(provider, model);
+      const preflightEffortSnapshot = pendingThinkingTiers
+        ? pendingThinkingTiers.tiers
+        : nextEffortSnapshot;
       if (setOpts?.effort) {
-        assertStartupEffortAllowed(nextEffortSnapshot, setOpts.effort);
+        assertStartupEffortAllowed(preflightEffortSnapshot, setOpts.effort);
       }
       // 远端启动快照保留会话内可切换的全部 native provider，但 SSH reverse-forward
       // 只为当前实际路由建立。否则用户仅仅登录过 xAI，就会在启动任意 Cindy/BYOM
@@ -7313,7 +7759,13 @@ export class PiAgent extends BaseAgent {
       // 所以这一步落的是**带 pending 标记的**新路由:内容已经就位(证明可写、内容可回滚),但
       // 扩展见到 `pending: true` 就拒绝派发。等待窗口里一个子进程都起不来,既不会用未确认的新
       // 路由、也不会用与父不一致的旧路由;确认后再清掉标记放行。
+// 上游把 no-op 判定改为只依赖 routeUnchanged + preview.action，不再比较
+      // oldReserve/targetReserve（压缩预留由 queryNativeRuntime 接管）。沿用上游判定，
+      // 保留本 PR 的思考档位落定点：同路由 no-op 不发 set_model，不会失败，是安全落定点。
       if (routeUnchanged && preview.action === 'hot' && !pendingCurrentDescriptorRefresh) {
+        // 同路由 no-op 不会失败(不发 set_model)，所以这里就是安全的落定点：否则
+        // 「声明完不动模型」永远对不上档位快照。
+        if (pendingThinkingTiers) activeEffortSnapshot = pendingThinkingTiers.tiers;
         if (!(await writeSubagentRuntimeFile({ model: wireModel, provider }))) {
           deps.logger.warn('pi: same-route setModel could not refresh subagent snapshot', { model, provider });
         }
@@ -7446,7 +7898,11 @@ export class PiAgent extends BaseAgent {
       pendingCurrentDescriptorRefresh = false;
       mutableWireModel = wireModel;
       mutablePiProviderId = provider;
-      activeEffortSnapshot = nextEffortSnapshot;
+      // 成功落定：优先用本次对账读到的目录结论（用户在会话期间声明过），
+      // 否则用本次切模解析出的档位。失败路径走不到这里，快照保持旧模型的值。
+      activeEffortSnapshot = pendingThinkingTiers
+        ? pendingThinkingTiers.tiers
+        : nextEffortSnapshot;
       mutableProviderId = effectiveProviderId;
       const explicitThinking =
         setOpts && typeof setOpts.thinkingEnabled === 'boolean' ? setOpts.thinkingEnabled : undefined;
@@ -7466,11 +7922,18 @@ export class PiAgent extends BaseAgent {
       // 都会掉回默认档，表现为「推理强度自动变低」。目标档 = 显式 effort > 会话原档位
       // （mutableEffort）；新模型不支持（含未声明档位）时保持 pi 默认，与能力收敛一致。
       const restoredEffort = setOpts?.effort ?? mutableEffort;
+      // 恢复门已确认的档位：尾部归一化段不再按过期快照重发/降档（2026-10-23 Greptile：
+      // 恢复门按落定快照发 max、归一化段按冻结快照改回首档的自相矛盾）。
+      let restoredConfirmed: Effort | null = null;
+      // 恢复门必须用**本次落定的** activeEffortSnapshot（含用户新声明），不能用启动时冻结的
+      // nextEffortSnapshot：目标模型启动时只有 low、用户随后声明了 max 并带 effort=max 切入时，
+      // 预检已放行、set_model 已成功，若这里仍按冻结快照判断就会静默不发 set_thinking_level，
+      // 实际跑在 Pi 自己的默认档上（切模请求本身没应用用户显式指定的档位）。
       if (
         !keepThinkingOff &&
         restoredEffort &&
-        nextEffortSnapshot && nextEffortSnapshot.length > 0 &&
-        nextEffortSnapshot.includes(restoredEffort)
+        activeEffortSnapshot && activeEffortSnapshot.length > 0 &&
+        activeEffortSnapshot.includes(restoredEffort)
       ) {
         // 恢复请求失败（被拒或超时/链路抛错）都不能让 setModel 失败：模型路由此刻已确认,
         // 抛错只会跳过下面的子代理快照收尾（pending 卡死委派），档位本身却救不回来。
@@ -7481,8 +7944,11 @@ export class PiAgent extends BaseAgent {
           });
           if (restore.success) {
             // mutableEffort 只跟随**已确认生效**的档位：它是对外汇报（getEffort）与后续
-            // 切换恢复的依据；被 Pi 拒绝的目标档不能当成已生效值。
+            // 切换恢复的依据；被 Pi 拒绝的目标档不能当成已生效值。应用具体档位同时意味着
+            // 思考是开着的 —— 关闭状态到此被覆盖，后续 reload 恢复的是这个档位而不是 off。
             mutableEffort = restoredEffort;
+            mutableThinkingEnabled = true;
+            restoredConfirmed = restoredEffort;
           } else {
             deps.logger.warn('pi: restore thinking level after model switch failed', {
               effort: restoredEffort,
@@ -7499,6 +7965,10 @@ export class PiAgent extends BaseAgent {
       autoReviewDecisionCache.clear();
       autoReviewUnavailableNotice.reset();
       autoReviewConfirmUndeliveredNotice.reset();
+      // 窗口/预留复核：采用上游新版实现（get_state 复核 verifiedWindow +
+      // reserveForWindow），它取代了旧版「读 set_model 响应 + switch_session 重载」：
+      // 旧路径依赖 set_model 响应里的窗口，而上游已把该响应降为不可信（目录与运行时
+      // 都可能与它分叉），复核必须以子进程自报的 get_state 为准。
       // The parent route and target-window settings are now confirmed. Clear the
       // pending marker last; failure keeps subagent delegation disabled but does
       // not misreport the parent switch itself.
@@ -7554,9 +8024,12 @@ export class PiAgent extends BaseAgent {
             model,
             provider,
           });
-        } else if (targetReasoning === true) {
-          const usable = nextEffortSnapshot && nextEffortSnapshot.length > 0
-            ? nextEffortSnapshot
+          } else if (targetReasoning === true) {
+          // 与恢复门同口径：用本次落定的 activeEffortSnapshot（含会话期间的新声明），
+          // 不能用启动冻结的 nextEffortSnapshot——目标模型刚声明的新档会被这里按旧快照
+          // 降档覆盖（恢复门发 max、归一化又改回首档）。
+          const usable = activeEffortSnapshot && activeEffortSnapshot.length > 0
+            ? activeEffortSnapshot
             : undefined;
           // 用户明确要开思考、目标也支持：即使这次 selection 没带档位（BYOM/无 effort
           // 目录的会话），也要按启动快照收敛——不能什么都不发，否则上一个模型的 off 会
@@ -7602,35 +8075,44 @@ export class PiAgent extends BaseAgent {
           }
         }
         if (level) {
-          // 拒绝可能是瞬时（例如 Pi 正在处理前一条档位请求）：重试一次再放弃。
-          // 仍失败只告警：模型路由已经确认，抛出会让上层回滚它。
-          for (let attempt = 0; attempt < 2; attempt += 1) {
-            try {
-              const resp = await proc.request({ type: 'set_thinking_level', level });
-              if (resp.success) {
-                if (appliedEffort) mutableEffort = appliedEffort;
-                // 意图标记只在 RPC 成功后更新：显式意图才算用户意图（显式 true+目标
-                // 不支持思考而下发的 off 不是用户意图，标记仍为开）；失败沿用原标记。
-                if (explicitThinking === true) thinkingExplicitlyOff = false;
-                else if (explicitThinking === false) thinkingExplicitlyOff = true;
-                break;
-              }
-              if (attempt === 1) {
-                deps.logger.warn('pi: set_thinking_level rejected after model switch', {
-                  model,
-                  provider,
-                  level,
-                  error: resp.error,
-                });
-              }
-            } catch (err) {
-              if (attempt === 1) {
-                deps.logger.warn('pi: set_thinking_level failed after model switch', {
-                  model,
-                  provider,
-                  level,
-                  message: err instanceof Error ? err.message : String(err),
-                });
+          if (restoredConfirmed && appliedEffort === restoredConfirmed) {
+            // 恢复门刚按落定快照确认过同档：只省重复 RPC，确认成功后的状态更新仍要做——
+            // 显式 thinkingEnabled 载体必须清掉 thinkingExplicitlyOff，否则随后的无载体
+            // 切模仍按「用户明确关过思考」跳过恢复，刚重新开启的档位被误判回关闭。
+            if (appliedEffort) mutableEffort = appliedEffort;
+            if (explicitThinking === true) thinkingExplicitlyOff = false;
+            else if (explicitThinking === false) thinkingExplicitlyOff = true;
+          } else {
+            // 拒绝可能是瞬时（例如 Pi 正在处理前一条档位请求）：重试一次再放弃。
+            // 仍失败只告警：模型路由已经确认，抛出会让上层回滚它。
+            for (let attempt = 0; attempt < 2; attempt += 1) {
+              try {
+                const resp = await proc.request({ type: 'set_thinking_level', level });
+                if (resp.success) {
+                  if (appliedEffort) mutableEffort = appliedEffort;
+                  // 意图标记只在 RPC 成功后更新：显式意图才算用户意图（显式 true+目标
+                  // 不支持思考而下发的 off 不是用户意图，标记仍为开）；失败沿用原标记。
+                  if (explicitThinking === true) thinkingExplicitlyOff = false;
+                  else if (explicitThinking === false) thinkingExplicitlyOff = true;
+                  break;
+                }
+                if (attempt === 1) {
+                  deps.logger.warn('pi: set_thinking_level rejected after model switch', {
+                    model,
+                    provider,
+                    level,
+                    error: resp.error,
+                  });
+                }
+              } catch (err) {
+                if (attempt === 1) {
+                  deps.logger.warn('pi: set_thinking_level failed after model switch', {
+                    model,
+                    provider,
+                    level,
+                    message: err instanceof Error ? err.message : String(err),
+                  });
+                }
               }
             }
           }
@@ -7728,7 +8210,9 @@ export class PiAgent extends BaseAgent {
           let { text, images } = await buildPiPrompt(message, { remote });
           const pinnedSkill = sendOpts?.[PINNED_SKILL_INVOCATION];
           rejectIfCancelled(sendOpts, 'send');
-          assertImageInputSupported(images);
+          await assertImageInputSupported(images);
+          // 对账可能 await 过(仅目录已声明的失败路径)，期间 signal 可能被撤：投递前再查一次。
+          rejectIfCancelled(sendOpts, 'send');
           setAutoReviewIntent(appendAutoReviewUserIntent(priorAutoReviewIntent(), message.content, sendOpts), { authority: autoReviewContext() });
           reviewIntentUpdated = true;
           const applyRequestedPlanMode = async () => {
@@ -8087,7 +8571,9 @@ export class PiAgent extends BaseAgent {
         let { text, images } = await buildPiPrompt(message, { remote });
         rejectIfCancelled(sendOpts, 'steer');
         rejectIfUnavailable();
-        assertImageInputSupported(images);
+        await assertImageInputSupported(images);
+        // 同上：对账 await 后重新确认未被取消。
+        rejectIfCancelled(sendOpts, 'steer');
         setAutoReviewIntent(appendAutoReviewUserIntent(priorAutoReviewIntent(), message.content, sendOpts));
         // A steered channel message can add confirmation requirements to the
         // running turn, but must not remove the current sender's restrictions.
@@ -8381,13 +8867,27 @@ export class PiAgent extends BaseAgent {
       async setEffort(effort: Effort): Promise<void> {
         if (reviewMode) return;
         assertStartupEffortAllowed(activeEffortSnapshot, effort);
-        if (activeEffortSnapshot?.length === 0) return;
+        if (activeEffortSnapshot?.length === 0) {
+          // 防御：该模型在本次启动快照里没有任何思考档位（目录声明不支持思考）。
+          // 静默 return 会让调用方以为「已应用」；而模型实际仍可能推理，推理文本只能走
+          // content 通道、被当成普通正文渲染。留一行 warn，让这类目录错报能在日志里定位，
+          // 而不是只能翻库看消息构成。
+          deps.logger.warn('pi set_thinking_level skipped: model declares no thinking tiers', {
+            model: opts.model,
+            providerId: opts.providerId ?? null,
+            requestedEffort: effort,
+          });
+          return;
+        }
         const resp = await runExclusivePiRpc(() => proc.request({
           type: 'set_thinking_level',
           level: effortToPiThinkingLevel(effort),
         }));
         if (!resp.success) throw new Error(`pi set_thinking_level failed: ${resp.error ?? 'unknown'}`);
+        // 应用具体档位同时意味着思考开着：覆盖可能的关闭记录，否则后续 reload 恢复会把
+        // 刚应用的档位用 off 关掉。
         mutableEffort = effort;
+        mutableThinkingEnabled = true;
       },
 
       getEffort() {
@@ -8400,6 +8900,7 @@ export class PiAgent extends BaseAgent {
         // 切模进行中用户关闭思考时，若这里不排队直接发，切模收尾仍会按更早的
         // thinkingEnabled:true 载体重新下发开启并清掉关闭标记——用户看到关闭成功、
         // 实际被覆盖（2026-10-17 Greptile P1）。串行后意图按时间序线性化，后发者胜。
+        // applyThinkingEnabled 已含「目标模型无思考档位时下发 off + 不动档位记忆」的防御。
         await runExclusivePiRpc(() => applyThinkingEnabled(enabled));
       },